[compile] Skip FX graph deserialiaztion on loading, further reducing warm compile time. (#40151)
Signed-off-by: zhxchen17 <zhxchen17@fb.com>
This commit is contained in:
@@ -23,6 +23,10 @@ from torch._logging._internal import trace_structured
|
||||
from torch.fx._lazy_graph_module import _use_lazy_graph_module
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.compilation.codegen import (
|
||||
compile_execution_fn,
|
||||
generate_execution_code,
|
||||
)
|
||||
from vllm.config import CompilationConfig, CUDAGraphMode, VllmConfig
|
||||
from vllm.config.compilation import DynamicShapesType
|
||||
from vllm.config.utils import Range, hash_factors
|
||||
@@ -1244,11 +1248,6 @@ class VllmBackend:
|
||||
original_split_gm if envs.VLLM_USE_MEGA_AOT_ARTIFACT else self.graph
|
||||
)
|
||||
|
||||
from vllm.compilation.codegen import (
|
||||
compile_execution_fn,
|
||||
generate_execution_code,
|
||||
)
|
||||
|
||||
execution_code, submod_names = generate_execution_code(self.split_gm)
|
||||
# Use getattr to get correct callables: __dict__ has PiecewiseBackend
|
||||
# instances (from PiecewiseCompileInterpreter), _modules has originals.
|
||||
|
||||
Reference in New Issue
Block a user