From 77d298c3c67e14965e47340ed0aaa22738e40497 Mon Sep 17 00:00:00 2001 From: Gary <1601978618@qq.com> Date: Sat, 5 Jul 2025 03:03:43 +0000 Subject: [PATCH] update --- inference.py | 20 +- model/dataset.py | 25 +- model/model_extra.py | 18 +- nohup.out | 2336 +++++++++++++++++++++++++++++++++---- stat_predicate_vocab.py | 27 + test.py | 243 ++++ train_extra_accelerate.py | 117 +- 7 files changed, 2472 insertions(+), 314 deletions(-) create mode 100644 stat_predicate_vocab.py create mode 100644 test.py diff --git a/inference.py b/inference.py index 0427d7d..39f4c9b 100644 --- a/inference.py +++ b/inference.py @@ -7,12 +7,14 @@ from transformers import AutoTokenizer from model.model_extra import MiniMindLM from model.LMConfig import LMConfig -def decode_triple(subject_logits, predicate_logits, object_logits, tokenizer): +PREDICATE_VOCAB_PATH = '/home/rwkv/RWKV-TS/RETRO_TEST/extract/predicate_vocab.json' +with open(PREDICATE_VOCAB_PATH, 'r', encoding='utf-8') as f: + PREDICATE_LIST = json.load(f) +print(len(PREDICATE_LIST)) +def decode_triple(subject_logits, predicate_logits, object_logits, tokenizer, predicate_cls_logits=None): # logits: [1, max_len, vocab_size] subject_ids = subject_logits.argmax(-1).squeeze(0).tolist() - predicate_ids = predicate_logits.argmax(-1).squeeze(0).tolist() object_ids = object_logits.argmax(-1).squeeze(0).tolist() - # 去除pad和eos def clean(ids): if isinstance(ids, int): ids = [ids] @@ -22,8 +24,14 @@ def decode_triple(subject_logits, predicate_logits, object_logits, tokenizer): ids = [i for i in ids if i != tokenizer.pad_token_id] return ids subject = tokenizer.decode(clean(subject_ids), skip_special_tokens=True).strip() - predicate = tokenizer.decode(clean(predicate_ids), skip_special_tokens=True).strip() object_ = tokenizer.decode(clean(object_ids), skip_special_tokens=True).strip() + # 谓词用分类结果 + if predicate_cls_logits is not None: + pred_id = predicate_cls_logits.argmax(-1).item() + predicate = PREDICATE_LIST[pred_id] if pred_id < len(PREDICATE_LIST) else "" + else: + predicate_ids = predicate_logits.argmax(-1).squeeze(0).tolist() + predicate = tokenizer.decode(clean(predicate_ids), skip_special_tokens=True).strip() return {"subject": subject, "predicate": predicate, "object": object_} def infer_triples(model, tokenizer, sentences, device): @@ -35,13 +43,13 @@ def infer_triples(model, tokenizer, sentences, device): input_ids = inputs["input_ids"].to(device) with torch.no_grad(): output = model(input_ids=input_ids) - triple = decode_triple(output.subject_logits, output.predicate_logits, output.object_logits, tokenizer) + triple = decode_triple(output.subject_logits, output.predicate_logits, output.object_logits, tokenizer, output.predicate_cls_logits) results.append({"input": sent, "output": [triple]}) return results def main(): parser = argparse.ArgumentParser(description="MiniMind 三元组抽取推理脚本") - parser.add_argument('--model_path', type=str, default='/home/rwkv/RWKV-TS/RETRO_TEST/Minimind/out/pretrain_512.pth') + parser.add_argument('--model_path', type=str, default='/home/rwkv/RWKV-TS/RETRO_TEST/Minimind/out/pretrain_cls512.pth') parser.add_argument('--tokenizer_path', type=str,default='/home/rwkv/RWKV-TS/RETRO_TEST/Minimind/model/minimind_tokenizer') parser.add_argument('--input_json', type=str,default='/home/rwkv/RWKV-TS/RETRO_TEST/extract/sample_1000.json') parser.add_argument('--output_dir', type=str,default='/home/rwkv/RWKV-TS/RETRO_TEST/Minimind/out', help='输出目录') diff --git a/model/dataset.py b/model/dataset.py index 6658eca..a8772cf 100644 --- a/model/dataset.py +++ b/model/dataset.py @@ -13,6 +13,11 @@ from tqdm import tqdm os.environ["TOKENIZERS_PARALLELISM"] = "true" +# 加载谓词类别（与train_extra_accelerate.py保持一致） +PREDICATE_VOCAB_PATH = '/home/rwkv/RWKV-TS/RETRO_TEST/extract/predicate_vocab.json' +with open(PREDICATE_VOCAB_PATH, 'r', encoding='utf-8') as f: + PREDICATE_LIST = json.load(f) +PREDICATE2ID = {p: i for i, p in enumerate(PREDICATE_LIST)} class PretrainDataset(Dataset): def __init__(self, data_path, tokenizer, max_length=512): @@ -302,9 +307,8 @@ class TriplePretrainDataset(Dataset): return f"{triple['subject']} {triple['predicate']} {triple['object']}" def __getitem__(self, index): - """返回数据，输入文本在运行时tokenize，目标已预tokenize""" + """返回数据，输入文本在运行时tokenize，目标已预tokenize，增加predicate_label字段""" sample = self.samples[index] - # 在运行时tokenize输入文本（用于语言建模） input_text = f"{self.tokenizer.bos_token}{sample['text']}{self.tokenizer.eos_token}" encoding = self.tokenizer( @@ -316,12 +320,22 @@ class TriplePretrainDataset(Dataset): ) input_ids = encoding.input_ids.squeeze() loss_mask = (input_ids != self.tokenizer.pad_token_id) - # 构建训练数据 X = input_ids[:-1] Y = input_ids[1:] loss_mask = loss_mask[1:] - + # 提取谓词label + # 先尝试从target_sentence中间取出谓词 + predicate_label = 0 # 默认0 + try: + # target_sentence格式：主语 谓语 宾语 + triple_str = sample['target_sentence'] + triple_parts = triple_str.strip().split() + if len(triple_parts) >= 3: + predicate = triple_parts[1] + predicate_label = PREDICATE2ID.get(predicate, 0) + except Exception: + predicate_label = 0 return { 'input_ids': X, 'labels': Y, @@ -329,7 +343,8 @@ class TriplePretrainDataset(Dataset): 'target_input_ids': sample['target_input_ids'], # 已经是tensor 'target_attention_mask': sample['target_attention_mask'], # 已经是tensor 'target_sentence': sample['target_sentence'], # 字符串，用于调试 - 'original_text': sample['text'] + 'original_text': sample['text'], + 'predicate_label': torch.tensor(predicate_label, dtype=torch.long) } diff --git a/model/model_extra.py b/model/model_extra.py index fca8c54..8d7571b 100644 --- a/model/model_extra.py +++ b/model/model_extra.py @@ -475,7 +475,7 @@ class MOEFeedForward(nn.Module): class TripleExtractionHead(nn.Module): """三元组提取任务头""" - def __init__(self, config: LMConfig): + def __init__(self, config: LMConfig, num_predicates=None): super().__init__() self.config = config @@ -506,6 +506,10 @@ class TripleExtractionHead(nn.Module): self.subject_output = nn.Linear(config.dim, self.max_subject_len * config.dim, bias=False) self.object_output = nn.Linear(config.dim, self.max_object_len * config.dim, bias=False) + # 分类头 + self.num_predicates = num_predicates if num_predicates is not None else 617 + self.predicate_cls = nn.Linear(config.dim, self.num_predicates) + print(f"三元组提取任务头配置:") print(f"- 主语最大长度: {self.max_subject_len}") print(f"- 谓语最大长度: {self.max_predicate_len}") @@ -520,6 +524,7 @@ class TripleExtractionHead(nn.Module): predicate_logits: [batch_size, seq_len, max_predicate_len, vocab_size] - 谓语序列预测 subject_logits: [batch_size, seq_len, max_subject_len, vocab_size] - 主语序列预测 object_logits: [batch_size, seq_len, max_object_len, vocab_size] - 宾语序列预测 + predicate_cls_logits: [batch_size, num_predicates] - 谓词分类logits """ batch_size, seq_len, dim = h.shape @@ -532,6 +537,8 @@ class TripleExtractionHead(nn.Module): predicate_features = predicate_features.mean(dim=1) predicate_raw = self.predicate_output(predicate_features) # [batch_size, max_predicate_len * vocab_size] predicate_logits = predicate_raw.view(batch_size, self.max_predicate_len, -1) + # 分类logits + predicate_cls_logits = self.predicate_cls(predicate_features) # [batch_size, num_predicates] # 3. h1通过交叉注意力（k,v都是h）得到h2 h2 = self.cross_attention_subject(h1, h) # query是h1，key和value都是h @@ -553,7 +560,7 @@ class TripleExtractionHead(nn.Module): object_raw = self.object_output(object_features) # [batch_size, max_object_len * vocab_size] object_logits = object_raw.view(batch_size, self.max_object_len, -1) - return predicate_logits, subject_logits, object_logits + return predicate_logits, subject_logits, object_logits, predicate_cls_logits class MiniMindBlock(nn.Module): @@ -586,7 +593,7 @@ class MiniMindBlock(nn.Module): class MiniMindLM(PreTrainedModel): config_class = LMConfig - def __init__(self, params: LMConfig = None,mode="triple"): + def __init__(self, params: LMConfig = None, mode="triple", num_predicates=None): self.params = params or LMConfig() super().__init__(self.params) self.vocab_size, self.n_layers = params.vocab_size, params.n_layers @@ -599,7 +606,7 @@ class MiniMindLM(PreTrainedModel): self.tok_embeddings.weight = self.output.weight # 添加三元组提取任务头（可训练） - self.triple_extraction_head = TripleExtractionHead(params) + self.triple_extraction_head = TripleExtractionHead(params, num_predicates=num_predicates) self.register_buffer("pos_cis", precompute_pos_cis(dim=params.dim // params.n_heads, theta=params.rope_theta), persistent=False) @@ -656,7 +663,7 @@ class MiniMindLM(PreTrainedModel): ) # 应用三元组提取任务头 - predicate_logits, subject_logits, object_logits = self.triple_extraction_head(h, pos_cis) + predicate_logits, subject_logits, object_logits, predicate_cls_logits = self.triple_extraction_head(h, pos_cis) predicate_logits = predicate_logits.reshape(input_ids.size(0)*self.params.max_predicate_len, -1) subject_logits = subject_logits.reshape(input_ids.size(0)*self.params.max_subject_len, -1) object_logits = object_logits.reshape(input_ids.size(0)*self.params.max_object_len, -1) @@ -685,6 +692,7 @@ class MiniMindLM(PreTrainedModel): output.predicate_logits = predicate_logits output.subject_logits = subject_logits output.object_logits = object_logits + output.predicate_cls_logits = predicate_cls_logits return output diff --git a/nohup.out b/nohup.out index 54143a4..fe5ab45 100644 --- a/nohup.out +++ b/nohup.out @@ -1,127 +1,127 @@ swanlab: \ Waiting for the swanlab cloud response. swanlab: swanlab version 0.6.4 is available! Upgrade: `pip install -U swanlab` swanlab: \ Getting project... swanlab: \ Creating experiment... swanlab: | Creating experiment... swanlab: Tracking run with swanlab version 0.6.3 -swanlab: Run data will be saved locally in /home/rwkv/RWKV-TS/RETRO_TEST/Minimind/swanlog/run-20250702_123051-d30a286e +swanlab: Run data will be saved locally in /home/rwkv/RWKV-TS/RETRO_TEST/Minimind/swanlog/run-20250704_161200-d30a286e swanlab: 👋 Hi Garylu, welcome to swanlab! swanlab: Syncing run MiniMind-TripleExtraction-Epoch-4-BatchSize-192-LearningRate-0.0002 to the cloud swanlab: 🏠 View project at https://swanlab.cn/@Garylu/MiniMind-TripleExtraction -swanlab: 🚀 View run at https://swanlab.cn/@Garylu/MiniMind-TripleExtraction/runs/pgnn4um8pb74vf4bpden3 -[2025-07-02 12:30:52] tokens_per_iter: 98304 -[2025-07-02 12:30:52] Configuration: -[2025-07-02 12:30:52] out_dir: out -[2025-07-02 12:30:52] epochs: 4 -[2025-07-02 12:30:52] embedding_epoch: 2 -[2025-07-02 12:30:52] batch_size: 192 -[2025-07-02 12:30:52] learning_rate: 0.0002 -[2025-07-02 12:30:52] dtype: bfloat16 -[2025-07-02 12:30:52] use_swanlab: True -[2025-07-02 12:30:52] swanlab_project: MiniMind-TripleExtraction -[2025-07-02 12:30:52] num_workers: 1 -[2025-07-02 12:30:52] accumulation_steps: 32 -[2025-07-02 12:30:52] grad_clip: 1.0 -[2025-07-02 12:30:52] warmup_iters: 0 -[2025-07-02 12:30:52] log_interval: 50 -[2025-07-02 12:30:52] save_interval: 10000 -[2025-07-02 12:30:52] dim: 512 -[2025-07-02 12:30:52] n_layers: 8 -[2025-07-02 12:30:52] max_seq_len: 512 -[2025-07-02 12:30:52] use_moe: False -[2025-07-02 12:30:52] disable_db: False -[2025-07-02 12:30:52] data_path: /home/rwkv/RWKV-TS/RETRO_TEST/extract/processed_trex_data.json -[2025-07-02 12:30:52] pretrained_embedding_path: None -[2025-07-02 12:30:52] profile: True -[2025-07-02 12:30:52] profile_interval: 10 -[2025-07-02 12:30:52] use_flash_attn: True -[2025-07-02 12:30:52] knowledge_num: 960400 -[2025-07-02 12:30:52] knowledge_length: 32 -[2025-07-02 12:30:52] database_init_path: ./dataset/combined_prepare.json -[2025-07-02 12:30:52] fast_clustering: True -[2025-07-02 12:30:52] cluster_cache_path: ./cache/cluster_tokens_single.pt -[2025-07-02 12:30:52] recompute_clusters: False -[2025-07-02 12:30:52] memory_monitor: False -[2025-07-02 12:30:52] memory_monitor_interval: 10 -[2025-07-02 12:30:52] max_targets: 5 -[2025-07-02 12:30:52] temperature: 1.0 -[2025-07-02 12:30:52] detailed_timing: True -[2025-07-02 12:30:52] save_dir: out -[2025-07-02 12:30:52] swanlab_run_name: MiniMind-TripleExtraction-Epoch-4-BatchSize-192-LearningRate-0.0002 -[2025-07-02 12:30:52] n_heads: 32 -[2025-07-02 12:30:52] n_kv_heads: 8 -[2025-07-02 12:30:52] vocab_size: 6400 -[2025-07-02 12:30:52] hidden_dim: None -[2025-07-02 12:30:52] multiple_of: 64 -[2025-07-02 12:30:52] norm_eps: 1e-05 -[2025-07-02 12:30:52] rope_theta: 1000000.0 -[2025-07-02 12:30:52] dropout: 0.0 -[2025-07-02 12:30:52] flash_attn: True -[2025-07-02 12:30:52] embeddings_epoch: 2 -[2025-07-02 12:30:52] num_experts_per_tok: 2 -[2025-07-02 12:30:52] n_routed_experts: 4 -[2025-07-02 12:30:52] n_shared_experts: True -[2025-07-02 12:30:52] scoring_func: softmax -[2025-07-02 12:30:52] aux_loss_alpha: 0.1 -[2025-07-02 12:30:52] seq_aux: True -[2025-07-02 12:30:52] norm_topk_prob: True -[2025-07-02 12:30:52] knowledge_dim: 128 -[2025-07-02 12:30:52] max_subject_len: 8 -[2025-07-02 12:30:52] max_predicate_len: 4 -[2025-07-02 12:30:52] max_object_len: 8 -[2025-07-02 12:30:52] return_dict: True -[2025-07-02 12:30:52] output_hidden_states: False -[2025-07-02 12:30:52] output_attentions: False -[2025-07-02 12:30:52] torchscript: False -[2025-07-02 12:30:52] torch_dtype: None -[2025-07-02 12:30:52] use_bfloat16: False -[2025-07-02 12:30:52] tf_legacy_loss: False -[2025-07-02 12:30:52] pruned_heads: {} -[2025-07-02 12:30:52] tie_word_embeddings: True -[2025-07-02 12:30:52] chunk_size_feed_forward: 0 -[2025-07-02 12:30:52] is_encoder_decoder: False -[2025-07-02 12:30:52] is_decoder: False -[2025-07-02 12:30:52] cross_attention_hidden_size: None -[2025-07-02 12:30:52] add_cross_attention: False -[2025-07-02 12:30:52] tie_encoder_decoder: False -[2025-07-02 12:30:52] max_length: 20 -[2025-07-02 12:30:52] min_length: 0 -[2025-07-02 12:30:52] do_sample: False -[2025-07-02 12:30:52] early_stopping: False -[2025-07-02 12:30:52] num_beams: 1 -[2025-07-02 12:30:52] num_beam_groups: 1 -[2025-07-02 12:30:52] diversity_penalty: 0.0 -[2025-07-02 12:30:52] top_k: 50 -[2025-07-02 12:30:52] top_p: 1.0 -[2025-07-02 12:30:52] typical_p: 1.0 -[2025-07-02 12:30:52] repetition_penalty: 1.0 -[2025-07-02 12:30:52] length_penalty: 1.0 -[2025-07-02 12:30:52] no_repeat_ngram_size: 0 -[2025-07-02 12:30:52] encoder_no_repeat_ngram_size: 0 -[2025-07-02 12:30:52] bad_words_ids: None -[2025-07-02 12:30:52] num_return_sequences: 1 -[2025-07-02 12:30:52] output_scores: False -[2025-07-02 12:30:52] return_dict_in_generate: False -[2025-07-02 12:30:52] forced_bos_token_id: None -[2025-07-02 12:30:52] forced_eos_token_id: None -[2025-07-02 12:30:52] remove_invalid_values: False -[2025-07-02 12:30:52] exponential_decay_length_penalty: None -[2025-07-02 12:30:52] suppress_tokens: None -[2025-07-02 12:30:52] begin_suppress_tokens: None -[2025-07-02 12:30:52] architectures: None -[2025-07-02 12:30:52] finetuning_task: None -[2025-07-02 12:30:52] id2label: {0: 'LABEL_0', 1: 'LABEL_1'} -[2025-07-02 12:30:52] label2id: {'LABEL_0': 0, 'LABEL_1': 1} -[2025-07-02 12:30:52] tokenizer_class: None -[2025-07-02 12:30:52] prefix: None -[2025-07-02 12:30:52] bos_token_id: None -[2025-07-02 12:30:52] pad_token_id: None -[2025-07-02 12:30:52] eos_token_id: None -[2025-07-02 12:30:52] sep_token_id: None -[2025-07-02 12:30:52] decoder_start_token_id: None -[2025-07-02 12:30:52] task_specific_params: None -[2025-07-02 12:30:52] problem_type: None -[2025-07-02 12:30:52] _name_or_path: -[2025-07-02 12:30:52] _commit_hash: None -[2025-07-02 12:30:52] _attn_implementation_internal: None -[2025-07-02 12:30:52] _attn_implementation_autoset: False -[2025-07-02 12:30:52] transformers_version: None +swanlab: 🚀 View run at https://swanlab.cn/@Garylu/MiniMind-TripleExtraction/runs/hin5oa7xx8oc4ae1q5e9i +[2025-07-04 16:12:01] tokens_per_iter: 98304 +[2025-07-04 16:12:01] Configuration: +[2025-07-04 16:12:01] out_dir: out +[2025-07-04 16:12:01] epochs: 4 +[2025-07-04 16:12:01] embedding_epoch: 2 +[2025-07-04 16:12:01] batch_size: 192 +[2025-07-04 16:12:01] learning_rate: 0.0002 +[2025-07-04 16:12:01] dtype: bfloat16 +[2025-07-04 16:12:01] use_swanlab: True +[2025-07-04 16:12:01] swanlab_project: MiniMind-TripleExtraction +[2025-07-04 16:12:01] num_workers: 1 +[2025-07-04 16:12:01] accumulation_steps: 32 +[2025-07-04 16:12:01] grad_clip: 1.0 +[2025-07-04 16:12:01] warmup_iters: 0 +[2025-07-04 16:12:01] log_interval: 50 +[2025-07-04 16:12:01] save_interval: 10000 +[2025-07-04 16:12:01] dim: 512 +[2025-07-04 16:12:01] n_layers: 8 +[2025-07-04 16:12:01] max_seq_len: 512 +[2025-07-04 16:12:01] use_moe: False +[2025-07-04 16:12:01] disable_db: False +[2025-07-04 16:12:01] data_path: /home/rwkv/RWKV-TS/RETRO_TEST/extract/processed_trex_data.json +[2025-07-04 16:12:01] pretrained_embedding_path: None +[2025-07-04 16:12:01] profile: True +[2025-07-04 16:12:01] profile_interval: 10 +[2025-07-04 16:12:01] use_flash_attn: True +[2025-07-04 16:12:01] knowledge_num: 960400 +[2025-07-04 16:12:01] knowledge_length: 32 +[2025-07-04 16:12:01] database_init_path: ./dataset/combined_prepare.json +[2025-07-04 16:12:01] fast_clustering: True +[2025-07-04 16:12:01] cluster_cache_path: ./cache/cluster_tokens_single.pt +[2025-07-04 16:12:01] recompute_clusters: False +[2025-07-04 16:12:01] memory_monitor: False +[2025-07-04 16:12:01] memory_monitor_interval: 10 +[2025-07-04 16:12:01] max_targets: 5 +[2025-07-04 16:12:01] temperature: 1.0 +[2025-07-04 16:12:01] detailed_timing: True +[2025-07-04 16:12:01] save_dir: out +[2025-07-04 16:12:01] swanlab_run_name: MiniMind-TripleExtraction-Epoch-4-BatchSize-192-LearningRate-0.0002 +[2025-07-04 16:12:01] n_heads: 32 +[2025-07-04 16:12:01] n_kv_heads: 8 +[2025-07-04 16:12:01] vocab_size: 6400 +[2025-07-04 16:12:01] hidden_dim: None +[2025-07-04 16:12:01] multiple_of: 64 +[2025-07-04 16:12:01] norm_eps: 1e-05 +[2025-07-04 16:12:01] rope_theta: 1000000.0 +[2025-07-04 16:12:01] dropout: 0.0 +[2025-07-04 16:12:01] flash_attn: True +[2025-07-04 16:12:01] embeddings_epoch: 2 +[2025-07-04 16:12:01] num_experts_per_tok: 2 +[2025-07-04 16:12:01] n_routed_experts: 4 +[2025-07-04 16:12:01] n_shared_experts: True +[2025-07-04 16:12:01] scoring_func: softmax +[2025-07-04 16:12:01] aux_loss_alpha: 0.1 +[2025-07-04 16:12:01] seq_aux: True +[2025-07-04 16:12:01] norm_topk_prob: True +[2025-07-04 16:12:01] knowledge_dim: 128 +[2025-07-04 16:12:01] max_subject_len: 8 +[2025-07-04 16:12:01] max_predicate_len: 4 +[2025-07-04 16:12:01] max_object_len: 8 +[2025-07-04 16:12:01] return_dict: True +[2025-07-04 16:12:01] output_hidden_states: False +[2025-07-04 16:12:01] output_attentions: False +[2025-07-04 16:12:01] torchscript: False +[2025-07-04 16:12:01] torch_dtype: None +[2025-07-04 16:12:01] use_bfloat16: False +[2025-07-04 16:12:01] tf_legacy_loss: False +[2025-07-04 16:12:01] pruned_heads: {} +[2025-07-04 16:12:01] tie_word_embeddings: True +[2025-07-04 16:12:01] chunk_size_feed_forward: 0 +[2025-07-04 16:12:01] is_encoder_decoder: False +[2025-07-04 16:12:01] is_decoder: False +[2025-07-04 16:12:01] cross_attention_hidden_size: None +[2025-07-04 16:12:01] add_cross_attention: False +[2025-07-04 16:12:01] tie_encoder_decoder: False +[2025-07-04 16:12:01] max_length: 20 +[2025-07-04 16:12:01] min_length: 0 +[2025-07-04 16:12:01] do_sample: False +[2025-07-04 16:12:01] early_stopping: False +[2025-07-04 16:12:01] num_beams: 1 +[2025-07-04 16:12:01] num_beam_groups: 1 +[2025-07-04 16:12:01] diversity_penalty: 0.0 +[2025-07-04 16:12:01] top_k: 50 +[2025-07-04 16:12:01] top_p: 1.0 +[2025-07-04 16:12:01] typical_p: 1.0 +[2025-07-04 16:12:01] repetition_penalty: 1.0 +[2025-07-04 16:12:01] length_penalty: 1.0 +[2025-07-04 16:12:01] no_repeat_ngram_size: 0 +[2025-07-04 16:12:01] encoder_no_repeat_ngram_size: 0 +[2025-07-04 16:12:01] bad_words_ids: None +[2025-07-04 16:12:01] num_return_sequences: 1 +[2025-07-04 16:12:01] output_scores: False +[2025-07-04 16:12:01] return_dict_in_generate: False +[2025-07-04 16:12:01] forced_bos_token_id: None +[2025-07-04 16:12:01] forced_eos_token_id: None +[2025-07-04 16:12:01] remove_invalid_values: False +[2025-07-04 16:12:01] exponential_decay_length_penalty: None +[2025-07-04 16:12:01] suppress_tokens: None +[2025-07-04 16:12:01] begin_suppress_tokens: None +[2025-07-04 16:12:01] architectures: None +[2025-07-04 16:12:01] finetuning_task: None +[2025-07-04 16:12:01] id2label: {0: 'LABEL_0', 1: 'LABEL_1'} +[2025-07-04 16:12:01] label2id: {'LABEL_0': 0, 'LABEL_1': 1} +[2025-07-04 16:12:01] tokenizer_class: None +[2025-07-04 16:12:01] prefix: None +[2025-07-04 16:12:01] bos_token_id: None +[2025-07-04 16:12:01] pad_token_id: None +[2025-07-04 16:12:01] eos_token_id: None +[2025-07-04 16:12:01] sep_token_id: None +[2025-07-04 16:12:01] decoder_start_token_id: None +[2025-07-04 16:12:01] task_specific_params: None +[2025-07-04 16:12:01] problem_type: None +[2025-07-04 16:12:01] _name_or_path: +[2025-07-04 16:12:01] _commit_hash: None +[2025-07-04 16:12:01] _attn_implementation_internal: None +[2025-07-04 16:12:01] _attn_implementation_autoset: False +[2025-07-04 16:12:01] transformers_version: None 三元组提取任务头配置: - 主语最大长度: 8 - 谓语最大长度: 4 @@ -133,70 +133,70 @@ - output - pos_cis 注意：triple_extraction_head 保持可训练状态 -[2025-07-02 12:30:53] Loading pretrained weights from /home/rwkv/RWKV-TS/RETRO_TEST/extract/Experiment_1_2_2_pretrain_512.pth -[2025-07-02 12:30:53] Successfully loaded pretrained state_dict with 143 parameters -[2025-07-02 12:30:53] Loaded 143 parameters from pretrained weights -[2025-07-02 12:30:53] Skipped 0 parameters -[2025-07-02 12:30:53] Key loaded parameters: -[2025-07-02 12:30:53] ✅ tok_embeddings.weight -[2025-07-02 12:30:53] ✅ knowledge_dataset.keys -[2025-07-02 12:30:53] ✅ knowledge_dataset.knowledge_dataset -[2025-07-02 12:30:53] ✅ knowledge_dataset.tok_embeddings.weight -[2025-07-02 12:30:53] ✅ knowledge_dataset.to_queries.0.weight -[2025-07-02 12:30:53] ... and 61 more -[2025-07-02 12:30:53] Database embeddings and sentences stored in model -[2025-07-02 12:30:53] LLM总参数量：14.486 百万 -[2025-07-02 12:30:53] 模型初始化完成 -[2025-07-02 12:30:53] 检测到pos_cis复数张量，将其设置为不参与分布式训练 -[2025-07-02 12:30:53] 三元组提取训练：使用 TriplePretrainDataset +[2025-07-04 16:12:02] Loading pretrained weights from /home/rwkv/RWKV-TS/RETRO_TEST/extract/Experiment_1_2_2_pretrain_512.pth +[2025-07-04 16:12:02] Successfully loaded pretrained state_dict with 143 parameters +[2025-07-04 16:12:02] Loaded 143 parameters from pretrained weights +[2025-07-04 16:12:02] Skipped 0 parameters +[2025-07-04 16:12:02] Key loaded parameters: +[2025-07-04 16:12:02] ✅ tok_embeddings.weight +[2025-07-04 16:12:02] ✅ knowledge_dataset.keys +[2025-07-04 16:12:02] ✅ knowledge_dataset.knowledge_dataset +[2025-07-04 16:12:02] ✅ knowledge_dataset.tok_embeddings.weight +[2025-07-04 16:12:02] ✅ knowledge_dataset.to_queries.0.weight +[2025-07-04 16:12:02] ... and 61 more +[2025-07-04 16:12:02] Database embeddings and sentences stored in model +[2025-07-04 16:12:02] LLM总参数量：14.802 百万 +[2025-07-04 16:12:02] 模型初始化完成 +[2025-07-04 16:12:02] 检测到pos_cis复数张量，将其设置为不参与分布式训练 +[2025-07-04 16:12:02] 三元组提取训练：使用 TriplePretrainDataset 🚀 开始加载和预处理三元组数据... 📂 加载原始数据... 📊 原始数据量: 3459987 个样本 🔍 验证数据格式并选择单个target... - 验证数据格式: 0%| | 0/3459987 [00:00 -[2025-07-02 12:39:26,868] [INFO] [logging.py:107:log_dist] [Rank 0] Creating torch.bfloat16 ZeRO stage 2 optimizer -[2025-07-02 12:39:26,869] [INFO] [stage_1_and_2.py:150:__init__] Reduce bucket size 500000000 -[2025-07-02 12:39:26,869] [INFO] [stage_1_and_2.py:151:__init__] Allgather bucket size 500000000 -[2025-07-02 12:39:26,869] [INFO] [stage_1_and_2.py:152:__init__] CPU Offload: False -[2025-07-02 12:39:26,869] [INFO] [stage_1_and_2.py:153:__init__] Round robin gradient partitioning: False -[2025-07-02 12:39:31,187] [INFO] [utils.py:781:see_memory_usage] Before initializing optimizer states -[2025-07-02 12:39:31,188] [INFO] [utils.py:782:see_memory_usage] MA 0.38 GB Max_MA 0.41 GB CA 0.43 GB Max_CA 0 GB -[2025-07-02 12:39:31,191] [INFO] [utils.py:789:see_memory_usage] CPU Virtual Memory: used = 60.17 GB, percent = 27.3% -[2025-07-02 12:39:33,966] [INFO] [utils.py:781:see_memory_usage] After initializing optimizer states -[2025-07-02 12:39:33,967] [INFO] [utils.py:782:see_memory_usage] MA 0.38 GB Max_MA 0.44 GB CA 0.49 GB Max_CA 0 GB -[2025-07-02 12:39:33,967] [INFO] [utils.py:789:see_memory_usage] CPU Virtual Memory: used = 60.21 GB, percent = 27.3% -[2025-07-02 12:39:33,968] [INFO] [stage_1_and_2.py:571:__init__] optimizer state initialized -[2025-07-02 12:39:36,078] [INFO] [utils.py:781:see_memory_usage] After initializing ZeRO optimizer -[2025-07-02 12:39:36,079] [INFO] [utils.py:782:see_memory_usage] MA 0.38 GB Max_MA 0.38 GB CA 0.49 GB Max_CA 0 GB -[2025-07-02 12:39:36,080] [INFO] [utils.py:789:see_memory_usage] CPU Virtual Memory: used = 60.21 GB, percent = 27.3% -[2025-07-02 12:39:36,082] [INFO] [logging.py:107:log_dist] [Rank 0] DeepSpeed Final Optimizer = DeepSpeedZeroOptimizer -[2025-07-02 12:39:36,082] [INFO] [logging.py:107:log_dist] [Rank 0] DeepSpeed using configured LR scheduler = None -[2025-07-02 12:39:36,082] [INFO] [logging.py:107:log_dist] [Rank 0] DeepSpeed LR Scheduler = None -[2025-07-02 12:39:36,083] [INFO] [logging.py:107:log_dist] [Rank 0] step=0, skipped=0, lr=[0.0], mom=[(0.9, 0.999)] -[2025-07-02 12:39:36,083] [INFO] [config.py:1014:print] DeepSpeedEngine configuration: -[2025-07-02 12:39:36,084] [INFO] [config.py:1018:print] activation_checkpointing_config { +[2025-07-04 16:20:50,149] [INFO] [logging.py:107:log_dist] [Rank 0] DeepSpeed Flops Profiler Enabled: False +[2025-07-04 16:20:50,150] [INFO] [logging.py:107:log_dist] [Rank 0] Using client Optimizer as basic optimizer +[2025-07-04 16:20:50,150] [INFO] [logging.py:107:log_dist] [Rank 0] Removing param_group that has no 'params' in the basic Optimizer +[2025-07-04 16:20:50,153] [INFO] [logging.py:107:log_dist] [Rank 0] DeepSpeed Basic Optimizer = AdamW +[2025-07-04 16:20:50,153] [INFO] [utils.py:59:is_zero_supported_optimizer] Checking ZeRO support for optimizer=AdamW type= +[2025-07-04 16:20:50,153] [INFO] [logging.py:107:log_dist] [Rank 0] Creating torch.bfloat16 ZeRO stage 2 optimizer +[2025-07-04 16:20:50,153] [INFO] [stage_1_and_2.py:150:__init__] Reduce bucket size 500000000 +[2025-07-04 16:20:50,153] [INFO] [stage_1_and_2.py:151:__init__] Allgather bucket size 500000000 +[2025-07-04 16:20:50,153] [INFO] [stage_1_and_2.py:152:__init__] CPU Offload: False +[2025-07-04 16:20:50,154] [INFO] [stage_1_and_2.py:153:__init__] Round robin gradient partitioning: False +[2025-07-04 16:20:55,774] [INFO] [utils.py:781:see_memory_usage] Before initializing optimizer states +[2025-07-04 16:20:55,775] [INFO] [utils.py:782:see_memory_usage] MA 0.38 GB Max_MA 0.41 GB CA 0.44 GB Max_CA 0 GB +[2025-07-04 16:20:55,778] [INFO] [utils.py:789:see_memory_usage] CPU Virtual Memory: used = 39.3 GB, percent = 17.8% +[2025-07-04 16:20:59,850] [INFO] [utils.py:781:see_memory_usage] After initializing optimizer states +[2025-07-04 16:20:59,851] [INFO] [utils.py:782:see_memory_usage] MA 0.38 GB Max_MA 0.44 GB CA 0.49 GB Max_CA 0 GB +[2025-07-04 16:20:59,852] [INFO] [utils.py:789:see_memory_usage] CPU Virtual Memory: used = 39.27 GB, percent = 17.8% +[2025-07-04 16:20:59,852] [INFO] [stage_1_and_2.py:571:__init__] optimizer state initialized +[2025-07-04 16:21:02,402] [INFO] [utils.py:781:see_memory_usage] After initializing ZeRO optimizer +[2025-07-04 16:21:02,403] [INFO] [utils.py:782:see_memory_usage] MA 0.38 GB Max_MA 0.38 GB CA 0.49 GB Max_CA 0 GB +[2025-07-04 16:21:02,404] [INFO] [utils.py:789:see_memory_usage] CPU Virtual Memory: used = 39.24 GB, percent = 17.8% +[2025-07-04 16:21:02,408] [INFO] [logging.py:107:log_dist] [Rank 0] DeepSpeed Final Optimizer = DeepSpeedZeroOptimizer +[2025-07-04 16:21:02,408] [INFO] [logging.py:107:log_dist] [Rank 0] DeepSpeed using configured LR scheduler = None +[2025-07-04 16:21:02,408] [INFO] [logging.py:107:log_dist] [Rank 0] DeepSpeed LR Scheduler = None +[2025-07-04 16:21:02,408] [INFO] [logging.py:107:log_dist] [Rank 0] step=0, skipped=0, lr=[0.0], mom=[(0.9, 0.999)] +[2025-07-04 16:21:02,409] [INFO] [config.py:1014:print] DeepSpeedEngine configuration: +[2025-07-04 16:21:02,409] [INFO] [config.py:1018:print] activation_checkpointing_config { "partition_activations": false, "contiguous_memory_optimization": false, "cpu_checkpointing": false, @@ -204,10 +204,10 @@ "synchronize_checkpoint_boundary": false, "profile": false } -[2025-07-02 12:39:36,084] [INFO] [config.py:1018:print] aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'intra_op_parallelism': 1, 'single_submit': False, 'overlap_events': True, 'use_gds': False} -[2025-07-02 12:39:36,084] [INFO] [config.py:1018:print] amp_enabled .................. False -[2025-07-02 12:39:36,084] [INFO] [config.py:1018:print] amp_params ................... False -[2025-07-02 12:39:36,085] [INFO] [config.py:1018:print] autotuning_config ............ { +[2025-07-04 16:21:02,409] [INFO] [config.py:1018:print] aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'intra_op_parallelism': 1, 'single_submit': False, 'overlap_events': True, 'use_gds': False} +[2025-07-04 16:21:02,409] [INFO] [config.py:1018:print] amp_enabled .................. False +[2025-07-04 16:21:02,409] [INFO] [config.py:1018:print] amp_params ................... False +[2025-07-04 16:21:02,410] [INFO] [config.py:1018:print] autotuning_config ............ { "enabled": false, "start_step": null, "end_step": null, @@ -232,33 +232,33 @@ "min_train_micro_batch_size_per_gpu": 1, "num_tuning_micro_batch_sizes": 3 } -[2025-07-02 12:39:36,085] [INFO] [config.py:1018:print] bfloat16_enabled ............. True -[2025-07-02 12:39:36,085] [INFO] [config.py:1018:print] bfloat16_immediate_grad_update True -[2025-07-02 12:39:36,085] [INFO] [config.py:1018:print] checkpoint_parallel_write_pipeline False -[2025-07-02 12:39:36,085] [INFO] [config.py:1018:print] checkpoint_tag_validation_enabled True -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] checkpoint_tag_validation_fail False -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] comms_config ................. -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] communication_data_type ...... None -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] compile_config ............... deepcompile=False free_activation=False offload_activation=False offload_opt_states=False double_buffer=True symmetric_memory=False debug_log=False offload_parameters=False sync_before_reduce=False sync_after_reduce=False sync_before_allgather=False sync_after_allgather=False -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] compression_config ........... {'weight_quantization': {'shared_parameters': {'enabled': False, 'quantizer_kernel': False, 'schedule_offset': 0, 'quantize_groups': 1, 'quantize_verbose': False, 'quantization_type': 'symmetric', 'quantize_weight_in_forward': False, 'rounding': 'nearest', 'fp16_mixed_quantize': False, 'quantize_change_ratio': 0.001}, 'different_groups': {}}, 'activation_quantization': {'shared_parameters': {'enabled': False, 'quantization_type': 'symmetric', 'range_calibration': 'dynamic', 'schedule_offset': 1000}, 'different_groups': {}}, 'sparse_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'row_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'head_pruning': {'shared_parameters': {'enabled': False, 'method': 'topk', 'schedule_offset': 1000}, 'different_groups': {}}, 'channel_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'layer_reduction': {'enabled': False}} -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] curriculum_enabled_legacy .... False -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] curriculum_params_legacy ..... False -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] data_efficiency_config ....... {'enabled': False, 'seed': 1234, 'data_sampling': {'enabled': False, 'num_epochs': 1000, 'num_workers': 0, 'pin_memory': False, 'curriculum_learning': {'enabled': False}, 'dynamic_batching': {'enabled': False, 'lr_scaling_method': 'linear', 'min_batch_size': 1, 'max_batch_size': None, 'sequence_picking_order': 'dataloader', 'verbose': False}}, 'data_routing': {'enabled': False, 'random_ltd': {'enabled': False, 'layer_token_lr_schedule': {'enabled': False}}}} -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] data_efficiency_enabled ...... False -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] dataloader_drop_last ......... False -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] disable_allgather ............ False -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] dump_state ................... False -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] dynamic_loss_scale_args ...... None -[2025-07-02 12:39:36,086] [INFO] [config.py:1018:print] eigenvalue_enabled ........... False -[2025-07-02 12:39:36,087] [INFO] [config.py:1018:print] eigenvalue_gas_boundary_resolution 1 -[2025-07-02 12:39:36,087] [INFO] [config.py:1018:print] eigenvalue_layer_name ........ bert.encoder.layer -[2025-07-02 12:39:36,087] [INFO] [config.py:1018:print] eigenvalue_layer_num ......... 0 -[2025-07-02 12:39:36,087] [INFO] [config.py:1018:print] eigenvalue_max_iter .......... 100 -[2025-07-02 12:39:36,088] [INFO] [config.py:1018:print] eigenvalue_stability ......... 1e-06 -[2025-07-02 12:39:36,088] [INFO] [config.py:1018:print] eigenvalue_tol ............... 0.01 -[2025-07-02 12:39:36,088] [INFO] [config.py:1018:print] eigenvalue_verbose ........... False -[2025-07-02 12:39:36,088] [INFO] [config.py:1018:print] elasticity_enabled ........... False -[2025-07-02 12:39:36,088] [INFO] [config.py:1018:print] flops_profiler_config ........ { +[2025-07-04 16:21:02,410] [INFO] [config.py:1018:print] bfloat16_enabled ............. True +[2025-07-04 16:21:02,410] [INFO] [config.py:1018:print] bfloat16_immediate_grad_update True +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] checkpoint_parallel_write_pipeline False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] checkpoint_tag_validation_enabled True +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] checkpoint_tag_validation_fail False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] comms_config ................. +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] communication_data_type ...... None +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] compile_config ............... deepcompile=False free_activation=False offload_activation=False offload_opt_states=False double_buffer=True symmetric_memory=False debug_log=False offload_parameters=False sync_before_reduce=False sync_after_reduce=False sync_before_allgather=False sync_after_allgather=False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] compression_config ........... {'weight_quantization': {'shared_parameters': {'enabled': False, 'quantizer_kernel': False, 'schedule_offset': 0, 'quantize_groups': 1, 'quantize_verbose': False, 'quantization_type': 'symmetric', 'quantize_weight_in_forward': False, 'rounding': 'nearest', 'fp16_mixed_quantize': False, 'quantize_change_ratio': 0.001}, 'different_groups': {}}, 'activation_quantization': {'shared_parameters': {'enabled': False, 'quantization_type': 'symmetric', 'range_calibration': 'dynamic', 'schedule_offset': 1000}, 'different_groups': {}}, 'sparse_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'row_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'head_pruning': {'shared_parameters': {'enabled': False, 'method': 'topk', 'schedule_offset': 1000}, 'different_groups': {}}, 'channel_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'layer_reduction': {'enabled': False}} +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] curriculum_enabled_legacy .... False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] curriculum_params_legacy ..... False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] data_efficiency_config ....... {'enabled': False, 'seed': 1234, 'data_sampling': {'enabled': False, 'num_epochs': 1000, 'num_workers': 0, 'pin_memory': False, 'curriculum_learning': {'enabled': False}, 'dynamic_batching': {'enabled': False, 'lr_scaling_method': 'linear', 'min_batch_size': 1, 'max_batch_size': None, 'sequence_picking_order': 'dataloader', 'verbose': False}}, 'data_routing': {'enabled': False, 'random_ltd': {'enabled': False, 'layer_token_lr_schedule': {'enabled': False}}}} +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] data_efficiency_enabled ...... False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] dataloader_drop_last ......... False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] disable_allgather ............ False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] dump_state ................... False +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] dynamic_loss_scale_args ...... None +[2025-07-04 16:21:02,411] [INFO] [config.py:1018:print] eigenvalue_enabled ........... False +[2025-07-04 16:21:02,412] [INFO] [config.py:1018:print] eigenvalue_gas_boundary_resolution 1 +[2025-07-04 16:21:02,412] [INFO] [config.py:1018:print] eigenvalue_layer_name ........ bert.encoder.layer +[2025-07-04 16:21:02,412] [INFO] [config.py:1018:print] eigenvalue_layer_num ......... 0 +[2025-07-04 16:21:02,412] [INFO] [config.py:1018:print] eigenvalue_max_iter .......... 100 +[2025-07-04 16:21:02,412] [INFO] [config.py:1018:print] eigenvalue_stability ......... 1e-06 +[2025-07-04 16:21:02,412] [INFO] [config.py:1018:print] eigenvalue_tol ............... 0.01 +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] eigenvalue_verbose ........... False +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] elasticity_enabled ........... False +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] flops_profiler_config ........ { "enabled": false, "recompute_fwd_factor": 0.0, "profile_step": 1, @@ -267,24 +267,24 @@ "detailed": true, "output_file": null } -[2025-07-02 12:39:36,088] [INFO] [config.py:1018:print] fp16_auto_cast ............... None -[2025-07-02 12:39:36,088] [INFO] [config.py:1018:print] fp16_enabled ................. False -[2025-07-02 12:39:36,088] [INFO] [config.py:1018:print] fp16_master_weights_and_gradients False -[2025-07-02 12:39:36,089] [INFO] [config.py:1018:print] global_rank .................. 0 -[2025-07-02 12:39:36,089] [INFO] [config.py:1018:print] grad_accum_dtype ............. None -[2025-07-02 12:39:36,089] [INFO] [config.py:1018:print] gradient_accumulation_steps .. 32 -[2025-07-02 12:39:36,089] [INFO] [config.py:1018:print] gradient_clipping ............ 1.0 -[2025-07-02 12:39:36,089] [INFO] [config.py:1018:print] gradient_predivide_factor .... 1.0 -[2025-07-02 12:39:36,089] [INFO] [config.py:1018:print] graph_harvesting ............. False -[2025-07-02 12:39:36,089] [INFO] [config.py:1018:print] hybrid_engine ................ enabled=False max_out_tokens=512 inference_tp_size=1 release_inference_cache=False pin_parameters=True tp_gather_partition_size=8 -[2025-07-02 12:39:36,089] [INFO] [config.py:1018:print] initial_dynamic_scale ........ 1 -[2025-07-02 12:39:36,090] [INFO] [config.py:1018:print] load_universal_checkpoint .... False -[2025-07-02 12:39:36,090] [INFO] [config.py:1018:print] loss_scale ................... 