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: resolve unexpected 'out of memory' issue in multi-GPU setup (#2373)
BienBoyc1e365fea09a
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex 08308a244..ec4e42013 100644--- a/funasr/auto/auto_model.py+++ b/funasr/auto/auto_model.py@@ -366,7 +366,8 @@ class AutoModel:if pbar:# pbar.update(1)pbar.set_description(f"rtf_avg: {time_escape_total/time_speech_total:0.3f}")- torch.cuda.empty_cache()+ with torch.cuda.device(next(model.parameters()).device):+ torch.cuda.empty_cache()return asr_result_listdef inference_with_vad(self, input, input_len=None, **cfg):diff --git a/funasr/bin/train.py b/funasr/bin/train.pyindex fcd763fa0..d0f154a0a 100644--- a/funasr/bin/train.py+++ b/funasr/bin/train.py@@ -221,7 +221,8 @@ def main(**kwargs):)trainer.start_step = 0- torch.cuda.empty_cache()+ with torch.cuda.device(kwargs["device"]):+ torch.cuda.empty_cache()time_escaped = (time.perf_counter() - time_slice_i) / 3600.0logging.info(diff --git a/funasr/bin/train_ds.py b/funasr/bin/train_ds.pyindex dc7fb4223..24e81f617 100644--- a/funasr/bin/train_ds.py+++ b/funasr/bin/train_ds.py@@ -184,7 +184,8 @@ def main(**kwargs):)trainer.start_step = 0- torch.cuda.empty_cache()+ with torch.cuda.device(kwargs["device"]):+ torch.cuda.empty_cache()time_escaped = (time.perf_counter() - time_slice_i) / 3600.0logging.info(diff --git a/funasr/models/language_model/rnn/decoders.py b/funasr/models/language_model/rnn/decoders.pyindex e7d35e92f..314d49f3b 100644--- a/funasr/models/language_model/rnn/decoders.py+++ b/funasr/models/language_model/rnn/decoders.py@@ -873,7 +873,8 @@ class Decoder(torch.nn.Module, ScorerInterface):ctc_state[idx], accum_best_ids)- torch.cuda.empty_cache()+ with torch.cuda.device(vscores.device):+ torch.cuda.empty_cache()dummy_hyps = [{"yseq": [self.sos, self.eos], "score": np.array([-float("inf")])}]ended_hyps = [