From 0b85f814edf686e7cec90f93f6dd2518861a46e3 Mon Sep 17 00:00:00 2001 From: jingyaogong Date: Thu, 7 May 2026 19:07:26 +0800 Subject: [PATCH] [fix] existing bugs --- model/model_minimind.py | 4 +--- model/model_omni.py | 11 +++-------- 2 files changed, 4 insertions(+), 11 deletions(-) diff --git a/model/model_minimind.py b/model/model_minimind.py index 7a82721..dc080e5 100755 --- a/model/model_minimind.py +++ b/model/model_minimind.py @@ -267,9 +267,7 @@ class MiniMindForCausalLM(PreTrainedModel, GenerationMixin): logits = outputs.logits[:, -1, :] / temperature if repetition_penalty != 1.0: for i in range(input_ids.shape[0]): - seen = torch.unique(input_ids[i]) - score = logits[i, seen] - logits[i, seen] = torch.where(score > 0, score / repetition_penalty, score * repetition_penalty) + seen = torch.unique(input_ids[i]); score = logits[i, seen]; logits[i, seen] = torch.where(score > 0, score / repetition_penalty, score * repetition_penalty) if top_k > 0: logits[logits < torch.topk(logits, top_k)[0][..., -1, None]] = -float('inf') if top_p < 1.0: diff --git a/model/model_omni.py b/model/model_omni.py index 754ac03..43c504f 100644 --- a/model/model_omni.py +++ b/model/model_omni.py @@ -339,7 +339,6 @@ class MiniMindOmni(MiniMindForCausalLM): if spk_emb is not None and fill_start > 0: audio_buffer[:, :, fill_start - 1] = self.audio_spk_token think_end_step, generated_tokens = None, ([] if args.get('open_thinking', False) else None) - seen_tokens = set(input_ids[0].tolist()) if rp != 1.0 else None while input_ids.shape[1] < start_pos + max_new_tokens: if past_kvs is None or not use_cache: out = self.forward(torch.cat((audio_buffer, input_ids.unsqueeze(1)), dim=1), past_key_values=past_kvs, use_cache=use_cache, **args) @@ -348,10 +347,8 @@ class MiniMindOmni(MiniMindForCausalLM): past_kvs = out.past_key_values logits = out.logits[0, -1, :].clone() / (temperature + 1e-9) - if seen_tokens is not None: - seen = list(seen_tokens) - score = logits[seen] - logits[seen] = torch.where(score > 0, score / rp, score * rp) + if rp != 1.0: + seen = list(set(input_ids[0].tolist())); score = logits[seen]; logits[seen] = torch.where(score > 0, score / rp, score * rp) if top_p and top_p < 1.0: sorted_l, sorted_i = torch.sort(logits, descending=True) mask = torch.cumsum(F.softmax(sorted_l, dim=-1), dim=-1) > top_p @@ -362,8 +359,6 @@ class MiniMindOmni(MiniMindForCausalLM): if text_finished: text_token = args.get('enter_token_id', 201) if first_finished else args.get('pad_token_id', 0) first_finished = False - if seen_tokens is not None: - seen_tokens.add(text_token) step = input_ids.shape[1] - start_pos # 已生成token数（0=首次，此时模型处理prompt末尾token） audio_step = step - 1 # 延迟1步：输出第1个text时无audio，输出第2个text时layer0开始 @@ -376,7 +371,7 @@ class MiniMindOmni(MiniMindForCausalLM): audio_codes[i].append(self.audio_pad_token) else: logits_i = al[0, -1, :].clone() / 0.2 - for prev_code in audio_codes[i][-3:]: logits_i[prev_code] /= 1.05 + for prev_code in audio_codes[i][-3:]: score = logits_i[prev_code]; logits_i[prev_code] = torch.where(score > 0, score / 1.05, score * 1.05) top_val, top_idx = logits_i.topk(50) code = top_idx[torch.multinomial(F.softmax(top_val, dim=-1), 1)].item() audio_codes[i].append(code)