1.0 -[2025-07-02 12:39:36,090] [INFO] [config.py:1018:print] memory_breakdown ............. False -[2025-07-02 12:39:36,090] [INFO] [config.py:1018:print] mics_hierarchial_params_gather False -[2025-07-02 12:39:36,090] [INFO] [config.py:1018:print] mics_shard_size .............. -1 -[2025-07-02 12:39:36,090] [INFO] [config.py:1018:print] monitor_config ............... tensorboard=TensorBoardConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') comet=CometConfig(enabled=False, samples_log_interval=100, project=None, workspace=None, api_key=None, experiment_name=None, experiment_key=None, online=None, mode=None) wandb=WandbConfig(enabled=False, group=None, team=None, project='deepspeed') csv_monitor=CSVConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') -[2025-07-02 12:39:36,090] [INFO] [config.py:1018:print] nebula_config ................ { +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] fp16_auto_cast ............... None +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] fp16_enabled ................. False +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] fp16_master_weights_and_gradients False +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] global_rank .................. 0 +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] grad_accum_dtype ............. None +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] gradient_accumulation_steps .. 32 +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] gradient_clipping ............ 1.0 +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] gradient_predivide_factor .... 1.0 +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] graph_harvesting ............. False +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] hybrid_engine ................ enabled=False max_out_tokens=512 inference_tp_size=1 release_inference_cache=False pin_parameters=True tp_gather_partition_size=8 +[2025-07-04 16:21:02,413] [INFO] [config.py:1018:print] initial_dynamic_scale ........ 1 +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] load_universal_checkpoint .... False +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] loss_scale ................... 1.0 +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] memory_breakdown ............. False +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] mics_hierarchial_params_gather False +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] mics_shard_size .............. -1 +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] monitor_config ............... tensorboard=TensorBoardConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') comet=CometConfig(enabled=False, samples_log_interval=100, project=None, workspace=None, api_key=None, experiment_name=None, experiment_key=None, online=None, mode=None) wandb=WandbConfig(enabled=False, group=None, team=None, project='deepspeed') csv_monitor=CSVConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] nebula_config ................ { "enabled": false, "persistent_storage_path": null, "persistent_time_interval": 100, @@ -292,34 +292,34 @@ "enable_nebula_load": true, "load_path": null } -[2025-07-02 12:39:36,090] [INFO] [config.py:1018:print] optimizer_legacy_fusion ...... False -[2025-07-02 12:39:36,091] [INFO] [config.py:1018:print] optimizer_name ............... None -[2025-07-02 12:39:36,091] [INFO] [config.py:1018:print] optimizer_params ............. None -[2025-07-02 12:39:36,091] [INFO] [config.py:1018:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0, 'pipe_partitioned': True, 'grad_partitioned': True} -[2025-07-02 12:39:36,091] [INFO] [config.py:1018:print] pld_enabled .................. False -[2025-07-02 12:39:36,091] [INFO] [config.py:1018:print] pld_params ................... False -[2025-07-02 12:39:36,091] [INFO] [config.py:1018:print] prescale_gradients ........... False -[2025-07-02 12:39:36,091] [INFO] [config.py:1018:print] scheduler_name ............... None -[2025-07-02 12:39:36,091] [INFO] [config.py:1018:print] scheduler_params ............. None -[2025-07-02 12:39:36,092] [INFO] [config.py:1018:print] seq_parallel_communication_data_type torch.float32 -[2025-07-02 12:39:36,092] [INFO] [config.py:1018:print] sparse_attention ............. None -[2025-07-02 12:39:36,092] [INFO] [config.py:1018:print] sparse_gradients_enabled ..... False -[2025-07-02 12:39:36,092] [INFO] [config.py:1018:print] steps_per_print .............. inf -[2025-07-02 12:39:36,092] [INFO] [config.py:1018:print] tensor_parallel_config ....... dtype=torch.float16 autotp_size=0 tp_overlap_comm=False tensor_parallel=TPConfig(tp_size=1, tp_grain_size=1, mpu=None, tp_group=None) injection_policy_tuple=None keep_module_on_host=False replace_with_kernel_inject=False -[2025-07-02 12:39:36,092] [INFO] [config.py:1018:print] timers_config ................ enabled=True synchronized=True -[2025-07-02 12:39:36,092] [INFO] [config.py:1018:print] train_batch_size ............. 6144 -[2025-07-02 12:39:36,092] [INFO] [config.py:1018:print] train_micro_batch_size_per_gpu 192 -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] use_data_before_expert_parallel_ False -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] use_node_local_storage ....... False -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] wall_clock_breakdown ......... False -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] weight_quantization_config ... None -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] world_size ................... 1 -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] zero_allow_untested_optimizer True -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] zero_config .................. stage=2 contiguous_gradients=True reduce_scatter=True reduce_bucket_size=500000000 use_multi_rank_bucket_allreduce=True allgather_partitions=True allgather_bucket_size=500000000 overlap_comm=False load_from_fp32_weights=True elastic_checkpoint=False offload_param=DeepSpeedZeroOffloadParamConfig(device='none', nvme_path=None, buffer_count=5, buffer_size=100000000, max_in_cpu=1000000000, pin_memory=False) offload_optimizer=DeepSpeedZeroOffloadOptimizerConfig(device='none', nvme_path=None, buffer_count=4, pin_memory=False, pipeline_read=False, pipeline_write=False, fast_init=False, ratio=1.0) sub_group_size=1000000000 cpu_offload_param=None cpu_offload_use_pin_memory=None cpu_offload=None prefetch_bucket_size=50000000 param_persistence_threshold=100000 model_persistence_threshold=9223372036854775807 max_live_parameters=1000000000 max_reuse_distance=1000000000 gather_16bit_weights_on_model_save=False module_granularity_threshold=0 use_all_reduce_for_fetch_params=False stage3_gather_fp16_weights_on_model_save=False ignore_unused_parameters=True legacy_stage1=False round_robin_gradients=False zero_hpz_partition_size=1 zero_quantized_weights=False zero_quantized_nontrainable_weights=False zero_quantized_gradients=False zeropp_loco_param=None mics_shard_size=-1 mics_hierarchical_params_gather=False memory_efficient_linear=True pipeline_loading_checkpoint=False override_module_apply=True log_trace_cache_warnings=False -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] zero_enabled ................. True -[2025-07-02 12:39:36,093] [INFO] [config.py:1018:print] zero_force_ds_cpu_optimizer .. True -[2025-07-02 12:39:36,094] [INFO] [config.py:1018:print] zero_optimization_stage ...... 2 -[2025-07-02 12:39:36,094] [INFO] [config.py:1004:print_user_config] json = { +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] optimizer_legacy_fusion ...... False +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] optimizer_name ............... None +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] optimizer_params ............. None +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0, 'pipe_partitioned': True, 'grad_partitioned': True} +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] pld_enabled .................. False +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] pld_params ................... False +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] prescale_gradients ........... False +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] scheduler_name ............... None +[2025-07-04 16:21:02,414] [INFO] [config.py:1018:print] scheduler_params ............. None +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] seq_parallel_communication_data_type torch.float32 +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] sparse_attention ............. None +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] sparse_gradients_enabled ..... False +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] steps_per_print .............. inf +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] tensor_parallel_config ....... dtype=torch.float16 autotp_size=0 tp_overlap_comm=False tensor_parallel=TPConfig(tp_size=1, tp_grain_size=1, mpu=None, tp_group=None) injection_policy_tuple=None keep_module_on_host=False replace_with_kernel_inject=False +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] timers_config ................ enabled=True synchronized=True +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] train_batch_size ............. 6144 +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] train_micro_batch_size_per_gpu 192 +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] use_data_before_expert_parallel_ False +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] use_node_local_storage ....... False +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] wall_clock_breakdown ......... False +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] weight_quantization_config ... None +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] world_size ................... 1 +[2025-07-04 16:21:02,415] [INFO] [config.py:1018:print] zero_allow_untested_optimizer True +[2025-07-04 16:21:02,416] [INFO] [config.py:1018:print] zero_config .................. stage=2 contiguous_gradients=True reduce_scatter=True reduce_bucket_size=500000000 use_multi_rank_bucket_allreduce=True allgather_partitions=True allgather_bucket_size=500000000 overlap_comm=False load_from_fp32_weights=True elastic_checkpoint=False offload_param=DeepSpeedZeroOffloadParamConfig(device='none', nvme_path=None, buffer_count=5, buffer_size=100000000, max_in_cpu=1000000000, pin_memory=False) offload_optimizer=DeepSpeedZeroOffloadOptimizerConfig(device='none', nvme_path=None, buffer_count=4, pin_memory=False, pipeline_read=False, pipeline_write=False, fast_init=False, ratio=1.0) sub_group_size=1000000000 cpu_offload_param=None cpu_offload_use_pin_memory=None cpu_offload=None prefetch_bucket_size=50000000 param_persistence_threshold=100000 model_persistence_threshold=9223372036854775807 max_live_parameters=1000000000 max_reuse_distance=1000000000 gather_16bit_weights_on_model_save=False module_granularity_threshold=0 use_all_reduce_for_fetch_params=False stage3_gather_fp16_weights_on_model_save=False ignore_unused_parameters=True legacy_stage1=False round_robin_gradients=False zero_hpz_partition_size=1 zero_quantized_weights=False zero_quantized_nontrainable_weights=False zero_quantized_gradients=False zeropp_loco_param=None mics_shard_size=-1 mics_hierarchical_params_gather=False memory_efficient_linear=True pipeline_loading_checkpoint=False override_module_apply=True log_trace_cache_warnings=False +[2025-07-04 16:21:02,416] [INFO] [config.py:1018:print] zero_enabled ................. True +[2025-07-04 16:21:02,416] [INFO] [config.py:1018:print] zero_force_ds_cpu_optimizer .. True +[2025-07-04 16:21:02,416] [INFO] [config.py:1018:print] zero_optimization_stage ...... 2 +[2025-07-04 16:21:02,416] [INFO] [config.py:1004:print_user_config] json = { "train_batch_size": 6.144000e+03, "train_micro_batch_size_per_gpu": 192, "gradient_accumulation_steps": 32, @@ -345,3 +345,1843 @@ }, "zero_allow_untested_optimizer": true } +[2025-07-04 16:21:02] 开始第1轮训练 +[2025-07-04 16:21:03] 三元组提取训练模式 +[2025-07-04 16:21:03] 使用预tokenized三元组目标数据 +[2025-07-04 16:21:46] Epoch 1/4, Step 50/18020, Loss(triple): 45.131344, Loss(predicate): 216.875000, LR: 0.000001, Speed: 110697.24 tokens/sec | Epoch Time Left: 4:25:58 | Total Time Left: 17:46:05 +[2025-07-04 16:22:28] Epoch 1/4, Step 100/18020, Loss(triple): 45.561543, Loss(predicate): 216.822922, LR: 0.000003, Speed: 117597.28 tokens/sec | Epoch Time Left: 4:17:26 | Total Time Left: 17:14:06 +[2025-07-04 16:23:10] Epoch 1/4, Step 150/18020, Loss(triple): 45.171806, Loss(predicate): 216.317703, LR: 0.000004, Speed: 118577.25 tokens/sec | Epoch Time Left: 4:13:27 | Total Time Left: 17:00:12 +[2025-07-04 16:23:51] Epoch 1/4, Step 200/18020, Loss(triple): 45.189026, Loss(predicate): 214.645828, LR: 0.000006, Speed: 119508.79 tokens/sec | Epoch Time Left: 4:10:38 | Total Time Left: 16:50:59 +[2025-07-04 16:24:32] Epoch 1/4, Step 250/18020, Loss(triple): 45.200523, Loss(predicate): 214.640625, LR: 0.000007, Speed: 118649.81 tokens/sec | Epoch Time Left: 4:09:01 | Total Time Left: 16:46:36 +[2025-07-04 16:25:13] Epoch 1/4, Step 300/18020, Loss(triple): 44.912113, Loss(predicate): 208.166672, LR: 0.000008, Speed: 118849.13 tokens/sec | Epoch Time Left: 4:07:38 | Total Time Left: 16:43:10 +[2025-07-04 16:25:55] Epoch 1/4, Step 350/18020, Loss(triple): 45.167187, Loss(predicate): 206.822922, LR: 0.000010, Speed: 118633.78 tokens/sec | Epoch Time Left: 4:06:31 | Total Time Left: 16:40:46 +[2025-07-04 16:26:36] Epoch 1/4, Step 400/18020, Loss(triple): 45.190811, Loss(predicate): 187.781250, LR: 0.000011, Speed: 120039.60 tokens/sec | Epoch Time Left: 4:05:10 | Total Time Left: 16:37:22 +[2025-07-04 16:27:17] Epoch 1/4, Step 450/18020, Loss(triple): 44.828392, Loss(predicate): 180.151047, LR: 0.000012, Speed: 119299.16 tokens/sec | Epoch Time Left: 4:04:07 | Total Time Left: 16:35:14 +[2025-07-04 16:27:58] === GPU性能分析 (平均每步) === +[2025-07-04 16:27:58] 前向传播: 63.19ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 16:27:58] GPU总时间: 65.14ms, 实际迭代时间: 828.68ms, GPU利用率: 7.9% +[2025-07-04 16:27:58] ================================================== +[2025-07-04 16:27:58] === 三元组预测示例 === +[2025-07-04 16:27:59] 样本1目标: Grabinek, Warmian-Masurian Voivodeship country Poland +[2025-07-04 16:27:59] 样本1预测: leumiesleoreicesurances terarconomore ajinessasiveduabversum +[2025-07-04 16:27:59] 样本2目标: Dichelopa anthracodelta taxon rank species +[2025-07-04 16:27:59] 样本2预测: veumiesandussiffdances formosadeore cententasictduakevers� +[2025-07-04 16:27:59] ================== +[2025-07-04 16:27:59] Epoch 1/4, Step 500/18020, Loss(triple): 44.883881, Loss(predicate): 177.859375, LR: 0.000014, Speed: 118626.51 tokens/sec | Epoch Time Left: 4:03:16 | Total Time Left: 16:33:56 +[2025-07-04 16:28:40] Epoch 1/4, Step 550/18020, Loss(triple): 44.452766, Loss(predicate): 161.770828, LR: 0.000015, Speed: 118544.62 tokens/sec | Epoch Time Left: 4:02:28 | Total Time Left: 16:32:49 +[2025-07-04 16:29:21] Epoch 1/4, Step 600/18020, Loss(triple): 44.755806, Loss(predicate): 156.583328, LR: 0.000017, Speed: 119429.58 tokens/sec | Epoch Time Left: 4:01:33 | Total Time Left: 16:31:10 +[2025-07-04 16:30:02] Epoch 1/4, Step 650/18020, Loss(triple): 44.717293, Loss(predicate): 135.052078, LR: 0.000018, Speed: 119505.12 tokens/sec | Epoch Time Left: 4:00:39 | Total Time Left: 16:29:37 +[2025-07-04 16:30:44] Epoch 1/4, Step 700/18020, Loss(triple): 44.283058, Loss(predicate): 128.705734, LR: 0.000019, Speed: 118780.09 tokens/sec | Epoch Time Left: 3:59:53 | Total Time Left: 16:28:37 +[2025-07-04 16:31:25] Epoch 1/4, Step 750/18020, Loss(triple): 44.472065, Loss(predicate): 118.117188, LR: 0.000021, Speed: 118188.27 tokens/sec | Epoch Time Left: 3:59:12 | Total Time Left: 16:27:59 +[2025-07-04 16:32:06] Epoch 1/4, Step 800/18020, Loss(triple): 44.302155, Loss(predicate): 115.382812, LR: 0.000022, Speed: 119395.03 tokens/sec | Epoch Time Left: 3:58:22 | Total Time Left: 16:26:43 +[2025-07-04 16:32:48] Epoch 1/4, Step 850/18020, Loss(triple): 43.987030, Loss(predicate): 93.483070, LR: 0.000024, Speed: 119365.38 tokens/sec | Epoch Time Left: 3:57:33 | Total Time Left: 16:25:32 +[2025-07-04 16:33:31] Epoch 1/4, Step 900/18020, Loss(triple): 43.786957, Loss(predicate): 64.125000, LR: 0.000025, Speed: 113585.18 tokens/sec | Epoch Time Left: 3:57:25 | Total Time Left: 16:27:10 +[2025-07-04 16:34:22] Epoch 1/4, Step 950/18020, Loss(triple): 43.298126, Loss(predicate): 62.707684, LR: 0.000026, Speed: 95845.84 tokens/sec | Epoch Time Left: 3:59:38 | Total Time Left: 16:38:33 +[2025-07-04 16:35:14] === GPU性能分析 (平均每步) === +[2025-07-04 16:35:14] 前向传播: 74.70ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 16:35:14] GPU总时间: 76.64ms, 实际迭代时间: 1046.35ms, GPU利用率: 7.3% +[2025-07-04 16:35:14] ================================================== +[2025-07-04 16:35:14] === 三元组预测示例 === +[2025-07-04 16:35:14] 样本1目标: Jim Crandall country of citizenship American +[2025-07-04 16:35:14] 样本1预测: avriakoreoreobakav oreav4ill obilavatingillifang� +[2025-07-04 16:35:14] 样本2目标: jeans made from material denim +[2025-07-04 16:35:14] 样本2预测: leavubandoreubakav oreavifir obilasokelapobum +[2025-07-04 16:35:14] ================== +[2025-07-04 16:35:14] Epoch 1/4, Step 1000/18020, Loss(triple): 43.106319, Loss(predicate): 36.057941, LR: 0.000028, Speed: 93949.80 tokens/sec | Epoch Time Left: 4:01:49 | Total Time Left: 16:49:56 +[2025-07-04 16:35:56] Epoch 1/4, Step 1050/18020, Loss(triple): 42.894886, Loss(predicate): 36.076824, LR: 0.000029, Speed: 117380.43 tokens/sec | Epoch Time Left: 4:00:54 | Total Time Left: 16:48:22 +[2025-07-04 16:36:37] Epoch 1/4, Step 1100/18020, Loss(triple): 41.748459, Loss(predicate): 31.330841, LR: 0.000031, Speed: 120668.82 tokens/sec | Epoch Time Left: 3:59:43 | Total Time Left: 16:45:40 +[2025-07-04 16:37:18] Epoch 1/4, Step 1150/18020, Loss(triple): 40.683487, Loss(predicate): 33.368099, LR: 0.000032, Speed: 119744.96 tokens/sec | Epoch Time Left: 3:58:39 | Total Time Left: 16:43:28 +[2025-07-04 16:37:59] Epoch 1/4, Step 1200/18020, Loss(triple): 39.217178, Loss(predicate): 39.034630, LR: 0.000033, Speed: 119007.37 tokens/sec | Epoch Time Left: 3:57:41 | Total Time Left: 16:41:38 +[2025-07-04 16:38:43] Epoch 1/4, Step 1250/18020, Loss(triple): 36.776688, Loss(predicate): 32.606281, LR: 0.000035, Speed: 113975.99 tokens/sec | Epoch Time Left: 3:57:08 | Total Time Left: 16:41:37 +[2025-07-04 16:39:24] Epoch 1/4, Step 1300/18020, Loss(triple): 35.117268, Loss(predicate): 31.107912, LR: 0.000036, Speed: 119089.17 tokens/sec | Epoch Time Left: 3:56:11 | Total Time Left: 16:39:52 +[2025-07-04 16:40:04] Epoch 1/4, Step 1350/18020, Loss(triple): 29.387671, Loss(predicate): 20.993273, LR: 0.000037, Speed: 120965.75 tokens/sec | Epoch Time Left: 3:55:07 | Total Time Left: 16:37:38 +[2025-07-04 16:40:45] Epoch 1/4, Step 1400/18020, Loss(triple): 27.587267, Loss(predicate): 40.426434, LR: 0.000039, Speed: 120373.42 tokens/sec | Epoch Time Left: 3:54:07 | Total Time Left: 16:35:41 +[2025-07-04 16:41:26] Epoch 1/4, Step 1450/18020, Loss(triple): 20.704500, Loss(predicate): 30.683535, LR: 0.000040, Speed: 121272.81 tokens/sec | Epoch Time Left: 3:53:05 | Total Time Left: 16:33:34 +[2025-07-04 16:42:09] === GPU性能分析 (平均每步) === +[2025-07-04 16:42:09] 前向传播: 64.47ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 16:42:09] GPU总时间: 66.42ms, 实际迭代时间: 870.24ms, GPU利用率: 7.6% +[2025-07-04 16:42:09] ================================================== +[2025-07-04 16:42:09] === 三元组预测示例 === +[2025-07-04 16:42:09] 样本1目标: Justin Brady occupation actor +[2025-07-04 16:42:09] 样本1预测: igieieandoreubakub ieilisore oreilstzelakandst +[2025-07-04 16:42:09] 样本2目标: Log Cabin Republicans field of work LGBT +[2025-07-04 16:42:09] 样本2预测: igieuboreoreubakub ieilisore oreilststeloreandum +[2025-07-04 16:42:09] ================== +[2025-07-04 16:42:09] Epoch 1/4, Step 1500/18020, Loss(triple): 18.886604, Loss(predicate): 28.259245, LR: 0.000042, Speed: 112962.06 tokens/sec | Epoch Time Left: 3:52:37 | Total Time Left: 16:33:54 +[2025-07-04 16:42:51] Epoch 1/4, Step 1550/18020, Loss(triple): 17.417841, Loss(predicate): 41.352993, LR: 0.000043, Speed: 116714.11 tokens/sec | Epoch Time Left: 3:51:54 | Total Time Left: 16:33:05 +[2025-07-04 16:43:33] Epoch 1/4, Step 1600/18020, Loss(triple): 17.146231, Loss(predicate): 46.373753, LR: 0.000044, Speed: 119339.91 tokens/sec | Epoch Time Left: 3:51:01 | Total Time Left: 16:31:37 +[2025-07-04 16:44:13] Epoch 1/4, Step 1650/18020, Loss(triple): 16.871035, Loss(predicate): 35.002216, LR: 0.000046, Speed: 121215.51 tokens/sec | Epoch Time Left: 3:50:02 | Total Time Left: 16:29:44 +[2025-07-04 16:44:54] Epoch 1/4, Step 1700/18020, Loss(triple): 16.763775, Loss(predicate): 42.194405, LR: 0.000047, Speed: 120663.17 tokens/sec | Epoch Time Left: 3:49:06 | Total Time Left: 16:28:02 +[2025-07-04 16:45:35] Epoch 1/4, Step 1750/18020, Loss(triple): 16.780495, Loss(predicate): 29.356102, LR: 0.000049, Speed: 119658.49 tokens/sec | Epoch Time Left: 3:48:14 | Total Time Left: 16:26:39 +[2025-07-04 16:46:16] Epoch 1/4, Step 1800/18020, Loss(triple): 15.295885, Loss(predicate): 35.041527, LR: 0.000050, Speed: 120679.49 tokens/sec | Epoch Time Left: 3:47:20 | Total Time Left: 16:25:03 +[2025-07-04 16:46:56] Epoch 1/4, Step 1850/18020, Loss(triple): 15.614405, Loss(predicate): 43.171726, LR: 0.000051, Speed: 120941.27 tokens/sec | Epoch Time Left: 3:46:26 | Total Time Left: 16:23:28 +[2025-07-04 16:47:37] Epoch 1/4, Step 1900/18020, Loss(triple): 15.281368, Loss(predicate): 53.407284, LR: 0.000053, Speed: 120606.29 tokens/sec | Epoch Time Left: 3:45:33 | Total Time Left: 16:22:00 +[2025-07-04 16:48:18] Epoch 1/4, Step 1950/18020, Loss(triple): 15.262808, Loss(predicate): 28.280325, LR: 0.000054, Speed: 120421.63 tokens/sec | Epoch Time Left: 3:44:42 | Total Time Left: 16:20:36 +[2025-07-04 16:48:59] === GPU性能分析 (平均每步) === +[2025-07-04 16:48:59] 前向传播: 55.86ms, 损失计算: 0.02ms, 反向传播: 1.92ms, 优化器: 0.00ms +[2025-07-04 16:48:59] GPU总时间: 57.80ms, 实际迭代时间: 818.91ms, GPU利用率: 7.1% +[2025-07-04 16:48:59] ================================================== +[2025-07-04 16:48:59] === 三元组预测示例 === +[2025-07-04 16:48:59] 样本1目标: 2008–09 Primera Divisió sport football +[2025-07-04 16:48:59] 样本1预测: )kinasinaninin ieilisie adachinininartolo +[2025-07-04 16:48:59] 样本2目标: Dutch Ussat date of birth April 11, 1904 +[2025-07-04 16:48:59] 样本2预测: )kinasinaninin ieilisie adachinininartolo +[2025-07-04 16:48:59] ================== +[2025-07-04 16:48:59] Epoch 1/4, Step 2000/18020, Loss(triple): 14.324287, Loss(predicate): 29.588079, LR: 0.000055, Speed: 120042.38 tokens/sec | Epoch Time Left: 3:43:52 | Total Time Left: 16:19:19 +[2025-07-04 16:49:39] Epoch 1/4, Step 2050/18020, Loss(triple): 14.262953, Loss(predicate): 29.560730, LR: 0.000057, Speed: 120890.16 tokens/sec | Epoch Time Left: 3:43:00 | Total Time Left: 16:17:54 +[2025-07-04 16:50:20] Epoch 1/4, Step 2100/18020, Loss(triple): 13.323330, Loss(predicate): 9.192417, LR: 0.000058, Speed: 121192.84 tokens/sec | Epoch Time Left: 3:42:08 | Total Time Left: 16:16:27 +[2025-07-04 16:51:01] Epoch 1/4, Step 2150/18020, Loss(triple): 12.619385, Loss(predicate): 26.274139, LR: 0.000060, Speed: 120375.06 tokens/sec | Epoch Time Left: 3:41:18 | Total Time Left: 16:15:12 +[2025-07-04 16:51:42] Epoch 1/4, Step 2200/18020, Loss(triple): 12.040726, Loss(predicate): 26.454088, LR: 0.000061, Speed: 120370.90 tokens/sec | Epoch Time Left: 3:40:29 | Total Time Left: 16:13:58 +[2025-07-04 16:52:23] Epoch 1/4, Step 2250/18020, Loss(triple): 12.375034, Loss(predicate): 20.576212, LR: 0.000062, Speed: 120017.35 tokens/sec | Epoch Time Left: 3:39:42 | Total Time Left: 16:12:50 +[2025-07-04 16:53:03] Epoch 1/4, Step 2300/18020, Loss(triple): 12.458298, Loss(predicate): 14.114086, LR: 0.000064, Speed: 120904.12 tokens/sec | Epoch Time Left: 3:38:52 | Total Time Left: 16:11:33 +[2025-07-04 16:53:44] Epoch 1/4, Step 2350/18020, Loss(triple): 11.816998, Loss(predicate): 17.101339, LR: 0.000065, Speed: 120954.51 tokens/sec | Epoch Time Left: 3:38:03 | Total Time Left: 16:10:18 +[2025-07-04 16:54:25] Epoch 1/4, Step 2400/18020, Loss(triple): 12.070099, Loss(predicate): 15.073262, LR: 0.000067, Speed: 119709.09 tokens/sec | Epoch Time Left: 3:37:16 | Total Time Left: 16:09:16 +[2025-07-04 16:55:06] Epoch 1/4, Step 2450/18020, Loss(triple): 11.754757, Loss(predicate): 16.771423, LR: 0.000068, Speed: 120651.73 tokens/sec | Epoch Time Left: 3:36:28 | Total Time Left: 16:08:06 +[2025-07-04 16:55:47] === GPU性能分析 (平均每步) === +[2025-07-04 16:55:47] 前向传播: 57.14ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 16:55:47] GPU总时间: 59.09ms, 实际迭代时间: 816.71ms, GPU利用率: 7.2% +[2025-07-04 16:55:47] ================================================== +[2025-07-04 16:55:47] === 三元组预测示例 === +[2025-07-04 16:55:47] 样本1目标: São Miguel do Fidalgo instance of municipality +[2025-07-04 16:55:47] 样本1预测: )umcladoameran ieilisie :artor ofmart Po +[2025-07-04 16:55:47] 样本2目标: Air Kasaï airline hub N'Dolo Airport +[2025-07-04 16:55:47] 样本2预测: )umcladoameran ieilisie :artor ofmart Po +[2025-07-04 16:55:47] ================== +[2025-07-04 16:55:47] Epoch 1/4, Step 2500/18020, Loss(triple): 11.397768, Loss(predicate): 12.955627, LR: 0.000069, Speed: 120365.60 tokens/sec | Epoch Time Left: 3:35:41 | Total Time Left: 16:07:00 +[2025-07-04 16:56:27] Epoch 1/4, Step 2550/18020, Loss(triple): 11.233477, Loss(predicate): 12.686275, LR: 0.000071, Speed: 121274.92 tokens/sec | Epoch Time Left: 3:34:52 | Total Time Left: 16:05:47 +[2025-07-04 16:57:08] Epoch 1/4, Step 2600/18020, Loss(triple): 11.599659, Loss(predicate): 9.364705, LR: 0.000072, Speed: 121175.88 tokens/sec | Epoch Time Left: 3:34:04 | Total Time Left: 16:04:36 +[2025-07-04 16:57:49] Epoch 1/4, Step 2650/18020, Loss(triple): 11.013382, Loss(predicate): 15.843562, LR: 0.000074, Speed: 119764.71 tokens/sec | Epoch Time Left: 3:33:19 | Total Time Left: 16:03:38 +[2025-07-04 16:58:30] Epoch 1/4, Step 2700/18020, Loss(triple): 11.520782, Loss(predicate): 15.870847, LR: 0.000075, Speed: 120057.64 tokens/sec | Epoch Time Left: 3:32:33 | Total Time Left: 16:02:38 +[2025-07-04 16:59:10] Epoch 1/4, Step 2750/18020, Loss(triple): 11.061575, Loss(predicate): 21.087545, LR: 0.000076, Speed: 120687.09 tokens/sec | Epoch Time Left: 3:31:47 | Total Time Left: 16:01:34 +[2025-07-04 16:59:51] Epoch 1/4, Step 2800/18020, Loss(triple): 10.555201, Loss(predicate): 12.945964, LR: 0.000078, Speed: 120872.08 tokens/sec | Epoch Time Left: 3:31:00 | Total Time Left: 16:00:29 +[2025-07-04 17:00:37] Epoch 1/4, Step 2850/18020, Loss(triple): 11.371559, Loss(predicate): 16.579773, LR: 0.000079, Speed: 107677.22 tokens/sec | Epoch Time Left: 3:30:40 | Total Time Left: 16:01:26 +[2025-07-04 17:01:27] Epoch 1/4, Step 2900/18020, Loss(triple): 11.440628, Loss(predicate): 8.388000, LR: 0.000080, Speed: 98160.73 tokens/sec | Epoch Time Left: 3:30:42 | Total Time Left: 16:04:05 +[2025-07-04 17:02:13] Epoch 1/4, Step 2950/18020, Loss(triple): 10.717131, Loss(predicate): 9.200276, LR: 0.000082, Speed: 105345.59 tokens/sec | Epoch Time Left: 3:30:25 | Total Time Left: 16:05:17 +[2025-07-04 17:03:02] === GPU性能分析 (平均每步) === +[2025-07-04 17:03:02] 前向传播: 62.73ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 17:03:02] GPU总时间: 64.67ms, 实际迭代时间: 976.86ms, GPU利用率: 6.6% +[2025-07-04 17:03:02] ================================================== +[2025-07-04 17:03:02] === 三元组预测示例 === +[2025-07-04 17:03:02] 样本1目标: Sam Spiegel award received Best Picture +[2025-07-04 17:03:02] 样本1预测: FAclicataoran ieolisug Eimorsmain Ro +[2025-07-04 17:03:02] 样本2目标: Wang Wei languages spoken, written or signed Chinese +[2025-07-04 17:03:02] 样本2预测: FAclicataoran zolisas Eimorsmain Ro +[2025-07-04 17:03:02] ================== +[2025-07-04 17:03:02] Epoch 1/4, Step 3000/18020, Loss(triple): 10.565201, Loss(predicate): 6.224538, LR: 0.000083, Speed: 100632.71 tokens/sec | Epoch Time Left: 3:30:18 | Total Time Left: 16:07:15 +[2025-07-04 17:03:44] Epoch 1/4, Step 3050/18020, Loss(triple): 10.845831, Loss(predicate): 12.588155, LR: 0.000085, Speed: 117240.59 tokens/sec | Epoch Time Left: 3:29:36 | Total Time Left: 16:06:31 +[2025-07-04 17:04:32] Epoch 1/4, Step 3100/18020, Loss(triple): 11.414492, Loss(predicate): 11.212677, LR: 0.000086, Speed: 102596.33 tokens/sec | Epoch Time Left: 3:29:22 | Total Time Left: 16:08:01 +[2025-07-04 17:05:18] Epoch 1/4, Step 3150/18020, Loss(triple): 10.651964, Loss(predicate): 11.292908, LR: 0.000087, Speed: 106700.77 tokens/sec | Epoch Time Left: 3:28:59 | Total Time Left: 16:08:45 +[2025-07-04 17:06:05] Epoch 1/4, Step 3200/18020, Loss(triple): 10.650085, Loss(predicate): 9.987854, LR: 0.000089, Speed: 106107.58 tokens/sec | Epoch Time Left: 3:28:36 | Total Time Left: 16:09:33 +[2025-07-04 17:06:47] Epoch 1/4, Step 3250/18020, Loss(triple): 10.667431, Loss(predicate): 13.620850, LR: 0.000090, Speed: 115158.93 tokens/sec | Epoch Time Left: 3:27:56 | Total Time Left: 16:09:00 +[2025-07-04 17:07:32] Epoch 1/4, Step 3300/18020, Loss(triple): 10.866062, Loss(predicate): 11.072453, LR: 0.000092, Speed: 110811.55 tokens/sec | Epoch Time Left: 3:27:23 | Total Time Left: 16:09:02 +[2025-07-04 17:08:14] Epoch 1/4, Step 3350/18020, Loss(triple): 10.789280, Loss(predicate): 10.153666, LR: 0.000093, Speed: 116468.18 tokens/sec | Epoch Time Left: 3:26:40 | Total Time Left: 16:08:18 +[2025-07-04 17:08:55] Epoch 1/4, Step 3400/18020, Loss(triple): 10.611526, Loss(predicate): 7.455292, LR: 0.000094, Speed: 118918.34 tokens/sec | Epoch Time Left: 3:25:54 | Total Time Left: 16:07:17 +[2025-07-04 17:09:36] Epoch 1/4, Step 3450/18020, Loss(triple): 10.356781, Loss(predicate): 12.921834, LR: 0.000096, Speed: 119990.50 tokens/sec | Epoch Time Left: 3:25:06 | Total Time Left: 16:06:09 +[2025-07-04 17:10:17] === GPU性能分析 (平均每步) === +[2025-07-04 17:10:17] 前向传播: 51.99ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 17:10:17] GPU总时间: 53.94ms, 实际迭代时间: 815.41ms, GPU利用率: 6.6% +[2025-07-04 17:10:17] ================================================== +[2025-07-04 17:10:17] === 三元组预测示例 === +[2025-07-04 17:10:17] 样本1目标: Sooneck Castle instance of castle +[2025-07-04 17:10:17] 样本1预测: FAneiconaoran isieyug Eateor ofmass S of +[2025-07-04 17:10:17] 样本2目标: Rothbury located in the administrative territorial entity Northumberland +[2025-07-04 17:10:17] 样本2预测: FAneiconaoran isieyav Eateor ofcass S of +[2025-07-04 17:10:17] ================== +[2025-07-04 17:10:17] Epoch 1/4, Step 3500/18020, Loss(triple): 10.918791, Loss(predicate): 9.729115, LR: 0.000097, Speed: 120557.14 tokens/sec | Epoch Time Left: 3:24:18 | Total Time Left: 16:04:58 +[2025-07-04 17:10:58] Epoch 1/4, Step 3550/18020, Loss(triple): 10.701302, Loss(predicate): 11.982961, LR: 0.000099, Speed: 120019.91 tokens/sec | Epoch Time Left: 3:23:31 | Total Time Left: 16:03:52 +[2025-07-04 17:11:38] Epoch 1/4, Step 3600/18020, Loss(triple): 10.953087, Loss(predicate): 15.946045, LR: 0.000100, Speed: 121063.80 tokens/sec | Epoch Time Left: 3:22:42 | Total Time Left: 16:02:39 +[2025-07-04 17:12:19] Epoch 1/4, Step 3650/18020, Loss(triple): 10.765022, Loss(predicate): 11.334468, LR: 0.000101, Speed: 120049.20 tokens/sec | Epoch Time Left: 3:21:55 | Total Time Left: 16:01:34 +[2025-07-04 17:13:00] Epoch 1/4, Step 3700/18020, Loss(triple): 9.940525, Loss(predicate): 10.785991, LR: 0.000103, Speed: 121209.71 tokens/sec | Epoch Time Left: 3:21:07 | Total Time Left: 16:00:22 +[2025-07-04 17:13:40] Epoch 1/4, Step 3750/18020, Loss(triple): 10.455500, Loss(predicate): 12.931915, LR: 0.000104, Speed: 121069.52 tokens/sec | Epoch Time Left: 3:20:19 | Total Time Left: 15:59:12 +[2025-07-04 17:14:21] Epoch 1/4, Step 3800/18020, Loss(triple): 10.663370, Loss(predicate): 13.389964, LR: 0.000105, Speed: 120051.48 tokens/sec | Epoch Time Left: 3:19:32 | Total Time Left: 15:58:09 +[2025-07-04 17:15:02] Epoch 1/4, Step 3850/18020, Loss(triple): 10.543655, Loss(predicate): 9.024343, LR: 0.000107, Speed: 120321.85 tokens/sec | Epoch Time Left: 3:18:46 | Total Time Left: 15:57:05 +[2025-07-04 17:15:43] Epoch 1/4, Step 3900/18020, Loss(triple): 10.120996, Loss(predicate): 10.471782, LR: 0.000108, Speed: 121062.04 tokens/sec | Epoch Time Left: 3:17:58 | Total Time Left: 15:55:57 +[2025-07-04 17:16:23] Epoch 1/4, Step 3950/18020, Loss(triple): 10.121040, Loss(predicate): 7.337962, LR: 0.000110, Speed: 121142.82 tokens/sec | Epoch Time Left: 3:17:11 | Total Time Left: 15:54:49 +[2025-07-04 17:17:04] === GPU性能分析 (平均每步) === +[2025-07-04 17:17:04] 前向传播: 53.07ms, 损失计算: 0.02ms, 反向传播: 1.97ms, 优化器: 0.00ms +[2025-07-04 17:17:04] GPU总时间: 55.06ms, 实际迭代时间: 816.51ms, GPU利用率: 6.7% +[2025-07-04 17:17:04] ================================================== +[2025-07-04 17:17:04] === 三元组预测示例 === +[2025-07-04 17:17:04] 样本1目标: 1999 WNBA All-Star Game instance of WNBA All-Star Game +[2025-07-04 17:17:04] 样本1预测: FAXitonaran isilyug EateX ofmr R of +[2025-07-04 17:17:04] 样本2目标: One for the Radio performer McFly +[2025-07-04 17:17:04] 样本2预测: FAXitonaran isieyug Eater ofmr R of +[2025-07-04 17:17:04] ================== +[2025-07-04 17:17:04] Epoch 1/4, Step 4000/18020, Loss(triple): 10.776836, Loss(predicate): 10.854706, LR: 0.000111, Speed: 120395.01 tokens/sec | Epoch Time Left: 3:16:24 | Total Time Left: 15:53:46 +[2025-07-04 17:17:45] Epoch 1/4, Step 4050/18020, Loss(triple): 10.671307, Loss(predicate): 10.347524, LR: 0.000112, Speed: 119442.16 tokens/sec | Epoch Time Left: 3:15:39 | Total Time Left: 15:52:50 +[2025-07-04 17:18:26] Epoch 1/4, Step 4100/18020, Loss(triple): 10.218933, Loss(predicate): 9.075938, LR: 0.000114, Speed: 119692.61 tokens/sec | Epoch Time Left: 3:14:54 | Total Time Left: 15:51:52 +[2025-07-04 17:19:07] Epoch 1/4, Step 4150/18020, Loss(triple): 10.460344, Loss(predicate): 13.110046, LR: 0.000115, Speed: 120352.43 tokens/sec | Epoch Time Left: 3:14:08 | Total Time Left: 15:50:51 +[2025-07-04 17:19:48] Epoch 1/4, Step 4200/18020, Loss(triple): 10.480438, Loss(predicate): 12.478027, LR: 0.000117, Speed: 120678.65 tokens/sec | Epoch Time Left: 3:13:22 | Total Time Left: 15:49:48 +[2025-07-04 17:20:29] Epoch 1/4, Step 4250/18020, Loss(triple): 10.306107, Loss(predicate): 7.799174, LR: 0.000118, Speed: 120179.90 tokens/sec | Epoch Time Left: 3:12:37 | Total Time Left: 15:48:49 +[2025-07-04 17:21:10] Epoch 1/4, Step 4300/18020, Loss(triple): 10.472752, Loss(predicate): 15.182027, LR: 0.000119, Speed: 119133.40 tokens/sec | Epoch Time Left: 3:11:52 | Total Time Left: 15:47:56 +[2025-07-04 17:21:51] Epoch 1/4, Step 4350/18020, Loss(triple): 9.949884, Loss(predicate): 8.771759, LR: 0.000121, Speed: 120178.44 tokens/sec | Epoch Time Left: 3:11:07 | Total Time Left: 15:46:58 +[2025-07-04 17:22:32] Epoch 1/4, Step 4400/18020, Loss(triple): 10.316746, Loss(predicate): 10.619761, LR: 0.000122, Speed: 120547.38 tokens/sec | Epoch Time Left: 3:10:22 | Total Time Left: 15:45:58 +[2025-07-04 17:23:13] Epoch 1/4, Step 4450/18020, Loss(triple): 10.424570, Loss(predicate): 13.574056, LR: 0.000123, Speed: 119490.34 tokens/sec | Epoch Time Left: 3:09:37 | Total Time Left: 15:45:04 +[2025-07-04 17:23:54] === GPU性能分析 (平均每步) === +[2025-07-04 17:23:54] 前向传播: 60.32ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 17:23:54] GPU总时间: 62.27ms, 实际迭代时间: 819.24ms, GPU利用率: 7.6% +[2025-07-04 17:23:54] ================================================== +[2025-07-04 17:23:54] === 三元组预测示例 === +[2025-07-04 17:23:54] 样本1目标: Undercurrent (Kenny Drew album) performer Kenny Drew +[2025-07-04 17:23:54] 样本1预测: FW Ititinaran iailyug Eanceory ofmth R of +[2025-07-04 17:23:54] 样本2目标: Tirana Bank headquarters location Tirana +[2025-07-04 17:23:54] 样本2预测: FW Ititinaran iailyug Eanceory ofmth R of +[2025-07-04 17:23:54] ================== +[2025-07-04 17:23:54] Epoch 1/4, Step 4500/18020, Loss(triple): 9.902962, Loss(predicate): 8.883037, LR: 0.000125, Speed: 119994.58 tokens/sec | Epoch Time Left: 3:08:52 | Total Time Left: 15:44:07 +[2025-07-04 17:24:35] Epoch 1/4, Step 4550/18020, Loss(triple): 10.594881, Loss(predicate): 14.881734, LR: 0.000126, Speed: 119163.02 tokens/sec | Epoch Time Left: 3:08:09 | Total Time Left: 15:43:15 +[2025-07-04 17:25:16] Epoch 1/4, Step 4600/18020, Loss(triple): 10.685036, Loss(predicate): 17.339050, LR: 0.000128, Speed: 120762.48 tokens/sec | Epoch Time Left: 3:07:23 | Total Time Left: 15:42:16 +[2025-07-04 17:25:57] Epoch 1/4, Step 4650/18020, Loss(triple): 9.765436, Loss(predicate): 12.562744, LR: 0.000129, Speed: 120988.28 tokens/sec | Epoch Time Left: 3:06:38 | Total Time Left: 15:41:16 +[2025-07-04 17:26:38] Epoch 1/4, Step 4700/18020, Loss(triple): 10.230135, Loss(predicate): 10.958598, LR: 0.000130, Speed: 119767.58 tokens/sec | Epoch Time Left: 3:05:53 | Total Time Left: 15:40:22 +[2025-07-04 17:27:18] Epoch 1/4, Step 4750/18020, Loss(triple): 10.419558, Loss(predicate): 14.578201, LR: 0.000132, Speed: 120619.16 tokens/sec | Epoch Time Left: 3:05:08 | Total Time Left: 15:39:24 +[2025-07-04 17:27:59] Epoch 1/4, Step 4800/18020, Loss(triple): 10.510899, Loss(predicate): 6.845978, LR: 0.000133, Speed: 120701.93 tokens/sec | Epoch Time Left: 3:04:23 | Total Time Left: 15:38:26 +[2025-07-04 17:28:40] Epoch 1/4, Step 4850/18020, Loss(triple): 9.912617, Loss(predicate): 10.131307, LR: 0.000135, Speed: 121220.72 tokens/sec | Epoch Time Left: 3:03:38 | Total Time Left: 15:37:26 +[2025-07-04 17:29:20] Epoch 1/4, Step 4900/18020, Loss(triple): 10.229767, Loss(predicate): 7.576660, LR: 0.000136, Speed: 120748.13 tokens/sec | Epoch Time Left: 3:02:53 | Total Time Left: 15:36:29 +[2025-07-04 17:30:07] Epoch 1/4, Step 4950/18020, Loss(triple): 10.137684, Loss(predicate): 12.544291, LR: 0.000137, Speed: 105246.73 tokens/sec | Epoch Time Left: 3:02:24 | Total Time Left: 15:36:54 +[2025-07-04 17:30:52] === GPU性能分析 (平均每步) === +[2025-07-04 17:30:52] 前向传播: 61.96ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 17:30:52] GPU总时间: 63.90ms, 实际迭代时间: 899.59ms, GPU利用率: 7.1% +[2025-07-04 17:30:52] ================================================== +[2025-07-04 17:30:52] === 三元组预测示例 === +[2025-07-04 17:30:52] 样本1目标: Antonov An-24 manufacturer Antonov Design Bureau +[2025-07-04 17:30:52] 样本1预测: FAlicitonaran iaieyug speciesist3 ofmr R of +[2025-07-04 17:30:52] 样本2目标: 87th Academy Awards point in time February 22, 2015 +[2025-07-04 17:30:52] 样本2预测: FWlicitonaram stieyug speciesist3 ofmr R of +[2025-07-04 17:30:52] ================== +[2025-07-04 17:30:52] Epoch 1/4, Step 5000/18020, Loss(triple): 10.441898, Loss(predicate): 13.429240, LR: 0.000139, Speed: 109276.88 tokens/sec | Epoch Time Left: 3:01:50 | Total Time Left: 15:36:53 +[2025-07-04 17:31:33] Epoch 1/4, Step 5050/18020, Loss(triple): 10.162394, Loss(predicate): 8.507548, LR: 0.000140, Speed: 119102.67 tokens/sec | Epoch Time Left: 3:01:07 | Total Time Left: 15:36:03 +[2025-07-04 17:32:15] Epoch 1/4, Step 5100/18020, Loss(triple): 9.855936, Loss(predicate): 10.993999, LR: 0.000142, Speed: 118917.67 tokens/sec | Epoch Time Left: 3:00:24 | Total Time Left: 