Skip to content

[MM] Keep device input normalization fused when encoder compilation is enabled #59195

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Isotr0py merged 3 commits into from
Sep 30, 2026
Merged

[MM] Keep device input normalization fused when encoder compilation is enabled #59195

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d4fbdf0
Keep Pixtral input normalization fused with encoder compilation
jzakrzew Sep 29, 2026
23d27b9
Always enable fused multimodal input normalization
jzakrzew Sep 29, 2026
77b1df6
Merge branch 'main' into fix/pixtral-mm-input-norm-compile
jzakrzew Sep 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 21 additions & 0 deletions tests/model_executor/layers/test_mm_input_norm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,8 @@
import pytest
import torch

from vllm.config import CompilationConfig, VllmConfig, set_current_vllm_config
from vllm.config.compilation import CompilationMode
from vllm.model_executor.layers.fusion.mm_input_norm import (
FusedMMInputNorm,
IdentityInputNorm,
Expand Down Expand Up @@ -263,6 +265,25 @@ def test_rejects_mismatched_output_shape(self):
class TestFusedMMInputNormConstruction:
"""Weight/bias buffer semantics at init time."""

def test_compiled_encoder_enables_only_requested_input_norm(self):
config = VllmConfig(
compilation_config=CompilationConfig(
mode=CompilationMode.VLLM_COMPILE,
compile_mm_encoder=True,
custom_ops=["none"],
)
)
with set_current_vllm_config(config):
FusedMMInputNorm(_RGB_MEAN, _RGB_STD, _RGB_RESCALE)
FusedMMInputNorm(_RGB_MEAN, _RGB_STD, _RGB_RESCALE, enforce_enable=True)
assert (
config.compilation_config.enabled_custom_ops["fused_mm_input_norm"] == 1
)
assert (
config.compilation_config.disabled_custom_ops["fused_mm_input_norm"]
== 1
)

def test_identity_config_buffers(self):
"""Numerically identity parameters still build plain weight/bias
buffers; the identity shortcut lives in ``from_model_config``."""
Expand Down
9 changes: 7 additions & 2 deletions vllm/model_executor/layers/fusion/mm_input_norm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -227,8 +227,10 @@ def __init__(
image_std: list[float],
rescale_factor: float,
channel: int = 3,
*,
enforce_enable: bool = False,
):
super().__init__()
super().__init__(enforce_enable=enforce_enable)

@Isotr0py Isotr0py Sep 29, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Let's always enable fusion for this ops directly like mhc:

vllm/vllm/model_executor/layers/mhc.py

Lines 118 to 120 in 998490c

@classmethod
def enabled(cls) -> bool:
return True

@jzakrzew jzakrzew Sep 29, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ok, modified this PR to enable it unconditionally


assert len(image_mean) == len(image_std) == channel, (
f"image_mean/image_std must have {channel} entries, "
Expand Down Expand Up @@ -332,7 +334,9 @@ def forward_oot(
return self.forward_native(pixel_values, visual_dtype)


def build_mm_input_norm(model_config: ModelConfig) -> nn.Module:
def build_mm_input_norm(
model_config: ModelConfig, *, enforce_enable: bool = False
) -> nn.Module:
"""Build the input normalisation module for a model.

Returns an ``IdentityInputNorm`` when device-side normalisation is
Expand All @@ -357,4 +361,5 @@ def build_mm_input_norm(model_config: ModelConfig) -> nn.Module:
image_std=params.image_std,
rescale_factor=params.rescale_factor,
channel=len(params.image_mean),
enforce_enable=enforce_enable,
)
5 changes: 4 additions & 1 deletion vllm/model_executor/models/mistral3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -479,7 +479,10 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
self.vision_tower = init_vision_tower_for_mistral3(
config,
quant_config=quant_config,
input_norm=build_mm_input_norm(vllm_config.model_config),
input_norm=build_mm_input_norm(
vllm_config.model_config,
enforce_enable=vllm_config.compilation_config.compile_mm_encoder,
),
require_post_norm=False,
prefix=maybe_prefix(prefix, "vision_tower"),
)
Expand Down
Loading