deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-04 17:39:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

add vllm inference

Browse Source
This commit is contained in:
lyuxiang.lx
2025-05-30 07:22:35 +00:00
parent 9f55c5af8f
commit 6dd68b9d5e
6 changed files with 105 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
cosyvoice/utils/common.py
View File

@@ -169,17 +169,18 @@ def mask_to_bias(mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
class TrtContextWrapper:
def __init__(self, trt_engine, trt_concurrent=1):
self.trt_context_pool = queue.Queue()
def __init__(self, trt_engine, trt_concurrent=1, device='cuda:0'):
self.trt_context_pool = queue.Queue(maxsize=trt_concurrent)
self.trt_engine = trt_engine
for _ in range(trt_concurrent):
trt_context = trt_engine.create_execution_context()
trt_stream = torch.cuda.stream(torch.cuda.Stream(device))
assert trt_context is not None, 'failed to create trt context, maybe not enough CUDA memory, try reduce current trt concurrent {}'.format(trt_concurrent)
self.trt_context_pool.put(trt_context)
self.trt_context_pool.put([trt_context, trt_stream])
assert self.trt_context_pool.empty() is False, 'no avaialbe estimator context'
def acquire_estimator(self):
return self.trt_context_pool.get(), self.trt_engine
def release_estimator(self, context):
self.trt_context_pool.put(context)
def release_estimator(self, context, stream):
self.trt_context_pool.put([context, stream])