15:35:14 +[2025-07-04 17:32:57] Epoch 1/4, Step 5150/18020, Loss(triple): 9.838099, Loss(predicate): 10.842387, LR: 0.000143, Speed: 115567.60 tokens/sec | Epoch Time Left: 2:59:43 | Total Time Left: 15:34:40 +[2025-07-04 17:33:38] Epoch 1/4, Step 5200/18020, Loss(triple): 9.755518, Loss(predicate): 11.545909, LR: 0.000144, Speed: 119230.42 tokens/sec | Epoch Time Left: 2:59:00 | Total Time Left: 15:33:50 +[2025-07-04 17:34:19] Epoch 1/4, Step 5250/18020, Loss(triple): 10.045525, Loss(predicate): 8.726949, LR: 0.000146, Speed: 120641.65 tokens/sec | Epoch Time Left: 2:58:15 | Total Time Left: 15:32:53 +[2025-07-04 17:35:00] Epoch 1/4, Step 5300/18020, Loss(triple): 10.477234, Loss(predicate): 8.290833, LR: 0.000147, Speed: 120362.27 tokens/sec | Epoch Time Left: 2:57:31 | Total Time Left: 15:31:58 +[2025-07-04 17:35:40] Epoch 1/4, Step 5350/18020, Loss(triple): 9.660015, Loss(predicate): 9.787333, LR: 0.000148, Speed: 121124.22 tokens/sec | Epoch Time Left: 2:56:46 | Total Time Left: 15:31:00 +[2025-07-04 17:36:21] Epoch 1/4, Step 5400/18020, Loss(triple): 9.812998, Loss(predicate): 12.247162, LR: 0.000150, Speed: 120948.86 tokens/sec | Epoch Time Left: 2:56:01 | Total Time Left: 15:30:03 +[2025-07-04 17:37:02] Epoch 1/4, Step 5450/18020, Loss(triple): 9.749969, Loss(predicate): 8.380157, LR: 0.000151, Speed: 119836.38 tokens/sec | Epoch Time Left: 2:55:17 | Total Time Left: 15:29:11 +[2025-07-04 17:37:43] === GPU性能分析 (平均每步) === +[2025-07-04 17:37:43] 前向传播: 60.01ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 17:37:43] GPU总时间: 61.96ms, 实际迭代时间: 815.26ms, GPU利用率: 7.6% +[2025-07-04 17:37:43] ================================================== +[2025-07-04 17:37:43] === 三元组预测示例 === +[2025-07-04 17:37:43] 样本1目标: Bülent Uzun place of birth Rize +[2025-07-04 17:37:43] 样本1预测: countryW aditonaere st�yas ativeistr ofmr S of +[2025-07-04 17:37:43] 样本2目标: Hippocrates occupation physician +[2025-07-04 17:37:43] 样本2预测: countryA aditonaeran stkyil ativeistr ofmr S of +[2025-07-04 17:37:43] ================== +[2025-07-04 17:37:43] Epoch 1/4, Step 5500/18020, Loss(triple): 10.221516, Loss(predicate): 14.944305, LR: 0.000153, Speed: 120580.46 tokens/sec | Epoch Time Left: 2:54:33 | Total Time Left: 15:28:17 +[2025-07-04 17:38:24] Epoch 1/4, Step 5550/18020, Loss(triple): 10.492153, Loss(predicate): 8.745341, LR: 0.000154, Speed: 121017.68 tokens/sec | Epoch Time Left: 2:53:48 | Total Time Left: 15:27:20 +[2025-07-04 17:39:04] Epoch 1/4, Step 5600/18020, Loss(triple): 10.460901, Loss(predicate): 10.011189, LR: 0.000155, Speed: 121047.54 tokens/sec | Epoch Time Left: 2:53:04 | Total Time Left: 15:26:24 +[2025-07-04 17:39:45] Epoch 1/4, Step 5650/18020, Loss(triple): 10.116673, Loss(predicate): 10.596375, LR: 0.000157, Speed: 120218.55 tokens/sec | Epoch Time Left: 2:52:20 | Total Time Left: 15:25:31 +[2025-07-04 17:40:26] Epoch 1/4, Step 5700/18020, Loss(triple): 9.967655, Loss(predicate): 7.076894, LR: 0.000158, Speed: 120068.33 tokens/sec | Epoch Time Left: 2:51:37 | Total Time Left: 15:24:40 +[2025-07-04 17:41:07] Epoch 1/4, Step 5750/18020, Loss(triple): 10.235773, Loss(predicate): 11.065231, LR: 0.000160, Speed: 121117.40 tokens/sec | Epoch Time Left: 2:50:52 | Total Time Left: 15:23:44 +[2025-07-04 17:41:47] Epoch 1/4, Step 5800/18020, Loss(triple): 10.298500, Loss(predicate): 13.692647, LR: 0.000161, Speed: 121083.07 tokens/sec | Epoch Time Left: 2:50:08 | Total Time Left: 15:22:49 +[2025-07-04 17:42:28] Epoch 1/4, Step 5850/18020, Loss(triple): 9.861221, Loss(predicate): 10.524312, LR: 0.000162, Speed: 119577.51 tokens/sec | Epoch Time Left: 2:49:25 | Total Time Left: 15:21:59 +[2025-07-04 17:43:09] Epoch 1/4, Step 5900/18020, Loss(triple): 9.823997, Loss(predicate): 11.631378, LR: 0.000164, Speed: 120546.46 tokens/sec | Epoch Time Left: 2:48:41 | Total Time Left: 15:21:06 +[2025-07-04 17:43:50] Epoch 1/4, Step 5950/18020, Loss(triple): 9.760220, Loss(predicate): 9.902832, LR: 0.000165, Speed: 120362.29 tokens/sec | Epoch Time Left: 2:47:57 | Total Time Left: 15:20:15 +[2025-07-04 17:44:30] === GPU性能分析 (平均每步) === +[2025-07-04 17:44:30] 前向传播: 50.32ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 17:44:30] GPU总时间: 52.27ms, 实际迭代时间: 809.77ms, GPU利用率: 6.5% +[2025-07-04 17:44:30] ================================================== +[2025-07-04 17:44:30] === 三元组预测示例 === +[2025-07-04 17:44:30] 样本1目标: Bruce North country Canada +[2025-07-04 17:44:30] 样本1预测: countryB aditonaaran iaad,as ativeistiz ofmr B of +[2025-07-04 17:44:30] 样本2目标: Saaristattus taxon rank species +[2025-07-04 17:44:30] 样本2预测: countryB aditonaaran umad,as ativeistiz ofmr B of +[2025-07-04 17:44:30] ================== +[2025-07-04 17:44:30] Epoch 1/4, Step 6000/18020, Loss(triple): 9.973810, Loss(predicate): 13.448375, LR: 0.000166, Speed: 121398.01 tokens/sec | Epoch Time Left: 2:47:13 | Total Time Left: 15:19:19 +[2025-07-04 17:45:11] Epoch 1/4, Step 6050/18020, Loss(triple): 10.363075, Loss(predicate): 10.635762, LR: 0.000168, Speed: 120667.83 tokens/sec | Epoch Time Left: 2:46:29 | Total Time Left: 15:18:26 +[2025-07-04 17:45:52] Epoch 1/4, Step 6100/18020, Loss(triple): 10.231865, Loss(predicate): 8.417500, LR: 0.000169, Speed: 119551.44 tokens/sec | Epoch Time Left: 2:45:46 | Total Time Left: 15:17:38 +[2025-07-04 17:46:33] Epoch 1/4, Step 6150/18020, Loss(triple): 9.583103, Loss(predicate): 9.016052, LR: 0.000171, Speed: 120642.39 tokens/sec | Epoch Time Left: 2:45:03 | Total Time Left: 15:16:46 +[2025-07-04 17:47:14] Epoch 1/4, Step 6200/18020, Loss(triple): 9.603821, Loss(predicate): 11.821355, LR: 0.000172, Speed: 121046.66 tokens/sec | Epoch Time Left: 2:44:19 | Total Time Left: 15:15:52 +[2025-07-04 17:47:54] Epoch 1/4, Step 6250/18020, Loss(triple): 10.033302, Loss(predicate): 12.067922, LR: 0.000173, Speed: 121015.29 tokens/sec | Epoch Time Left: 2:43:35 | Total Time Left: 15:14:59 +[2025-07-04 17:48:35] Epoch 1/4, Step 6300/18020, Loss(triple): 10.276566, Loss(predicate): 10.052612, LR: 0.000175, Speed: 120308.15 tokens/sec | Epoch Time Left: 2:42:52 | Total Time Left: 15:14:08 +[2025-07-04 17:49:16] Epoch 1/4, Step 6350/18020, Loss(triple): 10.069454, Loss(predicate): 9.968059, LR: 0.000176, Speed: 119610.37 tokens/sec | Epoch Time Left: 2:42:09 | Total Time Left: 15:13:21 +[2025-07-04 17:49:57] Epoch 1/4, Step 6400/18020, Loss(triple): 10.237320, Loss(predicate): 16.239542, LR: 0.000178, Speed: 121046.17 tokens/sec | Epoch Time Left: 2:41:25 | Total Time Left: 15:12:28 +[2025-07-04 17:50:37] Epoch 1/4, Step 6450/18020, Loss(triple): 9.901354, Loss(predicate): 10.547999, LR: 0.000179, Speed: 121017.72 tokens/sec | Epoch Time Left: 2:40:42 | Total Time Left: 15:11:35 +[2025-07-04 17:51:18] === GPU性能分析 (平均每步) === +[2025-07-04 17:51:18] 前向传播: 59.68ms, 损失计算: 0.02ms, 反向传播: 1.92ms, 优化器: 0.00ms +[2025-07-04 17:51:18] GPU总时间: 61.62ms, 实际迭代时间: 820.45ms, GPU利用率: 7.5% +[2025-07-04 17:51:18] ================================================== +[2025-07-04 17:51:18] === 三元组预测示例 === +[2025-07-04 17:51:18] 样本1目标: Lacey Township School District located in the administrative territorial entity New Jersey +[2025-07-04 17:51:18] 样本1预测: countryB adronaaran stie,om ativeanceiz ofmr B of +[2025-07-04 17:51:18] 样本2目标: Spang instance of village +[2025-07-04 17:51:18] 样本2预测: countryB aditonaaran stin,om ativeanceiz ofmr B of +[2025-07-04 17:51:18] ================== +[2025-07-04 17:51:18] Epoch 1/4, Step 6500/18020, Loss(triple): 10.270206, Loss(predicate): 18.829416, LR: 0.000180, Speed: 119817.01 tokens/sec | Epoch Time Left: 2:39:59 | Total Time Left: 15:10:47 +[2025-07-04 17:51:59] Epoch 1/4, Step 6550/18020, Loss(triple): 9.899246, Loss(predicate): 15.716624, LR: 0.000182, Speed: 120531.90 tokens/sec | Epoch Time Left: 2:39:16 | Total Time Left: 15:09:57 +[2025-07-04 17:52:40] Epoch 1/4, Step 6600/18020, Loss(triple): 10.114685, Loss(predicate): 10.694133, LR: 0.000183, Speed: 120621.52 tokens/sec | Epoch Time Left: 2:38:33 | Total Time Left: 15:09:06 +[2025-07-04 17:53:20] Epoch 1/4, Step 6650/18020, Loss(triple): 10.036005, Loss(predicate): 9.943257, LR: 0.000185, Speed: 121047.30 tokens/sec | Epoch Time Left: 2:37:49 | Total Time Left: 15:08:14 +[2025-07-04 17:54:01] Epoch 1/4, Step 6700/18020, Loss(triple): 9.882097, Loss(predicate): 11.535024, LR: 0.000186, Speed: 120788.83 tokens/sec | Epoch Time Left: 2:37:06 | Total Time Left: 15:07:23 +[2025-07-04 17:54:42] Epoch 1/4, Step 6750/18020, Loss(triple): 9.874130, Loss(predicate): 9.336141, LR: 0.000187, Speed: 119644.41 tokens/sec | Epoch Time Left: 2:36:23 | Total Time Left: 15:06:36 +[2025-07-04 17:55:23] Epoch 1/4, Step 6800/18020, Loss(triple): 10.434809, Loss(predicate): 11.201874, LR: 0.000189, Speed: 120108.35 tokens/sec | Epoch Time Left: 2:35:41 | Total Time Left: 15:05:48 +[2025-07-04 17:56:04] Epoch 1/4, Step 6850/18020, Loss(triple): 9.582499, Loss(predicate): 10.151944, LR: 0.000190, Speed: 121075.35 tokens/sec | Epoch Time Left: 2:34:57 | Total Time Left: 15:04:56 +[2025-07-04 17:56:44] Epoch 1/4, Step 6900/18020, Loss(triple): 9.761810, Loss(predicate): 7.500651, LR: 0.000191, Speed: 121174.08 tokens/sec | Epoch Time Left: 2:34:14 | Total Time Left: 15:04:04 +[2025-07-04 17:57:25] Epoch 1/4, Step 6950/18020, Loss(triple): 9.709805, Loss(predicate): 14.930949, LR: 0.000193, Speed: 120415.16 tokens/sec | Epoch Time Left: 2:33:31 | Total Time Left: 15:03:15 +[2025-07-04 17:58:06] === GPU性能分析 (平均每步) === +[2025-07-04 17:58:06] 前向传播: 58.02ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 17:58:06] GPU总时间: 59.97ms, 实际迭代时间: 821.79ms, GPU利用率: 7.3% +[2025-07-04 17:58:06] ================================================== +[2025-07-04 17:58:06] === 三元组预测示例 === +[2025-07-04 17:58:06] 样本1目标: Pol Theis date of birth February 10, 1968 +[2025-07-04 17:58:06] 样本1预测: countryB adronaran ia�,ug ialanceiz ofmax B of +[2025-07-04 17:58:06] 样本2目标: ward country Tanzania +[2025-07-04 17:58:06] 样本2预测: countryB adronaran osin,as ialanceiz ofmax B r +[2025-07-04 17:58:06] ================== +[2025-07-04 17:58:06] Epoch 1/4, Step 7000/18020, Loss(triple): 9.690159, Loss(predicate): 12.905477, LR: 0.000194, Speed: 119622.13 tokens/sec | Epoch Time Left: 2:32:49 | Total Time Left: 15:02:29 +[2025-07-04 17:58:47] Epoch 1/4, Step 7050/18020, Loss(triple): 9.551682, Loss(predicate): 13.167491, LR: 0.000196, Speed: 120546.78 tokens/sec | Epoch Time Left: 2:32:06 | Total Time Left: 15:01:40 +[2025-07-04 17:59:28] Epoch 1/4, Step 7100/18020, Loss(triple): 10.208935, Loss(predicate): 10.393825, LR: 0.000197, Speed: 121150.81 tokens/sec | Epoch Time Left: 2:31:23 | Total Time Left: 15:00:49 +[2025-07-04 18:00:08] Epoch 1/4, Step 7150/18020, Loss(triple): 9.867996, Loss(predicate): 7.808573, LR: 0.000198, Speed: 120529.45 tokens/sec | Epoch Time Left: 2:30:40 | Total Time Left: 15:00:00 +[2025-07-04 18:00:49] Epoch 1/4, Step 7200/18020, Loss(triple): 9.983902, Loss(predicate): 11.484883, LR: 0.000200, Speed: 120531.29 tokens/sec | Epoch Time Left: 2:29:57 | Total Time Left: 14:59:11 +[2025-07-04 18:01:30] Epoch 1/4, Step 7250/18020, Loss(triple): 9.712841, Loss(predicate): 7.862488, LR: 0.000200, Speed: 119858.05 tokens/sec | Epoch Time Left: 2:29:15 | Total Time Left: 14:58:24 +[2025-07-04 18:02:11] Epoch 1/4, Step 7300/18020, Loss(triple): 10.227303, Loss(predicate): 6.062093, LR: 0.000200, Speed: 120912.61 tokens/sec | Epoch Time Left: 2:28:32 | Total Time Left: 14:57:35 +[2025-07-04 18:02:51] Epoch 1/4, Step 7350/18020, Loss(triple): 9.806328, Loss(predicate): 6.199829, LR: 0.000200, Speed: 121051.83 tokens/sec | Epoch Time Left: 2:27:49 | Total Time Left: 14:56:45 +[2025-07-04 18:03:32] Epoch 1/4, Step 7400/18020, Loss(triple): 9.864021, Loss(predicate): 7.967885, LR: 0.000200, Speed: 119876.16 tokens/sec | Epoch Time Left: 2:27:06 | Total Time Left: 14:55:58 +[2025-07-04 18:04:13] Epoch 1/4, Step 7450/18020, Loss(triple): 9.659626, Loss(predicate): 8.590596, LR: 0.000200, Speed: 120859.16 tokens/sec | Epoch Time Left: 2:26:23 | Total Time Left: 14:55:09 +[2025-07-04 18:04:54] === GPU性能分析 (平均每步) === +[2025-07-04 18:04:54] 前向传播: 56.71ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 18:04:54] GPU总时间: 58.65ms, 实际迭代时间: 818.93ms, GPU利用率: 7.2% +[2025-07-04 18:04:54] ================================================== +[2025-07-04 18:04:54] === 三元组预测示例 === +[2025-07-04 18:04:54] 样本1目标: Sir George Byng occupation statesman +[2025-07-04 18:04:54] 样本1预测: countryK adoonaran itinyom ativeanceiz ofminal B A +[2025-07-04 18:04:54] 样本2目标: Cypress Hills Cemetery instance of cemetery +[2025-07-04 18:04:54] 样本2预测: countryB adoonaran iakyil ialanceiz ofminal B A +[2025-07-04 18:04:54] ================== +[2025-07-04 18:04:54] Epoch 1/4, Step 7500/18020, Loss(triple): 9.847832, Loss(predicate): 9.687408, LR: 0.000200, Speed: 120040.17 tokens/sec | Epoch Time Left: 2:25:41 | Total Time Left: 14:54:22 +[2025-07-04 18:05:35] Epoch 1/4, Step 7550/18020, Loss(triple): 9.657124, Loss(predicate): 10.669475, LR: 0.000200, Speed: 121160.84 tokens/sec | Epoch Time Left: 2:24:58 | Total Time Left: 14:53:32 +[2025-07-04 18:06:15] Epoch 1/4, Step 7600/18020, Loss(triple): 9.934193, Loss(predicate): 9.915466, LR: 0.000200, Speed: 120727.57 tokens/sec | Epoch Time Left: 2:24:15 | Total Time Left: 14:52:44 +[2025-07-04 18:06:56] Epoch 1/4, Step 7650/18020, Loss(triple): 9.840143, Loss(predicate): 9.997543, LR: 0.000200, Speed: 119640.00 tokens/sec | Epoch Time Left: 2:23:33 | Total Time Left: 14:51:58 +[2025-07-04 18:07:37] Epoch 1/4, Step 7700/18020, Loss(triple): 10.059061, Loss(predicate): 7.298025, LR: 0.000200, Speed: 120107.49 tokens/sec | Epoch Time Left: 2:22:51 | Total Time Left: 14:51:12 +[2025-07-04 18:08:18] Epoch 1/4, Step 7750/18020, Loss(triple): 9.645359, Loss(predicate): 8.234909, LR: 0.000200, Speed: 120988.96 tokens/sec | Epoch Time Left: 2:22:08 | Total Time Left: 14:50:23 +[2025-07-04 18:08:59] Epoch 1/4, Step 7800/18020, Loss(triple): 9.917780, Loss(predicate): 10.031860, LR: 0.000200, Speed: 121065.86 tokens/sec | Epoch Time Left: 2:21:26 | Total Time Left: 14:49:33 +[2025-07-04 18:09:39] Epoch 1/4, Step 7850/18020, Loss(triple): 9.878229, Loss(predicate): 12.149506, LR: 0.000200, Speed: 120346.11 tokens/sec | Epoch Time Left: 2:20:43 | Total Time Left: 14:48:46 +[2025-07-04 18:10:20] Epoch 1/4, Step 7900/18020, Loss(triple): 9.804041, Loss(predicate): 9.192668, LR: 0.000200, Speed: 119569.47 tokens/sec | Epoch Time Left: 2:20:01 | Total Time Left: 14:48:02 +[2025-07-04 18:11:01] Epoch 1/4, Step 7950/18020, Loss(triple): 9.559181, Loss(predicate): 9.996623, LR: 0.000200, Speed: 120540.41 tokens/sec | Epoch Time Left: 2:19:19 | Total Time Left: 14:47:14 +[2025-07-04 18:11:42] === GPU性能分析 (平均每步) === +[2025-07-04 18:11:42] 前向传播: 50.73ms, 损失计算: 0.02ms, 反向传播: 1.97ms, 优化器: 0.00ms +[2025-07-04 18:11:42] GPU总时间: 52.72ms, 实际迭代时间: 812.93ms, GPU利用率: 6.5% +[2025-07-04 18:11:42] ================================================== +[2025-07-04 18:11:42] === 三元组预测示例 === +[2025-07-04 18:11:42] 样本1目标: Eavestone instance of civil parish +[2025-07-04 18:11:42] 样本1预测: countryB adtonaran iain,as ialistiz ofmax B r +[2025-07-04 18:11:42] 样本2目标: Guillaume Durand (nephew) languages spoken, written or signed French +[2025-07-04 18:11:42] 样本2预测: countryB adoonaran iakyil ialitiz ofmax B r +[2025-07-04 18:11:42] ================== +[2025-07-04 18:11:42] Epoch 1/4, Step 8000/18020, Loss(triple): 10.065239, Loss(predicate): 13.941609, LR: 0.000200, Speed: 120925.74 tokens/sec | Epoch Time Left: 2:18:36 | Total Time Left: 14:46:26 +[2025-07-04 18:12:23] Epoch 1/4, Step 8050/18020, Loss(triple): 10.026138, Loss(predicate): 9.818950, LR: 0.000200, Speed: 120408.42 tokens/sec | Epoch Time Left: 2:17:54 | Total Time Left: 14:45:39 +[2025-07-04 18:13:03] Epoch 1/4, Step 8100/18020, Loss(triple): 9.924509, Loss(predicate): 8.684509, LR: 0.000200, Speed: 120606.71 tokens/sec | Epoch Time Left: 2:17:11 | Total Time Left: 14:44:52 +[2025-07-04 18:13:44] Epoch 1/4, Step 8150/18020, Loss(triple): 10.116272, Loss(predicate): 9.217610, LR: 0.000200, Speed: 119950.54 tokens/sec | Epoch Time Left: 2:16:29 | Total Time Left: 14:44:06 +[2025-07-04 18:14:25] Epoch 1/4, Step 8200/18020, Loss(triple): 9.879707, Loss(predicate): 7.179321, LR: 0.000200, Speed: 120541.71 tokens/sec | Epoch Time Left: 2:15:47 | Total Time Left: 14:43:19 +[2025-07-04 18:15:06] Epoch 1/4, Step 8250/18020, Loss(triple): 9.428799, Loss(predicate): 14.371358, LR: 0.000200, Speed: 120925.88 tokens/sec | Epoch Time Left: 2:15:04 | Total Time Left: 14:42:31 +[2025-07-04 18:15:47] Epoch 1/4, Step 8300/18020, Loss(triple): 9.914383, Loss(predicate): 10.364522, LR: 0.000200, Speed: 120380.75 tokens/sec | Epoch Time Left: 2:14:22 | Total Time Left: 14:41:45 +[2025-07-04 18:16:28] Epoch 1/4, Step 8350/18020, Loss(triple): 9.763493, Loss(predicate): 8.180043, LR: 0.000200, Speed: 120283.52 tokens/sec | Epoch Time Left: 2:13:40 | Total Time Left: 14:40:58 +[2025-07-04 18:17:09] Epoch 1/4, Step 8400/18020, Loss(triple): 9.659592, Loss(predicate): 10.473918, LR: 0.000200, Speed: 120090.81 tokens/sec | Epoch Time Left: 2:12:58 | Total Time Left: 14:40:13 +[2025-07-04 18:17:49] Epoch 1/4, Step 8450/18020, Loss(triple): 9.427729, Loss(predicate): 11.965709, LR: 0.000200, Speed: 120719.12 tokens/sec | Epoch Time Left: 2:12:16 | Total Time Left: 14:39:26 +[2025-07-04 18:18:30] === GPU性能分析 (平均每步) === +[2025-07-04 18:18:30] 前向传播: 50.38ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 18:18:30] GPU总时间: 52.33ms, 实际迭代时间: 811.74ms, GPU利用率: 6.4% +[2025-07-04 18:18:30] ================================================== +[2025-07-04 18:18:30] === 三元组预测示例 === +[2025-07-04 18:18:30] 样本1目标: Marievale Bird Sanctuary country South Africa +[2025-07-04 18:18:30] 样本1预测: countryM adoonaran iain,as ialistiz ofmax B r +[2025-07-04 18:18:30] 样本2目标: Sampan (film) director Terry Bourke +[2025-07-04 18:18:30] 样本2预测: countryM adoonaran osinusid ritanceiz ofmth M A +[2025-07-04 18:18:30] ================== +[2025-07-04 18:18:30] Epoch 1/4, Step 8500/18020, Loss(triple): 9.821871, Loss(predicate): 10.988770, LR: 0.000200, Speed: 121103.20 tokens/sec | Epoch Time Left: 2:11:33 | Total Time Left: 14:38:38 +[2025-07-04 18:19:11] Epoch 1/4, Step 8550/18020, Loss(triple): 9.768009, Loss(predicate): 12.120219, LR: 0.000200, Speed: 120137.49 tokens/sec | Epoch Time Left: 2:10:51 | Total Time Left: 14:37:52 +[2025-07-04 18:19:51] Epoch 1/4, Step 8600/18020, Loss(triple): 9.461374, Loss(predicate): 16.318155, LR: 0.000200, Speed: 120563.36 tokens/sec | Epoch Time Left: 2:10:09 | Total Time Left: 14:37:06 +[2025-07-04 18:20:32] Epoch 1/4, Step 8650/18020, Loss(triple): 9.969330, Loss(predicate): 11.650289, LR: 0.000200, Speed: 120540.26 tokens/sec | Epoch Time Left: 2:09:27 | Total Time Left: 14:36:19 +[2025-07-04 18:21:13] Epoch 1/4, Step 8700/18020, Loss(triple): 10.243027, Loss(predicate): 16.237732, LR: 0.000200, Speed: 121231.24 tokens/sec | Epoch Time Left: 2:08:44 | Total Time Left: 14:35:31 +[2025-07-04 18:21:53] Epoch 1/4, Step 8750/18020, Loss(triple): 9.521667, Loss(predicate): 9.933757, LR: 0.000200, Speed: 121267.01 tokens/sec | Epoch Time Left: 2:08:02 | Total Time Left: 14:34:43 +[2025-07-04 18:22:34] Epoch 1/4, Step 8800/18020, Loss(triple): 9.387680, Loss(predicate): 9.745463, LR: 0.000200, Speed: 120252.71 tokens/sec | Epoch Time Left: 2:07:20 | Total Time Left: 14:33:58 +[2025-07-04 18:23:15] Epoch 1/4, Step 8850/18020, Loss(triple): 9.938408, Loss(predicate): 12.367778, LR: 0.000200, Speed: 120638.06 tokens/sec | Epoch Time Left: 2:06:38 | Total Time Left: 14:33:11 +[2025-07-04 18:23:56] Epoch 1/4, Step 8900/18020, Loss(triple): 9.768990, Loss(predicate): 15.417287, LR: 0.000200, Speed: 120375.36 tokens/sec | Epoch Time Left: 2:05:56 | Total Time Left: 14:32:26 +[2025-07-04 18:24:36] Epoch 1/4, Step 8950/18020, Loss(triple): 9.583206, Loss(predicate): 5.179026, LR: 0.000200, Speed: 121518.65 tokens/sec | Epoch Time Left: 2:05:13 | Total Time Left: 14:31:37 +[2025-07-04 18:25:17] === GPU性能分析 (平均每步) === +[2025-07-04 18:25:17] 前向传播: 50.29ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 18:25:17] GPU总时间: 52.23ms, 实际迭代时间: 810.56ms, GPU利用率: 6.4% +[2025-07-04 18:25:17] ================================================== +[2025-07-04 18:25:17] === 三元组预测示例 === +[2025-07-04 18:25:17] 样本1目标: Shorn Cliff and Caswell Woods heritage designation SSSI +[2025-07-04 18:25:17] 样本1预测: countryP adiconaran ol�yil ativeanceiz ofmax M t +[2025-07-04 18:25:17] 样本2目标: SHOUTcast developer Nullsoft +[2025-07-04 18:25:17] 样本2预测: countryK adtonaran stkyil ativeitiz ofmth M A +[2025-07-04 18:25:17] ================== +[2025-07-04 18:25:17] Epoch 1/4, Step 9000/18020, Loss(triple): 10.200615, Loss(predicate): 12.519938, LR: 0.000200, Speed: 121279.34 tokens/sec | Epoch Time Left: 2:04:31 | Total Time Left: 14:30:50 +[2025-07-04 18:25:58] Epoch 1/4, Step 9050/18020, Loss(triple): 9.476456, Loss(predicate): 9.402608, LR: 0.000200, Speed: 119975.76 tokens/sec | Epoch Time Left: 2:03:49 | Total Time Left: 14:30:05 +[2025-07-04 18:26:38] Epoch 1/4, Step 9100/18020, Loss(triple): 9.792971, Loss(predicate): 11.886333, LR: 0.000200, Speed: 120701.84 tokens/sec | Epoch Time Left: 2:03:07 | Total Time Left: 14:29:19 +[2025-07-04 18:27:19] Epoch 1/4, Step 9150/18020, Loss(triple): 9.519321, Loss(predicate): 11.238667, LR: 0.000200, Speed: 120856.69 tokens/sec | Epoch Time Left: 2:02:25 | Total Time Left: 14:28:32 +[2025-07-04 18:28:00] Epoch 1/4, Step 9200/18020, Loss(triple): 9.934258, Loss(predicate): 8.380391, LR: 0.000200, Speed: 121242.02 tokens/sec | Epoch Time Left: 2:01:43 | Total Time Left: 14:27:45 +[2025-07-04 18:28:40] Epoch 1/4, Step 9250/18020, Loss(triple): 9.739586, Loss(predicate): 5.735209, LR: 0.000200, Speed: 120891.08 tokens/sec | Epoch Time Left: 2:01:00 | Total Time Left: 14:26:59 +[2025-07-04 18:29:21] Epoch 1/4, Step 9300/18020, Loss(triple): 9.383194, Loss(predicate): 6.955119, LR: 0.000199, Speed: 119967.62 tokens/sec | Epoch Time Left: 2:00:19 | Total Time Left: 14:26:14 +[2025-07-04 18:30:02] Epoch 1/4, Step 9350/18020, Loss(triple): 9.944231, Loss(predicate): 11.585887, LR: 0.000199, Speed: 120531.41 tokens/sec | Epoch Time Left: 1:59:37 | Total Time Left: 14:25:29 +[2025-07-04 18:30:43] Epoch 1/4, Step 9400/18020, Loss(triple): 9.931763, Loss(predicate): 10.766704, LR: 0.000199, Speed: 121270.41 tokens/sec | Epoch Time Left: 1:58:55 | Total Time Left: 14:24:42 +[2025-07-04 18:31:32] Epoch 1/4, Step 9450/18020, Loss(triple): 9.705708, Loss(predicate): 10.502869, LR: 0.000199, Speed: 100168.79 tokens/sec | Epoch Time Left: 1:58:20 | Total Time Left: 14:24:51 +[2025-07-04 18:32:13] === GPU性能分析 (平均每步) === +[2025-07-04 18:32:13] 前向传播: 59.45ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 18:32:13] GPU总时间: 61.39ms, 实际迭代时间: 832.86ms, GPU利用率: 7.4% +[2025-07-04 18:32:13] ================================================== +[2025-07-04 18:32:13] === 三元组预测示例 === +[2025-07-04 18:32:13] 样本1目标: Guelph country Canada +[2025-07-04 18:32:13] 样本1预测: countryW admmaoran allie,om ialistiz ofmax M t +[2025-07-04 18:32:13] 样本2目标: Robert Sadler occupation politician +[2025-07-04 18:32:13] 样本2预测: countryK admonaran allie,om ativeate3 ofmth M A +[2025-07-04 18:32:13] ================== +[2025-07-04 18:32:13] Epoch 1/4, Step 9500/18020, Loss(triple): 9.937683, Loss(predicate): 13.784953, LR: 0.000199, Speed: 118032.18 tokens/sec | Epoch Time Left: 1:57:39 | Total Time Left: 14:24:11 +[2025-07-04 18:32:54] Epoch 1/4, Step 9550/18020, Loss(triple): 9.039652, Loss(predicate): 7.524007, LR: 0.000199, Speed: 120143.18 tokens/sec | Epoch Time Left: 1:56:57 | Total Time Left: 14:23:27 +[2025-07-04 18:33:35] Epoch 1/4, Step 9600/18020, Loss(triple): 9.366877, Loss(predicate): 10.467204, LR: 0.000199, Speed: 119553.05 tokens/sec | Epoch Time Left: 1:56:15 | Total Time Left: 14:22:43 +[2025-07-04 18:34:16] Epoch 1/4, Step 9650/18020, Loss(triple): 9.371677, Loss(predicate): 14.415181, LR: 0.000199, Speed: 120191.05 tokens/sec | Epoch Time Left: 1:55:33 | Total Time Left: 14:21:58 +[2025-07-04 18:34:57] Epoch 1/4, Step 9700/18020, Loss(triple): 9.593929, Loss(predicate): 14.552510, LR: 0.000199, Speed: 119830.69 tokens/sec | Epoch Time Left: 1:54:52 | Total Time Left: 14:21:14 +[2025-07-04 18:35:38] Epoch 1/4, Step 9750/18020, Loss(triple): 9.865295, Loss(predicate): 8.613770, LR: 0.000199, Speed: 119266.53 tokens/sec | Epoch Time Left: 1:54:10 | Total Time Left: 14:20:31 +[2025-07-04 18:36:20] Epoch 1/4, Step 9800/18020, Loss(triple): 9.283487, Loss(predicate): 8.658508, LR: 0.000199, Speed: 118967.48 tokens/sec | Epoch Time Left: 1:53:29 | Total Time Left: 14:19:49 +[2025-07-04 18:37:01] Epoch 1/4, Step 9850/18020, Loss(triple): 9.561876, Loss(predicate): 11.518921, LR: 0.000199, Speed: 120099.57 tokens/sec | Epoch Time Left: 1:52:47 | Total Time Left: 14:19:05 +[2025-07-04 18:37:42] Epoch 1/4, Step 9900/18020, Loss(triple): 9.152651, Loss(predicate): 10.772522, LR: 0.000199, Speed: 120487.92 tokens/sec | Epoch Time Left: 1:52:05 | Total Time Left: 14:18:20 +[2025-07-04 18:38:23] Epoch 1/4, Step 9950/18020, Loss(triple): 9.775759, Loss(predicate): 8.695415, LR: 0.000199, Speed: 118965.72 tokens/sec | Epoch Time Left: 1:51:23 | Total Time Left: 14:17:38 +[2025-07-04 18:39:04] === GPU性能分析 (平均每步) === +[2025-07-04 18:39:04] 前向传播: 59.16ms, 损失计算: 0.02ms, 反向传播: 1.92ms, 优化器: 0.00ms +[2025-07-04 18:39:04] GPU总时间: 61.10ms, 实际迭代时间: 824.73ms, GPU利用率: 7.4% +[2025-07-04 18:39:04] ================================================== +[2025-07-04 18:39:04] === 三元组预测示例 === +[2025-07-04 18:39:04] 样本1目标: Applied Mechanics Reviews instance of scientific journal +[2025-07-04 18:39:04] 样本1预测: countryW birmmaeran opPore ativeateiz ofbth M F +[2025-07-04 18:39:04] 样本2目标: Min Shin Saw father Sithu I +[2025-07-04 18:39:04] 样本2预测: countryM admonaran ilkyas ativeisteg ofmth B t +[2025-07-04 18:39:04] ================== +[2025-07-04 18:39:04] Epoch 1/4, Step 10000/18020, Loss(triple): 9.607828, Loss(predicate): 10.054118, LR: 0.000199, Speed: 119195.41 tokens/sec | Epoch Time Left: 1:50:42 | Total Time Left: 14:16:55 +[2025-07-04 18:39:45] Epoch 1/4, Step 10050/18020, Loss(triple): 9.220058, Loss(predicate): 5.778463, LR: 0.000199, Speed: 120240.88 tokens/sec | Epoch Time Left: 1:50:00 | Total Time Left: 14:16:10 +[2025-07-04 18:40:26] Epoch 1/4, Step 10100/18020, Loss(triple): 9.433174, Loss(predicate): 14.427856, LR: 0.000199, Speed: 120624.36 tokens/sec | Epoch Time Left: 1:49:18 | Total Time Left: 14:15:25 +[2025-07-04 18:41:07] Epoch 1/4, Step 10150/18020, Loss(triple): 9.496689, Loss(predicate): 9.825917, LR: 0.000199, Speed: 119242.99 tokens/sec | Epoch Time Left: 1:48:36 | Total Time Left: 14:14:42 +[2025-07-04 18:41:48] Epoch 1/4, Step 10200/18020, Loss(triple): 9.632065, Loss(predicate): 7.746358, LR: 0.000199, Speed: 119644.92 tokens/sec | Epoch Time Left: 1:47:55 | Total Time Left: 14:13:59 +[2025-07-04 18:42:29] Epoch 1/4, Step 10250/18020, Loss(triple): 9.622002, Loss(predicate): 11.633769, LR: 0.000199, Speed: 120648.45 tokens/sec | Epoch Time Left: 1:47:13 | Total Time Left: 14:13:14 +[2025-07-04 18:43:10] Epoch 1/4, Step 10300/18020, Loss(triple): 9.879902, Loss(predicate): 10.101766, LR: 0.000199, Speed: 120226.46 tokens/sec | Epoch Time Left: 1:46:31 | Total Time Left: 14:12:29 +[2025-07-04 18:43:51] Epoch 1/4, Step 10350/18020, Loss(triple): 9.219288, Loss(predicate): 8.918030, LR: 0.000199, Speed: 119368.05 tokens/sec | Epoch Time Left: 1:45:50 | Total Time Left: 14:11:46 +[2025-07-04 18:44:32] Epoch 1/4, Step 10400/18020, Loss(triple): 9.310328, Loss(predicate): 15.123454, LR: 0.000199, Speed: 119567.11 tokens/sec | Epoch Time Left: 1:45:08 | Total Time Left: 14:11:03 +[2025-07-04 18:45:13] Epoch 1/4, Step 10450/18020, Loss(triple): 9.215229, Loss(predicate): 6.576437, LR: 0.000199, Speed: 120563.52 tokens/sec | Epoch Time Left: 1:44:26 | Total Time Left: 14:10:18 +[2025-07-04 18:45:53] === GPU性能分析 (平均每步) === +[2025-07-04 18:45:53] 前向传播: 51.81ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 18:45:53] GPU总时间: 53.75ms, 实际迭代时间: 815.32ms, GPU利用率: 6.6% +[2025-07-04 18:45:53] ================================================== +[2025-07-04 18:45:53] === 三元组预测示例 === +[2025-07-04 18:45:53] 样本1目标: Native American country USA +[2025-07-04 18:45:53] 样本1预测: countryK admmaoran itin,re ativeist species ofianres the P +[2025-07-04 18:45:53] 样本2目标: Setteri instance of village +[2025-07-04 18:45:53] 样本2预测: countryM admonaran zinusul ativeance species ofianax the t +[2025-07-04 18:45:53] ================== +[2025-07-04 18:45:53] Epoch 1/4, Step 10500/18020, Loss(triple): 9.380817, Loss(predicate): 12.881541, LR: 0.000199, Speed: 120570.50 tokens/sec | Epoch Time Left: 1:43:44 | Total Time Left: 14:09:33 +[2025-07-04 18:46:35] Epoch 1/4, Step 10550/18020, Loss(triple): 9.548615, Loss(predicate): 12.155701, LR: 0.000199, Speed: 119172.87 tokens/sec | Epoch Time Left: 1:43:03 | Total Time Left: 14:08:51 +[2025-07-04 18:47:16] Epoch 1/4, Step 10600/18020, Loss(triple): 9.402115, Loss(predicate): 12.618001, LR: 0.000199, Speed: 119887.44 tokens/sec | Epoch Time Left: 1:42:21 | Total Time Left: 14:08:07 +[2025-07-04 18:47:56] Epoch 1/4, Step 10650/18020, Loss(triple): 8.983889, Loss(predicate): 8.075459, LR: 0.000199, Speed: 120540.98 tokens/sec | Epoch Time Left: 1:41:39 | Total Time Left: 14:07:22 +[2025-07-04 18:48:37] Epoch 1/4, Step 10700/18020, Loss(triple): 9.781038, Loss(predicate): 11.332397, LR: 0.000199, Speed: 120992.83 tokens/sec | Epoch Time Left: 1:40:57 | Total Time Left: 14:06:37 +[2025-07-04 18:49:18] Epoch 1/4, Step 10750/18020, Loss(triple): 9.411701, Loss(predicate): 9.457530, LR: 0.000199, Speed: 119874.84 tokens/sec | Epoch Time Left: 1:40:16 | Total Time Left: 14:05:53 +[2025-07-04 18:49:59] Epoch 1/4, Step 10800/18020, Loss(triple): 9.604862, Loss(predicate): 13.134583, LR: 0.000198, Speed: 119091.24 tokens/sec | Epoch Time Left: 1:39:34 | Total Time Left: 14:05:11 +[2025-07-04 18:50:40] Epoch 1/4, Step 10850/18020, Loss(triple): 9.047724, Loss(predicate): 6.875570, LR: 0.000198, Speed: 119666.66 tokens/sec | Epoch Time Left: 1:38:53 | Total Time Left: 14:04:28 +[2025-07-04 18:51:21] Epoch 1/4, Step 10900/18020, Loss(triple): 9.800251, Loss(predicate): 7.655945, LR: 0.000198, Speed: 120294.90 tokens/sec | Epoch Time Left: 1:38:11 | Total Time Left: 14:03:44 +[2025-07-04 18:52:02] Epoch 1/4, Step 10950/18020, Loss(triple): 9.539423, Loss(predicate): 13.610107, LR: 0.000198, Speed: 119606.56 tokens/sec | Epoch Time Left: 1:37:29 | Total Time Left: 14:03:01 +[2025-07-04 18:52:43] === GPU性能分析 (平均每步) === +[2025-07-04 18:52:43] 前向传播: 60.58ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 18:52:43] GPU总时间: 62.52ms, 实际迭代时间: 819.64ms, GPU利用率: 7.6% +[2025-07-04 18:52:43] ================================================== +[2025-07-04 18:52:43] === 三元组预测示例 === +[2025-07-04 18:52:43] 样本1目标: USS Elfin (SP-965) operator United States Navy +[2025-07-04 18:52:43] 样本1预测: countryW entmmaran olinfre 6anceiz ofmth M A +[2025-07-04 18:52:43] 样本2目标: Rochelle Park School District located in the administrative territorial entity New Jersey +[2025-07-04 18:52:43] 样本2预测: countryW admmaran allC,om ialityater ofbrit the P +[2025-07-04 18:52:43] ================== +[2025-07-04 18:52:43] Epoch 1/4, Step 11000/18020, Loss(triple): 8.963814, Loss(predicate): 8.372498, LR: 0.000198, Speed: 119934.90 tokens/sec | Epoch Time Left: 1:36:48 | Total Time Left: 14:02:17 +[2025-07-04 18:53:24] Epoch 1/4, Step 11050/18020, Loss(triple): 9.341265, Loss(predicate): 14.434692, LR: 0.000198, Speed: 119593.59 tokens/sec | Epoch Time Left: 1:36:06 | Total Time Left: 14:01:34 +[2025-07-04 18:54:05] Epoch 1/4, Step 11100/18020, Loss(triple): 9.049828, Loss(predicate): 6.417521, LR: 0.000198, Speed: 120465.43 tokens/sec | Epoch Time Left: 1:35:25 | Total Time Left: 14:00:50 +[2025-07-04 18:54:46] Epoch 1/4, Step 11150/18020, Loss(triple): 9.260849, Loss(predicate): 7.415995, LR: 0.000198, Speed: 120081.25 tokens/sec | Epoch Time Left: 1:34:43 | Total Time Left: 14:00:06 +[2025-07-04 18:55:28] Epoch 1/4, Step 11200/18020, Loss(triple): 9.354084, Loss(predicate): 6.387472, LR: 0.000198, Speed: 118926.08 tokens/sec | Epoch Time Left: 1:34:02 | Total Time Left: 13:59:25 +[2025-07-04 18:56:09] Epoch 1/4, Step 11250/18020, Loss(triple): 9.107239, Loss(predicate): 9.502706, LR: 0.000198, Speed: 118963.38 tokens/sec | Epoch Time Left: 1:33:20 | Total Time Left: 13:58:43 +[2025-07-04 18:56:51] Epoch 1/4, Step 11300/18020, Loss(triple): 9.188892, Loss(predicate): 9.862285, LR: 0.000198, Speed: 115546.05 tokens/sec | Epoch Time Left: 1:32:40 | Total Time Left: 13:58:08 +[2025-07-04 18:57:41] Epoch 1/4, Step 11350/18020, Loss(triple): 9.627649, Loss(predicate): 9.340962, LR: 0.000198, Speed: 99952.09 tokens/sec | Epoch Time Left: 1:32:03 | Total Time Left: 13:58:08 +[2025-07-04 18:58:23] Epoch 1/4, Step 11400/18020, Loss(triple): 9.475908, Loss(predicate): 9.003448, LR: 0.000198, Speed: 117151.05 tokens/sec | Epoch Time Left: 1:31:22 | Total Time Left: 13:57:30 +[2025-07-04 18:59:04] Epoch 1/4, Step 11450/18020, Loss(triple): 9.090679, Loss(predicate): 12.635010, LR: 0.000198, Speed: 119199.66 tokens/sec | Epoch Time Left: 1:30:40 | Total Time Left: 13:56:48 +[2025-07-04 18:59:45] === GPU性能分析 (平均每步) === +[2025-07-04 18:59:45] 前向传播: 62.24ms, 损失计算: 0.02ms, 反向传播: 1.92ms, 优化器: 0.00ms +[2025-07-04 18:59:45] GPU总时间: 64.18ms, 实际迭代时间: 822.43ms, GPU利用率: 7.8% +[2025-07-04 18:59:45] ================================================== +[2025-07-04 18:59:45] === 三元组预测示例 === +[2025-07-04 18:59:45] 样本1目标: Barbara Lynn Ozen genre rhythm +[2025-07-04 18:59:45] 样本1预测: countryW biryhearan ilin,re ialistiz ofbrit B C +[2025-07-04 18:59:45] 样本2目标: Ramona Marquez place of birth London +[2025-07-04 18:59:45] 样本2预测: countryM entyharu il�yag 6ateiz ofbth M A +[2025-07-04 18:59:45] ================== +[2025-07-04 18:59:45] Epoch 1/4, Step 11500/18020, Loss(triple): 9.235483, Loss(predicate): 7.569560, LR: 0.000198, Speed: 119528.23 tokens/sec | Epoch Time Left: 1:29:59 | Total Time Left: 13:56:05 +[2025-07-04 19:00:26] Epoch 1/4, Step 11550/18020, Loss(triple): 9.047998, Loss(predicate): 9.189229, LR: 0.000198, Speed: 120116.21 tokens/sec | Epoch Time Left: 1:29:17 | Total Time Left: 13:55:21 +[2025-07-04 19:01:06] Epoch 1/4, Step 11600/18020, Loss(triple): 9.136301, Loss(predicate): 7.123617, LR: 0.000198, Speed: 120786.16 tokens/sec | Epoch Time Left: 1:28:35 | Total Time Left: 13:54:36 +[2025-07-04 19:01:47] Epoch 1/4, Step 11650/18020, Loss(triple): 9.018000, Loss(predicate): 8.029561, LR: 0.000198, Speed: 121340.31 tokens/sec | Epoch Time Left: 1:27:53 | Total Time Left: 13:53:50 +[2025-07-04 19:02:28] Epoch 1/4, Step 11700/18020, Loss(triple): 9.035583, Loss(predicate): 10.683330, LR: 0.000198, Speed: 120877.77 tokens/sec | Epoch Time Left: 1:27:11 | Total Time Left: 13:53:04 +[2025-07-04 19:03:09] Epoch 1/4, Step 11750/18020, Loss(triple): 9.061756, Loss(predicate): 6.973592, LR: 0.000198, Speed: 119585.49 tokens/sec | Epoch Time Left: 1:26:30 | Total Time Left: 13:52:22 +[2025-07-04 19:03:50] Epoch 1/4, Step 11800/18020, Loss(triple): 9.258392, Loss(predicate): 12.440145, LR: 0.000198, Speed: 120552.32 tokens/sec | Epoch Time Left: 1:25:48 | Total Time Left: 13:51:37 +[2025-07-04 19:04:30] Epoch 1/4, Step 11850/18020, Loss(triple): 9.055679, Loss(predicate): 10.732991, LR: 0.000197, Speed: 121202.74 tokens/sec | Epoch Time Left: 1:25:06 | Total Time Left: 13:50:51 +[2025-07-04 19:05:11] Epoch 1/4, Step 11900/18020, Loss(triple): 9.183863, Loss(predicate): 7.363892, LR: 0.000197, Speed: 120762.64 tokens/sec | Epoch Time Left: 1:24:25 | Total Time Left: 13:50:07 +[2025-07-04 19:05:52] Epoch 1/4, Step 11950/18020, Loss(triple): 9.288776, Loss(predicate): 11.410716, LR: 0.000197, Speed: 120676.17 tokens/sec | Epoch Time Left: 1:23:43 | Total Time Left: 13:49:22 +[2025-07-04 19:06:32] === GPU性能分析 (平均每步) === +[2025-07-04 19:06:32] 前向传播: 55.27ms, 损失计算: 0.02ms, 反向传播: 1.97ms, 优化器: 0.00ms +[2025-07-04 19:06:32] GPU总时间: 57.26ms, 实际迭代时间: 819.62ms, GPU利用率: 7.0% +[2025-07-04 19:06:32] ================================================== +[2025-07-04 19:06:32] === 三元组预测示例 === +[2025-07-04 19:06:33] 样本1目标: Embassy of Sweden, Prague country Czech Republic +[2025-07-04 19:06:33] 样本1预测: countryK admhaoru zia,ak ialist ter ofinarit the C +[2025-07-04 19:06:33] 样本2目标: Jakob Green Jensen country of citizenship Danish +[2025-07-04 19:06:33] 样本2预测: countryW biryheaeran all�,ah 6itiz ofbth M F +[2025-07-04 19:06:33] ================== +[2025-07-04 19:06:33] Epoch 1/4, Step 12000/18020, Loss(triple): 9.287348, Loss(predicate): 11.163696, LR: 0.000197, Speed: 119938.38 tokens/sec | Epoch Time Left: 1:23:01 | Total Time Left: 13:48:39 +[2025-07-04 19:07:13] Epoch 1/4, Step 12050/18020, Loss(triple): 8.940887, Loss(predicate): 10.625712, LR: 0.000197, Speed: 120577.35 tokens/sec | Epoch Time Left: 1:22:20 | Total Time Left: 13:47:54 +[2025-07-04 19:07:54] Epoch 1/4, Step 12100/18020, Loss(triple): 8.969437, Loss(predicate): 7.335083, LR: 0.000197, Speed: 121096.18 tokens/sec | Epoch Time Left: 1:21:38 | Total Time Left: 13:47:09 +[2025-07-04 19:08:35] Epoch 1/4, Step 12150/18020, Loss(triple): 8.893038, Loss(predicate): 10.795857, LR: 0.000197, Speed: 120494.76 tokens/sec | Epoch Time Left: 1:20:56 | Total Time Left: 13:46:25 +[2025-07-04 19:09:16] Epoch 1/4, Step 12200/18020, Loss(triple): 9.221699, Loss(predicate): 10.020020, LR: 0.000197, Speed: 120160.20 tokens/sec | Epoch Time Left: 1:20:15 | Total Time Left: 13:45:41 +[2025-07-04 19:09:56] Epoch 1/4, Step 12250/18020, Loss(triple): 9.346409, Loss(predicate): 8.396953, LR: 0.000197, Speed: 120205.14 tokens/sec | Epoch Time Left: 1:19:33 | Total Time Left: 13:44:57 +[2025-07-04 19:10:37] Epoch 1/4, Step 12300/18020, Loss(triple): 8.975611, Loss(predicate): 10.310283, LR: 0.000197, Speed: 121103.12 tokens/sec | Epoch Time Left: 1:18:51 | Total Time Left: 13:44:12 +[2025-07-04 19:11:18] Epoch 1/4, Step 12350/18020, Loss(triple): 9.218979, Loss(predicate): 13.260701, LR: 0.000197, Speed: 121145.52 tokens/sec | Epoch Time Left: 1:18:10 | Total Time Left: 13:43:27 +[2025-07-04 19:11:59] Epoch 1/4, Step 12400/18020, Loss(triple): 9.304636, Loss(predicate): 9.372151, LR: 0.000197, Speed: 119692.73 tokens/sec | Epoch Time Left: 1:17:28 | Total Time Left: 13:42:44 +[2025-07-04 19:12:39] Epoch 1/4, Step 12450/18020, Loss(triple): 9.142529, Loss(predicate): 19.259094, LR: 0.000197, Speed: 120388.91 tokens/sec | Epoch Time Left: 1:16:47 | Total Time Left: 13:42:00 +[2025-07-04 19:13:20] === GPU性能分析 (平均每步) === +[2025-07-04 19:13:20] 前向传播: 54.49ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 19:13:20] GPU总时间: 56.43ms, 实际迭代时间: 812.11ms, GPU利用率: 6.9% +[2025-07-04 19:13:20] ================================================== +[2025-07-04 19:13:20] === 三元组预测示例 === +[2025-07-04 19:13:20] 样本1目标: Vadim Karpov sport football +[2025-07-04 19:13:20] 样本1预测: countryK adyharu ilkyak umorthelersbert Ber +[2025-07-04 19:13:20] 样本2目标: Irundialba waorani taxon rank species +[2025-07-04 19:13:20] 