xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-04-01 04:27:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

Clean code

Browse Source 
Signed-off-by: Yu Gong <yu3.gong@gmail.com>
This commit is contained in:
Yu Gong 2025-12-18 21:09:47 +00:00
parent ea1a26d276
commit aa7917aaaa
2 changed files with 1 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/v1/cudagraph_dispatcher.py
View File

@ -109,14 +109,6 @@ class CudagraphDispatcher:
max_loras = self.vllm_config.lora_config.max_loras
# When speculative decoding is enabled, only capture with max_loras
# to avoid torch.compile conflicts during CUDA graph capture
if self.vllm_config.speculative_config is not None:
lora_cases = [(True, max_loras)]
if self.compilation_config.cudagraph_specialize_lora:
lora_cases.append((False, 0))
return lora_cases
# Capture for each num_active_loras from 1 to max_loras
lora_cases = [(True, n) for n in range(1, max_loras + 1)]
# Also capture the no-lora case

20
vllm/v1/worker/gpu_model_runner.py
View File

@ -4598,24 +4598,6 @@ class GPUModelRunner(
self.encoder_cache.clear()
gc.collect()
def _get_lora_capture_cases(self) -> list[tuple[bool, int]]:
"""
Returns list of (has_lora, num_active_loras) tuples for CUDA graph capture.
Returns cases for each num_active_loras from 1 to max_loras.
If cudagraph_specialize_lora is True, also includes the no-lora case.
"""
if not self.lora_config:
return [(False, 0)]
max_loras = self.lora_config.max_loras
# Capture for each num_active_loras from 1 to max_loras
lora_cases = [(True, n) for n in range(1, max_loras + 1)]
# Also capture the no-lora case if cudagraph_specialize_lora is True
if self.compilation_config.cudagraph_specialize_lora:
lora_cases.append((False, 0))
return lora_cases
def capture_model(self) -> int:
if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
logger.warning(
@ -4654,7 +4636,7 @@ class GPUModelRunner(
assert cudagraph_mode is not None
# Build LoRA cases: list of (has_lora, num_active_loras) tuples
lora_cases = self._get_lora_capture_cases()
lora_cases = self.cudagraph_dispatcher._get_lora_cases()
if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
cudagraph_runtime_mode = cudagraph_mode.mixed_mode()