[Multimodal][torch.compile] Add compilation config field for turning off ViT/MM compile (#28242)

Signed-off-by: Lucas Kabela <lucaskabela@meta.com>
This commit is contained in:
Lucas Kabela 2025-11-06 16:16:03 -08:00 committed by GitHub
parent 59b453eaa2
commit 4bf56c79cc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 60 additions and 3 deletions

View File

@ -3,10 +3,17 @@
import pytest
from vllm.compilation.counter import compilation_counter
from vllm.config import VllmConfig
from vllm.config.compilation import CompilationMode
from vllm.platforms import current_platform
def test_compile():
vllm_config = VllmConfig()
# Default configuration compiles mm encoder
assert vllm_config.compilation_config.compile_mm_encoder
# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
@pytest.mark.forked
@pytest.mark.skipif(not current_platform.is_cuda(), reason="Skip if not cuda")
@ -31,8 +38,33 @@ def test_qwen2_5_vl_compilation(vllm_runner, monkeypatch):
vllm_runner(
"Qwen/Qwen2.5-VL-3B-Instruct",
max_model_len=2048,
gpu_memory_utilization=0.7,
gpu_memory_utilization=0.8,
compilation_config={"mode": CompilationMode.VLLM_COMPILE},
) as _,
):
pass
# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
@pytest.mark.forked
@pytest.mark.skipif(not current_platform.is_cuda(), reason="Skip if not cuda")
def test_qwen2_5_vl_no_vit_compilation(vllm_runner, monkeypatch):
"""Test that Qwen2.5-VL vision submodules are not compiled when the
config is passed off
"""
# Disable multiprocessing so that the counter is in the same process
monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
with (
compilation_counter.expect(num_models_seen=1),
vllm_runner(
"Qwen/Qwen2.5-VL-3B-Instruct",
max_model_len=2048,
gpu_memory_utilization=0.8,
compilation_config={
"mode": CompilationMode.VLLM_COMPILE,
"compile_mm_encoder": False,
},
) as _,
):
pass

View File

@ -150,6 +150,7 @@ class CompilationConfig:
- [`backend`][vllm.config.CompilationConfig.backend]
- [`custom_ops`][vllm.config.CompilationConfig.custom_ops]
- [`splitting_ops`][vllm.config.CompilationConfig.splitting_ops]
- [`compile_mm_encoder`][vllm.config.CompilationConfig.compile_mm_encoder]
- CudaGraph capture:
- [`use_cudagraph`][vllm.config.CompilationConfig.use_cudagraph]
- [`cudagraph_mode`][vllm.config.CompilationConfig.cudagraph_mode]
@ -250,6 +251,13 @@ class CompilationConfig:
disabled when running with Inductor: mode>=VLLM_COMPILE and use_inductor=True.
Inductor generates (fused) Triton kernels for disabled custom ops."""
splitting_ops: list[str] | None = None
"""
Provide control over whether to compile the multimodal encoder
such as Qwen2_5_vl
"""
compile_mm_encoder: bool = True
"""A list of ops to exclude from cudagraphs, used in piecewise compilation.
The behavior depends on use_inductor_graph_partition:

View File

@ -67,6 +67,9 @@ from vllm.model_executor.layers.linear import (
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.module_mapping import MultiModelKeys
from vllm.model_executor.models.transformers.utils import (
should_torch_compile_mm_vit,
)
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.evs import (
compute_mrope_for_media,
@ -464,6 +467,7 @@ class Qwen2_5_VisionAttention(nn.Module):
"seqlens": 0,
},
mark_unbacked_dims={"seqlens": 0},
enable_if=should_torch_compile_mm_vit,
)
class Qwen2_5_VisionBlock(nn.Module):
def __init__(
@ -529,7 +533,8 @@ class Qwen2_5_VisionBlock(nn.Module):
@support_torch_compile(
dynamic_arg_dims={
"x": 0,
}
},
enable_if=should_torch_compile_mm_vit,
)
class Qwen2_5_VisionPatchEmbed(nn.Module):
def __init__(
@ -560,7 +565,8 @@ class Qwen2_5_VisionPatchEmbed(nn.Module):
@support_torch_compile(
dynamic_arg_dims={
"x": 0,
}
},
enable_if=should_torch_compile_mm_vit,
)
class Qwen2_5_VisionPatchMerger(nn.Module):
def __init__(

View File

@ -205,3 +205,14 @@ def can_enable_torch_compile(vllm_config: "VllmConfig") -> bool:
# Dynamic rope scaling is not compatible with torch.compile
rope_scaling: dict = getattr(text_config, "rope_scaling", None) or {}
return rope_scaling.get("rope_type") != "dynamic"
def should_torch_compile_mm_vit(vllm_config: "VllmConfig") -> bool:
"""
Callable to be passed to `@support_torch_compile`'s `enable_if` argument.
Defaults to `True` but is disabled in the following situations:
- The model uses dynamic rope scaling.
"""
return vllm_config.compilation_config.compile_mm_encoder