样本2预测: countryP admonare imin,as ritiss species rankax species t +[2025-07-04 19:13:20] ================== +[2025-07-04 19:13:20] Epoch 1/4, Step 12500/18020, Loss(triple): 9.625343, Loss(predicate): 14.778534, LR: 0.000197, Speed: 121047.81 tokens/sec | Epoch Time Left: 1:16:05 | Total Time Left: 13:41:15 +[2025-07-04 19:14:01] Epoch 1/4, Step 12550/18020, Loss(triple): 9.652880, Loss(predicate): 9.354634, LR: 0.000197, Speed: 121091.84 tokens/sec | Epoch Time Left: 1:15:23 | Total Time Left: 13:40:30 +[2025-07-04 19:14:42] Epoch 1/4, Step 12600/18020, Loss(triple): 9.175247, Loss(predicate): 15.741506, LR: 0.000197, Speed: 120191.32 tokens/sec | Epoch Time Left: 1:14:42 | Total Time Left: 13:39:47 +[2025-07-04 19:15:23] Epoch 1/4, Step 12650/18020, Loss(triple): 9.054499, Loss(predicate): 7.423116, LR: 0.000197, Speed: 119494.22 tokens/sec | Epoch Time Left: 1:14:00 | Total Time Left: 13:39:05 +[2025-07-04 19:16:03] Epoch 1/4, Step 12700/18020, Loss(triple): 9.373735, Loss(predicate): 9.339559, LR: 0.000196, Speed: 120729.35 tokens/sec | Epoch Time Left: 1:13:19 | Total Time Left: 13:38:20 +[2025-07-04 19:16:44] Epoch 1/4, Step 12750/18020, Loss(triple): 9.369547, Loss(predicate): 10.522552, LR: 0.000196, Speed: 121011.39 tokens/sec | Epoch Time Left: 1:12:37 | Total Time Left: 13:37:35 +[2025-07-04 19:17:25] Epoch 1/4, Step 12800/18020, Loss(triple): 9.004053, Loss(predicate): 15.008301, LR: 0.000196, Speed: 120087.74 tokens/sec | Epoch Time Left: 1:11:55 | Total Time Left: 13:36:52 +[2025-07-04 19:18:06] Epoch 1/4, Step 12850/18020, Loss(triple): 9.295528, Loss(predicate): 10.106934, LR: 0.000196, Speed: 120930.31 tokens/sec | Epoch Time Left: 1:11:14 | Total Time Left: 13:36:08 +[2025-07-04 19:18:54] Epoch 1/4, Step 12900/18020, Loss(triple): 9.031687, Loss(predicate): 10.317444, LR: 0.000196, Speed: 100826.75 tokens/sec | Epoch Time Left: 1:10:35 | Total Time Left: 13:36:00 +[2025-07-04 19:19:41] Epoch 1/4, Step 12950/18020, Loss(triple): 9.211246, Loss(predicate): 10.439392, LR: 0.000196, Speed: 104300.87 tokens/sec | Epoch Time Left: 1:09:56 | Total Time Left: 13:35:45 +[2025-07-04 19:20:24] === GPU性能分析 (平均每步) === +[2025-07-04 19:20:24] 前向传播: 54.68ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 19:20:24] GPU总时间: 56.63ms, 实际迭代时间: 850.53ms, GPU利用率: 6.7% +[2025-07-04 19:20:24] ================================================== +[2025-07-04 19:20:24] === 三元组预测示例 === +[2025-07-04 19:20:24] 样本1目标: Fukushima contains the administrative territorial entity Onahama +[2025-07-04 19:20:24] 样本1预测: countryK admharan ardiayom ialist ter ofancerit B C +[2025-07-04 19:20:24] 样本2目标: Frederic Muspratt occupation chemist +[2025-07-04 19:20:24] 样本2预测: countryK adyharan oninous 6itiz of cth M A +[2025-07-04 19:20:24] ================== +[2025-07-04 19:20:24] Epoch 1/4, Step 13000/18020, Loss(triple): 9.274521, Loss(predicate): 10.593221, LR: 0.000196, Speed: 115579.45 tokens/sec | Epoch Time Left: 1:09:15 | Total Time Left: 13:35:09 +[2025-07-04 19:21:05] Epoch 1/4, Step 13050/18020, Loss(triple): 9.093704, Loss(predicate): 9.126892, LR: 0.000196, Speed: 120637.71 tokens/sec | Epoch Time Left: 1:08:34 | Total Time Left: 13:34:25 +[2025-07-04 19:21:46] Epoch 1/4, Step 13100/18020, Loss(triple): 9.100220, Loss(predicate): 7.473022, LR: 0.000196, Speed: 119244.13 tokens/sec | Epoch Time Left: 1:07:52 | Total Time Left: 13:33:43 +[2025-07-04 19:22:27] Epoch 1/4, Step 13150/18020, Loss(triple): 9.108910, Loss(predicate): 8.327504, LR: 0.000196, Speed: 120413.26 tokens/sec | Epoch Time Left: 1:07:11 | Total Time Left: 13:32:59 +[2025-07-04 19:23:08] Epoch 1/4, Step 13200/18020, Loss(triple): 9.564722, Loss(predicate): 11.673055, LR: 0.000196, Speed: 120561.70 tokens/sec | Epoch Time Left: 1:06:29 | Total Time Left: 13:32:15 +[2025-07-04 19:23:49] Epoch 1/4, Step 13250/18020, Loss(triple): 9.126728, Loss(predicate): 11.749084, LR: 0.000196, Speed: 119281.75 tokens/sec | Epoch Time Left: 1:05:48 | Total Time Left: 13:31:32 +[2025-07-04 19:24:34] Epoch 1/4, Step 13300/18020, Loss(triple): 9.076303, Loss(predicate): 11.469014, LR: 0.000196, Speed: 109355.72 tokens/sec | Epoch Time Left: 1:05:07 | Total Time Left: 13:31:07 +[2025-07-04 19:25:23] Epoch 1/4, Step 13350/18020, Loss(triple): 8.868416, Loss(predicate): 9.458130, LR: 0.000196, Speed: 99448.12 tokens/sec | Epoch Time Left: 1:04:29 | Total Time Left: 13:31:01 +[2025-07-04 19:26:11] Epoch 1/4, Step 13400/18020, Loss(triple): 9.127522, Loss(predicate): 13.377757, LR: 0.000196, Speed: 102812.64 tokens/sec | Epoch Time Left: 1:03:50 | Total Time Left: 13:30:47 +[2025-07-04 19:27:03] Epoch 1/4, Step 13450/18020, Loss(triple): 9.007828, Loss(predicate): 5.621439, LR: 0.000195, Speed: 94816.81 tokens/sec | Epoch Time Left: 1:03:12 | Total Time Left: 13:30:51 +[2025-07-04 19:27:47] === GPU性能分析 (平均每步) === +[2025-07-04 19:27:47] 前向传播: 53.20ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 19:27:47] GPU总时间: 55.14ms, 实际迭代时间: 885.90ms, GPU利用率: 6.2% +[2025-07-04 19:27:47] ================================================== +[2025-07-04 19:27:47] === 三元组预测示例 === +[2025-07-04 19:27:47] 样本1目标: Honfleur instance of commune +[2025-07-04 19:27:47] 样本1预测: countryT entyonard eninoor ativeance ter ofanceax- t +[2025-07-04 19:27:47] 样本2目标: Mount Hope (Eternity Range) continent Antarctica +[2025-07-04 19:27:47] 样本2预测: countryP adyharan enin,us iality ter ofancerit the A +[2025-07-04 19:27:47] ================== +[2025-07-04 19:27:47] Epoch 1/4, Step 13500/18020, Loss(triple): 8.805393, Loss(predicate): 13.013733, LR: 0.000195, Speed: 110965.29 tokens/sec | Epoch Time Left: 1:02:31 | Total Time Left: 13:30:22 +[2025-07-04 19:28:29] Epoch 1/4, Step 13550/18020, Loss(triple): 9.543716, Loss(predicate): 10.449219, LR: 0.000195, Speed: 117580.65 tokens/sec | Epoch Time Left: 1:01:50 | Total Time Left: 13:29:41 +[2025-07-04 19:29:10] Epoch 1/4, Step 13600/18020, Loss(triple): 8.859369, Loss(predicate): 11.318237, LR: 0.000195, Speed: 120475.73 tokens/sec | Epoch Time Left: 1:01:08 | Total Time Left: 13:28:57 +[2025-07-04 19:29:50] Epoch 1/4, Step 13650/18020, Loss(triple): 8.824024, Loss(predicate): 8.941508, LR: 0.000195, Speed: 120428.24 tokens/sec | Epoch Time Left: 1:00:26 | Total Time Left: 13:28:12 +[2025-07-04 19:30:32] Epoch 1/4, Step 13700/18020, Loss(triple): 8.890594, Loss(predicate): 7.589701, LR: 0.000195, Speed: 118963.55 tokens/sec | Epoch Time Left: 0:59:45 | Total Time Left: 13:27:30 +[2025-07-04 19:31:13] Epoch 1/4, Step 13750/18020, Loss(triple): 8.988419, Loss(predicate): 7.946635, LR: 0.000195, Speed: 120321.30 tokens/sec | Epoch Time Left: 0:59:03 | Total Time Left: 13:26:46 +[2025-07-04 19:31:54] Epoch 1/4, Step 13800/18020, Loss(triple): 9.125483, Loss(predicate): 9.198608, LR: 0.000195, Speed: 119805.88 tokens/sec | Epoch Time Left: 0:58:21 | Total Time Left: 13:26:03 +[2025-07-04 19:32:35] Epoch 1/4, Step 13850/18020, Loss(triple): 9.145077, Loss(predicate): 8.104472, LR: 0.000195, Speed: 120358.80 tokens/sec | Epoch Time Left: 0:57:40 | Total Time Left: 13:25:18 +[2025-07-04 19:33:15] Epoch 1/4, Step 13900/18020, Loss(triple): 8.794399, Loss(predicate): 8.265808, LR: 0.000195, Speed: 120572.09 tokens/sec | Epoch Time Left: 0:56:58 | Total Time Left: 13:24:34 +[2025-07-04 19:33:57] Epoch 1/4, Step 13950/18020, Loss(triple): 9.083780, Loss(predicate): 11.230550, LR: 0.000195, Speed: 118967.10 tokens/sec | Epoch Time Left: 0:56:16 | Total Time Left: 13:23:52 +[2025-07-04 19:34:37] === GPU性能分析 (平均每步) === +[2025-07-04 19:34:37] 前向传播: 61.34ms, 损失计算: 0.02ms, 反向传播: 1.91ms, 优化器: 0.00ms +[2025-07-04 19:34:37] GPU总时间: 63.26ms, 实际迭代时间: 816.19ms, GPU利用率: 7.8% +[2025-07-04 19:34:37] ================================================== +[2025-07-04 19:34:37] === 三元组预测示例 === +[2025-07-04 19:34:37] 样本1目标: 2003 contest participant Rita Guerra +[2025-07-04 19:34:37] 样本1预测: countryF entyharan onCoel 7anceiz ofbth B A +[2025-07-04 19:34:37] 样本2目标: Shian pari instance of village +[2025-07-04 19:34:37] 样本2预测: SS adyonare ol�kak iality terinancerit in C +[2025-07-04 19:34:37] ================== +[2025-07-04 19:34:37] Epoch 1/4, Step 14000/18020, Loss(triple): 9.171421, Loss(predicate): 13.866048, LR: 0.000195, Speed: 120442.75 tokens/sec | Epoch Time Left: 0:55:35 | Total Time Left: 13:23:07 +[2025-07-04 19:35:19] Epoch 1/4, Step 14050/18020, Loss(triple): 9.207249, Loss(predicate): 10.190562, LR: 0.000195, Speed: 119558.46 tokens/sec | Epoch Time Left: 0:54:53 | Total Time Left: 13:22:24 +[2025-07-04 19:35:59] Epoch 1/4, Step 14100/18020, Loss(triple): 8.911564, Loss(predicate): 9.924602, LR: 0.000194, Speed: 120454.01 tokens/sec | Epoch Time Left: 0:54:12 | Total Time Left: 13:21:40 +[2025-07-04 19:36:40] Epoch 1/4, Step 14150/18020, Loss(triple): 10.023376, Loss(predicate): 8.672536, LR: 0.000194, Speed: 120751.25 tokens/sec | Epoch Time Left: 0:53:30 | Total Time Left: 13:20:55 +[2025-07-04 19:37:21] Epoch 1/4, Step 14200/18020, Loss(triple): 8.725475, Loss(predicate): 7.015543, LR: 0.000194, Speed: 119597.18 tokens/sec | Epoch Time Left: 0:52:48 | Total Time Left: 13:20:12 +[2025-07-04 19:38:02] Epoch 1/4, Step 14250/18020, Loss(triple): 9.711849, Loss(predicate): 10.321095, LR: 0.000194, Speed: 120276.94 tokens/sec | Epoch Time Left: 0:52:07 | Total Time Left: 13:19:28 +[2025-07-04 19:38:43] Epoch 1/4, Step 14300/18020, Loss(triple): 8.774158, Loss(predicate): 10.954997, LR: 0.000194, Speed: 119692.84 tokens/sec | Epoch Time Left: 0:51:25 | Total Time Left: 13:18:45 +[2025-07-04 19:39:24] Epoch 1/4, Step 14350/18020, Loss(triple): 8.533970, Loss(predicate): 6.812449, LR: 0.000194, Speed: 120176.36 tokens/sec | Epoch Time Left: 0:50:43 | Total Time Left: 13:18:01 +[2025-07-04 19:40:05] Epoch 1/4, Step 14400/18020, Loss(triple): 9.384430, Loss(predicate): 9.761333, LR: 0.000194, Speed: 120300.62 tokens/sec | Epoch Time Left: 0:50:02 | Total Time Left: 13:17:17 +[2025-07-04 19:40:46] Epoch 1/4, Step 14450/18020, Loss(triple): 8.942486, Loss(predicate): 9.750671, LR: 0.000194, Speed: 119097.20 tokens/sec | Epoch Time Left: 0:49:20 | Total Time Left: 13:16:35 +[2025-07-04 19:41:27] === GPU性能分析 (平均每步) === +[2025-07-04 19:41:27] 前向传播: 61.39ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 19:41:27] GPU总时间: 63.33ms, 实际迭代时间: 818.08ms, GPU利用率: 7.7% +[2025-07-04 19:41:27] ================================================== +[2025-07-04 19:41:27] === 三元组预测示例 === +[2025-07-04 19:41:27] 样本1目标: United States contains the administrative territorial entity California +[2025-07-04 19:41:27] 样本1预测: countryW entyhlran itS,us ialist ter ofbrit the C +[2025-07-04 19:41:27] 样本2目标: Baños Canton instance of canton +[2025-07-04 19:41:27] 样本2预测: countryB adtharu ian�,ah ialist ter ofancerit in C +[2025-07-04 19:41:27] ================== +[2025-07-04 19:41:27] Epoch 1/4, Step 14500/18020, Loss(triple): 9.167391, Loss(predicate): 5.639760, LR: 0.000194, Speed: 120164.59 tokens/sec | Epoch Time Left: 0:48:39 | Total Time Left: 13:15:52 +[2025-07-04 19:42:08] Epoch 1/4, Step 14550/18020, Loss(triple): 9.225260, Loss(predicate): 7.526276, LR: 0.000194, Speed: 119402.71 tokens/sec | Epoch Time Left: 0:47:57 | Total Time Left: 13:15:09 +[2025-07-04 19:42:49] Epoch 1/4, Step 14600/18020, Loss(triple): 8.883280, Loss(predicate): 12.222107, LR: 0.000194, Speed: 120513.05 tokens/sec | Epoch Time Left: 0:47:16 | Total Time Left: 13:14:25 +[2025-07-04 19:43:30] Epoch 1/4, Step 14650/18020, Loss(triple): 8.994547, Loss(predicate): 15.026917, LR: 0.000194, Speed: 120468.46 tokens/sec | Epoch Time Left: 0:46:34 | Total Time Left: 13:13:41 +[2025-07-04 19:44:11] Epoch 1/4, Step 14700/18020, Loss(triple): 9.127834, Loss(predicate): 8.994029, LR: 0.000193, Speed: 118713.97 tokens/sec | Epoch Time Left: 0:45:52 | Total Time Left: 13:12:59 +[2025-07-04 19:44:52] Epoch 1/4, Step 14750/18020, Loss(triple): 8.925653, Loss(predicate): 9.893585, LR: 0.000193, Speed: 119962.85 tokens/sec | Epoch Time Left: 0:45:11 | Total Time Left: 13:12:16 +[2025-07-04 19:45:33] Epoch 1/4, Step 14800/18020, Loss(triple): 8.867905, Loss(predicate): 10.836772, LR: 0.000193, Speed: 120049.57 tokens/sec | Epoch Time Left: 0:44:29 | Total Time Left: 13:11:32 +[2025-07-04 19:46:14] Epoch 1/4, Step 14850/18020, Loss(triple): 9.089184, Loss(predicate): 10.053670, LR: 0.000193, Speed: 120454.30 tokens/sec | Epoch Time Left: 0:43:48 | Total Time Left: 13:10:48 +[2025-07-04 19:46:55] Epoch 1/4, Step 14900/18020, Loss(triple): 8.909328, Loss(predicate): 9.668091, LR: 0.000193, Speed: 120534.56 tokens/sec | Epoch Time Left: 0:43:06 | Total Time Left: 13:10:04 +[2025-07-04 19:47:36] Epoch 1/4, Step 14950/18020, Loss(triple): 8.424021, Loss(predicate): 9.288045, LR: 0.000193, Speed: 118351.18 tokens/sec | Epoch Time Left: 0:42:25 | Total Time Left: 13:09:23 +[2025-07-04 19:48:17] === GPU性能分析 (平均每步) === +[2025-07-04 19:48:17] 前向传播: 65.41ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 19:48:17] GPU总时间: 67.36ms, 实际迭代时间: 824.80ms, GPU利用率: 8.2% +[2025-07-04 19:48:17] ================================================== +[2025-07-04 19:48:17] === 三元组预测示例 === +[2025-07-04 19:48:17] 样本1目标: Bolivia shares border with Peru +[2025-07-04 19:48:17] 样本1预测: SB adyonaala ilisusul riton species rankaxi t +[2025-07-04 19:48:17] 样本2目标: Joseph C. Strasser military branch United States Navy +[2025-07-04 19:48:17] 样本2预测: countryK entyhiran all�,ag 4ateiz of cth M A +[2025-07-04 19:48:17] ================== +[2025-07-04 19:48:17] Epoch 1/4, Step 15000/18020, Loss(triple): 8.930805, Loss(predicate): 11.765819, LR: 0.000193, Speed: 119185.58 tokens/sec | Epoch Time Left: 0:41:43 | Total Time Left: 13:08:41 +[2025-07-04 19:48:58] Epoch 1/4, Step 15050/18020, Loss(triple): 9.137657, Loss(predicate): 10.223775, LR: 0.000193, Speed: 119778.69 tokens/sec | Epoch Time Left: 0:41:02 | Total Time Left: 13:07:58 +[2025-07-04 19:49:39] Epoch 1/4, Step 15100/18020, Loss(triple): 8.702028, Loss(predicate): 9.425923, LR: 0.000193, Speed: 119930.10 tokens/sec | Epoch Time Left: 0:40:20 | Total Time Left: 13:07:15 +[2025-07-04 19:50:20] Epoch 1/4, Step 15150/18020, Loss(triple): 8.929050, Loss(predicate): 10.109589, LR: 0.000193, Speed: 119872.91 tokens/sec | Epoch Time Left: 0:39:39 | Total Time Left: 13:06:31 +[2025-07-04 19:51:02] Epoch 1/4, Step 15200/18020, Loss(triple): 9.213692, Loss(predicate): 11.569946, LR: 0.000193, Speed: 118326.49 tokens/sec | Epoch Time Left: 0:38:57 | Total Time Left: 13:05:50 +[2025-07-04 19:51:43] Epoch 1/4, Step 15250/18020, Loss(triple): 8.931734, Loss(predicate): 7.402669, LR: 0.000193, Speed: 119846.34 tokens/sec | Epoch Time Left: 0:38:16 | Total Time Left: 13:05:07 +[2025-07-04 19:52:24] Epoch 1/4, Step 15300/18020, Loss(triple): 8.456211, Loss(predicate): 9.763997, LR: 0.000192, Speed: 120082.41 tokens/sec | Epoch Time Left: 0:37:34 | Total Time Left: 13:04:24 +[2025-07-04 19:53:05] Epoch 1/4, Step 15350/18020, Loss(triple): 9.101830, Loss(predicate): 8.342977, LR: 0.000192, Speed: 119284.78 tokens/sec | Epoch Time Left: 0:36:53 | Total Time Left: 13:03:42 +[2025-07-04 19:53:46] Epoch 1/4, Step 15400/18020, Loss(triple): 8.652719, Loss(predicate): 11.837178, LR: 0.000192, Speed: 120512.68 tokens/sec | Epoch Time Left: 0:36:11 | Total Time Left: 13:02:58 +[2025-07-04 19:54:27] Epoch 1/4, Step 15450/18020, Loss(triple): 9.082726, Loss(predicate): 12.033386, LR: 0.000192, Speed: 119007.89 tokens/sec | Epoch Time Left: 0:35:30 | Total Time Left: 13:02:16 +[2025-07-04 19:55:08] === GPU性能分析 (平均每步) === +[2025-07-04 19:55:08] 前向传播: 58.52ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 19:55:08] GPU总时间: 60.47ms, 实际迭代时间: 818.11ms, GPU利用率: 7.4% +[2025-07-04 19:55:08] ================================================== +[2025-07-04 19:55:08] === 三元组预测示例 === +[2025-07-04 19:55:08] 样本1目标: The Cement Garden author Ian McEwan +[2025-07-04 19:55:08] 样本1预测: countryIn enty 200l. L enm,r ialist ter ofiaat B A +[2025-07-04 19:55:08] 样本2目标: Son of Aladdin original language of film or TV show English +[2025-07-04 19:55:08] 样本2预测: countryS entyhara enador ialance ter ofmrit B A +[2025-07-04 19:55:08] ================== +[2025-07-04 19:55:08] Epoch 1/4, Step 15500/18020, Loss(triple): 8.993525, Loss(predicate): 16.260956, LR: 0.000192, Speed: 120159.73 tokens/sec | Epoch Time Left: 0:34:48 | Total Time Left: 13:01:32 +[2025-07-04 19:55:49] Epoch 1/4, Step 15550/18020, Loss(triple): 8.261044, Loss(predicate): 7.806325, LR: 0.000192, Speed: 120499.58 tokens/sec | Epoch Time Left: 0:34:07 | Total Time Left: 13:00:49 +[2025-07-04 19:56:30] Epoch 1/4, Step 15600/18020, Loss(triple): 8.802261, Loss(predicate): 15.593547, LR: 0.000192, Speed: 119132.52 tokens/sec | Epoch Time Left: 0:33:25 | Total Time Left: 13:00:06 +[2025-07-04 19:57:11] Epoch 1/4, Step 15650/18020, Loss(triple): 8.930307, Loss(predicate): 4.596899, LR: 0.000192, Speed: 119753.56 tokens/sec | Epoch Time Left: 0:32:44 | Total Time Left: 12:59:24 +[2025-07-04 19:57:52] Epoch 1/4, Step 15700/18020, Loss(triple): 8.862690, Loss(predicate): 7.271627, LR: 0.000192, Speed: 119245.75 tokens/sec | Epoch Time Left: 0:32:02 | Total Time Left: 12:58:41 +[2025-07-04 19:58:34] Epoch 1/4, Step 15750/18020, Loss(triple): 8.936640, Loss(predicate): 12.181294, LR: 0.000192, Speed: 118864.67 tokens/sec | Epoch Time Left: 0:31:21 | Total Time Left: 12:58:00 +[2025-07-04 19:59:21] Epoch 1/4, Step 15800/18020, Loss(triple): 9.039698, Loss(predicate): 8.965363, LR: 0.000191, Speed: 103145.01 tokens/sec | Epoch Time Left: 0:30:40 | Total Time Left: 12:57:40 +[2025-07-04 20:00:07] Epoch 1/4, Step 15850/18020, Loss(triple): 8.838381, Loss(predicate): 7.877991, LR: 0.000191, Speed: 107331.63 tokens/sec | Epoch Time Left: 0:29:59 | Total Time Left: 12:57:14 +[2025-07-04 20:00:50] Epoch 1/4, Step 15900/18020, Loss(triple): 8.518801, Loss(predicate): 9.433034, LR: 0.000191, Speed: 115716.70 tokens/sec | Epoch Time Left: 0:29:18 | Total Time Left: 12:56:36 +[2025-07-04 20:01:38] Epoch 1/4, Step 15950/18020, Loss(triple): 8.706264, Loss(predicate): 10.768474, LR: 0.000191, Speed: 102195.94 tokens/sec | Epoch Time Left: 0:28:37 | Total Time Left: 12:56:18 +[2025-07-04 20:02:20] === GPU性能分析 (平均每步) === +[2025-07-04 20:02:20] 前向传播: 68.21ms, 损失计算: 0.02ms, 反向传播: 1.96ms, 优化器: 0.00ms +[2025-07-04 20:02:20] GPU总时间: 70.18ms, 实际迭代时间: 834.33ms, GPU利用率: 8.4% +[2025-07-04 20:02:20] ================================================== +[2025-07-04 20:02:20] === 三元组预测示例 === +[2025-07-04 20:02:20] 样本1目标: Brazil lowest point Atlantic Ocean +[2025-07-04 20:02:20] 样本1预测: countryF adyonara inilyus ritist species ofiaass in t +[2025-07-04 20:02:20] 样本2目标: Ținutul Suceava capital Cernăuți +[2025-07-04 20:02:20] 样本2预测: countryK adyharu ziakah upationci ofianth P P +[2025-07-04 20:02:20] ================== +[2025-07-04 20:02:20] Epoch 1/4, Step 16000/18020, Loss(triple): 8.906605, Loss(predicate): 11.710774, LR: 0.000191, Speed: 117823.33 tokens/sec | Epoch Time Left: 0:27:56 | Total Time Left: 12:55:38 +[2025-07-04 20:03:00] Epoch 1/4, Step 16050/18020, Loss(triple): 8.300142, Loss(predicate): 8.168294, LR: 0.000191, Speed: 120616.25 tokens/sec | Epoch Time Left: 0:27:14 | Total Time Left: 12:54:53 +[2025-07-04 20:03:41] Epoch 1/4, Step 16100/18020, Loss(triple): 8.521149, Loss(predicate): 8.806233, LR: 0.000191, Speed: 120211.71 tokens/sec | Epoch Time Left: 0:26:33 | Total Time Left: 12:54:10 +[2025-07-04 20:04:22] Epoch 1/4, Step 16150/18020, Loss(triple): 9.253815, Loss(predicate): 16.751181, LR: 0.000191, Speed: 120291.33 tokens/sec | Epoch Time Left: 0:25:51 | Total Time Left: 12:53:26 +[2025-07-04 20:05:03] Epoch 1/4, Step 16200/18020, Loss(triple): 9.175037, Loss(predicate): 8.948812, LR: 0.000191, Speed: 119339.49 tokens/sec | Epoch Time Left: 0:25:10 | Total Time Left: 12:52:44 +[2025-07-04 20:05:44] Epoch 1/4, Step 16250/18020, Loss(triple): 8.588884, Loss(predicate): 8.103282, LR: 0.000191, Speed: 119978.25 tokens/sec | Epoch Time Left: 0:24:28 | Total Time Left: 12:52:00 +[2025-07-04 20:06:25] Epoch 1/4, Step 16300/18020, Loss(triple): 8.640347, Loss(predicate): 8.214375, LR: 0.000190, Speed: 120780.55 tokens/sec | Epoch Time Left: 0:23:46 | Total Time Left: 12:51:16 +[2025-07-04 20:07:06] Epoch 1/4, Step 16350/18020, Loss(triple): 8.763826, Loss(predicate): 7.698771, LR: 0.000190, Speed: 120224.65 tokens/sec | Epoch Time Left: 0:23:05 | Total Time Left: 12:50:33 +[2025-07-04 20:07:47] Epoch 1/4, Step 16400/18020, Loss(triple): 8.727791, Loss(predicate): 14.253764, LR: 0.000190, Speed: 120431.45 tokens/sec | Epoch Time Left: 0:22:23 | Total Time Left: 12:49:49 +[2025-07-04 20:08:28] Epoch 1/4, Step 16450/18020, Loss(triple): 9.005106, Loss(predicate): 10.628296, LR: 0.000190, Speed: 119873.07 tokens/sec | Epoch Time Left: 0:21:42 | Total Time Left: 12:49:06 +[2025-07-04 20:09:10] === GPU性能分析 (平均每步) === +[2025-07-04 20:09:10] 前向传播: 73.86ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 20:09:10] GPU总时间: 75.80ms, 实际迭代时间: 841.50ms, GPU利用率: 9.0% +[2025-07-04 20:09:10] ================================================== +[2025-07-04 20:09:10] === 三元组预测示例 === +[2025-07-04 20:09:10] 样本1目标: Velika Ludina country Croatia +[2025-07-04 20:09:10] 样本1预测: countryK adyharu iania,ah ialist ter ofiar in C +[2025-07-04 20:09:10] 样本2目标: Marisa Siketa occupation actress +[2025-07-04 20:09:10] 样本2预测: AmericanSensy Maena ianjoak upitiz of cth Man +[2025-07-04 20:09:10] ================== +[2025-07-04 20:09:10] Epoch 1/4, Step 16500/18020, Loss(triple): 8.594322, Loss(predicate): 7.013092, LR: 0.000190, Speed: 116820.65 tokens/sec | Epoch Time Left: 0:21:00 | Total Time Left: 12:48:26 +[2025-07-04 20:09:54] Epoch 1/4, Step 16550/18020, Loss(triple): 8.495684, Loss(predicate): 8.812785, LR: 0.000190, Speed: 111749.76 tokens/sec | Epoch Time Left: 0:20:19 | Total Time Left: 12:47:53 +[2025-07-04 20:10:35] Epoch 1/4, Step 16600/18020, Loss(triple): 8.695988, Loss(predicate): 8.189128, LR: 0.000190, Speed: 119198.46 tokens/sec | Epoch Time Left: 0:19:38 | Total Time Left: 12:47:11 +[2025-07-04 20:11:16] Epoch 1/4, Step 16650/18020, Loss(triple): 8.591362, Loss(predicate): 9.316701, LR: 0.000190, Speed: 119847.96 tokens/sec | Epoch Time Left: 0:18:56 | Total Time Left: 12:46:28 +[2025-07-04 20:11:57] Epoch 1/4, Step 16700/18020, Loss(triple): 8.786194, Loss(predicate): 9.523112, LR: 0.000190, Speed: 119788.08 tokens/sec | Epoch Time Left: 0:18:15 | Total Time Left: 12:45:45 +[2025-07-04 20:12:38] Epoch 1/4, Step 16750/18020, Loss(triple): 8.374924, Loss(predicate): 10.371613, LR: 0.000190, Speed: 120247.56 tokens/sec | Epoch Time Left: 0:17:33 | Total Time Left: 12:45:02 +[2025-07-04 20:13:19] Epoch 1/4, Step 16800/18020, Loss(triple): 8.666180, Loss(predicate): 6.731893, LR: 0.000189, Speed: 120195.84 tokens/sec | Epoch Time Left: 0:16:52 | Total Time Left: 12:44:18 +[2025-07-04 20:14:00] Epoch 1/4, Step 16850/18020, Loss(triple): 8.782671, Loss(predicate): 7.100159, LR: 0.000189, Speed: 119201.61 tokens/sec | Epoch Time Left: 0:16:10 | Total Time Left: 12:43:36 +[2025-07-04 20:14:41] Epoch 1/4, Step 16900/18020, Loss(triple): 8.412064, Loss(predicate): 8.199788, LR: 0.000189, Speed: 119791.90 tokens/sec | Epoch Time Left: 0:15:29 | Total Time Left: 12:42:53 +[2025-07-04 20:15:22] Epoch 1/4, Step 16950/18020, Loss(triple): 8.791641, Loss(predicate): 9.594646, LR: 0.000189, Speed: 120159.67 tokens/sec | Epoch Time Left: 0:14:47 | Total Time Left: 12:42:10 +[2025-07-04 20:16:03] === GPU性能分析 (平均每步) === +[2025-07-04 20:16:03] 前向传播: 51.11ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 20:16:03] GPU总时间: 53.06ms, 实际迭代时间: 813.00ms, GPU利用率: 6.5% +[2025-07-04 20:16:03] ================================================== +[2025-07-04 20:16:03] === 三元组预测示例 === +[2025-07-04 20:16:03] 样本1目标: Fiji national rugby league team sport rugby league +[2025-07-04 20:16:03] 样本1预测: countryF biryhiran il�yag 7all States ofpion the A +[2025-07-04 20:16:03] 样本2目标: Steriphoma macranthum taxon rank species +[2025-07-04 20:16:03] 样本2预测: SM adoonaala ianiausr riton species rankaxi t +[2025-07-04 20:16:03] ================== +[2025-07-04 20:16:03] Epoch 1/4, Step 17000/18020, Loss(triple): 8.720161, Loss(predicate): 15.562083, LR: 0.000189, Speed: 120915.25 tokens/sec | Epoch Time Left: 0:14:06 | Total Time Left: 12:41:25 +[2025-07-04 20:16:43] Epoch 1/4, Step 17050/18020, Loss(triple): 8.400400, Loss(predicate): 11.605408, LR: 0.000189, Speed: 120238.83 tokens/sec | Epoch Time Left: 0:13:24 | Total Time Left: 12:40:42 +[2025-07-04 20:17:24] Epoch 1/4, Step 17100/18020, Loss(triple): 9.080421, Loss(predicate): 8.132812, LR: 0.000189, Speed: 119586.54 tokens/sec | Epoch Time Left: 0:12:43 | Total Time Left: 12:39:59 +[2025-07-04 20:18:05] Epoch 1/4, Step 17150/18020, Loss(triple): 8.503525, Loss(predicate): 7.226191, LR: 0.000189, Speed: 120622.44 tokens/sec | Epoch Time Left: 0:12:01 | Total Time Left: 12:39:16 +[2025-07-04 20:18:46] Epoch 1/4, Step 17200/18020, Loss(triple): 8.515068, Loss(predicate): 8.639485, LR: 0.000189, Speed: 120073.56 tokens/sec | Epoch Time Left: 0:11:20 | Total Time Left: 12:38:32 +[2025-07-04 20:19:27] Epoch 1/4, Step 17250/18020, Loss(triple): 8.952669, Loss(predicate): 7.927938, LR: 0.000188, Speed: 120861.57 tokens/sec | Epoch Time Left: 0:10:38 | Total Time Left: 12:37:48 +[2025-07-04 20:20:07] Epoch 1/4, Step 17300/18020, Loss(triple): 8.915699, Loss(predicate): 8.085561, LR: 0.000188, Speed: 120915.17 tokens/sec | Epoch Time Left: 0:09:57 | Total Time Left: 12:37:04 +[2025-07-04 20:20:49] Epoch 1/4, Step 17350/18020, Loss(triple): 8.607685, Loss(predicate): 9.469075, LR: 0.000188, Speed: 119621.45 tokens/sec | Epoch Time Left: 0:09:15 | Total Time Left: 12:36:22 +[2025-07-04 20:21:29] Epoch 1/4, Step 17400/18020, Loss(triple): 8.730839, Loss(predicate): 13.258586, LR: 0.000188, Speed: 120641.49 tokens/sec | Epoch Time Left: 0:08:34 | Total Time Left: 12:35:38 +[2025-07-04 20:22:10] Epoch 1/4, Step 17450/18020, Loss(triple): 8.663599, Loss(predicate): 12.430226, LR: 0.000188, Speed: 120244.59 tokens/sec | Epoch Time Left: 0:07:52 | Total Time Left: 12:34:55 +[2025-07-04 20:22:51] === GPU性能分析 (平均每步) === +[2025-07-04 20:22:51] 前向传播: 51.09ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 20:22:51] GPU总时间: 53.04ms, 实际迭代时间: 813.10ms, GPU利用率: 6.5% +[2025-07-04 20:22:51] ================================================== +[2025-07-04 20:22:51] === 三元组预测示例 === +[2025-07-04 20:22:51] 样本1目标: Kennedy Space Center headquarters location Merritt Island +[2025-07-04 20:22:51] 样本1预测: countryF entym (rd erCnun iality locatedoriarit the A +[2025-07-04 20:22:51] 样本2目标: The Marshall Tucker Band (album) performer The Marshall Tucker Band +[2025-07-04 20:22:51] 样本2预测: countryF biryh (ran eror,r umall foot ofbs Ber +[2025-07-04 20:22:51] ================== +[2025-07-04 20:22:51] Epoch 1/4, Step 17500/18020, Loss(triple): 8.495554, Loss(predicate): 9.365784, LR: 0.000188, Speed: 120900.59 tokens/sec | Epoch Time Left: 0:07:11 | Total Time Left: 12:34:11 +[2025-07-04 20:23:32] Epoch 1/4, Step 17550/18020, Loss(triple): 8.408411, Loss(predicate): 8.358754, LR: 0.000188, Speed: 120535.03 tokens/sec | Epoch Time Left: 0:06:29 | Total Time Left: 12:33:27 +[2025-07-04 20:24:13] Epoch 1/4, Step 17600/18020, Loss(triple): 9.068531, Loss(predicate): 13.037944, LR: 0.000188, Speed: 119638.26 tokens/sec | Epoch Time Left: 0:05:48 | Total Time Left: 12:32:45 +[2025-07-04 20:24:54] Epoch 1/4, Step 17650/18020, Loss(triple): 8.379213, Loss(predicate): 10.866414, LR: 0.000187, Speed: 120350.72 tokens/sec | Epoch Time Left: 0:05:06 | Total Time Left: 12:32:01 +[2025-07-04 20:25:34] Epoch 1/4, Step 17700/18020, Loss(triple): 8.349634, Loss(predicate): 5.620087, LR: 0.000187, Speed: 121192.94 tokens/sec | Epoch Time Left: 0:04:25 | Total Time Left: 12:31:17 +[2025-07-04 20:26:16] Epoch 1/4, Step 17750/18020, Loss(triple): 8.628513, Loss(predicate): 14.615926, LR: 0.000187, Speed: 116895.49 tokens/sec | Epoch Time Left: 0:03:43 | Total Time Left: 12:30:38 +[2025-07-04 20:26:58] Epoch 1/4, Step 17800/18020, Loss(triple): 8.949341, Loss(predicate): 8.466919, LR: 0.000187, Speed: 117112.32 tokens/sec | Epoch Time Left: 0:03:02 | Total Time Left: 12:29:58 +[2025-07-04 20:27:40] Epoch 1/4, Step 17850/18020, Loss(triple): 8.708130, Loss(predicate): 7.317667, LR: 0.000187, Speed: 118538.64 tokens/sec | Epoch Time Left: 0:02:20 | Total Time Left: 12:29:16 +[2025-07-04 20:28:20] Epoch 1/4, Step 17900/18020, Loss(triple): 8.909815, Loss(predicate): 6.787028, LR: 0.000187, Speed: 120860.37 tokens/sec | Epoch Time Left: 0:01:39 | Total Time Left: 12:28:32 +[2025-07-04 20:29:01] Epoch 1/4, Step 17950/18020, Loss(triple): 8.440165, Loss(predicate): 13.451497, LR: 0.000187, Speed: 120957.63 tokens/sec | Epoch Time Left: 0:00:58 | Total Time Left: 12:27:49 +[2025-07-04 20:29:42] === GPU性能分析 (平均每步) === +[2025-07-04 20:29:42] 前向传播: 61.84ms, 损失计算: 0.02ms, 反向传播: 1.93ms, 优化器: 0.00ms +[2025-07-04 20:29:42] GPU总时间: 63.78ms, 实际迭代时间: 824.16ms, GPU利用率: 7.7% +[2025-07-04 20:29:42] ================================================== +[2025-07-04 20:29:42] === 三元组预测示例 === +[2025-07-04 20:29:42] 样本1目标: 8 cm FK M. 5 instance of field gun +[2025-07-04 20:29:42] 样本1预测: countryF entymaru inCDT wayance ter ofclass M A +[2025-07-04 20:29:42] 样本2目标: Namangan region located in the administrative territorial entity Uzbekistan +[2025-07-04 20:29:42] 样本2预测: countryK adyinara iania,ay ialist ter ofiarit the C +[2025-07-04 20:29:42] ================== +[2025-07-04 20:29:42] Epoch 1/4, Step 18000/18020, Loss(triple): 8.523537, Loss(predicate): 8.124187, LR: 0.000187, Speed: 119278.52 tokens/sec | Epoch Time Left: 0:00:16 | Total Time Left: 12:27:06 +[2025-07-04 20:29:59] 第1轮训练完成，进行内存清理 +[2025-07-04 20:30:01] [Memory Monitor] Epoch 1 completed - System RSS: 27057.54MB, CUDA allocated: 550.62MB, CUDA reserved: 1310.00MB +[2025-07-04 20:30:01] 开始第2轮训练 +[2025-07-04 20:30:02] 三元组提取训练模式 +[2025-07-04 20:30:02] 使用预tokenized三元组目标数据 +[2025-07-04 20:30:42] Epoch 2/4, Step 50/18020, Loss(triple): 9.013447, Loss(predicate): 9.568695, LR: 0.000186, Speed: 120159.74 tokens/sec | Epoch Time Left: 4:05:01 | Total Time Left: 12:26:15 +[2025-07-04 20:31:23] Epoch 2/4, Step 100/18020, Loss(triple): 8.370209, Loss(predicate): 9.269063, LR: 0.000186, Speed: 120403.62 tokens/sec | Epoch Time Left: 4:04:05 | Total Time Left: 12:25:31 +[2025-07-04 20:32:04] Epoch 2/4, Step 150/18020, Loss(triple): 8.844728, Loss(predicate): 9.306030, LR: 0.000186, Speed: 119036.96 tokens/sec | Epoch Time Left: 4:04:15 | Total Time Left: 12:24:49 +[2025-07-04 20:32:49] Epoch 2/4, Step 200/18020, Loss(triple): 8.662569, Loss(predicate): 7.968180, LR: 0.000186, Speed: 110995.25 tokens/sec | Epoch Time Left: 4:08:26 | Total Time Left: 12:24:16 +[2025-07-04 20:33:30] Epoch 2/4, Step 250/18020, Loss(triple): 8.244606, Loss(predicate): 7.987101, LR: 0.000186, Speed: 118979.52 tokens/sec | Epoch Time Left: 4:07:08 | Total Time Left: 12:23:34 +[2025-07-04 20:34:11] Epoch 2/4, Step 300/18020, Loss(triple): 7.996712, Loss(predicate): 8.821788, LR: 0.000186, Speed: 120472.89 tokens/sec | Epoch Time Left: 4:05:32 | Total Time Left: 12:22:51 +[2025-07-04 20:34:52] Epoch 2/4, Step 350/18020, Loss(triple): 8.222561, Loss(predicate): 11.031250, LR: 0.000186, Speed: 118826.40 tokens/sec | Epoch Time Left: 4:04:40 | Total Time Left: 12:22:09 +[2025-07-04 20:35:33] Epoch 2/4, Step 400/18020, Loss(triple): 8.755869, Loss(predicate): 11.709889, LR: 0.000186, Speed: 119467.04 tokens/sec | Epoch Time Left: 4:03:41 | Total Time Left: 12:21:27 +[2025-07-04 20:36:23] Epoch 2/4, Step 450/18020, Loss(triple): 8.684849, Loss(predicate): 12.871765, LR: 0.000185, Speed: 99496.14 tokens/sec | Epoch Time Left: 4:08:08 | Total Time Left: 12:21:09 +[2025-07-04 20:37:05] === GPU性能分析 (平均每步) === +[2025-07-04 20:37:05] 前向传播: 8.09ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 20:37:05] GPU总时间: 10.01ms, 实际迭代时间: 847.04ms, GPU利用率: 1.2% +[2025-07-04 20:37:05] ================================================== +[2025-07-04 20:37:05] === 三元组预测示例 === +[2025-07-04 20:37:05] 样本1目标: Figueirópolis d'Oeste instance of municipality +[2025-07-04 20:37:05] 样本1预测: countryF instinasara il�kag wayance ter ofationrit- A +[2025-07-04 20:37:05] 样本2目标: British Columbia country Canada +[2025-07-04 20:37:05] 样本2预测: countryF entyhar L on�nag ativeist ter ofiaance in C +[2025-07-04 20:37:05] ================== +[2025-07-04 20:37:05] Epoch 2/4, Step 500/18020, Loss(triple): 9.004381, Loss(predicate): 8.449504, LR: 0.000185, Speed: 116056.02 tokens/sec | Epoch Time Left: 4:07:25 | Total Time Left: 12:20:30 +[2025-07-04 20:37:46] Epoch 2/4, Step 550/18020, Loss(triple): 8.346695, Loss(predicate): 8.182790, LR: 0.000185, Speed: 119379.80 tokens/sec | Epoch Time Left: 4:06:05 | Total Time Left: 12:19:47 +[2025-07-04 20:38:27] Epoch 2/4, Step 600/18020, Loss(triple): 8.139980, Loss(predicate): 9.233073, LR: 0.000185, Speed: 121233.58 tokens/sec | Epoch Time Left: 4:04:33 | Total Time Left: 12:19:03 +[2025-07-04 20:39:07] Epoch 2/4, Step 650/18020, Loss(triple): 8.496647, Loss(predicate): 9.264171, LR: 0.000185, Speed: 121092.44 tokens/sec | Epoch Time Left: 4:03:10 | Total Time Left: 12:18:19 +[2025-07-04 20:39:48] Epoch 2/4, Step 700/18020, Loss(triple): 8.998301, Loss(predicate): 9.329224, LR: 0.000185, Speed: 119755.90 tokens/sec | Epoch Time Left: 4:02:04 | Total Time Left: 12:17:36 +[2025-07-04 20:40:29] Epoch 2/4, Step 750/18020, Loss(triple): 8.853563, Loss(predicate): 12.656830, LR: 0.000185, Speed: 120329.34 tokens/sec | Epoch Time Left: 4:00:57 | Total Time Left: 12:16:53 +[2025-07-04 20:41:15] Epoch 2/4, Step 800/18020, Loss(triple): 8.521412, Loss(predicate): 9.992065, LR: 0.000185, Speed: 106401.37 tokens/sec | Epoch Time Left: 4:01:49 | Total Time Left: 12:16:25 +[2025-07-04 20:42:00] Epoch 2/4, Step 850/18020, Loss(triple): 8.485966, Loss(predicate): 8.273682, LR: 0.000184, Speed: 109246.86 tokens/sec | Epoch Time Left: 4:02:04 | Total Time Left: 12:15:53 +[2025-07-04 20:42:42] Epoch 2/4, Step 900/18020, Loss(triple): 8.573277, Loss(predicate): 6.368449, LR: 0.000184, Speed: 119278.36 tokens/sec | Epoch Time Left: 4:01:01 | Total Time Left: 12:15:11 +[2025-07-04 20:43:22] Epoch 2/4, Step 950/18020, Loss(triple): 8.367649, Loss(predicate): 10.788910, LR: 0.000184, Speed: 120113.29 tokens/sec | Epoch Time Left: 3:59:56 | Total Time Left: 12:14:28 +[2025-07-04 20:44:04] === GPU性能分析 (平均每步) === +[2025-07-04 20:44:04] 前向传播: 7.96ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 20:44:04] GPU总时间: 9.88ms, 实际迭代时间: 823.51ms, GPU利用率: 1.2% +[2025-07-04 20:44:04] ================================================== +[2025-07-04 20:44:04] === 三元组预测示例 === +[2025-07-04 20:44:04] 样本1目标: Pao-Lu Hsu date of birth September 1, 1910 +[2025-07-04 20:44:04] 样本1预测: countryM biristhaaru ilia del upationohan occth Pan +[2025-07-04 20:44:04] 样本2目标: Pseudosesia taxon rank genus +[2025-07-04 20:44:04] 样本2预测: instC instoonisea usisusus axon species rankax gen t +[2025-07-04 20:44:04] ================== +[2025-07-04 20:44:04] Epoch 2/4, Step 1000/18020, Loss(triple): 9.116312, Loss(predicate): 7.910197, LR: 0.000184, Speed: 119371.37 tokens/sec | Epoch Time Left: 3:58:57 | Total Time Left: 12:13:46 +[2025-07-04 20:44:44] Epoch 2/4, Step 1050/18020, Loss(triple): 8.589977, Loss(predicate): 7.441111, LR: 0.000184, Speed: 120397.53 tokens/sec | Epoch Time Left: 3:57:53 | Total Time Left: 12:13:02 +[2025-07-04 20:45:25] Epoch 2/4, Step 1100/18020, Loss(triple): 8.769566, Loss(predicate): 12.450480, LR: 0.000184, Speed: 120695.52 tokens/sec | Epoch Time Left: 3:56:51 | Total Time Left: 12:12:19 +[2025-07-04 20:46:06] Epoch 2/4, Step 1150/18020, Loss(triple): 8.645527, Loss(predicate): 12.111969, LR: 0.000184, Speed: 120757.23 tokens/sec | Epoch Time Left: 3:55:50 | Total Time Left: 12:11:35 +[2025-07-04 20:46:47] Epoch 2/4, Step 1200/18020, Loss(triple): 8.455103, Loss(predicate): 11.613312, LR: 0.000184, Speed: 120665.92 tokens/sec | Epoch Time Left: 3:54:51 | Total Time Left: 12:10:52 +[2025-07-04 20:47:28] Epoch 2/4, Step 1250/18020, Loss(triple): 8.703501, Loss(predicate): 12.212169, LR: 0.000183, Speed: 119606.99 tokens/sec | Epoch Time Left: 3:53:59 | Total Time Left: 12:10:09 +[2025-07-04 20:48:08] Epoch 2/4, Step 1300/18020, Loss(triple): 8.654804, Loss(predicate): 10.064992, LR: 0.000183, Speed: 120912.67 tokens/sec | Epoch Time Left: 3:53:01 | Total Time Left: 12:09:25 +[2025-07-04 20:48:49] Epoch 2/4, Step 1350/18020, Loss(triple): 8.604593, Loss(predicate): 12.331807, LR: 0.000183, Speed: 120981.49 tokens/sec | Epoch Time Left: 3:52:05 | Total Time Left: 12:08:42 +[2025-07-04 20:49:30] Epoch 2/4, Step 1400/18020, Loss(triple): 8.680607, Loss(predicate): 9.925923, LR: 0.000183, Speed: 120014.67 tokens/sec | Epoch Time Left: 3:51:13 | Total Time Left: 12:07:59 +[2025-07-04 20:50:11] Epoch 2/4, Step 1450/18020, Loss(triple): 8.330475, Loss(predicate): 11.688680, LR: 0.000183, Speed: 120945.11 tokens/sec | Epoch Time Left: 3:50:19 | Total Time Left: 12:07:15 +[2025-07-04 20:50:52] === GPU性能分析 (平均每步) === +[2025-07-04 20:50:52] 前向传播: 8.02ms, 损失计算: 0.02ms, 反向传播: 1.95ms, 优化器: 0.00ms +[2025-07-04 20:50:52] GPU总时间: 9.98ms, 实际迭代时间: 818.36ms, GPU利用率: 1.2% +[2025-07-04 20:50:52] ================================================== +[2025-07-04 20:50:52] === 三元组预测示例 === +[2025-07-04 20:50:52] 样本1目标: Kasthuri occupation actress +[2025-07-04 20:50:52] 样本1预测: countryF biryhaeu oniaoet ortation States ofmth B C +[2025-07-04 20:50:52] 样本2目标: Ivo Georgiev sport football +[2025-07-04 20:50:52] 样本2预测: countryW birurhaaran an�kel ortall footerbth ofver +[2025-07-04 20:50:52] ================== +[2025-07-04 20:50:52] Epoch 2/4, Step 1500/18020, Loss(triple): 8.851183, Loss(predicate): 8.194051, LR: 0.000183, Speed: 120123.33 tokens/sec | Epoch Time Left: 3:49:29 | Total Time Left: 12:06:32 +[2025-07-04 20:51:32] Epoch 2/4, Step 1550/18020, Loss(triple): 8.560959, Loss(predicate): 7.942159, LR: 0.000183, Speed: 121005.85 tokens/sec | Epoch Time Left: 3:48:36 | Total Time Left: 12:05:48 +[2025-07-04 20:52:13] Epoch 2/4, Step 1600/18020, Loss(triple): 8.640007, Loss(predicate): 9.806783, LR: 0.000182, Speed: 121086.31 tokens/sec | Epoch Time Left: 3:47:43 | Total Time Left: 12:05:04 +[2025-07-04 20:52:54] Epoch 2/4, Step 1650/18020, Loss(triple): 8.405960, Loss(predicate): 6.098379, LR: 0.000182, Speed: 119807.58 tokens/sec | Epoch Time Left: 3:46:56 | Total Time Left: 12:04:22 +[2025-07-04 20:53:34] Epoch 2/4, Step 1700/18020, Loss(triple): 8.799953, Loss(predicate): 6.256042, LR: 0.000182, Speed: 120731.05 tokens/sec | Epoch Time Left: 3:46:06 | Total Time Left: 12:03:38 +[2025-07-04 20:54:15] Epoch 2/4, Step 1750/18020, Loss(triple): 8.248829, Loss(predicate): 9.081218, LR: 0.000182, Speed: 120371.43 tokens/sec | Epoch Time Left: 3:45:18 | Total Time Left: 12:02:55 +[2025-07-04 20:54:56] Epoch 2/4, Step 1800/18020, Loss(triple): 8.005016, Loss(predicate): 9.364502, LR: 0.000182, Speed: 121225.91 tokens/sec | Epoch Time Left: 3:44:27 | Total Time Left: 12:02:11 +[2025-07-04 20:55:37] Epoch 2/4, Step 1850/18020, Loss(triple): 8.420399, Loss(predicate): 