[Frontend][torch.compile] CompilationConfig Overhaul (#20283): name change compilation level to compilation mode, deprecation compilation level (#26355)

Signed-off-by: morrison-turnansky <mturnans@redhat.com>
Signed-off-by: Morrison Turnansky <mturnans@redhat.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
Morrison Turnansky
2025-10-15 02:51:16 +00:00
committed by GitHub
co-authored by Luka Govedič
parent e66d787bce
commit 96b9aa5aa0
42 changed files with 270 additions and 248 deletions
+8 -7
View File
@@ -25,7 +25,7 @@ from vllm.compilation.counter import compilation_counter
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
from vllm.config import (
CompilationLevel,
CompilationMode,
CUDAGraphMode,
VllmConfig,
get_layers_from_vllm_config,
@@ -2927,14 +2927,15 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
)
if (
self.vllm_config.compilation_config.level == CompilationLevel.DYNAMO_AS_IS
self.vllm_config.compilation_config.mode
== CompilationMode.STOCK_TORCH_COMPILE
and supports_dynamo()
):
backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
compilation_counter.dynamo_as_is_count += 1
compilation_counter.stock_torch_compile_count += 1
self.model.compile(fullgraph=True, backend=backend)
return
# for other compilation levels, cudagraph behavior is controlled by
# for other compilation modes, cudagraph behavior is controlled by
# CudagraphWraper and CudagraphDispatcher of vllm.
# wrap the model with full cudagraph wrapper if needed.
@@ -3985,7 +3986,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
# if not supported any full cudagraphs, just raise it.
msg += (
"; please try cudagraph_mode=PIECEWISE, and "
"make sure compilation level is piecewise"
"make sure compilation mode is VLLM_COMPILE"
)
raise ValueError(msg)
@@ -4012,7 +4013,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
f"with {min_cg_builder_name} backend (support: "
f"{min_cg_support})"
)
if self.compilation_config.level == CompilationLevel.PIECEWISE and (
if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
self.compilation_config.splitting_ops_contain_attention()
or self.compilation_config.use_inductor_graph_partition
):
@@ -4068,7 +4069,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
f"supported with {min_cg_builder_name} backend ("
f"support:{min_cg_support}) "
"; please try cudagraph_mode=PIECEWISE, "
"and make sure compilation level is piecewise"
"and make sure compilation mode is VLLM_COMPILE"
)
# Trigger cudagraph dispatching keys initialization here (after