🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
update generate
jingyaogong9e6779839751
diff --git a/model/model.py b/model/model.pyindex 06f791d..d53b57a 100644--- a/model/model.py+++ b/model/model.py@@ -357,9 +357,7 @@ class MiniMindLM(PreTrainedModel):for seq in generated]output = torch.cat(generated, dim=0)- res = output.view(input_ids.size(0), num_return_sequences, -1)- res = res.squeeze(0) if input_ids.size(0) == 1 else res- res = res.squeeze(1) if num_return_sequences == 1 else res+ res = output.view(input_ids.size(0) * num_return_sequences, -1)return resdef _stream(self, input_ids, eos_token_id, max_new_tokens, temperature, top_p, rp, use_cache, **args):