9.052388, LR: 0.000182, Speed: 120739.71 tokens/sec | Epoch Time Left: 3:43:39 | Total Time Left: 12:01:28 +[2025-07-04 20:56:18] Epoch 2/4, Step 1900/18020, Loss(triple): 8.117638, Loss(predicate): 10.915629, LR: 0.000182, Speed: 120021.54 tokens/sec | Epoch Time Left: 3:42:53 | Total Time Left: 12:00:45 +[2025-07-04 20:56:58] Epoch 2/4, Step 1950/18020, Loss(triple): 8.188829, Loss(predicate): 8.430797, LR: 0.000182, Speed: 120889.29 tokens/sec | Epoch Time Left: 3:42:04 | Total Time Left: 12:00:02 +[2025-07-04 20:57:39] === GPU性能分析 (平均每步) === +[2025-07-04 20:57:39] 前向传播: 8.02ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 20:57:39] GPU总时间: 9.93ms, 实际迭代时间: 814.69ms, GPU利用率: 1.2% +[2025-07-04 20:57:39] ================================================== +[2025-07-04 20:57:39] === 三元组预测示例 === +[2025-07-04 20:57:39] 样本1目标: Kyrgyzstani parliamentary election, 2010 country Kyrgyzstan +[2025-07-04 20:57:39] 样本1预测: countryF entyas (ran onmor 6ate3 ofmion in 19 +[2025-07-04 20:57:39] 样本2目标: Peggys Cove, Nova Scotia located in the administrative territorial entity Halifax Regional Municipality +[2025-07-04 20:57:39] 样本2预测: countryK adyinutr H anin,r ativeist ter ofiarit the C +[2025-07-04 20:57:39] ================== +[2025-07-04 20:57:39] Epoch 2/4, Step 2000/18020, Loss(triple): 8.634449, Loss(predicate): 9.601634, LR: 0.000181, Speed: 120663.89 tokens/sec | Epoch Time Left: 3:41:17 | Total Time Left: 11:59:18 +[2025-07-04 20:58:25] Epoch 2/4, Step 2050/18020, Loss(triple): 8.308929, Loss(predicate): 8.736796, LR: 0.000181, Speed: 106057.20 tokens/sec | Epoch Time Left: 3:41:14 | Total Time Left: 11:58:50 +[2025-07-04 20:59:09] Epoch 2/4, Step 2100/18020, Loss(triple): 8.979057, Loss(predicate): 8.056244, LR: 0.000181, Speed: 112691.98 tokens/sec | Epoch Time Left: 3:40:48 | Total Time Left: 11:58:14 +[2025-07-04 20:59:50] Epoch 2/4, Step 2150/18020, Loss(triple): 8.554138, Loss(predicate): 9.117085, LR: 0.000181, Speed: 118798.70 tokens/sec | Epoch Time Left: 3:40:04 | Total Time Left: 11:57:32 +[2025-07-04 21:00:31] Epoch 2/4, Step 2200/18020, Loss(triple): 8.803385, Loss(predicate): 10.986745, LR: 0.000181, Speed: 121099.49 tokens/sec | Epoch Time Left: 3:39:16 | Total Time Left: 11:56:48 +[2025-07-04 21:01:11] Epoch 2/4, Step 2250/18020, Loss(triple): 8.231403, Loss(predicate): 10.361318, LR: 0.000181, Speed: 121042.12 tokens/sec | Epoch Time Left: 3:38:27 | Total Time Left: 11:56:05 +[2025-07-04 21:01:55] Epoch 2/4, Step 2300/18020, Loss(triple): 8.534189, Loss(predicate): 11.536702, LR: 0.000181, Speed: 113830.97 tokens/sec | Epoch Time Left: 3:37:57 | Total Time Left: 11:55:27 +[2025-07-04 21:02:49] Epoch 2/4, Step 2350/18020, Loss(triple): 8.058960, Loss(predicate): 8.236755, LR: 0.000180, Speed: 90119.74 tokens/sec | Epoch Time Left: 3:38:41 | Total Time Left: 11:55:19 +[2025-07-04 21:03:36] Epoch 2/4, Step 2400/18020, Loss(triple): 8.644007, Loss(predicate): 8.609019, LR: 0.000180, Speed: 104736.35 tokens/sec | Epoch Time Left: 3:38:32 | Total Time Left: 11:54:51 +[2025-07-04 21:04:20] Epoch 2/4, Step 2450/18020, Loss(triple): 8.937584, Loss(predicate): 14.971232, LR: 0.000180, Speed: 111967.66 tokens/sec | Epoch Time Left: 3:38:03 | Total Time Left: 11:54:16 +[2025-07-04 21:05:02] === GPU性能分析 (平均每步) === +[2025-07-04 21:05:02] 前向传播: 7.99ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 21:05:02] GPU总时间: 9.91ms, 实际迭代时间: 834.65ms, GPU利用率: 1.2% +[2025-07-04 21:05:02] ================================================== +[2025-07-04 21:05:02] === 三元组预测示例 === +[2025-07-04 21:05:02] 样本1目标: Alexandra Sokoloff country of citizenship American +[2025-07-04 21:05:02] 样本1预测: countryK biryh (.d erk Dre Americanitiz of cide countryens +[2025-07-04 21:05:02] 样本2目标: New Lebanon, Ohio located in the administrative territorial entity Montgomery County, Ohio +[2025-07-04 21:05:02] 样本2预测: countryK entymutor H eriz,ay iality locatedmianrit, C +[2025-07-04 21:05:02] ================== +[2025-07-04 21:05:02] Epoch 2/4, Step 2500/18020, Loss(triple): 8.255405, Loss(predicate): 11.109843, LR: 0.000180, Speed: 117778.12 tokens/sec | Epoch Time Left: 3:37:19 | Total Time Left: 11:53:35 +[2025-07-04 21:05:43] Epoch 2/4, Step 2550/18020, Loss(triple): 8.006861, Loss(predicate): 12.472137, LR: 0.000180, Speed: 119521.58 tokens/sec | Epoch Time Left: 3:36:32 | Total Time Left: 11:52:52 +[2025-07-04 21:06:24] Epoch 2/4, Step 2600/18020, Loss(triple): 8.370079, Loss(predicate): 8.968079, LR: 0.000180, Speed: 118452.37 tokens/sec | Epoch Time Left: 3:35:47 | Total Time Left: 11:52:11 +[2025-07-04 21:07:05] Epoch 2/4, Step 2650/18020, Loss(triple): 8.353691, Loss(predicate): 7.085063, LR: 0.000179, Speed: 119626.01 tokens/sec | Epoch Time Left: 3:34:59 | Total Time Left: 11:51:28 +[2025-07-04 21:07:46] Epoch 2/4, Step 2700/18020, Loss(triple): 8.092991, Loss(predicate): 8.119568, LR: 0.000179, Speed: 120106.80 tokens/sec | Epoch Time Left: 3:34:12 | Total Time Left: 11:50:45 +[2025-07-04 21:08:28] Epoch 2/4, Step 2750/18020, Loss(triple): 8.252014, Loss(predicate): 14.257874, LR: 0.000179, Speed: 118920.56 tokens/sec | Epoch Time Left: 3:33:26 | Total Time Left: 11:50:03 +[2025-07-04 21:09:09] Epoch 2/4, Step 2800/18020, Loss(triple): 8.575520, Loss(predicate): 11.986522, LR: 0.000179, Speed: 119410.47 tokens/sec | Epoch Time Left: 3:32:40 | Total Time Left: 11:49:21 +[2025-07-04 21:09:50] Epoch 2/4, Step 2850/18020, Loss(triple): 8.764040, Loss(predicate): 8.258159, LR: 0.000179, Speed: 118373.28 tokens/sec | Epoch Time Left: 3:31:56 | Total Time Left: 11:48:39 +[2025-07-04 21:10:31] Epoch 2/4, Step 2900/18020, Loss(triple): 8.320721, Loss(predicate): 9.439900, LR: 0.000179, Speed: 120007.12 tokens/sec | Epoch Time Left: 3:31:09 | Total Time Left: 11:47:57 +[2025-07-04 21:11:12] Epoch 2/4, Step 2950/18020, Loss(triple): 8.372698, Loss(predicate): 9.258016, LR: 0.000179, Speed: 120195.27 tokens/sec | Epoch Time Left: 3:30:22 | Total Time Left: 11:47:14 +[2025-07-04 21:11:54] === GPU性能分析 (平均每步) === +[2025-07-04 21:11:54] 前向传播: 7.97ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 21:11:54] GPU总时间: 9.89ms, 实际迭代时间: 827.48ms, GPU利用率: 1.2% +[2025-07-04 21:11:54] ================================================== +[2025-07-04 21:11:54] === 三元组预测示例 === +[2025-07-04 21:11:54] 样本1目标: Bita Farrahi place of birth Tehran +[2025-07-04 21:11:54] 样本1预测: RB birohiaru an�nun upationoh of occth Pan +[2025-07-04 21:11:54] 样本2目标: 2013 World Aquatics Championships location Palau Sant Jordi +[2025-07-04 21:11:54] 样本2预测: countryI insturasaM G onA,T 2um States ofpion- A +[2025-07-04 21:11:54] ================== +[2025-07-04 21:11:54] Epoch 2/4, Step 3000/18020, Loss(triple): 8.484852, Loss(predicate): 11.520070, LR: 0.000178, Speed: 118799.77 tokens/sec | Epoch Time Left: 3:29:38 | Total Time Left: 11:46:32 +[2025-07-04 21:12:35] Epoch 2/4, Step 3050/18020, Loss(triple): 8.548538, Loss(predicate): 6.266398, LR: 0.000178, Speed: 119701.76 tokens/sec | Epoch Time Left: 3:28:52 | Total Time Left: 11:45:49 +[2025-07-04 21:13:16] Epoch 2/4, Step 3100/18020, Loss(triple): 8.624470, Loss(predicate): 10.461406, LR: 0.000178, Speed: 120096.84 tokens/sec | Epoch Time Left: 3:28:06 | Total Time Left: 11:45:06 +[2025-07-04 21:13:56] Epoch 2/4, Step 3150/18020, Loss(triple): 8.607914, Loss(predicate): 10.775630, LR: 0.000178, Speed: 120388.02 tokens/sec | Epoch Time Left: 3:27:19 | Total Time Left: 11:44:23 +[2025-07-04 21:14:41] Epoch 2/4, Step 3200/18020, Loss(triple): 8.133745, Loss(predicate): 13.339798, LR: 0.000178, Speed: 110594.91 tokens/sec | Epoch Time Left: 3:26:49 | Total Time Left: 11:43:49 +[2025-07-04 21:15:23] Epoch 2/4, Step 3250/18020, Loss(triple): 8.417213, Loss(predicate): 9.901001, LR: 0.000178, Speed: 117368.34 tokens/sec | Epoch Time Left: 3:26:07 | Total Time Left: 11:43:08 +[2025-07-04 21:16:04] Epoch 2/4, Step 3300/18020, Loss(triple): 8.025570, Loss(predicate): 12.910014, LR: 0.000178, Speed: 118955.16 tokens/sec | Epoch Time Left: 3:25:23 | Total Time Left: 11:42:26 +[2025-07-04 21:16:45] Epoch 2/4, Step 3350/18020, Loss(triple): 8.437693, Loss(predicate): 11.229401, LR: 0.000177, Speed: 119996.77 tokens/sec | Epoch Time Left: 3:24:37 | Total Time Left: 11:41:43 +[2025-07-04 21:17:26] Epoch 2/4, Step 3400/18020, Loss(triple): 8.117462, Loss(predicate): 10.227966, LR: 0.000177, Speed: 119741.01 tokens/sec | Epoch Time Left: 3:23:52 | Total Time Left: 11:41:00 +[2025-07-04 21:18:08] Epoch 2/4, Step 3450/18020, Loss(triple): 8.099857, Loss(predicate): 9.501465, LR: 0.000177, Speed: 118299.98 tokens/sec | Epoch Time Left: 3:23:09 | Total Time Left: 11:40:19 +[2025-07-04 21:18:49] === GPU性能分析 (平均每步) === +[2025-07-04 21:18:49] 前向传播: 7.94ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 21:18:49] GPU总时间: 9.86ms, 实际迭代时间: 823.40ms, GPU利用率: 1.2% +[2025-07-04 21:18:49] ================================================== +[2025-07-04 21:18:49] === 三元组预测示例 === +[2025-07-04 21:18:49] 样本1目标: Idaho Falls Idaho Temple located in the administrative territorial entity Idaho +[2025-07-04 21:18:49] 样本1预测: countryF adyhard in�nre ialist ter ofiaance in C +[2025-07-04 21:18:49] 样本2目标: Xi'an University of Science and Technology country China +[2025-07-04 21:18:49] 样本2预测: countryK entyhara ianiaoah ialist ter ofiarit the C +[2025-07-04 21:18:49] ================== +[2025-07-04 21:18:49] Epoch 2/4, Step 3500/18020, Loss(triple): 8.337078, Loss(predicate): 13.324738, LR: 0.000177, Speed: 119387.90 tokens/sec | Epoch Time Left: 3:22:24 | Total Time Left: 11:39:37 +[2025-07-04 21:19:37] Epoch 2/4, Step 3550/18020, Loss(triple): 8.008625, Loss(predicate): 8.165172, LR: 0.000177, Speed: 101303.10 tokens/sec | Epoch Time Left: 3:22:10 | Total Time Left: 11:39:12 +[2025-07-04 21:20:30] Epoch 2/4, Step 3600/18020, Loss(triple): 8.311920, Loss(predicate): 7.598724, LR: 0.000177, Speed: 93814.13 tokens/sec | Epoch Time Left: 3:22:10 | Total Time Left: 11:38:55 +[2025-07-04 21:21:20] Epoch 2/4, Step 3650/18020, Loss(triple): 8.135990, Loss(predicate): 5.286845, LR: 0.000176, Speed: 98584.96 tokens/sec | Epoch Time Left: 3:21:59 | Total Time Left: 11:38:33 +[2025-07-04 21:22:08] Epoch 2/4, Step 3700/18020, Loss(triple): 8.454994, Loss(predicate): 10.558594, LR: 0.000176, Speed: 100554.90 tokens/sec | Epoch Time Left: 3:21:42 | Total Time Left: 11:38:09 +[2025-07-04 21:22:55] Epoch 2/4, Step 3750/18020, Loss(triple): 7.943956, Loss(predicate): 10.666809, LR: 0.000176, Speed: 104819.52 tokens/sec | Epoch Time Left: 3:21:18 | Total Time Left: 11:37:39 +[2025-07-04 21:23:37] Epoch 2/4, Step 3800/18020, Loss(triple): 8.006273, Loss(predicate): 9.411479, LR: 0.000176, Speed: 119135.59 tokens/sec | Epoch Time Left: 3:20:31 | Total Time Left: 11:36:57 +[2025-07-04 21:24:18] Epoch 2/4, Step 3850/18020, Loss(triple): 8.572220, Loss(predicate): 14.488739, LR: 0.000176, Speed: 119759.55 tokens/sec | Epoch Time Left: 3:19:44 | Total Time Left: 11:36:14 +[2025-07-04 21:25:06] Epoch 2/4, Step 3900/18020, Loss(triple): 8.427958, Loss(predicate): 7.756348, LR: 0.000176, Speed: 102069.97 tokens/sec | Epoch Time Left: 3:19:23 | Total Time Left: 11:35:47 +[2025-07-04 21:25:52] Epoch 2/4, Step 3950/18020, Loss(triple): 8.146957, Loss(predicate): 10.111969, LR: 0.000176, Speed: 105414.82 tokens/sec | Epoch Time Left: 3:18:56 | Total Time Left: 11:35:17 +[2025-07-04 21:26:35] === GPU性能分析 (平均每步) === +[2025-07-04 21:26:35] 前向传播: 8.01ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 21:26:35] GPU总时间: 9.92ms, 实际迭代时间: 847.34ms, GPU利用率: 1.2% +[2025-07-04 21:26:35] ================================================== +[2025-07-04 21:26:35] === 三元组预测示例 === +[2025-07-04 21:26:35] 样本1目标: Rosyam Nor occupation actor +[2025-07-04 21:26:35] 样本1预测: countryS birohiaran on�or upationoh of,th 2 19 +[2025-07-04 21:26:35] 样本2目标: Tony Ambrose sport rally driver +[2025-07-04 21:26:35] 样本2预测: AmericanThe birfvent (.an il� del 7atebor of,th 2 19 +[2025-07-04 21:26:35] ================== +[2025-07-04 21:26:35] Epoch 2/4, Step 4000/18020, Loss(triple): 8.632111, Loss(predicate): 13.162292, LR: 0.000175, Speed: 116014.62 tokens/sec | Epoch Time Left: 3:18:14 | Total Time Left: 11:34:37 +[2025-07-04 21:27:19] Epoch 2/4, Step 4050/18020, Loss(triple): 8.927500, Loss(predicate): 8.327952, LR: 0.000175, Speed: 111372.55 tokens/sec | Epoch Time Left: 3:17:37 | Total Time Left: 11:34:01 +[2025-07-04 21:28:01] Epoch 2/4, Step 4100/18020, Loss(triple): 8.794840, Loss(predicate): 12.205302, LR: 0.000175, Speed: 116192.15 tokens/sec | Epoch Time Left: 3:16:54 | Total Time Left: 11:33:21 +[2025-07-04 21:28:43] Epoch 2/4, Step 4150/18020, Loss(triple): 8.865608, Loss(predicate): 10.106282, LR: 0.000175, Speed: 118386.57 tokens/sec | Epoch Time Left: 3:16:09 | Total Time Left: 11:32:39 +[2025-07-04 21:29:24] Epoch 2/4, Step 4200/18020, Loss(triple): 8.213358, Loss(predicate): 13.779735, LR: 0.000175, Speed: 120354.83 tokens/sec | Epoch Time Left: 3:15:21 | Total Time Left: 11:31:56 +[2025-07-04 21:30:04] Epoch 2/4, Step 4250/18020, Loss(triple): 8.454355, Loss(predicate): 7.992635, LR: 0.000175, Speed: 120543.41 tokens/sec | Epoch Time Left: 3:14:33 | Total Time Left: 11:31:12 +[2025-07-04 21:30:45] Epoch 2/4, Step 4300/18020, Loss(triple): 8.372293, Loss(predicate): 6.601410, LR: 0.000174, Speed: 120205.88 tokens/sec | Epoch Time Left: 3:13:46 | Total Time Left: 11:30:29 +[2025-07-04 21:31:26] Epoch 2/4, Step 4350/18020, Loss(triple): 8.468719, Loss(predicate): 13.031199, LR: 0.000174, Speed: 120422.55 tokens/sec | Epoch Time Left: 3:12:59 | Total Time Left: 11:29:46 +[2025-07-04 21:32:07] Epoch 2/4, Step 4400/18020, Loss(triple): 8.920040, Loss(predicate): 12.723694, LR: 0.000174, Speed: 119403.78 tokens/sec | Epoch Time Left: 3:12:13 | Total Time Left: 11:29:03 +[2025-07-04 21:32:48] Epoch 2/4, Step 4450/18020, Loss(triple): 8.462666, Loss(predicate): 9.312286, LR: 0.000174, Speed: 120582.06 tokens/sec | Epoch Time Left: 3:11:26 | Total Time Left: 11:28:19 +[2025-07-04 21:33:29] === GPU性能分析 (平均每步) === +[2025-07-04 21:33:29] 前向传播: 7.97ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 21:33:29] GPU总时间: 9.89ms, 实际迭代时间: 814.42ms, GPU利用率: 1.2% +[2025-07-04 21:33:29] ================================================== +[2025-07-04 21:33:29] === 三元组预测示例 === +[2025-07-04 21:33:29] 样本1目标: Megacephala coerulea taxon rank species +[2025-07-04 21:33:29] 样本1预测: countryS entoonaara iciausus riton species rankaxph t +[2025-07-04 21:33:29] 样本2目标: Kimnyole ethnic group Nandi +[2025-07-04 21:33:29] 样本2预测: countryF entomgaea oninous ialance ter ofmr Ser +[2025-07-04 21:33:29] ================== +[2025-07-04 21:33:29] Epoch 2/4, Step 4500/18020, Loss(triple): 8.525742, Loss(predicate): 10.815969, LR: 0.000174, Speed: 120704.54 tokens/sec | Epoch Time Left: 3:10:39 | Total Time Left: 11:27:36 +[2025-07-04 21:34:10] Epoch 2/4, Step 4550/18020, Loss(triple): 8.326700, Loss(predicate): 8.281769, LR: 0.000174, Speed: 119805.33 tokens/sec | Epoch Time Left: 3:09:53 | Total Time Left: 11:26:53 +[2025-07-04 21:34:51] Epoch 2/4, Step 4600/18020, Loss(triple): 8.520350, Loss(predicate): 7.534586, LR: 0.000173, Speed: 120505.88 tokens/sec | Epoch Time Left: 3:09:06 | Total Time Left: 11:26:09 +[2025-07-04 21:35:32] Epoch 2/4, Step 4650/18020, Loss(triple): 8.520273, Loss(predicate): 7.652161, LR: 0.000173, Speed: 119700.46 tokens/sec | Epoch Time Left: 3:08:20 | Total Time Left: 11:25:27 +[2025-07-04 21:36:12] Epoch 2/4, Step 4700/18020, Loss(triple): 8.306551, Loss(predicate): 9.467448, LR: 0.000173, Speed: 120279.08 tokens/sec | Epoch Time Left: 3:07:34 | Total Time Left: 11:24:43 +[2025-07-04 21:36:53] Epoch 2/4, Step 4750/18020, Loss(triple): 8.374012, Loss(predicate): 6.353312, LR: 0.000173, Speed: 120263.43 tokens/sec | Epoch Time Left: 3:06:48 | Total Time Left: 11:24:00 +[2025-07-04 21:37:34] Epoch 2/4, Step 4800/18020, Loss(triple): 8.801388, Loss(predicate): 7.156748, LR: 0.000173, Speed: 119488.31 tokens/sec | Epoch Time Left: 3:06:03 | Total Time Left: 11:23:17 +[2025-07-04 21:38:15] Epoch 2/4, Step 4850/18020, Loss(triple): 7.868347, Loss(predicate): 10.487122, LR: 0.000173, Speed: 120038.12 tokens/sec | Epoch Time Left: 3:05:17 | Total Time Left: 11:22:34 +[2025-07-04 21:38:57] Epoch 2/4, Step 4900/18020, Loss(triple): 8.032501, Loss(predicate): 9.071055, LR: 0.000172, Speed: 118868.55 tokens/sec | Epoch Time Left: 3:04:32 | Total Time Left: 11:21:52 +[2025-07-04 21:39:38] Epoch 2/4, Step 4950/18020, Loss(triple): 7.950357, Loss(predicate): 8.527924, LR: 0.000172, Speed: 120112.27 tokens/sec | Epoch Time Left: 3:03:47 | Total Time Left: 11:21:09 +[2025-07-04 21:40:19] === GPU性能分析 (平均每步) === +[2025-07-04 21:40:19] 前向传播: 7.97ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 21:40:19] GPU总时间: 9.89ms, 实际迭代时间: 817.18ms, GPU利用率: 1.2% +[2025-07-04 21:40:19] ================================================== +[2025-07-04 21:40:19] === 三元组预测示例 === +[2025-07-04 21:40:19] 样本1目标: Megacyllene designata taxon rank species +[2025-07-04 21:40:19] 样本1预测: itC instusoniee anpusul axon species rankaxaris t +[2025-07-04 21:40:19] 样本2目标: Réjaumont, Gers instance of commune +[2025-07-04 21:40:19] 样本2预测: instC insttDaee oniaoay Eance commun ofance-- of +[2025-07-04 21:40:19] ================== +[2025-07-04 21:40:19] Epoch 2/4, Step 5000/18020, Loss(triple): 8.406944, Loss(predicate): 7.111033, LR: 0.000172, Speed: 120296.21 tokens/sec | Epoch Time Left: 3:03:01 | Total Time Left: 11:20:26 +[2025-07-04 21:41:00] Epoch 2/4, Step 5050/18020, Loss(triple): 7.862865, Loss(predicate): 14.859599, LR: 0.000172, Speed: 119769.97 tokens/sec | Epoch Time Left: 3:02:16 | Total Time Left: 11:19:43 +[2025-07-04 21:41:40] Epoch 2/4, Step 5100/18020, Loss(triple): 8.177170, Loss(predicate): 10.910268, LR: 0.000172, Speed: 120314.97 tokens/sec | Epoch Time Left: 3:01:31 | Total Time Left: 11:19:00 +[2025-07-04 21:42:22] Epoch 2/4, Step 5150/18020, Loss(triple): 8.289856, Loss(predicate): 13.923787, LR: 0.000172, Speed: 119423.26 tokens/sec | Epoch Time Left: 3:00:46 | Total Time Left: 11:18:17 +[2025-07-04 21:43:02] Epoch 2/4, Step 5200/18020, Loss(triple): 8.471497, Loss(predicate): 5.621287, LR: 0.000171, Speed: 120599.03 tokens/sec | Epoch Time Left: 3:00:00 | Total Time Left: 11:17:34 +[2025-07-04 21:43:43] Epoch 2/4, Step 5250/18020, Loss(triple): 8.248138, Loss(predicate): 8.880452, LR: 0.000171, Speed: 120553.98 tokens/sec | Epoch Time Left: 2:59:15 | Total Time Left: 11:16:51 +[2025-07-04 21:44:24] Epoch 2/4, Step 5300/18020, Loss(triple): 8.091782, Loss(predicate): 9.125854, LR: 0.000171, Speed: 119800.94 tokens/sec | Epoch Time Left: 2:58:30 | Total Time Left: 11:16:08 +[2025-07-04 21:45:05] Epoch 2/4, Step 5350/18020, Loss(triple): 8.417795, Loss(predicate): 10.144847, LR: 0.000171, Speed: 120703.03 tokens/sec | Epoch Time Left: 2:57:45 | Total Time Left: 11:15:24 +[2025-07-04 21:45:46] Epoch 2/4, Step 5400/18020, Loss(triple): 8.010019, Loss(predicate): 8.267517, LR: 0.000171, Speed: 119813.60 tokens/sec | Epoch Time Left: 2:57:00 | Total Time Left: 11:14:42 +[2025-07-04 21:46:27] Epoch 2/4, Step 5450/18020, Loss(triple): 8.585361, Loss(predicate): 10.467641, LR: 0.000171, Speed: 120379.11 tokens/sec | Epoch Time Left: 2:56:15 | Total Time Left: 11:13:58 +[2025-07-04 21:47:07] === GPU性能分析 (平均每步) === +[2025-07-04 21:47:07] 前向传播: 7.96ms, 损失计算: 0.02ms, 反向传播: 1.95ms, 优化器: 0.00ms +[2025-07-04 21:47:07] GPU总时间: 9.92ms, 实际迭代时间: 814.69ms, GPU利用率: 1.2% +[2025-07-04 21:47:07] ================================================== +[2025-07-04 21:47:07] === 三元组预测示例 === +[2025-07-04 21:47:07] 样本1目标: Zdeňka Vejnarová place of birth Jilemnice +[2025-07-04 21:47:07] 样本1预测: countryK biryhaeu an�nil upate bir of occth S 19 +[2025-07-04 21:47:07] 样本2目标: Barry Wood (cricketer) place of birth Ossett, Yorkshire +[2025-07-04 21:47:07] 样本2预测: RJ birll H (rd oniz,et 7ate bir of cth 2 19 +[2025-07-04 21:47:07] ================== +[2025-07-04 21:47:07] Epoch 2/4, Step 5500/18020, Loss(triple): 8.268330, Loss(predicate): 7.701864, LR: 0.000170, Speed: 120664.36 tokens/sec | Epoch Time Left: 2:55:30 | Total Time Left: 11:13:15 +[2025-07-04 21:47:48] Epoch 2/4, Step 5550/18020, Loss(triple): 8.187983, Loss(predicate): 12.246857, LR: 0.000170, Speed: 119885.42 tokens/sec | Epoch Time Left: 2:54:46 | Total Time Left: 11:12:32 +[2025-07-04 21:48:29] Epoch 2/4, Step 5600/18020, Loss(triple): 8.172272, Loss(predicate): 16.944529, LR: 0.000170, Speed: 120425.81 tokens/sec | Epoch Time Left: 2:54:01 | Total Time Left: 11:11:49 +[2025-07-04 21:49:10] Epoch 2/4, Step 5650/18020, Loss(triple): 8.109528, Loss(predicate): 12.209859, LR: 0.000170, Speed: 119383.34 tokens/sec | Epoch Time Left: 2:53:17 | Total Time Left: 11:11:07 +[2025-07-04 21:49:51] Epoch 2/4, Step 5700/18020, Loss(triple): 7.722672, Loss(predicate): 9.426178, LR: 0.000170, Speed: 120513.46 tokens/sec | Epoch Time Left: 2:52:32 | Total Time Left: 11:10:23 +[2025-07-04 21:50:32] Epoch 2/4, Step 5750/18020, Loss(triple): 7.946430, Loss(predicate): 7.622660, LR: 0.000170, Speed: 120467.21 tokens/sec | Epoch Time Left: 2:51:48 | Total Time Left: 11:09:40 +[2025-07-04 21:51:13] Epoch 2/4, Step 5800/18020, Loss(triple): 8.255585, Loss(predicate): 13.990509, LR: 0.000169, Speed: 119632.24 tokens/sec | Epoch Time Left: 2:51:04 | Total Time Left: 11:08:58 +[2025-07-04 21:51:54] Epoch 2/4, Step 5850/18020, Loss(triple): 7.739468, Loss(predicate): 9.538300, LR: 0.000169, Speed: 120479.94 tokens/sec | Epoch Time Left: 2:50:19 | Total Time Left: 11:08:14 +[2025-07-04 21:52:35] Epoch 2/4, Step 5900/18020, Loss(triple): 8.243860, Loss(predicate): 10.220764, LR: 0.000169, Speed: 119694.11 tokens/sec | Epoch Time Left: 2:49:36 | Total Time Left: 11:07:32 +[2025-07-04 21:53:16] Epoch 2/4, Step 5950/18020, Loss(triple): 8.052269, Loss(predicate): 14.128489, LR: 0.000169, Speed: 120770.56 tokens/sec | Epoch Time Left: 2:48:51 | Total Time Left: 11:06:48 +[2025-07-04 21:53:56] === GPU性能分析 (平均每步) === +[2025-07-04 21:53:56] 前向传播: 7.97ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 21:53:56] GPU总时间: 9.88ms, 实际迭代时间: 815.47ms, GPU利用率: 1.2% +[2025-07-04 21:53:56] ================================================== +[2025-07-04 21:53:56] === 三元组预测示例 === +[2025-07-04 21:53:56] 样本1目标: Ronnie Peterson date of birth 14 February 1944 +[2025-07-04 21:53:56] 样本1预测: countryK biryharu aill dom 7ate bir of,th M 19 +[2025-07-04 21:53:56] 样本2目标: Subdromomeryx taxon rank genus +[2025-07-04 21:53:56] 样本2预测: countryM entoonaca alatusus axonus rankax gen t +[2025-07-04 21:53:56] ================== +[2025-07-04 21:53:56] Epoch 2/4, Step 6000/18020, Loss(triple): 8.052507, Loss(predicate): 10.549977, LR: 0.000169, Speed: 120549.22 tokens/sec | Epoch Time Left: 2:48:07 | Total Time Left: 11:06:05 +[2025-07-04 21:54:38] Epoch 2/4, Step 6050/18020, Loss(triple): 7.947998, Loss(predicate): 9.545868, LR: 0.000168, Speed: 119341.00 tokens/sec | Epoch Time Left: 2:47:23 | Total Time Left: 11:05:23 +[2025-07-04 21:55:18] Epoch 2/4, Step 6100/18020, Loss(triple): 8.587585, Loss(predicate): 12.585979, LR: 0.000168, Speed: 120768.59 tokens/sec | Epoch Time Left: 2:46:39 | Total Time Left: 11:04:39 +[2025-07-04 21:55:59] Epoch 2/4, Step 6150/18020, Loss(triple): 8.520365, Loss(predicate): 7.384817, LR: 0.000168, Speed: 119650.67 tokens/sec | Epoch Time Left: 2:45:55 | Total Time Left: 11:03:57 +[2025-07-04 21:56:40] Epoch 2/4, Step 6200/18020, Loss(triple): 8.246775, Loss(predicate): 11.411754, LR: 0.000168, Speed: 120816.92 tokens/sec | Epoch Time Left: 2:45:11 | Total Time Left: 11:03:14 +[2025-07-04 21:57:21] Epoch 2/4, Step 6250/18020, Loss(triple): 7.799244, Loss(predicate): 6.431081, LR: 0.000168, Speed: 120247.01 tokens/sec | Epoch Time Left: 2:44:27 | Total Time Left: 11:02:31 +[2025-07-04 21:58:02] Epoch 2/4, Step 6300/18020, Loss(triple): 8.584316, Loss(predicate): 7.998240, LR: 0.000168, Speed: 119266.43 tokens/sec | Epoch Time Left: 2:43:44 | Total Time Left: 11:01:48 +[2025-07-04 21:58:43] Epoch 2/4, Step 6350/18020, Loss(triple): 8.074280, Loss(predicate): 7.010376, LR: 0.000167, Speed: 120041.50 tokens/sec | Epoch Time Left: 2:43:00 | Total Time Left: 11:01:06 +[2025-07-04 21:59:24] Epoch 2/4, Step 6400/18020, Loss(triple): 8.111292, Loss(predicate): 12.882050, LR: 0.000167, Speed: 119288.81 tokens/sec | Epoch Time Left: 2:42:17 | Total Time Left: 11:00:23 +[2025-07-04 22:00:05] Epoch 2/4, Step 6450/18020, Loss(triple): 8.492296, Loss(predicate): 11.626332, LR: 0.000167, Speed: 120468.69 tokens/sec | Epoch Time Left: 2:41:33 | Total Time Left: 10:59:40 +[2025-07-04 22:00:46] === GPU性能分析 (平均每步) === +[2025-07-04 22:00:46] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 22:00:46] GPU总时间: 9.92ms, 实际迭代时间: 819.34ms, GPU利用率: 1.2% +[2025-07-04 22:00:46] ================================================== +[2025-07-04 22:00:46] === 三元组预测示例 === +[2025-07-04 22:00:46] 样本1目标: Man on the Train (2011 film) director Mary McGuckian +[2025-07-04 22:00:46] 样本1预测: countryM entygaaran erill,T filin bir ofm or B) +[2025-07-04 22:00:46] 样本2目标: Aksella Luts occupation actress +[2025-07-04 22:00:46] 样本2预测: GKensond Ha.u oniz Dom upation bir of occth J 19 +[2025-07-04 22:00:46] ================== +[2025-07-04 22:00:46] Epoch 2/4, Step 6500/18020, Loss(triple): 7.967758, Loss(predicate): 6.296356, LR: 0.000167, Speed: 119978.89 tokens/sec | Epoch Time Left: 2:40:49 | Total Time Left: 10:58:57 +[2025-07-04 22:01:27] Epoch 2/4, Step 6550/18020, Loss(triple): 8.340937, Loss(predicate): 8.920670, LR: 0.000167, Speed: 119169.04 tokens/sec | Epoch Time Left: 2:40:06 | Total Time Left: 10:58:15 +[2025-07-04 22:02:08] Epoch 2/4, Step 6600/18020, Loss(triple): 8.218981, Loss(predicate): 8.955403, LR: 0.000167, Speed: 120640.50 tokens/sec | Epoch Time Left: 2:39:23 | Total Time Left: 10:57:32 +[2025-07-04 22:02:49] Epoch 2/4, Step 6650/18020, Loss(triple): 8.609011, Loss(predicate): 7.074717, LR: 0.000166, Speed: 119778.77 tokens/sec | Epoch Time Left: 2:38:39 | Total Time Left: 10:56:50 +[2025-07-04 22:03:30] Epoch 2/4, Step 6700/18020, Loss(triple): 8.093994, Loss(predicate): 7.151388, LR: 0.000166, Speed: 120545.59 tokens/sec | Epoch Time Left: 2:37:56 | Total Time Left: 10:56:07 +[2025-07-04 22:04:11] Epoch 2/4, Step 6750/18020, Loss(triple): 8.100895, Loss(predicate): 8.357747, LR: 0.000166, Speed: 120734.58 tokens/sec | Epoch Time Left: 2:37:12 | Total Time Left: 10:55:23 +[2025-07-04 22:04:52] Epoch 2/4, Step 6800/18020, Loss(triple): 7.966522, Loss(predicate): 7.707245, LR: 0.000166, Speed: 119341.01 tokens/sec | Epoch Time Left: 2:36:29 | Total Time Left: 10:54:41 +[2025-07-04 22:05:36] Epoch 2/4, Step 6850/18020, Loss(triple): 8.062540, Loss(predicate): 8.878540, LR: 0.000166, Speed: 111950.19 tokens/sec | Epoch Time Left: 2:35:50 | Total Time Left: 10:54:04 +[2025-07-04 22:06:17] Epoch 2/4, Step 6900/18020, Loss(triple): 7.866695, Loss(predicate): 8.976257, LR: 0.000165, Speed: 117852.80 tokens/sec | Epoch Time Left: 2:35:08 | Total Time Left: 10:53:23 +[2025-07-04 22:06:58] Epoch 2/4, Step 6950/18020, Loss(triple): 8.501055, Loss(predicate): 11.795552, LR: 0.000165, Speed: 120115.45 tokens/sec | Epoch Time Left: 2:34:25 | Total Time Left: 10:52:40 +[2025-07-04 22:07:39] === GPU性能分析 (平均每步) === +[2025-07-04 22:07:39] 前向传播: 7.97ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 22:07:39] GPU总时间: 9.89ms, 实际迭代时间: 814.86ms, GPU利用率: 1.2% +[2025-07-04 22:07:39] ================================================== +[2025-07-04 22:07:39] === 三元组预测示例 === +[2025-07-04 22:07:39] 样本1目标: 2006 Asian Games location Doha +[2025-07-04 22:07:39] 样本1预测: country20 instisasaM G owA,ay 9ass3 ofquion- A +[2025-07-04 22:07:39] 样本2目标: Frank Cady country of citizenship American +[2025-07-04 22:07:39] 样本2预测: RJ6y Hi.am ickyom Americanitiz of chip countryens +[2025-07-04 22:07:39] ================== +[2025-07-04 22:07:39] Epoch 2/4, Step 7000/18020, Loss(triple): 8.450624, Loss(predicate): 12.168223, LR: 0.000165, Speed: 120638.71 tokens/sec | Epoch Time Left: 2:33:41 | Total Time Left: 10:51:57 +[2025-07-04 22:08:20] Epoch 2/4, Step 7050/18020, Loss(triple): 8.498459, Loss(predicate): 7.040894, LR: 0.000165, Speed: 120023.24 tokens/sec | Epoch Time Left: 2:32:58 | Total Time Left: 10:51:14 +[2025-07-04 22:09:01] Epoch 2/4, Step 7100/18020, Loss(triple): 8.354044, Loss(predicate): 9.014445, LR: 0.000165, Speed: 118973.70 tokens/sec | Epoch Time Left: 2:32:15 | Total Time Left: 10:50:32 +[2025-07-04 22:09:42] Epoch 2/4, Step 7150/18020, Loss(triple): 8.681934, Loss(predicate): 10.832997, LR: 0.000164, Speed: 120059.23 tokens/sec | Epoch Time Left: 2:31:32 | Total Time Left: 10:49:49 +[2025-07-04 22:10:23] Epoch 2/4, Step 7200/18020, Loss(triple): 8.037760, Loss(predicate): 7.479513, LR: 0.000164, Speed: 120538.41 tokens/sec | Epoch Time Left: 2:30:49 | Total Time Left: 10:49:06 +[2025-07-04 22:11:04] Epoch 2/4, Step 7250/18020, Loss(triple): 8.219398, Loss(predicate): 10.181447, LR: 0.000164, Speed: 120461.15 tokens/sec | Epoch Time Left: 2:30:06 | Total Time Left: 10:48:23 +[2025-07-04 22:11:45] Epoch 2/4, Step 7300/18020, Loss(triple): 8.142139, Loss(predicate): 17.086874, LR: 0.000164, Speed: 119806.29 tokens/sec | Epoch Time Left: 2:29:23 | Total Time Left: 10:47:41 +[2025-07-04 22:12:26] Epoch 2/4, Step 7350/18020, Loss(triple): 8.045017, Loss(predicate): 11.698364, LR: 0.000164, Speed: 119190.31 tokens/sec | Epoch Time Left: 2:28:40 | Total Time Left: 10:46:59 +[2025-07-04 22:13:07] Epoch 2/4, Step 7400/18020, Loss(triple): 7.609299, Loss(predicate): 10.508514, LR: 0.000164, Speed: 120024.12 tokens/sec | Epoch Time Left: 2:27:57 | Total Time Left: 10:46:16 +[2025-07-04 22:13:48] Epoch 2/4, Step 7450/18020, Loss(triple): 8.098988, Loss(predicate): 11.879730, LR: 0.000163, Speed: 120461.79 tokens/sec | Epoch Time Left: 2:27:14 | Total Time Left: 10:45:33 +[2025-07-04 22:14:28] === GPU性能分析 (平均每步) === +[2025-07-04 22:14:28] 前向传播: 8.03ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 22:14:28] GPU总时间: 9.95ms, 实际迭代时间: 811.58ms, GPU利用率: 1.2% +[2025-07-04 22:14:28] ================================================== +[2025-07-04 22:14:28] === 三元组预测示例 === +[2025-07-04 22:14:28] 样本1目标: Nitin Sawhney occupation composer +[2025-07-04 22:14:28] 样本1预测: countryP origoas (ro ianjzay upation bir of occth Mici +[2025-07-04 22:14:28] 样本2目标: Cagnicourt instance of commune +[2025-07-04 22:14:28] 样本2预测: instC insttun (ee ict-et (ance commun ofance- inst of +[2025-07-04 22:14:28] ================== +[2025-07-04 22:14:28] Epoch 2/4, Step 7500/18020, Loss(triple): 8.338722, Loss(predicate): 10.212708, LR: 0.000163, Speed: 121126.85 tokens/sec | Epoch Time Left: 2:26:30 | Total Time Left: 10:44:50 +[2025-07-04 22:15:09] Epoch 2/4, Step 7550/18020, Loss(triple): 8.154774, Loss(predicate): 8.661885, LR: 0.000163, Speed: 120308.90 tokens/sec | Epoch Time Left: 2:25:47 | Total Time Left: 10:44:07 +[2025-07-04 22:15:50] Epoch 2/4, Step 7600/18020, Loss(triple): 8.201502, Loss(predicate): 10.925679, LR: 0.000163, Speed: 119471.82 tokens/sec | Epoch Time Left: 2:25:05 | Total Time Left: 10:43:25 +[2025-07-04 22:16:31] Epoch 2/4, Step 7650/18020, Loss(triple): 7.958811, Loss(predicate): 5.880808, LR: 0.000163, Speed: 120095.36 tokens/sec | Epoch Time Left: 2:24:22 | Total Time Left: 10:42:42 +[2025-07-04 22:17:12] Epoch 2/4, Step 7700/18020, Loss(triple): 7.790806, Loss(predicate): 11.038716, LR: 0.000162, Speed: 120612.79 tokens/sec | Epoch Time Left: 2:23:38 | Total Time Left: 10:41:59 +[2025-07-04 22:17:53] Epoch 2/4, Step 7750/18020, Loss(triple): 7.888111, Loss(predicate): 10.967163, LR: 0.000162, Speed: 121323.28 tokens/sec | Epoch Time Left: 2:22:55 | Total Time Left: 10:41:16 +[2025-07-04 22:18:33] Epoch 2/4, Step 7800/18020, Loss(triple): 8.199253, Loss(predicate): 11.808309, LR: 0.000162, Speed: 120468.13 tokens/sec | Epoch Time Left: 2:22:12 | Total Time Left: 10:40:33 +[2025-07-04 22:19:15] Epoch 2/4, Step 7850/18020, Loss(triple): 8.285934, Loss(predicate): 9.970683, LR: 0.000162, Speed: 118903.15 tokens/sec | Epoch Time Left: 2:21:30 | Total Time Left: 10:39:51 +[2025-07-04 22:19:56] Epoch 2/4, Step 7900/18020, Loss(triple): 8.383104, Loss(predicate): 5.321503, LR: 0.000162, Speed: 119957.80 tokens/sec | Epoch Time Left: 2:20:47 | Total Time Left: 10:39:08 +[2025-07-04 22:20:37] Epoch 2/4, Step 7950/18020, Loss(triple): 8.255207, Loss(predicate): 9.399231, LR: 0.000161, Speed: 119873.34 tokens/sec | Epoch Time Left: 2:20:04 | Total Time Left: 10:38:26 +[2025-07-04 22:21:17] === GPU性能分析 (平均每步) === +[2025-07-04 22:21:17] 前向传播: 7.96ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 22:21:17] GPU总时间: 9.88ms, 实际迭代时间: 813.82ms, GPU利用率: 1.2% +[2025-07-04 22:21:17] ================================================== +[2025-07-04 22:21:17] === 三元组预测示例 === +[2025-07-04 22:21:17] 样本1目标: Allan Agar place of birth Pontefract +[2025-07-04 22:21:17] 样本1预测: placeJ birllh (rd onill Del ortall footerbth sper +[2025-07-04 22:21:17] 样本2目标: Holmes à Court Gallery instance of art +[2025-07-04 22:21:17] 样本2预测: country20 adomhialMam onill Del Americaniver States ofiaance country Can +[2025-07-04 22:21:17] ================== +[2025-07-04 22:21:17] Epoch 2/4, Step 8000/18020, Loss(triple): 8.494432, Loss(predicate): 6.344879, LR: 0.000161, Speed: 120793.70 tokens/sec | Epoch Time Left: 2:19:21 | Total Time Left: 10:37:43 +[2025-07-04 22:21:58] Epoch 2/4, Step 8050/18020, Loss(triple): 8.117237, Loss(predicate): 11.126454, LR: 0.000161, Speed: 120391.63 tokens/sec | Epoch Time Left: 2:18:39 | Total Time Left: 10:37:00 +[2025-07-04 22:22:40] Epoch 2/4, Step 8100/18020, Loss(triple): 8.282297, Loss(predicate): 8.741821, LR: 0.000161, Speed: 118911.21 tokens/sec | Epoch Time Left: 2:17:56 | Total Time Left: 10:36:18 +[2025-07-04 22:23:20] Epoch 2/4, Step 8150/18020, Loss(triple): 8.003555, Loss(predicate): 9.161366, LR: 0.000161, Speed: 120211.31 tokens/sec | Epoch Time Left: 2:17:14 | Total Time Left: 10:35:35 +[2025-07-04 22:24:02] Epoch 2/4, Step 8200/18020, Loss(triple): 8.023142, Loss(predicate): 11.180491, LR: 0.000161, Speed: 119529.24 tokens/sec | Epoch Time Left: 2:16:31 | Total Time Left: 10:34:53 +[2025-07-04 22:24:54] Epoch 2/4, Step 8250/18020, Loss(triple): 8.283457, Loss(predicate): 12.748851, LR: 0.000160, Speed: 94337.00 tokens/sec | Epoch Time Left: 2:16:02 | Total Time Left: 10:34:30 +[2025-07-04 22:25:47] Epoch 2/4, Step 8300/18020, Loss(triple): 7.833214, Loss(predicate): 9.514689, LR: 0.000160, Speed: 92592.33 tokens/sec | Epoch Time Left: 2:15:33 | Total Time Left: 10:34:09 +[2025-07-04 22:26:37] Epoch 2/4, Step 8350/18020, Loss(triple): 8.099985, Loss(predicate): 17.359629, LR: 0.000160, Speed: 98171.50 tokens/sec | Epoch Time Left: 2:15:01 | Total Time Left: 10:33:42 +[2025-07-04 22:27:26] Epoch 2/4, Step 8400/18020, Loss(triple): 7.863472, Loss(predicate): 11.186046, LR: 0.000160, Speed: 99080.77 tokens/sec | Epoch Time Left: 2:14:28 | Total Time Left: 10:33:14 +[2025-07-04 22:28:09] Epoch 2/4, Step 8450/18020, Loss(triple): 8.098429, Loss(predicate): 10.348724, LR: 0.000160, Speed: 116445.77 tokens/sec | Epoch Time Left: 2:13:46 | Total Time Left: 10:32:33 +[2025-07-04 22:28:50] === GPU性能分析 (平均每步) === +[2025-07-04 22:28:50] 前向传播: 7.94ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 22:28:50] GPU总时间: 9.86ms, 实际迭代时间: 832.64ms, GPU利用率: 1.2% +[2025-07-04 22:28:50] ================================================== +[2025-07-04 22:28:50] === 三元组预测示例 === +[2025-07-04 22:28:50] 样本1目标: Zaio country Morocco +[2025-07-04 22:28:50] 样本1预测: countryO entistgara enin,ay alityance ter oficance- of +[2025-07-04 22:28:50] 样本2目标: The Memory Keeper's Daughter country of origin American +[2025-07-04 22:28:50] 样本2预测: countryThe origanceg (ean eror Dr songin L ofgeance Mer +[2025-07-04 22:28:50] ================== +[2025-07-04 22:28:50] Epoch 2/4, Step 8500/18020, Loss(triple): 8.000183, Loss(predicate): 9.035420, LR: 0.000159, Speed: 118062.69 tokens/sec | Epoch Time Left: 2:13:04 | Total Time Left: 10:31:52 +[2025-07-04 22:29:36] Epoch 2/4, Step 8550/18020, Loss(triple): 8.057587, Loss(predicate): 8.391225, LR: 0.000159, Speed: 106662.82 tokens/sec | Epoch Time Left: 2:12:27 | Total Time Left: 10:31:18 +[2025-07-04 22:30:20] Epoch 2/4, Step 8600/18020, Loss(triple): 8.121286, Loss(predicate): 14.099172, LR: 0.000159, Speed: 112850.67 tokens/sec | Epoch Time Left: 2:11:46 | Total Time Left: 10:30:40 +[2025-07-04 22:31:07] Epoch 2/4, Step 8650/18020, Loss(triple): 8.271515, Loss(predicate): 10.167775, LR: 0.000159, Speed: 104165.48 tokens/sec | Epoch Time Left: 2:11:10 | Total Time Left: 10:30:08 +[2025-07-04 22:31:51] Epoch 2/4, Step 8700/18020, Loss(triple): 8.254005, Loss(predicate): 7.449158, LR: 0.000159, Speed: 111013.03 tokens/sec | Epoch Time Left: 2:10:31 | Total Time Left: 10:29:30 +[2025-07-04 22:32:35] Epoch 2/4, Step 8750/18020, Loss(triple): 7.576557, Loss(predicate): 13.456838, LR: 0.000158, Speed: 112716.12 tokens/sec | Epoch Time Left: 2:09:50 | Total Time Left: 10:28:52 +[2025-07-04 22:33:25] Epoch 2/4, Step 8800/18020, Loss(triple): 8.012300, Loss(predicate): 11.404897, LR: 0.000158, Speed: 97436.82 tokens/sec | Epoch Time Left: 2:09:17 | Total Time Left: 10:28:25 +[2025-07-04 22:34:09] Epoch 2/4, Step 8850/18020, Loss(triple): 7.986557, Loss(predicate): 8.999858, LR: 0.000158, Speed: 112536.32 tokens/sec | Epoch Time Left: 2:08:37 | Total Time Left: 10:27:47 +[2025-07-04 22:34:51] Epoch 2/4, Step 8900/18020, Loss(triple): 8.856043, Loss(predicate): 20.649668, LR: 0.000158, Speed: 118602.52 tokens/sec | Epoch Time Left: 2:07:54 | Total Time Left: 10:27:05 +[2025-07-04 22:35:31] Epoch 2/4, Step 8950/18020, Loss(triple): 8.093033, Loss(predicate): 11.085052, LR: 0.000158, Speed: 120731.49 tokens/sec | Epoch Time Left: 2:07:10 | Total Time Left: 10:26:22 +[2025-07-04 22:36:12] === GPU性能分析 (平均每步) === +[2025-07-04 22:36:12] 前向传播: 7.95ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 22:36:12] GPU总时间: 9.87ms, 实际迭代时间: 814.22ms, GPU利用率: 1.2% +[2025-07-04 22:36:12] ================================================== +[2025-07-04 22:36:12] === 三元组预测示例 === +[2025-07-04 22:36:12] 样本1目标: Marajul located in the administrative territorial entity Bakeshluchay Rural District +[2025-07-04 22:36:12] 样本1预测: countryM entyharan an� Dah ialers Proversian whole, P +[2025-07-04 22:36:12] 样本2目标: Estádio Juca Ribeiro country Brazil +[2025-07-04 22:36:12] 样本2预测: GB insttgaeu zkkak ialance ter ofiar in D +[2025-07-04 22:36:12] ================== +[2025-07-04 22:36:12] Epoch 2/4, Step 9000/18020, Loss(triple): 8.304035, Loss(predicate): 11.769450, LR: 0.000157, Speed: 120733.60 tokens/sec | Epoch Time Left: 2:06:27 | Total Time Left: 10:25:39 +[2025-07-04 22:36:53] Epoch 2/4, Step 9050/18020, Loss(triple): 8.462641, Loss(predicate): 11.677562, LR: 0.000157, Speed: 119869.49 tokens/sec | Epoch Time Left: 2:05:44 | Total Time Left: 10:24:56 +[2025-07-04 22:37:38] Epoch 2/4, Step 9100/18020, Loss(triple): 8.081135, Loss(predicate): 8.798920, LR: 0.000157, Speed: 109496.11 tokens/sec | Epoch Time Left: 2:05:05 | Total Time Left: 10:24:20 +[2025-07-04 22:38:19] Epoch 2/4, Step 9150/18020, Loss(triple): 