Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
fix: avoid deep-copying the checkpoint in load_pretrained_model (#3729)
YYDH9e7f6af8c30b
diff --git a/funasr/train_utils/load_pretrained_model.py b/funasr/train_utils/load_pretrained_model.pyindex 383c22a4b..042887d8d 100644--- a/funasr/train_utils/load_pretrained_model.py+++ b/funasr/train_utils/load_pretrained_model.py@@ -8,7 +8,6 @@ import torchimport torch.nnimport torch.optimimport pdb-import copydef load_pretrained_model(@@ -41,7 +40,10 @@ def load_pretrained_model(buffer = BytesIO(oss_bucket.get_object(path).read())ori_state = torch.load(buffer, map_location=map_location)- src_state = copy.deepcopy(ori_state)+ # ori_state is created by torch.load inside this function, is not shared+ # with the caller, and is not read again below; src_state is only read.+ # Deep-copying every tensor here doubles peak load memory with no effect.+ src_state = ori_statesrc_state = src_state["state_dict"] if "state_dict" in src_state else src_statesrc_state = src_state["model_state_dict"] if "model_state_dict" in src_state else src_statesrc_state = src_state["model"] if "model" in src_state else src_state