[Bugfix] VLLM_V1 supports passing other compilation levels (#19340)
Signed-off-by: Richard Zou <zou3519@gmail.com>
This commit is contained in:
@@ -22,6 +22,7 @@ import vllm.envs as envs
|
||||
from vllm.attention import AttentionMetadata, get_attn_backend
|
||||
from vllm.attention.backends.abstract import AttentionState
|
||||
from vllm.attention.backends.utils import CommonAttentionState
|
||||
from vllm.compilation.counter import compilation_counter
|
||||
from vllm.config import CompilationLevel, VllmConfig
|
||||
from vllm.core.scheduler import SchedulerOutputs
|
||||
from vllm.distributed import broadcast_tensor_dict, get_pp_group
|
||||
@@ -1121,6 +1122,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
|
||||
CompilationLevel.DYNAMO_AS_IS and supports_dynamo():
|
||||
backend = self.vllm_config.compilation_config.init_backend(
|
||||
self.vllm_config)
|
||||
compilation_counter.dynamo_as_is_count += 1
|
||||
self.model = torch.compile(
|
||||
self.model,
|
||||
fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
|
||||
|
||||
Reference in New Issue
Block a user