7.554981, Loss(predicate): 8.770009, LR: 0.000157, Speed: 118864.29 tokens/sec | Epoch Time Left: 2:04:22 | Total Time Left: 10:23:37 +[2025-07-04 22:39:00] Epoch 2/4, Step 9200/18020, Loss(triple): 8.402651, Loss(predicate): 7.326508, LR: 0.000157, Speed: 120605.81 tokens/sec | Epoch Time Left: 2:03:39 | Total Time Left: 10:22:54 +[2025-07-04 22:39:52] Epoch 2/4, Step 9250/18020, Loss(triple): 7.866333, Loss(predicate): 6.556895, LR: 0.000156, Speed: 94435.15 tokens/sec | Epoch Time Left: 2:03:06 | Total Time Left: 10:22:30 +[2025-07-04 22:40:41] Epoch 2/4, Step 9300/18020, Loss(triple): 8.512156, Loss(predicate): 12.526225, LR: 0.000156, Speed: 100579.88 tokens/sec | Epoch Time Left: 2:02:30 | Total Time Left: 10:22:00 +[2025-07-04 22:41:22] Epoch 2/4, Step 9350/18020, Loss(triple): 7.940994, Loss(predicate): 11.918762, LR: 0.000156, Speed: 119618.41 tokens/sec | Epoch Time Left: 2:01:47 | Total Time Left: 10:21:17 +[2025-07-04 22:42:03] Epoch 2/4, Step 9400/18020, Loss(triple): 8.048634, Loss(predicate): 9.375478, LR: 0.000156, Speed: 120752.41 tokens/sec | Epoch Time Left: 2:01:04 | Total Time Left: 10:20:34 +[2025-07-04 22:42:44] Epoch 2/4, Step 9450/18020, Loss(triple): 8.028854, Loss(predicate): 8.939860, LR: 0.000156, Speed: 119519.78 tokens/sec | Epoch Time Left: 2:00:21 | Total Time Left: 10:19:51 +[2025-07-04 22:43:25] === GPU性能分析 (平均每步) === +[2025-07-04 22:43:25] 前向传播: 7.94ms, 损失计算: 0.02ms, 反向传播: 1.95ms, 优化器: 0.00ms +[2025-07-04 22:43:25] GPU总时间: 9.90ms, 实际迭代时间: 816.54ms, GPU利用率: 1.2% +[2025-07-04 22:43:25] ================================================== +[2025-07-04 22:43:25] === 三元组预测示例 === +[2025-07-04 22:43:25] 样本1目标: Neftchi Baku league or competition Azerbaijan Premier League +[2025-07-04 22:43:25] 样本1预测: countryThe�omhiMu onk-ak ortort teradiar Ber +[2025-07-04 22:43:25] 样本2目标: Urth 4 author Peter Stone +[2025-07-04 22:43:25] 样本2预测: countryI instyil (rd ert,et songance bir ofbth Ser +[2025-07-04 22:43:25] ================== +[2025-07-04 22:43:25] Epoch 2/4, Step 9500/18020, Loss(triple): 7.590509, Loss(predicate): 8.242554, LR: 0.000155, Speed: 120391.22 tokens/sec | Epoch Time Left: 1:59:37 | Total Time Left: 10:19:08 +[2025-07-04 22:44:05] Epoch 2/4, Step 9550/18020, Loss(triple): 7.699036, Loss(predicate): 6.447367, LR: 0.000155, Speed: 120967.17 tokens/sec | Epoch Time Left: 1:58:54 | Total Time Left: 10:18:25 +[2025-07-04 22:44:46] Epoch 2/4, Step 9600/18020, Loss(triple): 8.089432, Loss(predicate): 5.771098, LR: 0.000155, Speed: 121401.52 tokens/sec | Epoch Time Left: 1:58:10 | Total Time Left: 10:17:41 +[2025-07-04 22:45:26] Epoch 2/4, Step 9650/18020, Loss(triple): 8.201698, Loss(predicate): 10.185506, LR: 0.000155, Speed: 121078.29 tokens/sec | Epoch Time Left: 1:57:27 | Total Time Left: 10:16:58 +[2025-07-04 22:46:07] Epoch 2/4, Step 9700/18020, Loss(triple): 8.065027, Loss(predicate): 8.409637, LR: 0.000155, Speed: 119672.77 tokens/sec | Epoch Time Left: 1:56:44 | Total Time Left: 10:16:15 +[2025-07-04 22:46:48] Epoch 2/4, Step 9750/18020, Loss(triple): 8.124405, Loss(predicate): 10.067769, LR: 0.000154, Speed: 120312.23 tokens/sec | Epoch Time Left: 1:56:01 | Total Time Left: 10:15:32 +[2025-07-04 22:47:29] Epoch 2/4, Step 9800/18020, Loss(triple): 7.700100, Loss(predicate): 9.331981, LR: 0.000154, Speed: 120977.96 tokens/sec | Epoch Time Left: 1:55:17 | Total Time Left: 10:14:49 +[2025-07-04 22:48:09] Epoch 2/4, Step 9850/18020, Loss(triple): 8.026321, Loss(predicate): 13.852916, LR: 0.000154, Speed: 121205.07 tokens/sec | Epoch Time Left: 1:54:34 | Total Time Left: 10:14:05 +[2025-07-04 22:48:50] Epoch 2/4, Step 9900/18020, Loss(triple): 8.057404, Loss(predicate): 11.541809, LR: 0.000154, Speed: 121039.12 tokens/sec | Epoch Time Left: 1:53:51 | Total Time Left: 10:13:22 +[2025-07-04 22:49:31] Epoch 2/4, Step 9950/18020, Loss(triple): 7.947105, Loss(predicate): 12.664388, LR: 0.000154, Speed: 119633.00 tokens/sec | Epoch Time Left: 1:53:08 | Total Time Left: 10:12:39 +[2025-07-04 22:50:12] === GPU性能分析 (平均每步) === +[2025-07-04 22:50:12] 前向传播: 7.97ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 22:50:12] GPU总时间: 9.89ms, 实际迭代时间: 814.62ms, GPU利用率: 1.2% +[2025-07-04 22:50:12] ================================================== +[2025-07-04 22:50:12] === 三元组预测示例 === +[2025-07-04 22:50:12] 样本1目标: Abbey, Saskatchewan country Canada +[2025-07-04 22:50:12] 样本1预测: countryK adyasaran aninnun ialist locatedorativer the O +[2025-07-04 22:50:12] 样本2目标: Tractor Sazi headquarters location Tabriz +[2025-07-04 22:50:12] 样本2预测: countryJensyhaed ankvil ortall footbb sp sper +[2025-07-04 22:50:12] ================== +[2025-07-04 22:50:12] Epoch 2/4, Step 10000/18020, Loss(triple): 7.657108, Loss(predicate): 8.410706, LR: 0.000153, Speed: 120674.96 tokens/sec | Epoch Time Left: 1:52:25 | Total Time Left: 10:11:56 +[2025-07-04 22:50:53] Epoch 2/4, Step 10050/18020, Loss(triple): 7.863647, Loss(predicate): 10.777568, LR: 0.000153, Speed: 121053.84 tokens/sec | Epoch Time Left: 1:51:42 | Total Time Left: 10:11:13 +[2025-07-04 22:51:33] Epoch 2/4, Step 10100/18020, Loss(triple): 7.709438, Loss(predicate): 10.726140, LR: 0.000153, Speed: 121041.91 tokens/sec | Epoch Time Left: 1:50:58 | Total Time Left: 10:10:30 +[2025-07-04 22:52:14] Epoch 2/4, Step 10150/18020, Loss(triple): 7.889843, Loss(predicate): 12.712845, LR: 0.000153, Speed: 121055.43 tokens/sec | Epoch Time Left: 1:50:15 | Total Time Left: 10:09:46 +[2025-07-04 22:52:55] Epoch 2/4, Step 10200/18020, Loss(triple): 7.883869, Loss(predicate): 12.483856, LR: 0.000153, Speed: 119580.72 tokens/sec | Epoch Time Left: 1:49:33 | Total Time Left: 10:09:04 +[2025-07-04 22:53:36] Epoch 2/4, Step 10250/18020, Loss(triple): 7.894627, Loss(predicate): 10.819142, LR: 0.000152, Speed: 120457.35 tokens/sec | Epoch Time Left: 1:48:50 | Total Time Left: 10:08:21 +[2025-07-04 22:54:16] Epoch 2/4, Step 10300/18020, Loss(triple): 8.241922, Loss(predicate): 16.848755, LR: 0.000152, Speed: 121218.87 tokens/sec | Epoch Time Left: 1:48:06 | Total Time Left: 10:07:37 +[2025-07-04 22:54:57] Epoch 2/4, Step 10350/18020, Loss(triple): 8.191399, Loss(predicate): 11.562184, LR: 0.000152, Speed: 121151.98 tokens/sec | Epoch Time Left: 1:47:23 | Total Time Left: 10:06:54 +[2025-07-04 22:55:37] Epoch 2/4, Step 10400/18020, Loss(triple): 7.864723, Loss(predicate): 9.373494, LR: 0.000152, Speed: 121170.41 tokens/sec | Epoch Time Left: 1:46:40 | Total Time Left: 10:06:11 +[2025-07-04 22:56:18] Epoch 2/4, Step 10450/18020, Loss(triple): 8.475662, Loss(predicate): 13.732320, LR: 0.000152, Speed: 120023.94 tokens/sec | Epoch Time Left: 1:45:58 | Total Time Left: 10:05:28 +[2025-07-04 22:56:59] === GPU性能分析 (平均每步) === +[2025-07-04 22:56:59] 前向传播: 8.01ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 22:56:59] GPU总时间: 9.93ms, 实际迭代时间: 813.85ms, GPU利用率: 1.2% +[2025-07-04 22:56:59] ================================================== +[2025-07-04 22:56:59] === 三元组预测示例 === +[2025-07-04 22:56:59] 样本1目标: Unisławice, Kuyavian-Pomeranian Voivodeship country Poland +[2025-07-04 22:56:59] 样本1预测: countryF entymares olzoand ialist locatedminar in P +[2025-07-04 22:56:59] 样本2目标: Rochelle School of the Arts instance of art school +[2025-07-04 22:56:59] 样本2预测: countryK entyhaeam onob Hil ialist Statesorger the C +[2025-07-04 22:56:59] ================== +[2025-07-04 22:56:59] Epoch 2/4, Step 10500/18020, Loss(triple): 7.796810, Loss(predicate): 10.516978, LR: 0.000151, Speed: 120789.21 tokens/sec | Epoch Time Left: 1:45:15 | Total Time Left: 10:04:45 +[2025-07-04 22:57:40] Epoch 2/4, Step 10550/18020, Loss(triple): 8.187946, Loss(predicate): 9.880595, LR: 0.000151, Speed: 120137.38 tokens/sec | Epoch Time Left: 1:44:32 | Total Time Left: 10:04:02 +[2025-07-04 22:58:20] Epoch 2/4, Step 10600/18020, Loss(triple): 8.065042, Loss(predicate): 6.518860, LR: 0.000151, Speed: 121397.03 tokens/sec | Epoch Time Left: 1:43:49 | Total Time Left: 10:03:19 +[2025-07-04 22:59:01] Epoch 2/4, Step 10650/18020, Loss(triple): 8.207420, Loss(predicate): 7.264771, LR: 0.000151, Speed: 121255.21 tokens/sec | Epoch Time Left: 1:43:06 | Total Time Left: 10:02:35 +[2025-07-04 22:59:42] Epoch 2/4, Step 10700/18020, Loss(triple): 7.763597, Loss(predicate): 12.655803, LR: 0.000150, Speed: 119747.87 tokens/sec | Epoch Time Left: 1:42:23 | Total Time Left: 10:01:53 +[2025-07-04 23:00:23] Epoch 2/4, Step 10750/18020, Loss(triple): 8.258379, Loss(predicate): 6.890564, LR: 0.000150, Speed: 121059.19 tokens/sec | Epoch Time Left: 1:41:40 | Total Time Left: 10:01:10 +[2025-07-04 23:01:04] Epoch 2/4, Step 10800/18020, Loss(triple): 7.922874, Loss(predicate): 9.743378, LR: 0.000150, Speed: 120042.12 tokens/sec | Epoch Time Left: 1:40:58 | Total Time Left: 10:00:27 +[2025-07-04 23:01:45] Epoch 2/4, Step 10850/18020, Loss(triple): 7.736689, Loss(predicate): 11.308716, LR: 0.000150, Speed: 119026.83 tokens/sec | Epoch Time Left: 1:40:15 | Total Time Left: 9:59:45 +[2025-07-04 23:02:31] Epoch 2/4, Step 10900/18020, Loss(triple): 7.873240, Loss(predicate): 9.949727, LR: 0.000150, Speed: 106494.05 tokens/sec | Epoch Time Left: 1:39:36 | Total Time Left: 9:59:10 +[2025-07-04 23:03:20] Epoch 2/4, Step 10950/18020, Loss(triple): 7.667736, Loss(predicate): 7.073944, LR: 0.000149, Speed: 100775.81 tokens/sec | Epoch Time Left: 1:38:59 | Total Time Left: 9:58:39 +[2025-07-04 23:04:04] === GPU性能分析 (平均每步) === +[2025-07-04 23:04:04] 前向传播: 7.97ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 23:04:04] GPU总时间: 9.89ms, 实际迭代时间: 887.30ms, GPU利用率: 1.1% +[2025-07-04 23:04:04] ================================================== +[2025-07-04 23:04:04] === 三元组预测示例 === +[2025-07-04 23:04:04] 样本1目标: Miguel Rimba place of birth Riberalta +[2025-07-04 23:04:04] 样本1预测: SMensyasa ofu z�kak ortall foot ofbth sp 19 +[2025-07-04 23:04:04] 样本2目标: Trout River (Quebec) located in the administrative territorial entity Quebec +[2025-07-04 23:04:04] 样本2预测: countryG entyhiram anillyy ialiver located ofiverance inst C +[2025-07-04 23:04:04] ================== +[2025-07-04 23:04:04] Epoch 2/4, Step 11000/18020, Loss(triple): 8.034172, Loss(predicate): 8.921275, LR: 0.000149, Speed: 110790.43 tokens/sec | Epoch Time Left: 1:38:18 | Total Time Left: 9:58:01 +[2025-07-04 23:04:46] Epoch 2/4, Step 11050/18020, Loss(triple): 8.005802, Loss(predicate): 11.195038, LR: 0.000149, Speed: 117714.41 tokens/sec | Epoch Time Left: 1:37:36 | Total Time Left: 9:57:20 +[2025-07-04 23:05:27] Epoch 2/4, Step 11100/18020, Loss(triple): 8.108360, Loss(predicate): 10.464946, LR: 0.000149, Speed: 119198.22 tokens/sec | Epoch Time Left: 1:36:53 | Total Time Left: 9:56:37 +[2025-07-04 23:06:08] Epoch 2/4, Step 11150/18020, Loss(triple): 8.226244, Loss(predicate): 7.971334, LR: 0.000149, Speed: 120711.01 tokens/sec | Epoch Time Left: 1:36:11 | Total Time Left: 9:55:54 +[2025-07-04 23:06:49] Epoch 2/4, Step 11200/18020, Loss(triple): 7.811922, Loss(predicate): 6.767538, LR: 0.000148, Speed: 120067.21 tokens/sec | Epoch Time Left: 1:35:28 | Total Time Left: 9:55:12 +[2025-07-04 23:07:30] Epoch 2/4, Step 11250/18020, Loss(triple): 7.609619, Loss(predicate): 6.115479, LR: 0.000148, Speed: 119958.05 tokens/sec | Epoch Time Left: 1:34:45 | Total Time Left: 9:54:29 +[2025-07-04 23:08:11] Epoch 2/4, Step 11300/18020, Loss(triple): 7.954014, Loss(predicate): 11.839284, LR: 0.000148, Speed: 119253.81 tokens/sec | Epoch Time Left: 1:34:03 | Total Time Left: 9:53:47 +[2025-07-04 23:08:52] Epoch 2/4, Step 11350/18020, Loss(triple): 8.189713, Loss(predicate): 7.862650, LR: 0.000148, Speed: 119616.56 tokens/sec | Epoch Time Left: 1:33:20 | Total Time Left: 9:53:04 +[2025-07-04 23:09:33] Epoch 2/4, Step 11400/18020, Loss(triple): 7.686579, Loss(predicate): 9.705211, LR: 0.000148, Speed: 120879.31 tokens/sec | Epoch Time Left: 1:32:38 | Total Time Left: 9:52:21 +[2025-07-04 23:10:14] Epoch 2/4, Step 11450/18020, Loss(triple): 7.864155, Loss(predicate): 8.209361, LR: 0.000147, Speed: 120309.56 tokens/sec | Epoch Time Left: 1:31:55 | Total Time Left: 9:51:38 +[2025-07-04 23:10:55] === GPU性能分析 (平均每步) === +[2025-07-04 23:10:55] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 23:10:55] GPU总时间: 9.92ms, 实际迭代时间: 819.41ms, GPU利用率: 1.2% +[2025-07-04 23:10:55] ================================================== +[2025-07-04 23:10:55] === 三元组预测示例 === +[2025-07-04 23:10:55] 样本1目标: Frederick Luther Fowke country of citizenship Canada +[2025-07-04 23:10:55] 样本1预测: GJ6y Hren. H onm Net 4ate bir of dth J 18 +[2025-07-04 23:10:55] 样本2目标: Jean-Guy Talbot sport ice hockey +[2025-07-04 23:10:55] 样本2预测: GJ birrglel erkyel 7ate bir of,th J 19 +[2025-07-04 23:10:55] ================== +[2025-07-04 23:10:55] Epoch 2/4, Step 11500/18020, Loss(triple): 7.565683, Loss(predicate): 12.593160, LR: 0.000147, Speed: 119969.72 tokens/sec | Epoch Time Left: 1:31:12 | Total Time Left: 9:50:55 +[2025-07-04 23:11:39] Epoch 2/4, Step 11550/18020, Loss(triple): 8.345154, Loss(predicate): 9.135590, LR: 0.000147, Speed: 109686.98 tokens/sec | Epoch Time Left: 1:30:32 | Total Time Left: 9:50:18 +[2025-07-04 23:12:24] Epoch 2/4, Step 11600/18020, Loss(triple): 8.083948, Loss(predicate): 13.580770, LR: 0.000147, Speed: 111043.86 tokens/sec | Epoch Time Left: 1:29:51 | Total Time Left: 9:49:40 +[2025-07-04 23:13:06] Epoch 2/4, Step 11650/18020, Loss(triple): 8.361141, Loss(predicate): 6.838328, LR: 0.000146, Speed: 115555.21 tokens/sec | Epoch Time Left: 1:29:10 | Total Time Left: 9:49:00 +[2025-07-04 23:13:48] Epoch 2/4, Step 11700/18020, Loss(triple): 7.577911, Loss(predicate): 6.789551, LR: 0.000146, Speed: 117045.43 tokens/sec | Epoch Time Left: 1:28:28 | Total Time Left: 9:48:19 +[2025-07-04 23:14:29] Epoch 2/4, Step 11750/18020, Loss(triple): 8.691294, Loss(predicate): 10.824574, LR: 0.000146, Speed: 119490.65 tokens/sec | Epoch Time Left: 1:27:45 | Total Time Left: 9:47:36 +[2025-07-04 23:15:10] Epoch 2/4, Step 11800/18020, Loss(triple): 7.821850, Loss(predicate): 10.115468, LR: 0.000146, Speed: 120640.95 tokens/sec | Epoch Time Left: 1:27:03 | Total Time Left: 9:46:53 +[2025-07-04 23:15:58] Epoch 2/4, Step 11850/18020, Loss(triple): 7.982014, Loss(predicate): 11.798676, LR: 0.000146, Speed: 102176.51 tokens/sec | Epoch Time Left: 1:26:24 | Total Time Left: 9:46:21 +[2025-07-04 23:16:40] Epoch 2/4, Step 11900/18020, Loss(triple): 7.257889, Loss(predicate): 5.500305, LR: 0.000145, Speed: 116084.76 tokens/sec | Epoch Time Left: 1:25:42 | Total Time Left: 9:45:40 +[2025-07-04 23:17:22] Epoch 2/4, Step 11950/18020, Loss(triple): 7.830774, Loss(predicate): 12.371572, LR: 0.000145, Speed: 117800.93 tokens/sec | Epoch Time Left: 1:25:00 | Total Time Left: 9:44:59 +[2025-07-04 23:18:03] === GPU性能分析 (平均每步) === +[2025-07-04 23:18:03] 前向传播: 7.96ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 23:18:03] GPU总时间: 9.88ms, 实际迭代时间: 826.02ms, GPU利用率: 1.2% +[2025-07-04 23:18:03] ================================================== +[2025-07-04 23:18:03] === 三元组预测示例 === +[2025-07-04 23:18:03] 样本1目标: Allan H. Dougall date of birth July 17, 1836 +[2025-07-04 23:18:03] 样本1预测: GB6y H.rd oniz,ay 4ate bir of dth 2 18 +[2025-07-04 23:18:03] 样本2目标: Dragon Zakura instance of manga +[2025-07-04 23:18:03] 样本2预测: countryDppohstenon on�kak upationz of occ language inst In +[2025-07-04 23:18:03] ================== +[2025-07-04 23:18:03] Epoch 2/4, Step 12000/18020, Loss(triple): 8.091021, Loss(predicate): 13.565125, LR: 0.000145, Speed: 119008.57 tokens/sec | Epoch Time Left: 1:24:17 | Total Time Left: 9:44:16 +[2025-07-04 23:18:44] Epoch 2/4, Step 12050/18020, Loss(triple): 7.582424, Loss(predicate): 8.231273, LR: 0.000145, Speed: 119815.08 tokens/sec | Epoch Time Left: 1:23:35 | Total Time Left: 9:43:34 +[2025-07-04 23:19:25] Epoch 2/4, Step 12100/18020, Loss(triple): 8.024155, Loss(predicate): 9.295949, LR: 0.000145, Speed: 119972.45 tokens/sec | Epoch Time Left: 1:22:52 | Total Time Left: 9:42:51 +[2025-07-04 23:20:07] Epoch 2/4, Step 12150/18020, Loss(triple): 7.729485, Loss(predicate): 7.713755, LR: 0.000144, Speed: 119742.29 tokens/sec | Epoch Time Left: 1:22:10 | Total Time Left: 9:42:09 +[2025-07-04 23:20:48] Epoch 2/4, Step 12200/18020, Loss(triple): 8.077518, Loss(predicate): 7.813802, LR: 0.000144, Speed: 117810.19 tokens/sec | Epoch Time Left: 1:21:28 | Total Time Left: 9:41:27 +[2025-07-04 23:21:34] Epoch 2/4, Step 12250/18020, Loss(triple): 7.663889, Loss(predicate): 7.333491, LR: 0.000144, Speed: 107792.79 tokens/sec | Epoch Time Left: 1:20:47 | Total Time Left: 9:40:51 +[2025-07-04 23:22:22] Epoch 2/4, Step 12300/18020, Loss(triple): 7.782146, Loss(predicate): 7.166199, LR: 0.000144, Speed: 101539.80 tokens/sec | Epoch Time Left: 1:20:08 | Total Time Left: 9:40:18 +[2025-07-04 23:23:07] Epoch 2/4, Step 12350/18020, Loss(triple): 8.040382, Loss(predicate): 9.618999, LR: 0.000143, Speed: 108934.05 tokens/sec | Epoch Time Left: 1:19:28 | Total Time Left: 9:39:41 +[2025-07-04 23:23:54] Epoch 2/4, Step 12400/18020, Loss(triple): 8.060242, Loss(predicate): 10.059036, LR: 0.000143, Speed: 106191.57 tokens/sec | Epoch Time Left: 1:18:48 | Total Time Left: 9:39:06 +[2025-07-04 23:24:40] Epoch 2/4, Step 12450/18020, Loss(triple): 7.997080, Loss(predicate): 8.217418, LR: 0.000143, Speed: 106245.67 tokens/sec | Epoch Time Left: 1:18:08 | Total Time Left: 9:38:30 +[2025-07-04 23:25:22] === GPU性能分析 (平均每步) === +[2025-07-04 23:25:22] 前向传播: 7.93ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 23:25:22] GPU总时间: 9.85ms, 实际迭代时间: 834.55ms, GPU利用率: 1.2% +[2025-07-04 23:25:22] ================================================== +[2025-07-04 23:25:22] === 三元组预测示例 === +[2025-07-04 23:25:22] 样本1目标: Apex Hill located in the administrative territorial entity Nunavut +[2025-07-04 23:25:22] 样本1预测: countryH entyasaores éjkak ativeity locatedadictr the D +[2025-07-04 23:25:22] 样本2目标: Cnesia taxon rank genus +[2025-07-04 23:25:22] 样本2预测: genC instosonisalmore iciausus axonus rankax gen t +[2025-07-04 23:25:22] ================== +[2025-07-04 23:25:22] Epoch 2/4, Step 12500/18020, Loss(triple): 7.846018, Loss(predicate): 7.979167, LR: 0.000143, Speed: 117793.09 tokens/sec | Epoch Time Left: 1:17:25 | Total Time Left: 9:37:49 +[2025-07-04 23:26:03] Epoch 2/4, Step 12550/18020, Loss(triple): 7.791775, Loss(predicate): 8.768107, LR: 0.000143, Speed: 119495.77 tokens/sec | Epoch Time Left: 1:16:43 | Total Time Left: 9:37:06 +[2025-07-04 23:26:44] Epoch 2/4, Step 12600/18020, Loss(triple): 8.528873, Loss(predicate): 9.820964, LR: 0.000142, Speed: 120445.65 tokens/sec | Epoch Time Left: 1:16:00 | Total Time Left: 9:36:23 +[2025-07-04 23:27:25] Epoch 2/4, Step 12650/18020, Loss(triple): 8.121965, Loss(predicate): 12.211548, LR: 0.000142, Speed: 119671.78 tokens/sec | Epoch Time Left: 1:15:18 | Total Time Left: 9:35:41 +[2025-07-04 23:28:06] Epoch 2/4, Step 12700/18020, Loss(triple): 7.609772, Loss(predicate): 11.199391, LR: 0.000142, Speed: 118540.83 tokens/sec | Epoch Time Left: 1:14:35 | Total Time Left: 9:34:59 +[2025-07-04 23:28:47] Epoch 2/4, Step 12750/18020, Loss(triple): 7.665354, Loss(predicate): 8.646617, LR: 0.000142, Speed: 119892.67 tokens/sec | Epoch Time Left: 1:13:53 | Total Time Left: 9:34:16 +[2025-07-04 23:29:28] Epoch 2/4, Step 12800/18020, Loss(triple): 8.291912, Loss(predicate): 11.670929, LR: 0.000141, Speed: 120044.33 tokens/sec | Epoch Time Left: 1:13:10 | Total Time Left: 9:33:33 +[2025-07-04 23:30:09] Epoch 2/4, Step 12850/18020, Loss(triple): 8.066141, Loss(predicate): 7.161051, LR: 0.000141, Speed: 121091.13 tokens/sec | Epoch Time Left: 1:12:28 | Total Time Left: 9:32:50 +[2025-07-04 23:30:50] Epoch 2/4, Step 12900/18020, Loss(triple): 8.019222, Loss(predicate): 9.912892, LR: 0.000141, Speed: 119571.32 tokens/sec | Epoch Time Left: 1:11:45 | Total Time Left: 9:32:08 +[2025-07-04 23:31:31] Epoch 2/4, Step 12950/18020, Loss(triple): 7.539806, Loss(predicate): 13.086497, LR: 0.000141, Speed: 118989.02 tokens/sec | Epoch Time Left: 1:11:03 | Total Time Left: 9:31:25 +[2025-07-04 23:32:12] === GPU性能分析 (平均每步) === +[2025-07-04 23:32:12] 前向传播: 7.97ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 23:32:12] GPU总时间: 9.89ms, 实际迭代时间: 821.70ms, GPU利用率: 1.2% +[2025-07-04 23:32:12] ================================================== +[2025-07-04 23:32:12] === 三元组预测示例 === +[2025-07-04 23:32:12] 样本1目标: Nannacara aureocephalus taxon rank species +[2025-07-04 23:32:12] 样本1预测: mP entoonaea alia-us iumon species rankaxes t +[2025-07-04 23:32:12] 样本2目标: Volga tributary Kama +[2025-07-04 23:32:12] 样本2预测: countryO entoasaMa alinoun ialist ter ofiaance B M +[2025-07-04 23:32:12] ================== +[2025-07-04 23:32:12] Epoch 2/4, Step 13000/18020, Loss(triple): 7.514309, Loss(predicate): 8.429688, LR: 0.000141, Speed: 119635.23 tokens/sec | Epoch Time Left: 1:10:20 | Total Time Left: 9:30:43 +[2025-07-04 23:32:53] Epoch 2/4, Step 13050/18020, Loss(triple): 8.277439, Loss(predicate): 13.291260, LR: 0.000140, Speed: 119812.76 tokens/sec | Epoch Time Left: 1:09:38 | Total Time Left: 9:30:00 +[2025-07-04 23:33:34] Epoch 2/4, Step 13100/18020, Loss(triple): 8.143707, Loss(predicate): 10.558044, LR: 0.000140, Speed: 120541.26 tokens/sec | Epoch Time Left: 1:08:56 | Total Time Left: 9:29:17 +[2025-07-04 23:34:15] Epoch 2/4, Step 13150/18020, Loss(triple): 7.735485, Loss(predicate): 13.237208, LR: 0.000140, Speed: 120177.82 tokens/sec | Epoch Time Left: 1:08:13 | Total Time Left: 9:28:35 +[2025-07-04 23:34:56] Epoch 2/4, Step 13200/18020, Loss(triple): 7.956226, Loss(predicate): 9.569631, LR: 0.000140, Speed: 119472.25 tokens/sec | Epoch Time Left: 1:07:31 | Total Time Left: 9:27:52 +[2025-07-04 23:35:37] Epoch 2/4, Step 13250/18020, Loss(triple): 7.695751, Loss(predicate): 7.839050, LR: 0.000139, Speed: 120145.88 tokens/sec | Epoch Time Left: 1:06:48 | Total Time Left: 9:27:10 +[2025-07-04 23:36:18] Epoch 2/4, Step 13300/18020, Loss(triple): 7.649317, Loss(predicate): 9.812317, LR: 0.000139, Speed: 119914.63 tokens/sec | Epoch Time Left: 1:06:06 | Total Time Left: 9:26:27 +[2025-07-04 23:36:58] Epoch 2/4, Step 13350/18020, Loss(triple): 7.776787, Loss(predicate): 9.139873, LR: 0.000139, Speed: 121106.15 tokens/sec | Epoch Time Left: 1:05:23 | Total Time Left: 9:25:44 +[2025-07-04 23:37:39] Epoch 2/4, Step 13400/18020, Loss(triple): 7.029007, Loss(predicate): 9.495931, LR: 0.000139, Speed: 120455.84 tokens/sec | Epoch Time Left: 1:04:41 | Total Time Left: 9:25:01 +[2025-07-04 23:38:20] Epoch 2/4, Step 13450/18020, Loss(triple): 7.655466, Loss(predicate): 5.345469, LR: 0.000139, Speed: 120309.19 tokens/sec | Epoch Time Left: 1:03:59 | Total Time Left: 9:24:18 +[2025-07-04 23:39:01] === GPU性能分析 (平均每步) === +[2025-07-04 23:39:01] 前向传播: 7.98ms, 损失计算: 0.02ms, 反向传播: 1.95ms, 优化器: 0.00ms +[2025-07-04 23:39:01] GPU总时间: 9.95ms, 实际迭代时间: 818.95ms, GPU利用率: 1.2% +[2025-07-04 23:39:01] ================================================== +[2025-07-04 23:39:01] === 三元组预测示例 === +[2025-07-04 23:39:01] 样本1目标: Copelatus striaticollis taxon rank species +[2025-07-04 23:39:01] 样本1预测: mM entyonaales ilisusus boon species rankaxes t +[2025-07-04 23:39:01] 样本2目标: Oued Rhiou instance of commune +[2025-07-04 23:39:01] 样本2预测: countryL instlhiaru ania,ay upance country ofitancea R +[2025-07-04 23:39:01] ================== +[2025-07-04 23:39:01] Epoch 2/4, Step 13500/18020, Loss(triple): 7.511158, Loss(predicate): 9.802063, LR: 0.000138, Speed: 120037.27 tokens/sec | Epoch Time Left: 1:03:16 | Total Time Left: 9:23:36 +[2025-07-04 23:39:42] Epoch 2/4, Step 13550/18020, Loss(triple): 7.749260, Loss(predicate): 7.594543, LR: 0.000138, Speed: 119201.07 tokens/sec | Epoch Time Left: 1:02:34 | Total Time Left: 9:22:53 +[2025-07-04 23:40:23] Epoch 2/4, Step 13600/18020, Loss(triple): 8.314156, Loss(predicate): 10.932454, LR: 0.000138, Speed: 120790.08 tokens/sec | Epoch Time Left: 1:01:52 | Total Time Left: 9:22:10 +[2025-07-04 23:41:04] Epoch 2/4, Step 13650/18020, Loss(triple): 8.033493, Loss(predicate): 10.761078, LR: 0.000138, Speed: 120609.72 tokens/sec | Epoch Time Left: 1:01:09 | Total Time Left: 9:21:27 +[2025-07-04 23:41:44] Epoch 2/4, Step 13700/18020, Loss(triple): 8.195343, Loss(predicate): 9.267568, LR: 0.000137, Speed: 120688.36 tokens/sec | Epoch Time Left: 1:00:27 | Total Time Left: 9:20:45 +[2025-07-04 23:42:25] Epoch 2/4, Step 13750/18020, Loss(triple): 7.922405, Loss(predicate): 8.356587, LR: 0.000137, Speed: 119833.93 tokens/sec | Epoch Time Left: 0:59:44 | Total Time Left: 9:20:02 +[2025-07-04 23:43:07] Epoch 2/4, Step 13800/18020, Loss(triple): 7.496399, Loss(predicate): 8.070577, LR: 0.000137, Speed: 118770.28 tokens/sec | Epoch Time Left: 0:59:02 | Total Time Left: 9:19:20 +[2025-07-04 23:43:48] Epoch 2/4, Step 13850/18020, Loss(triple): 8.387598, Loss(predicate): 12.032887, LR: 0.000137, Speed: 120164.70 tokens/sec | Epoch Time Left: 0:58:20 | Total Time Left: 9:18:37 +[2025-07-04 23:44:29] Epoch 2/4, Step 13900/18020, Loss(triple): 7.985039, Loss(predicate): 12.092072, LR: 0.000137, Speed: 119930.30 tokens/sec | Epoch Time Left: 0:57:38 | Total Time Left: 9:17:55 +[2025-07-04 23:45:10] Epoch 2/4, Step 13950/18020, Loss(triple): 8.372314, Loss(predicate): 12.644470, LR: 0.000136, Speed: 120462.13 tokens/sec | Epoch Time Left: 0:56:55 | Total Time Left: 9:17:12 +[2025-07-04 23:45:51] === GPU性能分析 (平均每步) === +[2025-07-04 23:45:51] 前向传播: 7.98ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 23:45:51] GPU总时间: 9.90ms, 实际迭代时间: 819.18ms, GPU利用率: 1.2% +[2025-07-04 23:45:51] ================================================== +[2025-07-04 23:45:51] === 三元组预测示例 === +[2025-07-04 23:45:51] 样本1目标: Kenesa part of Karaite +[2025-07-04 23:45:51] 样本1预测: countryK entistonaara alkhr filon species ofcl ora t +[2025-07-04 23:45:51] 样本2目标: HMS Express (1896) instance of torpedo boat destroyer +[2025-07-04 23:45:51] 样本2预测: GH entyin (mon enP-re 3ist3adcS- A +[2025-07-04 23:45:51] ================== +[2025-07-04 23:45:51] Epoch 2/4, Step 14000/18020, Loss(triple): 7.514154, Loss(predicate): 8.636821, LR: 0.000136, Speed: 120002.88 tokens/sec | Epoch Time Left: 0:56:13 | Total Time Left: 9:16:29 +[2025-07-04 23:46:32] Epoch 2/4, Step 14050/18020, Loss(triple): 8.068256, Loss(predicate): 11.103302, LR: 0.000136, Speed: 118742.80 tokens/sec | Epoch Time Left: 0:55:31 | Total Time Left: 9:15:47 +[2025-07-04 23:47:13] Epoch 2/4, Step 14100/18020, Loss(triple): 7.832493, Loss(predicate): 8.284404, LR: 0.000136, Speed: 120141.51 tokens/sec | Epoch Time Left: 0:54:49 | Total Time Left: 9:15:05 +[2025-07-04 23:47:54] Epoch 2/4, Step 14150/18020, Loss(triple): 8.262905, Loss(predicate): 7.330399, LR: 0.000135, Speed: 119885.13 tokens/sec | Epoch Time Left: 0:54:07 | Total Time Left: 9:14:22 +[2025-07-04 23:48:35] Epoch 2/4, Step 14200/18020, Loss(triple): 7.443470, Loss(predicate): 8.113912, LR: 0.000135, Speed: 119887.48 tokens/sec | Epoch Time Left: 0:53:24 | Total Time Left: 9:13:40 +[2025-07-04 23:49:16] Epoch 2/4, Step 14250/18020, Loss(triple): 7.593729, Loss(predicate): 10.220317, LR: 0.000135, Speed: 120259.49 tokens/sec | Epoch Time Left: 0:52:42 | Total Time Left: 9:12:57 +[2025-07-04 23:49:57] Epoch 2/4, Step 14300/18020, Loss(triple): 7.439579, Loss(predicate): 10.430211, LR: 0.000135, Speed: 119568.10 tokens/sec | Epoch Time Left: 0:52:00 | Total Time Left: 9:12:15 +[2025-07-04 23:50:37] Epoch 2/4, Step 14350/18020, Loss(triple): 8.155693, Loss(predicate): 7.090551, LR: 0.000135, Speed: 120783.57 tokens/sec | Epoch Time Left: 0:51:18 | Total Time Left: 9:11:32 +[2025-07-04 23:51:18] Epoch 2/4, Step 14400/18020, Loss(triple): 7.615465, Loss(predicate): 8.513875, LR: 0.000134, Speed: 120523.97 tokens/sec | Epoch Time Left: 0:50:36 | Total Time Left: 9:10:49 +[2025-07-04 23:51:59] Epoch 2/4, Step 14450/18020, Loss(triple): 7.961060, Loss(predicate): 9.439769, LR: 0.000134, Speed: 120858.49 tokens/sec | Epoch Time Left: 0:49:53 | Total Time Left: 9:10:06 +[2025-07-04 23:52:40] === GPU性能分析 (平均每步) === +[2025-07-04 23:52:40] 前向传播: 7.96ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 23:52:40] GPU总时间: 9.88ms, 实际迭代时间: 819.41ms, GPU利用率: 1.2% +[2025-07-04 23:52:40] ================================================== +[2025-07-04 23:52:40] === 三元组预测示例 === +[2025-07-04 23:52:40] 样本1目标: IntraHealth International instance of non-profit organization +[2025-07-04 23:52:40] 样本1预测: countryI electuras (man ékyov Americanize States ofiaational country In +[2025-07-04 23:52:40] 样本2目标: Lucy Woodward country of citizenship American +[2025-07-04 23:52:40] 样本2预测: placeD7ondritasu enmyre Americanitiz of chip countryens +[2025-07-04 23:52:40] ================== +[2025-07-04 23:52:40] Epoch 2/4, Step 14500/18020, Loss(triple): 7.731236, Loss(predicate): 10.009155, LR: 0.000134, Speed: 119969.79 tokens/sec | Epoch Time Left: 0:49:11 | Total Time Left: 9:09:24 +[2025-07-04 23:53:21] Epoch 2/4, Step 14550/18020, Loss(triple): 7.924221, Loss(predicate): 9.111867, LR: 0.000134, Speed: 119289.72 tokens/sec | Epoch Time Left: 0:48:29 | Total Time Left: 9:08:42 +[2025-07-04 23:54:02] Epoch 2/4, Step 14600/18020, Loss(triple): 7.801149, Loss(predicate): 10.879750, LR: 0.000133, Speed: 120233.26 tokens/sec | Epoch Time Left: 0:47:47 | Total Time Left: 9:07:59 +[2025-07-04 23:54:43] Epoch 2/4, Step 14650/18020, Loss(triple): 7.571888, Loss(predicate): 9.971608, LR: 0.000133, Speed: 120547.84 tokens/sec | Epoch Time Left: 0:47:05 | Total Time Left: 9:07:16 +[2025-07-04 23:55:23] Epoch 2/4, Step 14700/18020, Loss(triple): 8.046452, Loss(predicate): 10.975891, LR: 0.000133, Speed: 120972.27 tokens/sec | Epoch Time Left: 0:46:22 | Total Time Left: 9:06:33 +[2025-07-04 23:56:05] Epoch 2/4, Step 14750/18020, Loss(triple): 7.943714, Loss(predicate): 7.772003, LR: 0.000133, Speed: 119302.29 tokens/sec | Epoch Time Left: 0:45:40 | Total Time Left: 9:05:51 +[2025-07-04 23:56:46] Epoch 2/4, Step 14800/18020, Loss(triple): 7.844486, Loss(predicate): 7.354614, LR: 0.000132, Speed: 119042.91 tokens/sec | Epoch Time Left: 0:44:58 | Total Time Left: 9:05:09 +[2025-07-04 23:57:27] Epoch 2/4, Step 14850/18020, Loss(triple): 7.796793, Loss(predicate): 8.363007, LR: 0.000132, Speed: 119968.92 tokens/sec | Epoch Time Left: 0:44:16 | Total Time Left: 9:04:26 +[2025-07-04 23:58:08] Epoch 2/4, Step 14900/18020, Loss(triple): 7.514994, Loss(predicate): 7.295003, LR: 0.000132, Speed: 120156.05 tokens/sec | Epoch Time Left: 0:43:34 | Total Time Left: 9:03:44 +[2025-07-04 23:58:49] Epoch 2/4, Step 14950/18020, Loss(triple): 7.685940, Loss(predicate): 11.771179, LR: 0.000132, Speed: 119983.42 tokens/sec | Epoch Time Left: 0:42:52 | Total Time Left: 9:03:01 +[2025-07-04 23:59:35] === GPU性能分析 (平均每步) === +[2025-07-04 23:59:35] 前向传播: 8.06ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-04 23:59:35] GPU总时间: 9.98ms, 实际迭代时间: 930.38ms, GPU利用率: 1.1% +[2025-07-04 23:59:35] ================================================== +[2025-07-04 23:59:35] === 三元组预测示例 === +[2025-07-04 23:59:35] 样本1目标: Richard P. Binzel employer MIT +[2025-07-04 23:59:35] 样本1预测: countryMensyelaea erillyet Americanit bir of cth countryens +[2025-07-04 23:59:35] 样本2目标: Epic: The Poetry of War instance of studio album +[2025-07-04 23:59:35] 样本2预测: placeTptyas (rd ict,om umance al ofb stud instio +[2025-07-04 23:59:35] ================== +[2025-07-04 23:59:35] Epoch 2/4, Step 15000/18020, Loss(triple): 7.544062, Loss(predicate): 8.281403, LR: 0.000132, Speed: 105659.89 tokens/sec | Epoch Time Left: 0:42:11 | Total Time Left: 9:02:26 +[2025-07-05 00:00:28] Epoch 2/4, Step 15050/18020, Loss(triple): 7.612118, Loss(predicate): 12.502004, LR: 0.000131, Speed: 92436.05 tokens/sec | Epoch Time Left: 0:41:31 | Total Time Left: 9:01:57 +[2025-07-05 00:01:13] Epoch 2/4, Step 15100/18020, Loss(triple): 7.971729, Loss(predicate): 8.608521, LR: 0.000131, Speed: 110563.30 tokens/sec | Epoch Time Left: 0:40:50 | Total Time Left: 9:01:19 +[2025-07-05 00:01:54] Epoch 2/4, Step 15150/18020, Loss(triple): 8.064777, Loss(predicate): 7.482819, LR: 0.000131, Speed: 119369.10 tokens/sec | Epoch Time Left: 0:40:08 | Total Time Left: 9:00:37 +[2025-07-05 00:02:35] Epoch 2/4, Step 15200/18020, Loss(triple): 7.966888, Loss(predicate): 10.170395, LR: 0.000131, Speed: 120422.46 tokens/sec | Epoch Time Left: 0:39:26 | Total Time Left: 8:59:54 +[2025-07-05 00:03:16] Epoch 2/4, Step 15250/18020, Loss(triple): 7.686213, Loss(predicate): 11.402608, LR: 0.000130, Speed: 120148.25 tokens/sec | Epoch Time Left: 0:38:43 | Total Time Left: 8:59:12 +[2025-07-05 00:03:57] Epoch 2/4, Step 15300/18020, Loss(triple): 8.119442, Loss(predicate): 12.471334, LR: 0.000130, Speed: 120516.93 tokens/sec | Epoch Time Left: 0:38:01 | Total Time Left: 8:58:29 +[2025-07-05 00:04:37] Epoch 2/4, Step 15350/18020, Loss(triple): 7.921381, Loss(predicate): 8.028046, LR: 0.000130, Speed: 121362.15 tokens/sec | Epoch Time Left: 0:37:19 | Total Time Left: 8:57:46 +[2025-07-05 00:05:18] Epoch 2/4, Step 15400/18020, Loss(triple): 8.528217, Loss(predicate): 14.383087, LR: 0.000130, Speed: 121288.19 tokens/sec | Epoch Time Left: 0:36:37 | Total Time Left: 8:57:03 +[2025-07-05 00:05:58] Epoch 2/4, Step 15450/18020, Loss(triple): 7.767677, Loss(predicate): 9.336060, LR: 0.000129, Speed: 120779.23 tokens/sec | Epoch Time Left: 0:35:55 | Total Time Left: 8:56:20 +[2025-07-05 00:06:39] === GPU性能分析 (平均每步) === +[2025-07-05 00:06:39] 前向传播: 8.03ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 00:06:39] GPU总时间: 9.95ms, 实际迭代时间: 818.58ms, GPU利用率: 1.2% +[2025-07-05 00:06:39] ================================================== +[2025-07-05 00:06:39] === 三元组预测示例 === +[2025-07-05 00:06:39] 样本1目标: Northern Ireland Assembly (1982) applies to jurisdiction Northern Ireland +[2025-07-05 00:06:39] 样本1预测: mThe instyil (Mm inkyak 6ia ter oflandion S C +[2025-07-05 00:06:39] 样本2目标: Hebden Bridge located in the administrative territorial entity England +[2025-07-05 00:06:39] 样本2预测: SH origyg (rb alt,es upation composre occ language inster +[2025-07-05 00:06:39] ================== +[2025-07-05 00:06:39] Epoch 2/4, Step 15500/18020, Loss(triple): 7.834194, Loss(predicate): 7.901733, LR: 0.000129, Speed: 120091.39 tokens/sec | Epoch Time Left: 0:35:13 | Total Time Left: 8:55:37 +[2025-07-05 00:07:20] Epoch 2/4, Step 15550/18020, Loss(triple): 8.076288, Loss(predicate): 12.119868, LR: 0.000129, Speed: 120380.61 tokens/sec | Epoch Time Left: 0:34:31 | Total Time Left: 8:54:55 +[2025-07-05 00:08:01] Epoch 2/4, Step 15600/18020, Loss(triple): 8.390747, Loss(predicate): 10.448354, LR: 0.000129, Speed: 120481.67 tokens/sec | Epoch Time Left: 0:33:49 | Total Time Left: 8:54:12 +[2025-07-05 00:08:41] Epoch 2/4, Step 15650/18020, Loss(triple): 8.004587, Loss(predicate): 9.647450, LR: 0.000129, Speed: 121025.90 tokens/sec | Epoch Time Left: 0:33:06 | Total Time Left: 8:53:29 +[2025-07-05 00:09:22] Epoch 2/4, Step 15700/18020, Loss(triple): 7.921543, Loss(predicate): 9.485504, LR: 0.000128, Speed: 120935.93 tokens/sec | Epoch Time Left: 0:32:24 | Total Time Left: 8:52:46 +[2025-07-05 00:10:03] Epoch 2/4, Step 15750/18020, Loss(triple): 7.545536, Loss(predicate): 8.981669, LR: 0.000128, Speed: 119956.30 tokens/sec | Epoch Time Left: 0:31:42 | Total Time Left: 8:52:04 +[2025-07-05 00:10:44] Epoch 2/4, Step 15800/18020, Loss(triple): 7.600677, Loss(predicate): 9.619415, LR: 0.000128, Speed: 120739.73 tokens/sec | Epoch Time Left: 0:31:00 | Total Time Left: 8:51:21 +[2025-07-05 00:11:25] Epoch 2/4, Step 15850/18020, Loss(triple): 7.520197, Loss(predicate): 5.952444, LR: 0.000128, Speed: 120758.20 tokens/sec | Epoch Time Left: 0:30:18 | Total Time Left: 8:50:38 +[2025-07-05 00:12:05] Epoch 2/4, Step 15900/18020, Loss(triple): 7.675272, Loss(predicate): 9.347310, LR: 0.000127, Speed: 121274.62 tokens/sec | Epoch Time Left: 0:29:36 | Total Time Left: 8:49:56 +[2025-07-05 00:12:46] Epoch 2/4, Step 15950/18020, Loss(triple): 7.757998, Loss(predicate): 9.519414, LR: 0.000127, Speed: 120974.49 tokens/sec | Epoch Time Left: 0:28:54 | Total Time Left: 8:49:13 +[2025-07-05 00:13:27] === GPU性能分析 (平均每步) === +[2025-07-05 00:13:27] 前向传播: 8.01ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 00:13:27] GPU总时间: 9.93ms, 实际迭代时间: 819.94ms, GPU利用率: 1.2% +[2025-07-05 00:13:27] ================================================== +[2025-07-05 00:13:27] === 三元组预测示例 === +[2025-07-05 00:13:27] 样本1目标: Venango Township, Crawford County, Pennsylvania located in the administrative territorial entity Crawford County, Pennsylvania +[2025-07-05 00:13:27] 样本1预测: GA entymaord owm,out ativeistountinnsrit, C +[2025-07-05 00:13:27] 样本2目标: Del Webb Explosion genre pop +[2025-07-05 00:13:27] 样本2预测: PoPensyil (aran erillyle upation bir of occth S 19 +[2025-07-05 00:13:27] ================== +[2025-07-05 00:13:27] Epoch 2/4, Step 16000/18020, Loss(triple): 7.421558, Loss(predicate): 10.665538, LR: 0.000127, Speed: 119891.03 tokens/sec | Epoch Time Left: 0:28:12 | Total Time Left: 8:48:30 +[2025-07-05 00:14:07] Epoch 2/4, Step 16050/18020, Loss(triple): 7.603165, Loss(predicate): 6.656138, LR: 0.000127, Speed: 120833.26 tokens/sec | Epoch Time Left: 0:27:30 | Total Time Left: 8:47:48 +[2025-07-05 00:14:48] Epoch 2/4, Step 16100/18020, Loss(triple): 7.949320, Loss(predicate): 12.416911, LR: 0.000126, Speed: 120135.60 tokens/sec | Epoch Time Left: 0:26:48 | Total Time Left: 8:47:05 +[2025-07-05 00:15:29] Epoch 2/4, Step 16150/18020, Loss(triple): 8.181160, Loss(predicate): 12.110026, LR: 0.000126, Speed: 120973.90 tokens/sec | Epoch Time Left: 0:26:06 | Total Time Left: 8:46:22 +[2025-07-05 00:16:09] Epoch 2/4, Step 16200/18020, Loss(triple): 7.701927, Loss(predicate): 7.921570, LR: 0.000126, Speed: 121007.22 tokens/sec | Epoch Time Left: 0:25:24 | Total Time Left: 8:45:39 +[2025-07-05 00:16:50] Epoch 2/4, Step 16250/18020, Loss(triple): 8.052013, Loss(predicate): 8.288524, LR: 0.000126, Speed: 120379.22 tokens/sec | Epoch Time Left: 0:24:42 | Total Time Left: 8:44:57 +[2025-07-05 00:17:31] Epoch 2/4, Step 16300/18020, Loss(triple): 7.988703, Loss(predicate): 11.710521, LR: 0.000125, Speed: 121169.20 tokens/sec | Epoch Time Left: 0:24:00 | Total Time Left: 8:44:14 +[2025-07-05 00:18:12] Epoch 2/4, Step 16350/18020, Loss(triple): 7.705505, Loss(predicate): 9.802165, LR: 0.000125, Speed: 120297.37 tokens/sec | Epoch Time Left: 0:23:18 | Total Time Left: 8:43:32 +[2025-07-05 00:18:52] Epoch 2/4, Step 16400/18020, Loss(triple): 8.022923, Loss(predicate): 8.584137, LR: 0.000125, Speed: 120803.99 tokens/sec | Epoch Time Left: 0:22:36 | Total Time Left: 8:42:49 +[2025-07-05 00:19:33] Epoch 2/4, Step 16450/18020, Loss(triple): 7.686981, Loss(predicate): 10.028931, LR: 0.000125, Speed: 120894.46 tokens/sec | Epoch Time Left: 0:21:54 | Total Time Left: 8:42:06 +[2025-07-05 00:20:14] === GPU性能分析 (平均每步) === +[2025-07-05 00:20:14] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 00:20:14] GPU总时间: 9.92ms, 实际迭代时间: 816.25ms, GPU利用率: 1.2% +[2025-07-05 00:20:14] ================================================== +[2025-07-05 00:20:14] === 三元组预测示例 === +[2025-07-05 00:20:14] 样本1目标: Reyshawn Terry country of citizenship American +[2025-07-05 00:20:14] 样本1预测: GD biryhberel en�yay Americanitiz of chip countryens +[2025-07-05 00:20:14] 