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

Signed-off-by: Jakub Zakrzewski <jzakrzewski@nvidia.com>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Jakub Zakrzewski
2026-09-30 11:08:24 +00:00
committed by GitHub
co-authored by Codex
parent 5225f1bacb
commit 5a4351ab09
2 changed files with 23 additions and 0 deletions
@@ -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,
@@ -263,6 +265,23 @@ class TestFusedMMInputNormKernel:
class TestFusedMMInputNormConstruction:
"""Weight/bias buffer semantics at init time."""
def test_input_norm_is_always_enabled(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)
assert (
config.compilation_config.enabled_custom_ops["fused_mm_input_norm"] == 1
)
assert not config.compilation_config.disabled_custom_ops[
"fused_mm_input_norm"
]
def test_identity_config_buffers(self):
"""Numerically identity parameters still build plain weight/bias
buffers; the identity shortcut lives in ``from_model_config``."""
@@ -246,6 +246,10 @@ class FusedMMInputNorm(CustomOp):
self.register_buffer("weight", (rescale_factor / std).to(device))
self.register_buffer("bias", (-mean / std).to(device))
@classmethod
def enabled(cls) -> bool:
return True
@property
def input_dtype(self) -> torch.dtype | None:
return torch.uint8