Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

oom fix

游雁604ae30fdbe9
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 60bfeff10..96c642e52 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -147,7 +147,9 @@ class AutoModel:
# if spk_model is not None, build spk model else None
spk_model = kwargs.get("spk_model", None)
spk_kwargs = {} if kwargs.get("spk_kwargs", {}) is None else kwargs.get("spk_kwargs", {})
- cb_kwargs = {} if spk_kwargs.get("cb_kwargs", {}) is None else spk_kwargs.get("cb_kwargs", {})
+ cb_kwargs = (
+ {} if spk_kwargs.get("cb_kwargs", {}) is None else spk_kwargs.get("cb_kwargs", {})
+ )
if spk_model is not None:
logging.info("Building SPK model.")
spk_kwargs["model"] = spk_model
@@ -368,8 +370,8 @@ class AutoModel:
pbar.set_description(f"rtf_avg: {time_escape_total/time_speech_total:0.3f}")
device = next(model.parameters()).device
- if device.type == 'cuda':
- with torch.cuda.device():
+ if device.type == "cuda":
+ with torch.cuda.device(device):
torch.cuda.empty_cache()
return asr_result_list
diff --git a/funasr/bin/train.py b/funasr/bin/train.py
index 45e23a8c6..c56d04752 100644
--- a/funasr/bin/train.py
+++ b/funasr/bin/train.py
@@ -221,10 +221,10 @@ def main(**kwargs):
)
trainer.start_step = 0
- # device = next(model.parameters()).device
- # if device.type == 'cuda':
- # with torch.cuda.device():
- # torch.cuda.empty_cache()
+ device = next(model.parameters()).device
+ if device.type == "cuda":
+ with torch.cuda.device(device):
+ torch.cuda.empty_cache()
time_escaped = (time.perf_counter() - time_slice_i) / 3600.0
logging.info(
diff --git a/funasr/bin/train_ds.py b/funasr/bin/train_ds.py
index b28752ba8..10a5d087f 100644
--- a/funasr/bin/train_ds.py
+++ b/funasr/bin/train_ds.py
@@ -184,10 +184,10 @@ def main(**kwargs):
)
trainer.start_step = 0
- # device = next(model.parameters()).device
- # if device.type == 'cuda':
- # with torch.cuda.device():
- # torch.cuda.empty_cache()
+ device = next(model.parameters()).device
+ if device.type == "cuda":
+ with torch.cuda.device(device):
+ torch.cuda.empty_cache()
time_escaped = (time.perf_counter() - time_slice_i) / 3600.0
logging.info(