样本2目标: Luka Žorić sport professional basketball player +[2025-07-05 00:20:14] 样本2预测: SLensyki.u an�yak 4ate bir of dth S 19 +[2025-07-05 00:20:14] ================== +[2025-07-05 00:20:14] Epoch 2/4, Step 16500/18020, Loss(triple): 7.831501, Loss(predicate): 8.903432, LR: 0.000125, Speed: 120433.83 tokens/sec | Epoch Time Left: 0:21:12 | Total Time Left: 8:41:24 +[2025-07-05 00:20:55] Epoch 2/4, Step 16550/18020, Loss(triple): 7.704432, Loss(predicate): 8.520314, LR: 0.000124, Speed: 120519.59 tokens/sec | Epoch Time Left: 0:20:30 | Total Time Left: 8:40:41 +[2025-07-05 00:21:36] Epoch 2/4, Step 16600/18020, Loss(triple): 8.040812, Loss(predicate): 11.511622, LR: 0.000124, Speed: 119726.21 tokens/sec | Epoch Time Left: 0:19:48 | Total Time Left: 8:39:59 +[2025-07-05 00:22:17] Epoch 2/4, Step 16650/18020, Loss(triple): 7.770401, Loss(predicate): 9.979645, LR: 0.000124, Speed: 120492.77 tokens/sec | Epoch Time Left: 0:19:06 | Total Time Left: 8:39:16 +[2025-07-05 00:22:57] Epoch 2/4, Step 16700/18020, Loss(triple): 8.068531, Loss(predicate): 5.650095, LR: 0.000124, Speed: 121037.22 tokens/sec | Epoch Time Left: 0:18:24 | Total Time Left: 8:38:33 +[2025-07-05 00:23:38] Epoch 2/4, Step 16750/18020, Loss(triple): 7.812708, Loss(predicate): 5.962179, LR: 0.000123, Speed: 121226.12 tokens/sec | Epoch Time Left: 0:17:42 | Total Time Left: 8:37:51 +[2025-07-05 00:24:18] Epoch 2/4, Step 16800/18020, Loss(triple): 8.064869, Loss(predicate): 7.518117, LR: 0.000123, Speed: 120987.89 tokens/sec | Epoch Time Left: 0:17:00 | Total Time Left: 8:37:08 +[2025-07-05 00:24:59] Epoch 2/4, Step 16850/18020, Loss(triple): 7.977564, Loss(predicate): 12.711568, LR: 0.000123, Speed: 119690.24 tokens/sec | Epoch Time Left: 0:16:18 | Total Time Left: 8:36:26 +[2025-07-05 00:25:40] Epoch 2/4, Step 16900/18020, Loss(triple): 8.289507, Loss(predicate): 7.593302, LR: 0.000123, Speed: 120468.15 tokens/sec | Epoch Time Left: 0:15:37 | Total Time Left: 8:35:43 +[2025-07-05 00:26:21] Epoch 2/4, Step 16950/18020, Loss(triple): 7.739807, Loss(predicate): 9.303752, LR: 0.000122, Speed: 120351.42 tokens/sec | Epoch Time Left: 0:14:55 | Total Time Left: 8:35:01 +[2025-07-05 00:27:02] === GPU性能分析 (平均每步) === +[2025-07-05 00:27:02] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 00:27:02] GPU总时间: 9.92ms, 实际迭代时间: 812.05ms, GPU利用率: 1.2% +[2025-07-05 00:27:02] ================================================== +[2025-07-05 00:27:02] === 三元组预测示例 === +[2025-07-05 00:27:02] 样本1目标: João Nunes (footballer, born 1995) sport footballer +[2025-07-05 00:27:02] 样本1预测: placeJensoelaeno on�nak ortall footerb sp spim +[2025-07-05 00:27:02] 样本2目标: Lennie Bush place of birth London +[2025-07-05 00:27:02] 样本2预测: placeDensyh (eon onob Met 4ate bir of dth 2 19 +[2025-07-05 00:27:02] ================== +[2025-07-05 00:27:02] Epoch 2/4, Step 17000/18020, Loss(triple): 7.874474, Loss(predicate): 11.438893, LR: 0.000122, Speed: 121055.89 tokens/sec | Epoch Time Left: 0:14:13 | Total Time Left: 8:34:18 +[2025-07-05 00:27:42] Epoch 2/4, Step 17050/18020, Loss(triple): 7.525713, Loss(predicate): 15.762441, LR: 0.000122, Speed: 120905.18 tokens/sec | Epoch Time Left: 0:13:31 | Total Time Left: 8:33:35 +[2025-07-05 00:28:23] Epoch 2/4, Step 17100/18020, Loss(triple): 8.001438, Loss(predicate): 9.401326, LR: 0.000122, Speed: 119691.73 tokens/sec | Epoch Time Left: 0:12:49 | Total Time Left: 8:32:53 +[2025-07-05 00:29:04] Epoch 2/4, Step 17150/18020, Loss(triple): 7.690424, Loss(predicate): 9.021525, LR: 0.000121, Speed: 120937.88 tokens/sec | Epoch Time Left: 0:12:07 | Total Time Left: 8:32:10 +[2025-07-05 00:29:45] Epoch 2/4, Step 17200/18020, Loss(triple): 7.861801, Loss(predicate): 13.436981, LR: 0.000121, Speed: 119801.16 tokens/sec | Epoch Time Left: 0:11:25 | Total Time Left: 8:31:28 +[2025-07-05 00:30:26] Epoch 2/4, Step 17250/18020, Loss(triple): 7.341709, Loss(predicate): 13.806691, LR: 0.000121, Speed: 121042.20 tokens/sec | Epoch Time Left: 0:10:43 | Total Time Left: 8:30:45 +[2025-07-05 00:31:06] Epoch 2/4, Step 17300/18020, Loss(triple): 8.313717, Loss(predicate): 6.730291, LR: 0.000121, Speed: 121067.69 tokens/sec | Epoch Time Left: 0:10:02 | Total Time Left: 8:30:03 +[2025-07-05 00:31:47] Epoch 2/4, Step 17350/18020, Loss(triple): 7.826641, Loss(predicate): 7.093414, LR: 0.000121, Speed: 120736.90 tokens/sec | Epoch Time Left: 0:09:20 | Total Time Left: 8:29:20 +[2025-07-05 00:32:28] Epoch 2/4, Step 17400/18020, Loss(triple): 7.802551, Loss(predicate): 12.014648, LR: 0.000120, Speed: 120451.00 tokens/sec | Epoch Time Left: 0:08:38 | Total Time Left: 8:28:38 +[2025-07-05 00:33:09] Epoch 2/4, Step 17450/18020, Loss(triple): 7.875462, Loss(predicate): 7.731425, LR: 0.000120, Speed: 120043.11 tokens/sec | Epoch Time Left: 0:07:56 | Total Time Left: 8:27:55 +[2025-07-05 00:33:49] === GPU性能分析 (平均每步) === +[2025-07-05 00:33:49] 前向传播: 8.02ms, 损失计算: 0.02ms, 反向传播: 1.95ms, 优化器: 0.00ms +[2025-07-05 00:33:49] GPU总时间: 9.98ms, 实际迭代时间: 815.85ms, GPU利用率: 1.2% +[2025-07-05 00:33:49] ================================================== +[2025-07-05 00:33:49] === 三元组预测示例 === +[2025-07-05 00:33:49] 样本1目标: Mantronik place of birth Jamaica +[2025-07-05 00:33:49] 样本1预测: countryBensyh (end ick Mom Americanitiz of chip countryens +[2025-07-05 00:33:49] 样本2目标: Margical History Tour part of the series The Simpsons +[2025-07-05 00:33:49] 样本2预测: LThe participosery (aran ent Des songing L ofgeance Mer +[2025-07-05 00:33:49] ================== +[2025-07-05 00:33:49] Epoch 2/4, Step 17500/18020, Loss(triple): 7.727058, Loss(predicate): 13.745534, LR: 0.000120, Speed: 120493.46 tokens/sec | Epoch Time Left: 0:07:14 | Total Time Left: 8:27:13 +[2025-07-05 00:34:30] Epoch 2/4, Step 17550/18020, Loss(triple): 7.498562, Loss(predicate): 6.582774, LR: 0.000120, Speed: 121182.30 tokens/sec | Epoch Time Left: 0:06:32 | Total Time Left: 8:26:30 +[2025-07-05 00:35:11] Epoch 2/4, Step 17600/18020, Loss(triple): 7.906548, Loss(predicate): 13.615001, LR: 0.000119, Speed: 121135.46 tokens/sec | Epoch Time Left: 0:05:51 | Total Time Left: 8:25:47 +[2025-07-05 00:35:51] Epoch 2/4, Step 17650/18020, Loss(triple): 8.291437, Loss(predicate): 8.351298, LR: 0.000119, Speed: 121068.83 tokens/sec | Epoch Time Left: 0:05:09 | Total Time Left: 8:25:05 +[2025-07-05 00:36:32] Epoch 2/4, Step 17700/18020, Loss(triple): 7.513285, Loss(predicate): 10.183736, LR: 0.000119, Speed: 120134.02 tokens/sec | Epoch Time Left: 0:04:27 | Total Time Left: 8:24:22 +[2025-07-05 00:37:13] Epoch 2/4, Step 17750/18020, Loss(triple): 7.237061, Loss(predicate): 8.899800, LR: 0.000119, Speed: 120734.53 tokens/sec | Epoch Time Left: 0:03:45 | Total Time Left: 8:23:40 +[2025-07-05 00:37:54] Epoch 2/4, Step 17800/18020, Loss(triple): 7.962360, Loss(predicate): 7.820760, LR: 0.000118, Speed: 120610.97 tokens/sec | Epoch Time Left: 0:03:03 | Total Time Left: 8:22:57 +[2025-07-05 00:38:34] Epoch 2/4, Step 17850/18020, Loss(triple): 7.619013, Loss(predicate): 9.827972, LR: 0.000118, Speed: 120808.85 tokens/sec | Epoch Time Left: 0:02:22 | Total Time Left: 8:22:15 +[2025-07-05 00:39:15] Epoch 2/4, Step 17900/18020, Loss(triple): 7.273239, Loss(predicate): 8.308766, LR: 0.000118, Speed: 121086.28 tokens/sec | Epoch Time Left: 0:01:40 | Total Time Left: 8:21:32 +[2025-07-05 00:39:55] Epoch 2/4, Step 17950/18020, Loss(triple): 7.653469, Loss(predicate): 11.286427, LR: 0.000118, Speed: 120998.64 tokens/sec | Epoch Time Left: 0:00:58 | Total Time Left: 8:20:50 +[2025-07-05 00:40:36] === GPU性能分析 (平均每步) === +[2025-07-05 00:40:36] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 00:40:36] GPU总时间: 9.92ms, 实际迭代时间: 815.08ms, GPU利用率: 1.2% +[2025-07-05 00:40:36] ================================================== +[2025-07-05 00:40:36] === 三元组预测示例 === +[2025-07-05 00:40:36] 样本1目标: Live for Loving You performer Gloria Estefan +[2025-07-05 00:40:36] 样本1预测: AmericanThe origory ()o erM,T songin performerm T Mer +[2025-07-05 00:40:36] 样本2目标: Israil Gurung sport footballer +[2025-07-05 00:40:36] 样本2预测: GCh entyilstenl erknov ortall footerb sp sp� +[2025-07-05 00:40:36] ================== +[2025-07-05 00:40:36] Epoch 2/4, Step 18000/18020, Loss(triple): 7.440418, Loss(predicate): 7.280324, LR: 0.000117, Speed: 120607.28 tokens/sec | Epoch Time Left: 0:00:16 | Total Time Left: 8:20:07 +[2025-07-05 00:40:56] 第2轮训练完成，进行内存清理 +[2025-07-05 00:40:58] [Memory Monitor] Epoch 2 completed - System RSS: 27059.48MB, CUDA allocated: 550.62MB, CUDA reserved: 1310.00MB +[2025-07-05 00:40:58] 开始第3轮训练 +[2025-07-05 00:40:58] Set freeze_embedding=True for epoch 2, step 0 +[2025-07-05 00:40:58] 三元组提取训练模式 +[2025-07-05 00:40:58] 使用预tokenized三元组目标数据 +[2025-07-05 00:41:00] Model saved to out/pretrain_cls512.pth +[2025-07-05 00:41:39] Epoch 3/4, Step 50/18020, Loss(triple): 7.863266, Loss(predicate): 11.413544, LR: 0.000117, Speed: 117893.78 tokens/sec | Epoch Time Left: 4:09:44 | Total Time Left: 8:19:14 +[2025-07-05 00:42:20] Epoch 3/4, Step 100/18020, Loss(triple): 7.298143, Loss(predicate): 9.506307, LR: 0.000117, Speed: 120301.16 tokens/sec | Epoch Time Left: 4:06:32 | Total Time Left: 8:18:31 +[2025-07-05 00:42:53] Model saved to out/pretrain_cls512.pth +[2025-07-05 00:43:02] Epoch 3/4, Step 150/18020, Loss(triple): 7.634363, Loss(predicate): 5.154765, LR: 0.000117, Speed: 117897.04 tokens/sec | Epoch Time Left: 4:06:41 | Total Time Left: 8:17:50 +[2025-07-05 00:43:43] Epoch 3/4, Step 200/18020, Loss(triple): 7.704468, Loss(predicate): 10.366506, LR: 0.000116, Speed: 120901.80 tokens/sec | Epoch Time Left: 4:04:52 | Total Time Left: 8:17:07 +[2025-07-05 00:44:23] Epoch 3/4, Step 250/18020, Loss(triple): 7.983997, Loss(predicate): 11.306987, LR: 0.000116, Speed: 121388.73 tokens/sec | Epoch Time Left: 4:03:18 | Total Time Left: 8:16:25 +[2025-07-05 00:45:04] Epoch 3/4, Step 300/18020, Loss(triple): 7.751877, Loss(predicate): 9.779103, LR: 0.000116, Speed: 121199.88 tokens/sec | Epoch Time Left: 4:02:06 | Total Time Left: 8:15:42 +[2025-07-05 00:45:45] Epoch 3/4, Step 350/18020, Loss(triple): 7.450825, Loss(predicate): 8.671733, LR: 0.000116, Speed: 120032.57 tokens/sec | Epoch Time Left: 4:01:23 | Total Time Left: 8:15:00 +[2025-07-05 00:46:25] Epoch 3/4, Step 400/18020, Loss(triple): 7.664288, Loss(predicate): 7.947459, LR: 0.000115, Speed: 121188.31 tokens/sec | Epoch Time Left: 4:00:24 | Total Time Left: 8:14:17 +[2025-07-05 00:47:06] Epoch 3/4, Step 450/18020, Loss(triple): 7.394943, Loss(predicate): 10.299759, LR: 0.000115, Speed: 120513.50 tokens/sec | Epoch Time Left: 3:59:37 | Total Time Left: 8:13:35 +[2025-07-05 00:47:47] === GPU性能分析 (平均每步) === +[2025-07-05 00:47:47] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.91ms, 优化器: 0.00ms +[2025-07-05 00:47:47] GPU总时间: 9.93ms, 实际迭代时间: 812.30ms, GPU利用率: 1.2% +[2025-07-05 00:47:47] ================================================== +[2025-07-05 00:47:47] === 三元组预测示例 === +[2025-07-05 00:47:47] 样本1目标: Harpalus apache taxon rank species +[2025-07-05 00:47:47] 样本1预测: mH adramisarr idpusus (on species rankax gen t +[2025-07-05 00:47:47] 样本2目标: Alex Goude place of birth Neuilly-sur-Seine +[2025-07-05 00:47:47] 样本2预测: GB biryryael onk-ov upation bir of occth Bor +[2025-07-05 00:47:47] ================== +[2025-07-05 00:47:47] Epoch 3/4, Step 500/18020, Loss(triple): 7.757477, Loss(predicate): 12.714192, LR: 0.000115, Speed: 121019.77 tokens/sec | Epoch Time Left: 3:58:46 | Total Time Left: 8:12:52 +[2025-07-05 00:48:27] Epoch 3/4, Step 550/18020, Loss(triple): 8.129456, Loss(predicate): 6.590566, LR: 0.000115, Speed: 121106.44 tokens/sec | Epoch Time Left: 3:57:55 | Total Time Left: 8:12:10 +[2025-07-05 00:48:34] Model saved to out/pretrain_cls512.pth +[2025-07-05 00:49:09] Epoch 3/4, Step 600/18020, Loss(triple): 7.360239, Loss(predicate): 8.906230, LR: 0.000114, Speed: 117928.20 tokens/sec | Epoch Time Left: 3:57:38 | Total Time Left: 8:11:28 +[2025-07-05 00:49:50] Epoch 3/4, Step 650/18020, Loss(triple): 7.699226, Loss(predicate): 11.280660, LR: 0.000114, Speed: 120779.77 tokens/sec | Epoch Time Left: 3:56:51 | Total Time Left: 8:10:46 +[2025-07-05 00:50:30] Epoch 3/4, Step 700/18020, Loss(triple): 7.474915, Loss(predicate): 10.960317, LR: 0.000114, Speed: 120298.85 tokens/sec | Epoch Time Left: 3:56:09 | Total Time Left: 8:10:03 +[2025-07-05 00:51:11] Epoch 3/4, Step 750/18020, Loss(triple): 8.191460, Loss(predicate): 10.494222, LR: 0.000114, Speed: 120817.28 tokens/sec | Epoch Time Left: 3:55:23 | Total Time Left: 8:09:21 +[2025-07-05 00:51:52] Epoch 3/4, Step 800/18020, Loss(triple): 7.190292, Loss(predicate): 12.814555, LR: 0.000114, Speed: 121047.37 tokens/sec | Epoch Time Left: 3:54:36 | Total Time Left: 8:08:38 +[2025-07-05 00:52:32] Epoch 3/4, Step 850/18020, Loss(triple): 7.627954, Loss(predicate): 9.557628, LR: 0.000113, Speed: 120903.37 tokens/sec | Epoch Time Left: 3:53:51 | Total Time Left: 8:07:56 +[2025-07-05 00:53:13] Epoch 3/4, Step 900/18020, Loss(triple): 7.594570, Loss(predicate): 8.305766, LR: 0.000113, Speed: 120342.98 tokens/sec | Epoch Time Left: 3:53:09 | Total Time Left: 8:07:13 +[2025-07-05 00:53:54] Epoch 3/4, Step 950/18020, Loss(triple): 7.636948, Loss(predicate): 7.551508, LR: 0.000113, Speed: 120375.68 tokens/sec | Epoch Time Left: 3:52:28 | Total Time Left: 8:06:31 +[2025-07-05 00:54:35] === GPU性能分析 (平均每步) === +[2025-07-05 00:54:35] 前向传播: 7.98ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 00:54:35] GPU总时间: 9.90ms, 实际迭代时间: 815.36ms, GPU利用率: 1.2% +[2025-07-05 00:54:35] ================================================== +[2025-07-05 00:54:35] === 三元组预测示例 === +[2025-07-05 00:54:35] 样本1目标: R v Beaulac applies to jurisdiction Canada +[2025-07-05 00:54:35] 样本1预测: SR entyilaall inM Ire Fionance ofcS S A +[2025-07-05 00:54:35] 样本2目标: Eija Krogerus place of birth Helsinki +[2025-07-05 00:54:35] 样本2预测: placeKensukiara ank,ay 4ate bir of dth 2 19 +[2025-07-05 00:54:35] ================== +[2025-07-05 00:54:35] Epoch 3/4, Step 1000/18020, Loss(triple): 7.855988, Loss(predicate): 9.252869, LR: 0.000113, Speed: 120564.52 tokens/sec | Epoch Time Left: 3:51:46 | Total Time Left: 8:05:49 +[2025-07-05 00:55:15] Epoch 3/4, Step 1050/18020, Loss(triple): 7.304949, Loss(predicate): 15.632507, LR: 0.000112, Speed: 120967.22 tokens/sec | Epoch Time Left: 3:51:01 | Total Time Left: 8:05:06 +[2025-07-05 00:55:56] Epoch 3/4, Step 1100/18020, Loss(triple): 7.563259, Loss(predicate): 10.233989, LR: 0.000112, Speed: 121297.56 tokens/sec | Epoch Time Left: 3:50:16 | Total Time Left: 8:04:24 +[2025-07-05 00:56:37] Epoch 3/4, Step 1150/18020, Loss(triple): 8.061487, Loss(predicate): 4.627279, LR: 0.000112, Speed: 120732.78 tokens/sec | Epoch Time Left: 3:49:33 | Total Time Left: 8:03:41 +[2025-07-05 00:57:18] Epoch 3/4, Step 1200/18020, Loss(triple): 7.631809, Loss(predicate): 12.142741, LR: 0.000112, Speed: 119926.54 tokens/sec | Epoch Time Left: 3:48:54 | Total Time Left: 8:02:59 +[2025-07-05 00:57:59] Epoch 3/4, Step 1250/18020, Loss(triple): 7.332247, Loss(predicate): 8.951457, LR: 0.000111, Speed: 119979.22 tokens/sec | Epoch Time Left: 3:48:16 | Total Time Left: 8:02:17 +[2025-07-05 00:58:39] Epoch 3/4, Step 1300/18020, Loss(triple): 7.463503, Loss(predicate): 11.425792, LR: 0.000111, Speed: 121122.34 tokens/sec | Epoch Time Left: 3:47:31 | Total Time Left: 8:01:34 +[2025-07-05 00:59:20] Epoch 3/4, Step 1350/18020, Loss(triple): 7.453012, Loss(predicate): 10.312912, LR: 0.000111, Speed: 121157.61 tokens/sec | Epoch Time Left: 3:46:47 | Total Time Left: 8:00:52 +[2025-07-05 01:00:00] Epoch 3/4, Step 1400/18020, Loss(triple): 7.818134, Loss(predicate): 10.849010, LR: 0.000111, Speed: 120965.31 tokens/sec | Epoch Time Left: 3:46:04 | Total Time Left: 8:00:09 +[2025-07-05 01:00:41] Epoch 3/4, Step 1450/18020, Loss(triple): 7.929598, Loss(predicate): 12.057770, LR: 0.000110, Speed: 119999.30 tokens/sec | Epoch Time Left: 3:45:25 | Total Time Left: 7:59:27 +[2025-07-05 01:01:22] === GPU性能分析 (平均每步) === +[2025-07-05 01:01:22] 前向传播: 8.08ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 01:01:22] GPU总时间: 10.00ms, 实际迭代时间: 814.65ms, GPU利用率: 1.2% +[2025-07-05 01:01:22] ================================================== +[2025-07-05 01:01:22] === 三元组预测示例 === +[2025-07-05 01:01:22] 样本1目标: Show Low located in the administrative territorial entity Navajo County, Arizona +[2025-07-05 01:01:22] 样本1预测: countryS adyhaalan on�,ay ialist Statesorityr, C +[2025-07-05 01:01:22] 样本2目标: Yuki Shimizu occupation manga artist +[2025-07-05 01:01:22] 样本2预测: countryThe�ukiiu aniezay filance orig ofitth inst Pro +[2025-07-05 01:01:22] ================== +[2025-07-05 01:01:22] Epoch 3/4, Step 1500/18020, Loss(triple): 7.616734, Loss(predicate): 9.620300, LR: 0.000110, Speed: 120670.23 tokens/sec | Epoch Time Left: 3:44:44 | Total Time Left: 7:58:45 +[2025-07-05 01:02:03] Epoch 3/4, Step 1550/18020, Loss(triple): 7.941622, Loss(predicate): 8.747182, LR: 0.000110, Speed: 120225.12 tokens/sec | Epoch Time Left: 3:44:04 | Total Time Left: 7:58:03 +[2025-07-05 01:02:44] Epoch 3/4, Step 1600/18020, Loss(triple): 7.883350, Loss(predicate): 8.705383, LR: 0.000110, Speed: 121218.33 tokens/sec | Epoch Time Left: 3:43:20 | Total Time Left: 7:57:20 +[2025-07-05 01:03:24] Epoch 3/4, Step 1650/18020, Loss(triple): 7.752230, Loss(predicate): 8.291860, LR: 0.000109, Speed: 121160.03 tokens/sec | Epoch Time Left: 3:42:37 | Total Time Left: 7:56:37 +[2025-07-05 01:04:05] Epoch 3/4, Step 1700/18020, Loss(triple): 7.968653, Loss(predicate): 10.731730, LR: 0.000109, Speed: 120261.55 tokens/sec | Epoch Time Left: 3:41:57 | Total Time Left: 7:55:55 +[2025-07-05 01:04:46] Epoch 3/4, Step 1750/18020, Loss(triple): 7.557671, Loss(predicate): 9.414876, LR: 0.000109, Speed: 120965.74 tokens/sec | Epoch Time Left: 3:41:14 | Total Time Left: 7:55:13 +[2025-07-05 01:05:26] Epoch 3/4, Step 1800/18020, Loss(triple): 7.911936, Loss(predicate): 7.959671, LR: 0.000109, Speed: 120322.21 tokens/sec | Epoch Time Left: 3:40:34 | Total Time Left: 7:54:31 +[2025-07-05 01:06:07] Epoch 3/4, Step 1850/18020, Loss(triple): 7.368202, Loss(predicate): 9.612956, LR: 0.000108, Speed: 120695.92 tokens/sec | Epoch Time Left: 3:39:53 | Total Time Left: 7:53:48 +[2025-07-05 01:06:48] Epoch 3/4, Step 1900/18020, Loss(triple): 7.904743, Loss(predicate): 8.824463, LR: 0.000108, Speed: 121015.36 tokens/sec | Epoch Time Left: 3:39:10 | Total Time Left: 7:53:06 +[2025-07-05 01:07:29] Epoch 3/4, Step 1950/18020, Loss(triple): 7.248446, Loss(predicate): 7.134776, LR: 0.000108, Speed: 120597.80 tokens/sec | Epoch Time Left: 3:38:29 | Total Time Left: 7:52:24 +[2025-07-05 01:08:09] === GPU性能分析 (平均每步) === +[2025-07-05 01:08:09] 前向传播: 8.01ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 01:08:09] GPU总时间: 9.93ms, 实际迭代时间: 814.93ms, GPU利用率: 1.2% +[2025-07-05 01:08:09] ================================================== +[2025-07-05 01:08:09] === 三元组预测示例 === +[2025-07-05 01:08:09] 样本1目标: Geelong located in the administrative territorial entity Victoria, Australia +[2025-07-05 01:08:09] 样本1预测: countryO instanceilstem onill-et iality teroriarit the v +[2025-07-05 01:08:09] 样本2目标: Mo Hussein place of birth Hackney +[2025-07-05 01:08:09] 样本2预测: placeM signog (aro onm Met izate bir of dth M 19 +[2025-07-05 01:08:09] ================== +[2025-07-05 01:08:09] Epoch 3/4, Step 2000/18020, Loss(triple): 7.740993, Loss(predicate): 10.247009, LR: 0.000108, Speed: 120628.34 tokens/sec | Epoch Time Left: 3:37:48 | Total Time Left: 7:51:41 +[2025-07-05 01:08:50] Epoch 3/4, Step 2050/18020, Loss(triple): 7.765945, Loss(predicate): 10.890605, LR: 0.000108, Speed: 120338.83 tokens/sec | Epoch Time Left: 3:37:08 | Total Time Left: 7:50:59 +[2025-07-05 01:09:31] Epoch 3/4, Step 2100/18020, Loss(triple): 7.639664, Loss(predicate): 9.403534, LR: 0.000107, Speed: 120549.95 tokens/sec | Epoch Time Left: 3:36:27 | Total Time Left: 7:50:17 +[2025-07-05 01:10:12] Epoch 3/4, Step 2150/18020, Loss(triple): 8.121918, Loss(predicate): 10.415477, LR: 0.000107, Speed: 121062.70 tokens/sec | Epoch Time Left: 3:35:45 | Total Time Left: 7:49:34 +[2025-07-05 01:10:52] Epoch 3/4, Step 2200/18020, Loss(triple): 7.545364, Loss(predicate): 6.301824, LR: 0.000107, Speed: 120618.23 tokens/sec | Epoch Time Left: 3:35:04 | Total Time Left: 7:48:52 +[2025-07-05 01:11:33] Epoch 3/4, Step 2250/18020, Loss(triple): 7.541048, Loss(predicate): 7.354167, LR: 0.000107, Speed: 120016.80 tokens/sec | Epoch Time Left: 3:34:24 | Total Time Left: 7:48:10 +[2025-07-05 01:12:14] Epoch 3/4, Step 2300/18020, Loss(triple): 8.319031, Loss(predicate): 12.617640, LR: 0.000106, Speed: 119204.73 tokens/sec | Epoch Time Left: 3:33:46 | Total Time Left: 7:47:28 +[2025-07-05 01:12:56] Epoch 3/4, Step 2350/18020, Loss(triple): 7.730831, Loss(predicate): 11.100057, LR: 0.000106, Speed: 119743.30 tokens/sec | Epoch Time Left: 3:33:07 | Total Time Left: 7:46:46 +[2025-07-05 01:13:44] Epoch 3/4, Step 2400/18020, Loss(triple): 7.897448, Loss(predicate): 8.734650, LR: 0.000106, Speed: 101321.85 tokens/sec | Epoch Time Left: 3:33:17 | Total Time Left: 7:46:11 +[2025-07-05 01:14:25] Epoch 3/4, Step 2450/18020, Loss(triple): 7.740786, Loss(predicate): 6.017965, LR: 0.000106, Speed: 118960.77 tokens/sec | Epoch Time Left: 3:32:38 | Total Time Left: 7:45:29 +[2025-07-05 01:15:14] === GPU性能分析 (平均每步) === +[2025-07-05 01:15:14] 前向传播: 7.98ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 01:15:14] GPU总时间: 9.90ms, 实际迭代时间: 967.86ms, GPU利用率: 1.0% +[2025-07-05 01:15:14] ================================================== +[2025-07-05 01:15:14] === 三元组预测示例 === +[2025-07-05 01:15:14] 样本1目标: Babrius languages spoken, written or signed Greek +[2025-07-05 01:15:14] 样本1预测: countryThe partyas (ron inol Ius ritist L ofgeance Mer +[2025-07-05 01:15:14] 样本2目标: Mridula Sinha date of birth 27 November 1942 +[2025-07-05 01:15:14] 样本2预测: countryMensthaara alatores upation bir of occth B 19 +[2025-07-05 01:15:14] ================== +[2025-07-05 01:15:14] Epoch 3/4, Step 2500/18020, Loss(triple): 7.791239, Loss(predicate): 8.804993, LR: 0.000105, Speed: 101568.21 tokens/sec | Epoch Time Left: 3:32:43 | Total Time Left: 7:44:53 +[2025-07-05 01:16:01] Epoch 3/4, Step 2550/18020, Loss(triple): 7.642288, Loss(predicate): 8.978749, LR: 0.000105, Speed: 102961.32 tokens/sec | Epoch Time Left: 3:32:42 | Total Time Left: 7:44:17 +[2025-07-05 01:16:43] Epoch 3/4, Step 2600/18020, Loss(triple): 7.665146, Loss(predicate): 8.497040, LR: 0.000105, Speed: 119215.89 tokens/sec | Epoch Time Left: 3:32:01 | Total Time Left: 7:43:35 +[2025-07-05 01:17:25] Epoch 3/4, Step 2650/18020, Loss(triple): 8.294504, Loss(predicate): 10.453003, LR: 0.000105, Speed: 115159.74 tokens/sec | Epoch Time Left: 3:31:28 | Total Time Left: 7:42:54 +[2025-07-05 01:18:12] Epoch 3/4, Step 2700/18020, Loss(triple): 7.957659, Loss(predicate): 8.634552, LR: 0.000104, Speed: 105545.64 tokens/sec | Epoch Time Left: 3:31:16 | Total Time Left: 7:42:17 +[2025-07-05 01:18:56] Epoch 3/4, Step 2750/18020, Loss(triple): 7.503223, Loss(predicate): 10.800624, LR: 0.000104, Speed: 112777.53 tokens/sec | Epoch Time Left: 3:30:47 | Total Time Left: 7:41:37 +[2025-07-05 01:19:36] Epoch 3/4, Step 2800/18020, Loss(triple): 7.517351, Loss(predicate): 12.189957, LR: 0.000104, Speed: 120912.51 tokens/sec | Epoch Time Left: 3:30:02 | Total Time Left: 7:40:55 +[2025-07-05 01:20:17] Epoch 3/4, Step 2850/18020, Loss(triple): 7.232981, Loss(predicate): 7.377706, LR: 0.000104, Speed: 119960.43 tokens/sec | Epoch Time Left: 3:29:18 | Total Time Left: 7:40:13 +[2025-07-05 01:20:58] Epoch 3/4, Step 2900/18020, Loss(triple): 7.841204, Loss(predicate): 8.920888, LR: 0.000103, Speed: 120855.69 tokens/sec | Epoch Time Left: 3:28:33 | Total Time Left: 7:39:30 +[2025-07-05 01:21:39] Epoch 3/4, Step 2950/18020, Loss(triple): 7.752213, Loss(predicate): 12.267944, LR: 0.000103, Speed: 119872.91 tokens/sec | Epoch Time Left: 3:27:50 | Total Time Left: 7:38:48 +[2025-07-05 01:22:19] === GPU性能分析 (平均每步) === +[2025-07-05 01:22:19] 前向传播: 8.01ms, 损失计算: 0.02ms, 反向传播: 1.95ms, 优化器: 0.00ms +[2025-07-05 01:22:19] GPU总时间: 9.97ms, 实际迭代时间: 811.23ms, GPU利用率: 1.2% +[2025-07-05 01:22:19] ================================================== +[2025-07-05 01:22:19] === 三元组预测示例 === +[2025-07-05 01:22:19] 样本1目标: Rhabdomantis taxon rank genus +[2025-07-05 01:22:19] 样本1预测: imP dyhihd opisusom axonus rankax gen t +[2025-07-05 01:22:19] 样本2目标: P. Ranganath Shenoy occupation politician +[2025-07-05 01:22:19] 样本2预测: countryL adelicha ofu ania Pov upation politan occhip countryici +[2025-07-05 01:22:19] ================== +[2025-07-05 01:22:19] Epoch 3/4, Step 3000/18020, Loss(triple): 7.456995, Loss(predicate): 9.545654, LR: 0.000103, Speed: 121179.03 tokens/sec | Epoch Time Left: 3:27:04 | Total Time Left: 7:38:06 +[2025-07-05 01:23:00] Epoch 3/4, Step 3050/18020, Loss(triple): 7.654345, Loss(predicate): 6.280263, LR: 0.000103, Speed: 120857.83 tokens/sec | Epoch Time Left: 3:26:19 | Total Time Left: 7:37:23 +[2025-07-05 01:23:41] Epoch 3/4, Step 3100/18020, Loss(triple): 7.811867, Loss(predicate): 6.436340, LR: 0.000102, Speed: 119822.71 tokens/sec | Epoch Time Left: 3:25:37 | Total Time Left: 7:36:41 +[2025-07-05 01:24:22] Epoch 3/4, Step 3150/18020, Loss(triple): 7.253643, Loss(predicate): 8.476003, LR: 0.000102, Speed: 120222.90 tokens/sec | Epoch Time Left: 3:24:53 | Total Time Left: 7:35:59 +[2025-07-05 01:25:03] Epoch 3/4, Step 3200/18020, Loss(triple): 7.624474, Loss(predicate): 11.424469, LR: 0.000102, Speed: 120636.22 tokens/sec | Epoch Time Left: 3:24:09 | Total Time Left: 7:35:17 +[2025-07-05 01:25:43] Epoch 3/4, Step 3250/18020, Loss(triple): 7.238461, Loss(predicate): 10.430969, LR: 0.000102, Speed: 121285.51 tokens/sec | Epoch Time Left: 3:23:24 | Total Time Left: 7:34:34 +[2025-07-05 01:26:24] Epoch 3/4, Step 3300/18020, Loss(triple): 7.739857, Loss(predicate): 12.021901, LR: 0.000101, Speed: 120894.73 tokens/sec | Epoch Time Left: 3:22:40 | Total Time Left: 7:33:52 +[2025-07-05 01:27:05] Epoch 3/4, Step 3350/18020, Loss(triple): 7.658184, Loss(predicate): 9.079021, LR: 0.000101, Speed: 119432.95 tokens/sec | Epoch Time Left: 3:21:58 | Total Time Left: 7:33:10 +[2025-07-05 01:27:46] Epoch 3/4, Step 3400/18020, Loss(triple): 7.268055, Loss(predicate): 10.681682, LR: 0.000101, Speed: 120115.90 tokens/sec | Epoch Time Left: 3:21:15 | Total Time Left: 7:32:28 +[2025-07-05 01:28:27] Epoch 3/4, Step 3450/18020, Loss(triple): 7.064732, Loss(predicate): 8.482646, LR: 0.000101, Speed: 120754.30 tokens/sec | Epoch Time Left: 3:20:31 | Total Time Left: 7:31:45 +[2025-07-05 01:29:07] === GPU性能分析 (平均每步) === +[2025-07-05 01:29:07] 前向传播: 8.01ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 01:29:07] GPU总时间: 9.93ms, 实际迭代时间: 813.38ms, GPU利用率: 1.2% +[2025-07-05 01:29:07] ================================================== +[2025-07-05 01:29:07] === 三元组预测示例 === +[2025-07-05 01:29:07] 样本1目标: Simona de Silvestro country of citizenship Swiss +[2025-07-05 01:29:07] 样本1预测: SS entoonaalon ilknir 9ate bir of dth B 19 +[2025-07-05 01:29:07] 样本2目标: Town Hill, Bermuda located in the administrative territorial entity Bermuda +[2025-07-05 01:29:07] 样本2预测: countryR entistonam K owk,ov ustrent States ofiaance country United +[2025-07-05 01:29:07] ================== +[2025-07-05 01:29:07] Epoch 3/4, Step 3500/18020, Loss(triple): 7.487009, Loss(predicate): 8.136485, LR: 0.000101, Speed: 120857.91 tokens/sec | Epoch Time Left: 3:19:47 | Total Time Left: 7:31:03 +[2025-07-05 01:29:48] Epoch 3/4, Step 3550/18020, Loss(triple): 7.751060, Loss(predicate): 9.274983, LR: 0.000100, Speed: 120301.73 tokens/sec | Epoch Time Left: 3:19:04 | Total Time Left: 7:30:21 +[2025-07-05 01:30:29] Epoch 3/4, Step 3600/18020, Loss(triple): 7.658543, Loss(predicate): 10.434184, LR: 0.000100, Speed: 119558.21 tokens/sec | Epoch Time Left: 3:18:22 | Total Time Left: 7:29:39 +[2025-07-05 01:31:10] Epoch 3/4, Step 3650/18020, Loss(triple): 7.635777, Loss(predicate): 7.556498, LR: 0.000100, Speed: 120581.51 tokens/sec | Epoch Time Left: 3:17:39 | Total Time Left: 7:28:57 +[2025-07-05 01:31:23] Model saved to out/pretrain_cls512.pth +[2025-07-05 01:31:52] Epoch 3/4, Step 3700/18020, Loss(triple): 7.319849, Loss(predicate): 11.299489, LR: 0.000100, Speed: 118434.19 tokens/sec | Epoch Time Left: 3:16:59 | Total Time Left: 7:28:15 +[2025-07-05 01:32:32] Epoch 3/4, Step 3750/18020, Loss(triple): 7.337112, Loss(predicate): 10.317871, LR: 0.000099, Speed: 121020.46 tokens/sec | Epoch Time Left: 3:16:15 | Total Time Left: 7:27:33 +[2025-07-05 01:33:13] Model saved to out/pretrain_cls512.pth +[2025-07-05 01:33:14] Epoch 3/4, Step 3800/18020, Loss(triple): 7.852848, Loss(predicate): 6.015829, LR: 0.000099, Speed: 117809.93 tokens/sec | Epoch Time Left: 3:15:35 | Total Time Left: 7:26:51 +[2025-07-05 01:33:55] Epoch 3/4, Step 3850/18020, Loss(triple): 7.725748, Loss(predicate): 5.981842, LR: 0.000099, Speed: 119637.42 tokens/sec | Epoch Time Left: 3:14:53 | Total Time Left: 7:26:09 +[2025-07-05 01:34:36] Epoch 3/4, Step 3900/18020, Loss(triple): 7.760180, Loss(predicate): 10.975372, LR: 0.000099, Speed: 120722.21 tokens/sec | Epoch Time Left: 3:14:10 | Total Time Left: 7:25:27 +[2025-07-05 01:35:16] Epoch 3/4, Step 3950/18020, Loss(triple): 7.266924, Loss(predicate): 8.359212, LR: 0.000098, Speed: 120837.13 tokens/sec | Epoch Time Left: 3:13:27 | Total Time Left: 7:24:45 +[2025-07-05 01:35:57] === GPU性能分析 (平均每步) === +[2025-07-05 01:35:57] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.91ms, 优化器: 0.00ms +[2025-07-05 01:35:57] GPU总时间: 9.93ms, 实际迭代时间: 815.98ms, GPU利用率: 1.2% +[2025-07-05 01:35:57] ================================================== +[2025-07-05 01:35:57] === 三元组预测示例 === +[2025-07-05 01:35:57] 样本1目标: Harry Pickett date of birth 26 March 1862 +[2025-07-05 01:35:57] 样本1预测: GD6yelael erk Mom 4ate bir of dth 2 19 +[2025-07-05 01:35:57] 样本2目标: Blue Front Cafe located in the administrative territorial entity Mississippi +[2025-07-05 01:35:57] 样本2预测: countryCh entom Haeu erk Hay iality teroriarit the C +[2025-07-05 01:35:57] ================== +[2025-07-05 01:35:57] Epoch 3/4, Step 4000/18020, Loss(triple): 7.203583, Loss(predicate): 10.284343, LR: 0.000098, Speed: 120473.60 tokens/sec | Epoch Time Left: 3:12:44 | Total Time Left: 7:24:02 +[2025-07-05 01:36:38] Epoch 3/4, Step 4050/18020, Loss(triple): 8.334499, Loss(predicate): 10.777659, LR: 0.000098, Speed: 120597.48 tokens/sec | Epoch Time Left: 3:12:01 | Total Time Left: 7:23:20 +[2025-07-05 01:37:19] Epoch 3/4, Step 4100/18020, Loss(triple): 7.765713, Loss(predicate): 8.813751, LR: 0.000098, Speed: 120056.36 tokens/sec | Epoch Time Left: 3:11:19 | Total Time Left: 7:22:38 +[2025-07-05 01:38:00] Epoch 3/4, Step 4150/18020, Loss(triple): 7.816029, Loss(predicate): 9.460744, LR: 0.000097, Speed: 121009.62 tokens/sec | Epoch Time Left: 3:10:36 | Total Time Left: 7:21:56 +[2025-07-05 01:38:40] Epoch 3/4, Step 4200/18020, Loss(triple): 7.646183, Loss(predicate): 9.875702, LR: 0.000097, Speed: 121029.33 tokens/sec | Epoch Time Left: 3:09:52 | Total Time Left: 7:21:13 +[2025-07-05 01:39:21] Epoch 3/4, Step 4250/18020, Loss(triple): 7.243958, Loss(predicate): 9.734100, LR: 0.000097, Speed: 119827.06 tokens/sec | Epoch Time Left: 3:09:10 | Total Time Left: 7:20:31 +[2025-07-05 01:40:02] Epoch 3/4, Step 4300/18020, Loss(triple): 7.691301, Loss(predicate): 4.494604, LR: 0.000097, Speed: 121099.47 tokens/sec | Epoch Time Left: 3:08:27 | Total Time Left: 7:19:49 +[2025-07-05 01:40:43] Epoch 3/4, Step 4350/18020, Loss(triple): 7.584068, Loss(predicate): 7.833476, LR: 0.000096, Speed: 119918.17 tokens/sec | Epoch Time Left: 3:07:45 | Total Time Left: 7:19:07 +[2025-07-05 01:41:23] Epoch 3/4, Step 4400/18020, Loss(triple): 7.232653, Loss(predicate): 7.384725, LR: 0.000096, Speed: 121314.61 tokens/sec | Epoch Time Left: 3:07:02 | Total Time Left: 7:18:25 +[2025-07-05 01:42:04] Epoch 3/4, Step 4450/18020, Loss(triple): 7.592621, Loss(predicate): 12.106730, LR: 0.000096, Speed: 121126.24 tokens/sec | Epoch Time Left: 3:06:19 | Total Time Left: 7:17:42 +[2025-07-05 01:42:45] === GPU性能分析 (平均每步) === +[2025-07-05 01:42:45] 前向传播: 8.05ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 01:42:45] GPU总时间: 9.97ms, 实际迭代时间: 819.07ms, GPU利用率: 1.2% +[2025-07-05 01:42:45] ================================================== +[2025-07-05 01:42:45] === 三元组预测示例 === +[2025-07-05 01:42:45] 样本1目标: Frank Cacciatore date of birth April 25, 1955 +[2025-07-05 01:42:45] 样本1预测: countryO enturasard inknil izate bir of,th Sens +[2025-07-05 01:42:45] 样本2目标: 2006 Africa Cup of Nations Final location Cairo International Stadium +[2025-07-05 01:42:45] 样本2预测: country20 basy 200aean erI Cay ortall footbb19 S� +[2025-07-05 01:42:45] ================== +[2025-07-05 01:42:45] Epoch 3/4, Step 4500/18020, Loss(triple): 7.489326, Loss(predicate): 11.421819, LR: 0.000096, Speed: 120019.68 tokens/sec | Epoch Time Left: 3:05:37 | Total Time Left: 7:17:00 +[2025-07-05 01:43:26] Epoch 3/4, Step 4550/18020, Loss(triple): 7.414871, Loss(predicate): 5.963094, LR: 0.000095, Speed: 120493.10 tokens/sec | Epoch Time Left: 3:04:55 | Total Time Left: 7:16:18 +[2025-07-05 01:44:06] Epoch 3/4, Step 4600/18020, Loss(triple): 7.783417, Loss(predicate): 10.520518, LR: 0.000095, Speed: 120788.68 tokens/sec | Epoch Time Left: 3:04:12 | Total Time Left: 7:15:36 +[2025-07-05 01:44:47] Epoch 3/4, Step 4650/18020, Loss(triple): 7.413513, Loss(predicate): 7.769506, LR: 0.000095, Speed: 121456.01 tokens/sec | Epoch Time Left: 3:03:29 | Total Time Left: 7:14:53 +[2025-07-05 01:45:27] Epoch 3/4, Step 4700/18020, Loss(triple): 7.224716, Loss(predicate): 9.806560, LR: 0.000095, Speed: 120993.44 tokens/sec | Epoch Time Left: 3:02:46 | Total Time Left: 7:14:11 +[2025-07-05 01:46:08] Epoch 3/4, Step 4750/18020, Loss(triple): 7.873400, Loss(predicate): 7.213969, LR: 0.000094, Speed: 120214.16 tokens/sec | Epoch Time Left: 3:02:04 | Total Time Left: 7:13:29 +[2025-07-05 01:46:49] Epoch 3/4, Step 4800/18020, Loss(triple): 7.625492, Loss(predicate): 5.394409, LR: 0.000094, Speed: 120692.30 tokens/sec | Epoch Time Left: 3:01:22 | Total Time Left: 7:12:47 +[2025-07-05 01:47:30] Epoch 3/4, Step 4850/18020, Loss(triple): 7.712082, Loss(predicate): 10.673889, LR: 0.000094, Speed: 121051.89 tokens/sec | Epoch Time Left: 3:00:39 | Total Time Left: 7:12:05 +[2025-07-05 01:48:10] Epoch 3/4, Step 4900/18020, Loss(triple): 7.980644, Loss(predicate): 7.161936, LR: 0.000094, Speed: 121267.47 tokens/sec | Epoch Time Left: 2:59:56 | Total Time Left: 7:11:22 +[2025-07-05 01:48:51] Epoch 3/4, Step 4950/18020, Loss(triple): 7.866302, Loss(predicate): 10.937637, LR: 0.000093, Speed: 120940.53 tokens/sec | Epoch Time Left: 2:59:14 | Total Time Left: 7:10:40 +[2025-07-05 01:49:32] === GPU性能分析 (平均每步) === +[2025-07-05 01:49:32] 前向传播: 7.99ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 01:49:32] GPU总时间: 9.91ms, 实际迭代时间: 818.20ms, GPU利用率: 1.2% +[2025-07-05 01:49:32] ================================================== +[2025-07-05 01:49:32] === 三元组预测示例 === +[2025-07-05 01:49:32] 样本1目标: Peeter Sauter place of birth Tallinn +[2025-07-05 01:49:32] 样本1预测: countryLensether (eer eror,ay 4ate bir of dth M 19 +[2025-07-05 01:49:32] 样本2目标: Anna Rüh place of birth Greifswald +[2025-07-05 01:49:32] 样本2预测: countryCh entulhadam an�nun 6ate bir of dth J 1 +[2025-07-05 01:49:32] ================== +[2025-07-05 01:49:32] Epoch 3/4, Step 5000/18020, Loss(triple): 7.313736, Loss(predicate): 10.591350, LR: 0.000093, Speed: 120146.20 tokens/sec | Epoch Time Left: 2:58:32 | Total Time Left: 7:09:58 +[2025-07-05 01:50:12] Epoch 3/4, Step 5050/18020, Loss(triple): 7.462969, Loss(predicate): 10.199391, LR: 0.000093, Speed: 120667.86 tokens/sec | Epoch Time Left: 2:57:50 | Total Time Left: 7:09:16 +[2025-07-05 01:50:53] Epoch 3/4, Step 5100/18020, Loss(triple): 7.254246, Loss(predicate): 14.128407, LR: 0.000093, Speed: 120576.26 tokens/sec | Epoch Time Left: 2:57:08 | Total Time Left: 7:08:34 +[2025-07-05 01:51:34] Epoch 3/4, Step 5150/18020, Loss(triple): 7.955555, Loss(predicate): 6.666311, LR: 0.000093, Speed: 121447.55 tokens/sec | Epoch Time Left: 2:56:25 | Total Time Left: 7:07:51 +[2025-07-05 01:52:14] Epoch 3/4, Step 5200/18020, Loss(triple): 8.004669, Loss(predicate): 9.587341, LR: 0.000092, Speed: 121044.91 tokens/sec | Epoch Time Left: 2:55:43 | Total Time Left: 7:07:09 +[2025-07-05 01:52:55] Epoch 3/4, Step 5250/18020, Loss(triple): 7.386559, Loss(predicate): 6.970006, LR: 0.000092, Speed: 119991.63 tokens/sec | Epoch Time Left: 2:55:01 | Total Time Left: 7:06:27 +[2025-07-05 01:53:36] Epoch 3/4, Step 5300/18020, Loss(triple): 7.937683, Loss(predicate): 14.769002, LR: 0.000092, Speed: 120787.10 tokens/sec | Epoch Time Left: 2:54:19 | Total Time Left: 7:05:45 +[2025-07-05 01:54:17] Epoch 3/4, Step 5350/18020, Loss(triple): 7.594044, Loss(predicate): 7.377309, LR: 0.000092, Speed: 120646.50 tokens/sec | Epoch Time Left: 2:53:37 | Total Time Left: 7:05:03 +[2025-07-05 01:54:57] Epoch 3/4, Step 5400/18020, Loss(triple): 7.360035, Loss(predicate): 7.606628, LR: 0.000091, Speed: 121350.50 tokens/sec | Epoch Time Left: 2:52:54 | Total Time Left: 7:04:20 +[2025-07-05 01:55:38] Epoch 3/4, Step 5450/18020, Loss(triple): 8.005432, Loss(predicate): 10.940460, LR: 0.000091, Speed: 120966.65 tokens/sec | Epoch Time Left: 2:52:12 | Total Time Left: 7:03:38 +[2025-07-05 01:56:19] === GPU性能分析 (平均每步) === +[2025-07-05 01:56:19] 前向传播: 7.98ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 01:56:19] GPU总时间: 9.91ms, 实际迭代时间: 820.39ms, GPU利用率: 1.2% +[2025-07-05 01:56:19] ================================================== +[2025-07-05 01:56:19] === 三元组预测示例 === +[2025-07-05 01:56:19] 样本1目标: Saint Petersburg twinned administrative body Saint Petersburg, Florida +[2025-07-05 01:56:19] 样本1预测: countryC entyinaarb erill Cir ialist locatedoriverance the C +[2025-07-05 01:56:19] 样本2目标: Robert Kendall (poet) country of citizenship United States +[2025-07-05 01:56:19] 样本2预测: GRensetry (ser ink Aes Americanitiz of chip countryens +[2025-07-05 01:56:19] ================== +[2025-07-05 01:56:19] Epoch 3/4, Step 5500/18020, Loss(triple): 7.247797, Loss(predicate): 6.897298, LR: 0.000091, Speed: 119825.68 tokens/sec | Epoch Time Left: 2:51:31 | Total Time Left: 7:02:56 +[2025-07-05 01:56:32] Model saved to out/pretrain_cls512.pth +[2025-07-05 01:57:01] Epoch 3/4, Step 5550/18020, Loss(triple): 7.264294, Loss(predicate): 8.187210, LR: 0.000091, Speed: 117781.23 tokens/sec | Epoch Time Left: 2:50:51 | Total Time Left: 7:02:15 +[2025-07-05 01:57:41] Epoch 3/4, Step 5600/18020, Loss(triple): 7.827484, Loss(predicate): 11.220856, LR: 0.000090, Speed: 121021.16 tokens/sec | Epoch Time Left: 2:50:09 | Total Time Left: 7:01:32 +[2025-07-05 01:58:22] Epoch 3/4, Step 5650/18020, Loss(triple): 7.216026, Loss(predicate): 13.455841, LR: 0.000090, Speed: 121291.48 tokens/sec | Epoch Time Left: 2:49:27 | Total Time Left: 7:00:50 +[2025-07-05 01:59:07] Epoch 3/4, Step 5700/18020, Loss(triple): 7.261200, Loss(predicate): 10.330098, LR: 0.000090, Speed: 107700.47 tokens/sec | Epoch Time Left: 2:48:55 | Total Time Left: 7:00:12 +[2025-07-05 