[MRV2] Miscellaneous code structure updates (#60754)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -1091,12 +1091,12 @@ class VllmBackend:
|
||||
# TODO(patchy): the V1 torch.compile ngram-gpu kernel causes vllm
|
||||
# torch compile cache errors. The V2 implementation is pure Triton and
|
||||
# does not need the cache disabled.
|
||||
is_ngram_gpu_enabled = (
|
||||
is_v1_ngram_gpu_enabled = (
|
||||
vllm_config.speculative_config is not None
|
||||
and vllm_config.speculative_config.use_ngram_gpu()
|
||||
and not vllm_config.use_v2_model_runner
|
||||
)
|
||||
disable_cache = disable_cache or is_ngram_gpu_enabled
|
||||
disable_cache = disable_cache or is_v1_ngram_gpu_enabled
|
||||
|
||||
if disable_cache:
|
||||
logger.info("%svLLM's torch.compile cache is disabled.", self._log_prefix())
|
||||
|
||||
Reference in New Issue
Block a user