01:59:50] Epoch 3/4, Step 5750/18020, Loss(triple): 7.834351, Loss(predicate): 14.985168, LR: 0.000090, Speed: 115832.11 tokens/sec | Epoch Time Left: 2:48:17 | Total Time Left: 6:59:31 +[2025-07-05 02:00:32] Epoch 3/4, Step 5800/18020, Loss(triple): 8.116213, Loss(predicate): 9.438619, LR: 0.000089, Speed: 116739.96 tokens/sec | Epoch Time Left: 2:47:38 | Total Time Left: 6:58:50 +[2025-07-05 02:01:15] Epoch 3/4, Step 5850/18020, Loss(triple): 7.865152, Loss(predicate): 3.936737, LR: 0.000089, Speed: 112850.06 tokens/sec | Epoch Time Left: 2:47:02 | Total Time Left: 6:58:09 +[2025-07-05 02:01:56] Epoch 3/4, Step 5900/18020, Loss(triple): 7.588633, Loss(predicate): 8.505321, LR: 0.000089, Speed: 119679.29 tokens/sec | Epoch Time Left: 2:46:20 | Total Time Left: 6:57:28 +[2025-07-05 02:02:37] Epoch 3/4, Step 5950/18020, Loss(triple): 7.421284, Loss(predicate): 11.577403, LR: 0.000089, Speed: 120941.34 tokens/sec | Epoch Time Left: 2:45:38 | Total Time Left: 6:56:45 +[2025-07-05 02:03:18] === GPU性能分析 (平均每步) === +[2025-07-05 02:03:18] 前向传播: 7.99ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 02:03:18] GPU总时间: 9.91ms, 实际迭代时间: 818.42ms, GPU利用率: 1.2% +[2025-07-05 02:03:18] ================================================== +[2025-07-05 02:03:18] === 三元组预测示例 === +[2025-07-05 02:03:18] 样本1目标: Adolf Schiffer instrument cellist +[2025-07-05 02:03:18] 样本1预测: GCh biryg (cd erch Met upation bir of occth Mici +[2025-07-05 02:03:18] 样本2目标: Bulbophyllum louisiadum parent taxon Bulbophyllum +[2025-07-05 02:03:18] 样本2预测: STptyllumcb opisusus boon speciesonankaxph t +[2025-07-05 02:03:18] ================== +[2025-07-05 02:03:18] Epoch 3/4, Step 6000/18020, Loss(triple): 8.435398, Loss(predicate): 11.926005, LR: 0.000088, Speed: 120113.85 tokens/sec | Epoch Time Left: 2:44:56 | Total Time Left: 6:56:03 +[2025-07-05 02:03:59] Epoch 3/4, Step 6050/18020, Loss(triple): 7.356251, Loss(predicate): 8.608826, LR: 0.000088, Speed: 120393.17 tokens/sec | Epoch Time Left: 2:44:15 | Total Time Left: 6:55:21 +[2025-07-05 02:04:39] Epoch 3/4, Step 6100/18020, Loss(triple): 8.249638, Loss(predicate): 9.484904, LR: 0.000088, Speed: 121069.03 tokens/sec | Epoch Time Left: 2:43:32 | Total Time Left: 6:54:39 +[2025-07-05 02:05:20] Epoch 3/4, Step 6150/18020, Loss(triple): 7.822681, Loss(predicate): 6.912058, LR: 0.000088, Speed: 121309.32 tokens/sec | Epoch Time Left: 2:42:50 | Total Time Left: 6:53:57 +[2025-07-05 02:06:01] Epoch 3/4, Step 6200/18020, Loss(triple): 7.529428, Loss(predicate): 9.917953, LR: 0.000087, Speed: 120085.26 tokens/sec | Epoch Time Left: 2:42:08 | Total Time Left: 6:53:15 +[2025-07-05 02:06:42] Epoch 3/4, Step 6250/18020, Loss(triple): 7.441029, Loss(predicate): 6.703695, LR: 0.000087, Speed: 119645.67 tokens/sec | Epoch Time Left: 2:41:27 | Total Time Left: 6:52:33 +[2025-07-05 02:07:23] Epoch 3/4, Step 6300/18020, Loss(triple): 7.318270, Loss(predicate): 7.028636, LR: 0.000087, Speed: 120308.87 tokens/sec | Epoch Time Left: 2:40:45 | Total Time Left: 6:51:51 +[2025-07-05 02:08:05] Epoch 3/4, Step 6350/18020, Loss(triple): 7.499249, Loss(predicate): 9.540751, LR: 0.000087, Speed: 117375.54 tokens/sec | Epoch Time Left: 2:40:06 | Total Time Left: 6:51:10 +[2025-07-05 02:08:48] Epoch 3/4, Step 6400/18020, Loss(triple): 7.376068, Loss(predicate): 11.114858, LR: 0.000087, Speed: 114320.27 tokens/sec | Epoch Time Left: 2:39:28 | Total Time Left: 6:50:29 +[2025-07-05 02:09:29] Epoch 3/4, Step 6450/18020, Loss(triple): 7.534164, Loss(predicate): 9.134868, LR: 0.000086, Speed: 119276.03 tokens/sec | Epoch Time Left: 2:38:47 | Total Time Left: 6:49:47 +[2025-07-05 02:10:10] === GPU性能分析 (平均每步) === +[2025-07-05 02:10:10] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 02:10:10] GPU总时间: 9.92ms, 实际迭代时间: 822.29ms, GPU利用率: 1.2% +[2025-07-05 02:10:10] ================================================== +[2025-07-05 02:10:10] === 三元组预测示例 === +[2025-07-05 02:10:10] 样本1目标: Bulbophyllum marginatum parent taxon Bulbophyllum +[2025-07-05 02:10:10] 样本1预测: SDptyllum ofer ialisusus boon speciesonankaxph t +[2025-07-05 02:10:10] 样本2目标: Charles Alexander, Duke of Württemberg date of birth 24 May 1684 +[2025-07-05 02:10:10] 样本2预测: GCh adoryisharan en�nun 7ate bir of dth J 1 +[2025-07-05 02:10:10] ================== +[2025-07-05 02:10:10] Epoch 3/4, Step 6500/18020, Loss(triple): 7.509144, Loss(predicate): 12.274608, LR: 0.000086, Speed: 119549.66 tokens/sec | Epoch Time Left: 2:38:05 | Total Time Left: 6:49:05 +[2025-07-05 02:10:51] Epoch 3/4, Step 6550/18020, Loss(triple): 7.770100, Loss(predicate): 9.637013, LR: 0.000086, Speed: 119460.03 tokens/sec | Epoch Time Left: 2:37:24 | Total Time Left: 6:48:24 +[2025-07-05 02:11:32] Epoch 3/4, Step 6600/18020, Loss(triple): 7.811792, Loss(predicate): 7.914973, LR: 0.000086, Speed: 121333.07 tokens/sec | Epoch Time Left: 2:36:42 | Total Time Left: 6:47:41 +[2025-07-05 02:12:12] Epoch 3/4, Step 6650/18020, Loss(triple): 7.465015, Loss(predicate): 7.954651, LR: 0.000085, Speed: 121154.76 tokens/sec | Epoch Time Left: 2:36:00 | Total Time Left: 6:46:59 +[2025-07-05 02:12:53] Epoch 3/4, Step 6700/18020, Loss(triple): 7.432281, Loss(predicate): 12.801564, LR: 0.000085, Speed: 119936.84 tokens/sec | Epoch Time Left: 2:35:18 | Total Time Left: 6:46:17 +[2025-07-05 02:13:34] Epoch 3/4, Step 6750/18020, Loss(triple): 7.325113, Loss(predicate): 8.537190, LR: 0.000085, Speed: 121290.02 tokens/sec | Epoch Time Left: 2:34:36 | Total Time Left: 6:45:35 +[2025-07-05 02:14:15] Epoch 3/4, Step 6800/18020, Loss(triple): 7.526657, Loss(predicate): 10.823578, LR: 0.000085, Speed: 119835.51 tokens/sec | Epoch Time Left: 2:33:55 | Total Time Left: 6:44:53 +[2025-07-05 02:14:55] Epoch 3/4, Step 6850/18020, Loss(triple): 7.360935, Loss(predicate): 9.616689, LR: 0.000084, Speed: 120820.72 tokens/sec | Epoch Time Left: 2:33:13 | Total Time Left: 6:44:11 +[2025-07-05 02:15:36] Epoch 3/4, Step 6900/18020, Loss(triple): 7.337078, Loss(predicate): 8.940669, LR: 0.000084, Speed: 120363.10 tokens/sec | Epoch Time Left: 2:32:31 | Total Time Left: 6:43:29 +[2025-07-05 02:16:18] Epoch 3/4, Step 6950/18020, Loss(triple): 7.415985, Loss(predicate): 7.596176, LR: 0.000084, Speed: 119141.23 tokens/sec | Epoch Time Left: 2:31:50 | Total Time Left: 6:42:47 +[2025-07-05 02:16:59] === GPU性能分析 (平均每步) === +[2025-07-05 02:16:59] 前向传播: 8.01ms, 损失计算: 0.02ms, 反向传播: 1.95ms, 优化器: 0.00ms +[2025-07-05 02:16:59] GPU总时间: 9.97ms, 实际迭代时间: 819.84ms, GPU利用率: 1.2% +[2025-07-05 02:16:59] ================================================== +[2025-07-05 02:16:59] === 三元组预测示例 === +[2025-07-05 02:16:59] 样本1目标: A. K. Gopalan place of death India +[2025-07-05 02:16:59] 样本1预测: GTensygienan itist Mom 4ate bir of dth 2 19 +[2025-07-05 02:16:59] 样本2目标: Jakob Denzinger date of birth June 29, 1924 +[2025-07-05 02:16:59] 样本2预测: countryJensish.ener erie Day 3ate States of dth country 19 +[2025-07-05 02:16:59] ================== +[2025-07-05 02:16:59] Epoch 3/4, Step 7000/18020, Loss(triple): 7.381241, Loss(predicate): 6.021261, LR: 0.000084, Speed: 119906.02 tokens/sec | Epoch Time Left: 2:31:09 | Total Time Left: 6:42:05 +[2025-07-05 02:17:41] Epoch 3/4, Step 7050/18020, Loss(triple): 7.383904, Loss(predicate): 4.526474, LR: 0.000083, Speed: 115133.93 tokens/sec | Epoch Time Left: 2:30:30 | Total Time Left: 6:41:24 +[2025-07-05 02:18:23] Epoch 3/4, Step 7100/18020, Loss(triple): 7.935829, Loss(predicate): 11.332774, LR: 0.000083, Speed: 116461.76 tokens/sec | Epoch Time Left: 2:29:50 | Total Time Left: 6:40:43 +[2025-07-05 02:19:05] Epoch 3/4, Step 7150/18020, Loss(triple): 7.296938, Loss(predicate): 7.381083, LR: 0.000083, Speed: 119435.54 tokens/sec | Epoch Time Left: 2:29:09 | Total Time Left: 6:40:02 +[2025-07-05 02:19:45] Epoch 3/4, Step 7200/18020, Loss(triple): 7.352798, Loss(predicate): 5.430277, LR: 0.000083, Speed: 120828.60 tokens/sec | Epoch Time Left: 2:28:27 | Total Time Left: 6:39:19 +[2025-07-05 02:20:30] Epoch 3/4, Step 7250/18020, Loss(triple): 7.608238, Loss(predicate): 10.247711, LR: 0.000082, Speed: 110408.54 tokens/sec | Epoch Time Left: 2:27:51 | Total Time Left: 6:38:40 +[2025-07-05 02:21:17] Epoch 3/4, Step 7300/18020, Loss(triple): 7.701981, Loss(predicate): 6.366760, LR: 0.000082, Speed: 103721.58 tokens/sec | Epoch Time Left: 2:27:19 | Total Time Left: 6:38:02 +[2025-07-05 02:21:59] Epoch 3/4, Step 7350/18020, Loss(triple): 7.799828, Loss(predicate): 13.665324, LR: 0.000082, Speed: 118024.19 tokens/sec | Epoch Time Left: 2:26:38 | Total Time Left: 6:37:21 +[2025-07-05 02:22:39] Epoch 3/4, Step 7400/18020, Loss(triple): 7.577137, Loss(predicate): 6.160080, LR: 0.000082, Speed: 120861.84 tokens/sec | Epoch Time Left: 2:25:56 | Total Time Left: 6:36:39 +[2025-07-05 02:23:20] Epoch 3/4, Step 7450/18020, Loss(triple): 7.326111, Loss(predicate): 7.889587, LR: 0.000081, Speed: 119962.96 tokens/sec | Epoch Time Left: 2:25:15 | Total Time Left: 6:35:57 +[2025-07-05 02:24:01] === GPU性能分析 (平均每步) === +[2025-07-05 02:24:01] 前向传播: 8.04ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 02:24:01] GPU总时间: 9.96ms, 实际迭代时间: 815.58ms, GPU利用率: 1.2% +[2025-07-05 02:24:01] ================================================== +[2025-07-05 02:24:01] === 三元组预测示例 === +[2025-07-05 02:24:01] 样本1目标: New France named after France +[2025-07-05 02:24:01] 样本1预测: countryI entyasaean onm Ist ialist countryadiaion the Can +[2025-07-05 02:24:01] 样本2目标: Ryue Nishizawa occupation architect +[2025-07-05 02:24:01] 样本2预测: placeRensyicha.a ian�yov upation bir of occth insted +[2025-07-05 02:24:01] ================== +[2025-07-05 02:24:01] Epoch 3/4, Step 7500/18020, Loss(triple): 7.390827, Loss(predicate): 12.461223, LR: 0.000081, Speed: 120532.39 tokens/sec | Epoch Time Left: 2:24:33 | Total Time Left: 6:35:15 +[2025-07-05 02:24:42] Epoch 3/4, Step 7550/18020, Loss(triple): 7.248285, Loss(predicate): 7.664703, LR: 0.000081, Speed: 119524.59 tokens/sec | Epoch Time Left: 2:23:51 | Total Time Left: 6:34:33 +[2025-07-05 02:25:23] Epoch 3/4, Step 7600/18020, Loss(triple): 7.213709, Loss(predicate): 7.781982, LR: 0.000081, Speed: 121218.38 tokens/sec | Epoch Time Left: 2:23:09 | Total Time Left: 6:33:51 +[2025-07-05 02:26:03] Epoch 3/4, Step 7650/18020, Loss(triple): 7.854647, Loss(predicate): 8.862473, LR: 0.000081, Speed: 121042.07 tokens/sec | Epoch Time Left: 2:22:27 | Total Time Left: 6:33:08 +[2025-07-05 02:26:44] Epoch 3/4, Step 7700/18020, Loss(triple): 7.317926, Loss(predicate): 6.089610, LR: 0.000080, Speed: 120654.42 tokens/sec | Epoch Time Left: 2:21:45 | Total Time Left: 6:32:26 +[2025-07-05 02:27:25] Epoch 3/4, Step 7750/18020, Loss(triple): 7.677830, Loss(predicate): 8.179738, LR: 0.000080, Speed: 120788.01 tokens/sec | Epoch Time Left: 2:21:04 | Total Time Left: 6:31:44 +[2025-07-05 02:28:06] Epoch 3/4, Step 7800/18020, Loss(triple): 7.030643, Loss(predicate): 10.451350, LR: 0.000080, Speed: 119825.23 tokens/sec | Epoch Time Left: 2:20:22 | Total Time Left: 6:31:02 +[2025-07-05 02:28:47] Epoch 3/4, Step 7850/18020, Loss(triple): 7.506516, Loss(predicate): 9.781880, LR: 0.000080, Speed: 121011.38 tokens/sec | Epoch Time Left: 2:19:40 | Total Time Left: 6:30:20 +[2025-07-05 02:29:27] Epoch 3/4, Step 7900/18020, Loss(triple): 7.629196, Loss(predicate): 7.802602, LR: 0.000079, Speed: 120640.65 tokens/sec | Epoch Time Left: 2:18:58 | Total Time Left: 6:29:38 +[2025-07-05 02:30:08] Epoch 3/4, Step 7950/18020, Loss(triple): 7.436535, Loss(predicate): 11.439301, LR: 0.000079, Speed: 119791.94 tokens/sec | Epoch Time Left: 2:18:17 | Total Time Left: 6:28:56 +[2025-07-05 02:30:49] === GPU性能分析 (平均每步) === +[2025-07-05 02:30:49] 前向传播: 8.05ms, 损失计算: 0.02ms, 反向传播: 1.91ms, 优化器: 0.00ms +[2025-07-05 02:30:49] GPU总时间: 9.97ms, 实际迭代时间: 816.10ms, GPU利用率: 1.2% +[2025-07-05 02:30:49] ================================================== +[2025-07-05 02:30:49] === 三元组预测示例 === +[2025-07-05 02:30:49] 样本1目标: Heliopolis (Scudamore novel) author James Scudamore +[2025-07-05 02:30:49] 样本1预测: countryH instyelicb anill Pak fil with country ofz language Ser +[2025-07-05 02:30:49] 样本2目标: Mayamalavagowla part of melakarta +[2025-07-05 02:30:49] 样本2预测: placeM entlgaea anab-ov Oance country ofanceort inst of +[2025-07-05 02:30:49] ================== +[2025-07-05 02:30:49] Epoch 3/4, Step 8000/18020, Loss(triple): 7.302195, Loss(predicate): 10.067769, LR: 0.000079, Speed: 120455.58 tokens/sec | Epoch Time Left: 2:17:35 | Total Time Left: 6:28:14 +[2025-07-05 02:31:30] Epoch 3/4, Step 8050/18020, Loss(triple): 7.505365, Loss(predicate): 9.850067, LR: 0.000079, Speed: 120017.53 tokens/sec | Epoch Time Left: 2:16:54 | Total Time Left: 6:27:32 +[2025-07-05 02:32:11] Epoch 3/4, Step 8100/18020, Loss(triple): 7.378160, Loss(predicate): 7.957662, LR: 0.000078, Speed: 121149.32 tokens/sec | Epoch Time Left: 2:16:12 | Total Time Left: 6:26:50 +[2025-07-05 02:32:51] Epoch 3/4, Step 8150/18020, Loss(triple): 7.446381, Loss(predicate): 10.746404, LR: 0.000078, Speed: 121025.57 tokens/sec | Epoch Time Left: 2:15:30 | Total Time Left: 6:26:08 +[2025-07-05 02:33:32] Epoch 3/4, Step 8200/18020, Loss(triple): 7.273523, Loss(predicate): 7.277120, LR: 0.000078, Speed: 119710.17 tokens/sec | Epoch Time Left: 2:14:49 | Total Time Left: 6:25:26 +[2025-07-05 02:34:13] Epoch 3/4, Step 8250/18020, Loss(triple): 7.764458, Loss(predicate): 10.619618, LR: 0.000078, Speed: 121017.08 tokens/sec | Epoch Time Left: 2:14:07 | Total Time Left: 6:24:44 +[2025-07-05 02:34:54] Epoch 3/4, Step 8300/18020, Loss(triple): 7.726954, Loss(predicate): 5.056534, LR: 0.000077, Speed: 119991.71 tokens/sec | Epoch Time Left: 2:13:25 | Total Time Left: 6:24:02 +[2025-07-05 02:35:34] Epoch 3/4, Step 8350/18020, Loss(triple): 7.201832, Loss(predicate): 5.951243, LR: 0.000077, Speed: 121332.48 tokens/sec | Epoch Time Left: 2:12:43 | Total Time Left: 6:23:20 +[2025-07-05 02:36:15] Epoch 3/4, Step 8400/18020, Loss(triple): 7.203646, Loss(predicate): 6.639547, LR: 0.000077, Speed: 120742.61 tokens/sec | Epoch Time Left: 2:12:02 | Total Time Left: 6:22:38 +[2025-07-05 02:36:56] Epoch 3/4, Step 8450/18020, Loss(triple): 7.185928, Loss(predicate): 10.918635, LR: 0.000077, Speed: 119574.73 tokens/sec | Epoch Time Left: 2:11:20 | Total Time Left: 6:21:56 +[2025-07-05 02:37:37] === GPU性能分析 (平均每步) === +[2025-07-05 02:37:37] 前向传播: 8.06ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 02:37:37] GPU总时间: 9.98ms, 实际迭代时间: 816.86ms, GPU利用率: 1.2% +[2025-07-05 02:37:37] ================================================== +[2025-07-05 02:37:37] === 三元组预测示例 === +[2025-07-05 02:37:37] 样本1目标: You've Got a Friend in Me performer Randy Newman +[2025-07-05 02:37:37] 样本1预测: GTheockyrit (e L manames Gay songight perform ofetame Mer +[2025-07-05 02:37:37] 样本2目标: Shama Sikander occupation actress +[2025-07-05 02:37:37] 样本2预测: GS signelhabam ank Pay upation act of occress Bor +[2025-07-05 02:37:37] ================== +[2025-07-05 02:37:37] Epoch 3/4, Step 8500/18020, Loss(triple): 7.252228, Loss(predicate): 14.169952, LR: 0.000077, Speed: 120343.78 tokens/sec | Epoch Time Left: 2:10:39 | Total Time Left: 6:21:14 +[2025-07-05 02:38:18] Epoch 3/4, Step 8550/18020, Loss(triple): 7.400818, Loss(predicate): 4.906413, LR: 0.000076, Speed: 120524.81 tokens/sec | Epoch Time Left: 2:09:57 | Total Time Left: 6:20:32 +[2025-07-05 02:38:58] Epoch 3/4, Step 8600/18020, Loss(triple): 7.383066, Loss(predicate): 11.695796, LR: 0.000076, Speed: 121387.66 tokens/sec | Epoch Time Left: 2:09:15 | Total Time Left: 6:19:50 +[2025-07-05 02:39:39] Epoch 3/4, Step 8650/18020, Loss(triple): 7.610371, Loss(predicate): 9.044571, LR: 0.000076, Speed: 120600.54 tokens/sec | Epoch Time Left: 2:08:34 | Total Time Left: 6:19:08 +[2025-07-05 02:40:20] Epoch 3/4, Step 8700/18020, Loss(triple): 7.655350, Loss(predicate): 11.188405, LR: 0.000076, Speed: 119362.75 tokens/sec | Epoch Time Left: 2:07:52 | Total Time Left: 6:18:26 +[2025-07-05 02:41:01] Epoch 3/4, Step 8750/18020, Loss(triple): 7.531357, Loss(predicate): 8.737569, LR: 0.000075, Speed: 120694.72 tokens/sec | Epoch Time Left: 2:07:11 | Total Time Left: 6:17:44 +[2025-07-05 02:41:42] Epoch 3/4, Step 8800/18020, Loss(triple): 8.150238, Loss(predicate): 10.587077, LR: 0.000075, Speed: 121068.18 tokens/sec | Epoch Time Left: 2:06:29 | Total Time Left: 6:17:02 +[2025-07-05 02:42:22] Epoch 3/4, Step 8850/18020, Loss(triple): 8.115883, Loss(predicate): 10.764201, LR: 0.000075, Speed: 121341.85 tokens/sec | Epoch Time Left: 2:05:47 | Total Time Left: 6:16:20 +[2025-07-05 02:43:03] Epoch 3/4, Step 8900/18020, Loss(triple): 7.164158, Loss(predicate): 8.918010, LR: 0.000075, Speed: 120634.29 tokens/sec | Epoch Time Left: 2:05:06 | Total Time Left: 6:15:38 +[2025-07-05 02:43:44] Epoch 3/4, Step 8950/18020, Loss(triple): 7.543808, Loss(predicate): 8.204448, LR: 0.000074, Speed: 119564.53 tokens/sec | Epoch Time Left: 2:04:24 | Total Time Left: 6:14:56 +[2025-07-05 02:44:34] === GPU性能分析 (平均每步) === +[2025-07-05 02:44:34] 前向传播: 7.96ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 02:44:34] GPU总时间: 9.88ms, 实际迭代时间: 1009.17ms, GPU利用率: 1.0% +[2025-07-05 02:44:34] ================================================== +[2025-07-05 02:44:34] === 三元组预测示例 === +[2025-07-05 02:44:34] 样本1目标: Philyra (plant) taxon rank species +[2025-07-05 02:44:34] 样本1预测: mPptygiaeb alisusus axonus rankax gen t +[2025-07-05 02:44:34] 样本2目标: Casper Asbjornson member of sports team Boston Red Sox +[2025-07-05 02:44:34] 样本2预测: GRensyil (sd onill,ak 7ate bir of dth 2 19 +[2025-07-05 02:44:34] ================== +[2025-07-05 02:44:34] Epoch 3/4, Step 9000/18020, Loss(triple): 7.828110, Loss(predicate): 11.073649, LR: 0.000074, Speed: 97410.99 tokens/sec | Epoch Time Left: 2:03:53 | Total Time Left: 6:14:20 +[2025-07-05 02:45:28] Epoch 3/4, Step 9050/18020, Loss(triple): 7.478359, Loss(predicate): 13.398656, LR: 0.000074, Speed: 92384.51 tokens/sec | Epoch Time Left: 2:03:23 | Total Time Left: 6:13:46 +[2025-07-05 02:46:20] Epoch 3/4, Step 9100/18020, Loss(triple): 7.473841, Loss(predicate): 6.615438, LR: 0.000074, Speed: 93054.27 tokens/sec | Epoch Time Left: 2:02:53 | Total Time Left: 6:13:11 +[2025-07-05 02:47:05] Epoch 3/4, Step 9150/18020, Loss(triple): 7.321939, Loss(predicate): 12.520203, LR: 0.000073, Speed: 109933.42 tokens/sec | Epoch Time Left: 2:02:15 | Total Time Left: 6:12:31 +[2025-07-05 02:47:07] Model saved to out/pretrain_cls512.pth +[2025-07-05 02:47:48] Epoch 3/4, Step 9200/18020, Loss(triple): 7.386345, Loss(predicate): 7.321152, LR: 0.000073, Speed: 113591.88 tokens/sec | Epoch Time Left: 2:01:36 | Total Time Left: 6:11:51 +[2025-07-05 02:48:29] Epoch 3/4, Step 9250/18020, Loss(triple): 7.755924, Loss(predicate): 6.169556, LR: 0.000073, Speed: 120690.81 tokens/sec | Epoch Time Left: 2:00:54 | Total Time Left: 6:11:09 +[2025-07-05 02:49:10] Epoch 3/4, Step 9300/18020, Loss(triple): 7.542915, Loss(predicate): 11.486247, LR: 0.000073, Speed: 121391.46 tokens/sec | Epoch Time Left: 2:00:12 | Total Time Left: 6:10:26 +[2025-07-05 02:49:50] Epoch 3/4, Step 9350/18020, Loss(triple): 7.581324, Loss(predicate): 12.960643, LR: 0.000073, Speed: 121531.26 tokens/sec | Epoch Time Left: 1:59:29 | Total Time Left: 6:09:44 +[2025-07-05 02:50:31] Epoch 3/4, Step 9400/18020, Loss(triple): 7.406649, Loss(predicate): 8.373485, LR: 0.000072, Speed: 120951.30 tokens/sec | Epoch Time Left: 1:58:47 | Total Time Left: 6:09:02 +[2025-07-05 02:51:12] Epoch 3/4, Step 9450/18020, Loss(triple): 7.507889, Loss(predicate): 7.998896, LR: 0.000072, Speed: 119830.28 tokens/sec | Epoch Time Left: 1:58:06 | Total Time Left: 6:08:20 +[2025-07-05 02:51:53] === GPU性能分析 (平均每步) === +[2025-07-05 02:51:53] 前向传播: 8.00ms, 损失计算: 0.02ms, 反向传播: 1.90ms, 优化器: 0.00ms +[2025-07-05 02:51:53] GPU总时间: 9.92ms, 实际迭代时间: 816.34ms, GPU利用率: 1.2% +[2025-07-05 02:51:53] ================================================== +[2025-07-05 02:51:53] === 三元组预测示例 === +[2025-07-05 02:51:53] 样本1目标: Yariminai Station instance of railway station +[2025-07-05 02:51:53] 样本1预测: countryH entalilianaran ick-ov wayance dam ofationation ra St +[2025-07-05 02:51:53] 样本2目标: Ga'ash instance of kibbutz +[2025-07-05 02:51:53] 样本2预测: countryM entthaara alk Mov izance bir of occth Sed +[2025-07-05 02:51:53] ================== diff --git a/stat_predicate_vocab.py b/stat_predicate_vocab.py new file mode 100644 index 0000000..34b7460 --- /dev/null +++ b/stat_predicate_vocab.py @@ -0,0 +1,27 @@ +import json +from collections import Counter + +input_path = '/home/rwkv/RWKV-TS/RETRO_TEST/extract/processed_trex_data.json' +output_path = '/home/rwkv/RWKV-TS/RETRO_TEST/extract/predicate_vocab.json' + +with open(input_path, 'r', encoding='utf-8') as f: + data = json.load(f) + +predicate_set = set() + +for item in data: + if 'target' in item and isinstance(item['target'], list): + # 用集合去重本条数据的谓词 + predicates_in_item = set() + for triple in item['target']: + if isinstance(triple, dict) and 'predicate' in triple: + predicates_in_item.add(triple['predicate']) + predicate_set.update(predicates_in_item) + +predicate_list = list(predicate_set) + +with open(output_path, 'w', encoding='utf-8') as f: + json.dump(predicate_list, f, ensure_ascii=False, indent=2) + +print(f'已统计{len(predicate_list)}个谓词，保存到 {output_path}') + diff --git a/test.py b/test.py new file mode 100644 index 0000000..5b30e2e --- /dev/null +++ b/test.py @@ -0,0 +1,243 @@ +import os +import json +import argparse +import torch +import numpy as np +from tqdm import tqdm +from transformers import AutoTokenizer +from model.model_extra import MiniMindLM +from model.LMConfig import LMConfig +from sklearn.metrics import accuracy_score, precision_recall_fscore_support, classification_report + +# 加载谓词词汇表 +PREDICATE_VOCAB_PATH = '/home/rwkv/RWKV-TS/RETRO_TEST/extract/predicate_vocab.json' +with open(PREDICATE_VOCAB_PATH, 'r', encoding='utf-8') as f: + PREDICATE_LIST = json.load(f) +PREDICATE2ID = {p: i for i, p in enumerate(PREDICATE_LIST)} +NUM_PREDICATES = len(PREDICATE_LIST) + +def evaluate_model(model, tokenizer, test_data, device): + """ + 评估模型性能 - 只关注谓词分类 + """ + model.eval() + results = [] + all_pred_predicates = [] + all_gold_predicates = [] + correct_predictions = 0 + total_predictions = 0 + + print("开始评估...") + + # 添加调试信息 + print(f"测试数据样本数量: {len(test_data)}") + if test_data: + print(f"第一个样本格式: {type(test_data[0])}") + print(f"第一个样本内容: {test_data[0]}") + if isinstance(test_data[0], dict): + print(f"第一个样本的键: {list(test_data[0].keys())}") + + for i, item in enumerate(tqdm(test_data, desc="评估进度")): + input_text = item["input"] + gold_triples = item.get("output", []) + + # 调试信息（前几个样本） + if i < 3: + print(f"\n样本 {i+1} 调试信息:") + print(f" 输入文本: {input_text[:100]}...") + print(f" 真值三元组数量: {len(gold_triples)}") + if gold_triples: + print(f" 真值三元组: {gold_triples[0]}") + + # 模型推理 + inputs = tokenizer(input_text, return_tensors="pt", truncation=True, max_length=512, padding='max_length') + input_ids = inputs["input_ids"].to(device) + + with torch.no_grad(): + output = model(input_ids=input_ids) + + # 获取谓词分类结果 + pred_predicate_id = output.predicate_cls_logits.argmax(-1).item() + pred_predicate = PREDICATE_LIST[pred_predicate_id] if pred_predicate_id < len(PREDICATE_LIST) else "" + + # 收集所有目标谓词 + target_predicates = [] + if gold_triples: + for triple in gold_triples: + if "predicate" in triple: + target_predicates.append(triple["predicate"]) + + # 检查预测是否正确（只要在目标谓词列表中就算正确） + is_correct = False + if target_predicates and pred_predicate in target_predicates: + is_correct = True + correct_predictions += 1 + total_predictions += 1 + + # 调试信息（前几个样本） + if i < 3: + print(f" 预测谓词: {pred_predicate}") + print(f" 目标谓词: {target_predicates}") + print(f" 是否正确: {is_correct}") + + # 收集谓词分类标签（用于详细分析） + if target_predicates: + # 取第一个目标谓词作为主要标签 + main_target = target_predicates[0] + if main_target in PREDICATE2ID: + all_gold_predicates.append(PREDICATE2ID[main_target]) + all_pred_predicates.append(pred_predicate_id) + + results.append({ + "input": input_text, + "predicted_predicate": pred_predicate, + "target_predicates": target_predicates, + "is_correct": is_correct + }) + + print(f"\n评估完成，总预测数: {total_predictions}, 正确数: {correct_predictions}") + return results, all_pred_predicates, all_gold_predicates, correct_predictions, total_predictions + +def print_evaluation_summary(results, pred_predicates, gold_predicates, correct_predictions, total_predictions): + """ + 打印评估结果摘要 - 只关注谓词分类 + """ + print("\n" + "="*60) + print("谓词分类评估结果摘要") + print("="*60) + + # 谓词分类准确率 + if total_predictions > 0: + predicate_accuracy = correct_predictions / total_predictions + print(f"谓词分类准确率: {predicate_accuracy:.4f} ({correct_predictions}/{total_predictions})") + else: + print("谓词分类准确率: 无法计算（没有有效预测）") + + # 详细的分类报告（如果有足够的标签数据） + if pred_predicates and gold_predicates and len(pred_predicates) > 10: + try: + print(f"\n谓词分类详细报告:") + print(classification_report(gold_predicates, pred_predicates, + target_names=PREDICATE_LIST[:10] + ["..."] if len(PREDICATE_LIST) > 10 else PREDICATE_LIST, + zero_division=0)) + except Exception as e: + print(f"\n谓词分类详细报告生成失败: {e}") + + # 样本预测示例 + print(f"\n预测示例 (前5个):") + for i, result in enumerate(results[:5]): + print(f"样本 {i+1}:") + print(f" 输入: {result['input'][:100]}...") + print(f" 预测谓词: {result['predicted_predicate']}") + print(f" 目标谓词: {result['target_predicates']}") + print(f" 是否正确: {'✓' if result['is_correct'] else '✗'}") + print() + +def main(): + parser = argparse.ArgumentParser(description="MiniMind 三元组抽取模型评估脚本") + parser.add_argument('--model_path', type=str, default='/home/rwkv/RWKV-TS/RETRO_TEST/Minimind/out/pretrain_cls512.pth') + parser.add_argument('--tokenizer_path', type=str, default='/home/rwkv/RWKV-TS/RETRO_TEST/Minimind/model/minimind_tokenizer') + parser.add_argument('--test_json', type=str, default='/home/rwkv/RWKV-TS/RETRO_TEST/extract/sample_1000.json') + parser.add_argument('--output_dir', type=str, default='/home/rwkv/RWKV-TS/RETRO_TEST/Minimind/out', help='输出目录') + parser.add_argument('--device', type=str, default='cuda', help='推理设备') + + # 模型配置参数 + parser.add_argument('--dim', default=512, type=int) + parser.add_argument('--n_layers', default=8, type=int) + parser.add_argument('--max_seq_len', default=512, type=int) + parser.add_argument('--use_moe', default=False, type=bool) + parser.add_argument('--disable_db', action='store_true', help="禁用数据库功能") + parser.add_argument('--flash_attn', action='store_true', default=True, help="启用FlashAttention") + parser.add_argument('--knowledge_num', type=int, default=960400, help="知识库的数据数目") + parser.add_argument('--knowledge_length', type=int, default=32, help="知识库的句子长度") + parser.add_argument('--embeddings_epoch', type=int, default=2, help="embedding训练的epoch数") + + args = parser.parse_args() + + os.makedirs(args.output_dir, exist_ok=True) + + # 加载模型和分词器 + print("加载模型和分词器...") + lm_config = LMConfig( + dim=args.dim, + n_layers=args.n_layers, + max_seq_len=args.max_seq_len, + use_moe=args.use_moe, + disable_db=args.disable_db, + flash_attn=args.flash_attn, + knowledge_num=args.knowledge_num, + knowledge_length=args.knowledge_length, + embeddings_epoch=args.embeddings_epoch + ) + + model = MiniMindLM(lm_config, mode="triple", num_predicates=NUM_PREDICATES) + + # 加载模型权重 + try: + state_dict = torch.load(args.model_path, map_location=args.device) + model.load_state_dict(state_dict, strict=False) + print(f"成功加载模型权重: {args.model_path}") + except Exception as e: + print(f"加载模型权重失败: {e}") + print("使用随机初始化的模型进行测试") + + model.to(args.device) + tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path) + + print(f"谓词词汇表大小: {len(PREDICATE_LIST)}") + + # 加载测试数据 + print(f"加载测试数据: {args.test_json}") + with open(args.test_json, 'r', encoding='utf-8') as f: + test_data = json.load(f) + + # 支持多种数据格式 + if isinstance(test_data[0], dict) and "text" in test_data[0]: + # 格式: [{"text": "...", "target": [...]}, ...] + test_data = [{"input": item["text"], "output": item.get("target", [])} for item in test_data] + elif isinstance(test_data[0], dict) and "input" in test_data[0]: + # 格式: [{"input": "...", "output": [...]}, ...] + pass + else: + # 格式: ["句子", ...] - 没有真值，只能做推理 + test_data = [{"input": text, "output": []} for text in test_data] + + print(f"测试样本数量: {len(test_data)}") + + # 评估模型 + results, pred_predicates, gold_predicates, correct_predictions, total_predictions = evaluate_model(model, tokenizer, test_data, args.device) + + # 打印评估结果 + print_evaluation_summary(results, pred_predicates, gold_predicates, correct_predictions, total_predictions) + + # 保存详细结果 + output_path = os.path.join(args.output_dir, 'evaluation_results.json') + with open(output_path, 'w', encoding='utf-8') as f: + json.dump(results, f, indent=2, ensure_ascii=False) + print(f"\n详细评估结果已保存到: {output_path}") + + # 保存准确率统计 + accuracy_stats = { + "total_predictions": total_predictions, + "correct_predictions": correct_predictions, + "accuracy": correct_predictions / total_predictions if total_predictions > 0 else 0.0, + "model_path": args.model_path, + "test_data_path": args.test_json, + "predicate_vocab_size": len(PREDICATE_LIST), + "evaluation_timestamp": str(np.datetime64('now')) + } + + accuracy_path = os.path.join(args.output_dir, 'accuracy_stats.json') + with open(accuracy_path, 'w', encoding='utf-8') as f: + json.dump(accuracy_stats, f, indent=2, ensure_ascii=False) + print(f"准确率统计已保存到: {accuracy_path}") + + # 保存预测结果 + predictions = [{"input": r["input"], "predicted_predicate": r["predicted_predicate"], "gold_predicates": r["target_predicates"]} for r in results] + pred_output_path = os.path.join(args.output_dir, 'predictions.json') + with open(pred_output_path, 'w', encoding='utf-8') as f: + json.dump(predictions, f, indent=2, ensure_ascii=False) + print(f"预测结果已保存到: {pred_output_path}") + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/train_extra_accelerate.py b/train_extra_accelerate.py index 2cd34d7..1c0b148 100644 --- a/train_extra_accelerate.py +++ b/train_extra_accelerate.py @@ -25,6 +25,7 @@ import swanlab # 替换wandb导入 import gc # 添加垃圾回收模块 import psutil # 添加系统资源监控模块 import os +import json os.environ['CUDA_VISIBLE_DEVICES']='2' from model.model_extra import MiniMindLM, RMSNorm # 使用model_extra @@ -208,7 +209,7 @@ def compute_cosine_similarity_batch(pred_embeddings, target_embeddings): return similarities -def triple_to_sentence(subject_logits, predicate_logits, object_logits, tokenizer): +def triple_to_sentence(subject_logits, predicate_logits, object_logits, tokenizer, predicate_cls_logits=None): """ 将三元组logits转换为句子 Args: @@ -216,54 +217,54 @@ def triple_to_sentence(subject_logits, predicate_logits, object_logits, tokenize predicate_logits: [batch_size, seq_len, max_predicate_len, vocab_size] object_logits: [batch_size, seq_len, max_object_len, vocab_size] tokenizer: 分词器 + predicate_cls_logits: [batch_size, num_predicates]，如果提供则用分类结果输出谓词 Returns: - List[List[str]]: 每个样本每个位置的三元组句子 + List[str]: 每个样本的三元组句子 """ batch_size = subject_logits.shape[0] - predicate_seq_len = predicate_logits.shape[1] + # 主语 subject_seq_len = subject_logits.shape[1] + subject_logits_ = subject_logits.reshape(batch_size * subject_seq_len, -1) + subject_ids = torch.argmax(subject_logits_, dim=-1) + subject_ids = subject_ids.reshape(batch_size, subject_seq_len) + # 宾语 object_seq_len = object_logits.shape[1] + object_logits_ = object_logits.reshape(batch_size * object_seq_len, -1) + object_ids = torch.argmax(object_logits_, dim=-1) + object_ids = object_ids.reshape(batch_size, object_seq_len) - predicate_logits = predicate_logits.reshape(batch_size*predicate_seq_len, -1) - subject_logits = subject_logits.reshape(batch_size*subject_seq_len, -1) - object_logits = object_logits.reshape(batch_size*object_seq_len, -1) + # 谓词 + predicate_texts = [] + if predicate_cls_logits is not None: + # 用分类结果输出谓词 + pred_ids = torch.argmax(predicate_cls_logits, dim=-1) # [batch_size] + for i in range(batch_size): + pred_id = pred_ids[i].item() + pred_text = PREDICATE_LIST[pred_id] if pred_id < len(PREDICATE_LIST) else "" + predicate_texts.append(pred_text) + else: + # 兼容原有行为：用序列生成的谓词 + predicate_seq_len = predicate_logits.shape[1] + predicate_logits_ = predicate_logits.reshape(batch_size * predicate_seq_len, -1) + predicate_ids = torch.argmax(predicate_logits_, dim=-1) + predicate_ids = predicate_ids.reshape(batch_size, predicate_seq_len) + predicate_texts = tokenizer.batch_decode(predicate_ids, skip_special_tokens=True) - predicate_logits = torch.argmax(predicate_logits, dim=-1) - subject_logits = torch.argmax(subject_logits, dim=-1) - object_logits = torch.argmax(object_logits, dim=-1) + # 主语和宾语文本 + subject_texts = tokenizer.batch_decode(subject_ids, skip_special_tokens=True) + object_texts = tokenizer.batch_decode(object_ids, skip_special_tokens=True) - predicate_logits = predicate_logits.reshape(batch_size, predicate_seq_len) - subject_logits = subject_logits.reshape(batch_size, subject_seq_len) - object_logits = object_logits.reshape(batch_size, object_seq_len) - - combined_logits = torch.cat([subject_logits, predicate_logits, object_logits], dim=1) - - sentences = tokenizer.batch_decode(combined_logits, skip_special_tokens=True) - - # sentences = [] - - # for batch_idx in range(batch_size): - # batch_sentences = [] - # for seq_idx in range(seq_len): - # # 获取预测的token ids - # subject_ids = torch.argmax(subject_logits[batch_idx, seq_idx], dim=-1) - # predicate_ids = torch.argmax(predicate_logits[batch_idx, seq_idx], dim=-1) - # object_ids = torch.argmax(object_logits[batch_idx, seq_idx], dim=-1) - - # # 转换为文本 - # subject_text = tokenizer.decode(subject_ids, skip_special_tokens=True).strip() - # predicate_text = tokenizer.decode(predicate_ids, skip_special_tokens=True).strip() - # object_text = tokenizer.decode(object_ids, skip_special_tokens=True).strip() - - # # 拼接为句子 (主语 + 谓语 + 宾语) - # if subject_text and predicate_text and object_text: - # sentence = f"{subject_text} {predicate_text} {object_text}" - # else: - # sentence = "" - - # batch_sentences.append(sentence) - # sentences.append(batch_sentences) - + # 拼接为三元组句子 + sentences = [] + for i in range(batch_size): + subject = subject_texts[i].strip() + predicate = predicate_texts[i].strip() if isinstance(predicate_texts[i], str) else str(predicate_texts[i]) + object_ = object_texts[i].strip() + if subject and predicate and object_: + sentence = f"{subject} {predicate} {object_}" + else: + sentence = "" + sentences.append(sentence) return sentences def compute_triple_rouge_loss_optimized(subject_logits, predicate_logits, object_logits, @@ -414,10 +415,17 @@ def get_lr(it, num_iters, learning_rate): # 余弦学习率衰减 return learning_rate * 0.5 * (1.0 + math.cos(math.pi * it / num_iters)) +# 加载谓词类别 +PREDICATE_VOCAB_PATH = '/home/rwkv/RWKV-TS/RETRO_TEST/extract/predicate_vocab.json' +with open(PREDICATE_VOCAB_PATH, 'r', encoding='utf-8') as f: + PREDICATE_LIST = json.load(f) +PREDICATE2ID = {p: i for i, p in enumerate(PREDICATE_LIST)} +NUM_PREDICATES = len(PREDICATE_LIST) + # 初始化模型函数 def init_model(lm_config, pretrained_embedding_path=None, database_init_path=None, args=None): tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer') - model = MiniMindLM(lm_config, mode="triple") # 设置为三元组模式 + model = MiniMindLM(lm_config, mode="triple", num_predicates=NUM_PREDICATES) # 加载预训练权重 pretrained_path = "/home/rwkv/RWKV-TS/RETRO_TEST/extract/Experiment_1_2_2_pretrain_512.pth" @@ -553,6 +561,7 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a last_log_time = epoch_start_time # 使用DataLoader内置的iterator，移除自定义预取 + criterion_predicate = nn.CrossEntropyLoss() for step, batch_data in enumerate(train_loader): # === 每个step开始 === @@ -611,7 +620,7 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a loss_start.record() # 计算优化后的嵌入余弦相似度损失 - loss = compute_triple_rouge_loss_optimized( + loss_triple = compute_triple_rouge_loss_optimized( res.subject_logits, res.predicate_logits, res.object_logits, target_input_ids, target_attention_mask, model.tok_embeddings, temperature=args.temperature ) @@ -624,8 +633,13 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a Logger(f"Error: ROUGE loss computation failed: {e}", accelerator) import traceback Logger(f"Traceback: {traceback.format_exc()}", accelerator) - loss = res.logits.sum() * 0.0 + 1.0 + loss_triple = res.logits.sum() * 0.0 + 1.0 + # 谓词分类loss + loss_predicate = criterion_predicate(res.predicate_cls_logits, batch_data['predicate_label'].to(accelerator.device)) + + # 总loss + loss = 0.99*loss_triple + 0.01*loss_predicate loss = loss / args.accumulation_steps # === 5. 反向传播 === @@ -686,7 +700,7 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a Logger("=" * 50, accelerator) Logger("=== 三元组预测示例 ===", accelerator) - predict_sentences = triple_to_sentence(res.subject_logits, res.predicate_logits, res.object_logits,tokenizer) + predict_sentences = triple_to_sentence(res.subject_logits, res.predicate_logits, res.object_logits, tokenizer) # 显示前2个样本的目标句子 for i, target_sentence in enumerate(target_sentences[:2]): Logger(f"样本{i+1}目标: {target_sentence}", accelerator) @@ -728,7 +742,8 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a # 基本训练信息 Logger(f"Epoch {epoch+1}/{args.epochs}, Step {step+1}/{total_steps_in_epoch}, " - f"Loss: {loss.item() * args.accumulation_steps:.6f}, " + f"Loss(triple): {loss_triple.item() * args.accumulation_steps:.6f}, " + f"Loss(predicate): {loss_predicate.item() * args.accumulation_steps:.6f}, " f"LR: {current_lr:.6f}, " f"Speed: {tokens_per_sec:.2f} tokens/sec | " f"Epoch Time Left: {format_time(epoch_remaining_time)} | " @@ -740,7 +755,8 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a "epoch": epoch + 1, "step": step + 1, "total_steps_in_epoch": total_steps_in_epoch, - "triple_embedding_cosine_loss": loss.item() * args.accumulation_steps, + "triple_embedding_cosine_loss": loss_triple.item() * args.accumulation_steps, + "predicate_cross_entropy_loss": loss_predicate.item() * args.accumulation_steps, "lr": current_lr, "tokens_per_sec": tokens_per_sec, "epoch_time_left_seconds": epoch_remaining_time, @@ -753,7 +769,7 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a loss_total = loss.item() * args.accumulation_steps if epoch > 1 and best_loss > loss_total and accelerator.is_main_process: best_loss = loss_total - ckp = f'{args.save_dir}/pretrain_{args.dim}{moe_path}.pth' + ckp = f'{args.save_dir}/pretrain_cls{args.dim}{moe_path}.pth' unwrapped_model = accelerator.unwrap_model(model) accelerator.save(unwrapped_model.state_dict(), ckp) Logger(f"Model saved to {ckp}", accelerator) @@ -945,14 +961,15 @@ def main(): target_input_ids = torch.stack([item['target_input_ids'] for item in batch]) target_attention_mask = torch.stack([item['target_attention_mask'] for item in batch]) target_sentences = [item['target_sentence'] for item in batch] # 用于调试 - + predicate_label = torch.stack([item['predicate_label'] for item in batch]) return { 'input_ids': input_ids, 'labels': labels, 'loss_mask': loss_mask, 'target_input_ids': target_input_ids, 'target_attention_mask': target_attention_mask, - 'target_sentences': target_sentences + 'target_sentences': target_sentences, + 'predicate_label': predicate_label } train_loader = DataLoader(