From 5a4351ab096268944bf29a9b804800066309baa9 Mon Sep 17 00:00:00 2001 From: Jakub Zakrzewski Date: Wed, 30 Sep 2026 13:08:24 +0200 Subject: [PATCH] [MM] Keep device input normalization fused when encoder compilation is enabled (#59195) Signed-off-by: Jakub Zakrzewski Co-authored-by: Codex --- .../layers/test_mm_input_norm.py | 19 +++++++++++++++++++ .../layers/fusion/mm_input_norm.py | 4 ++++ 2 files changed, 23 insertions(+) diff --git a/tests/model_executor/layers/test_mm_input_norm.py b/tests/model_executor/layers/test_mm_input_norm.py index fc114eccd0..e3df570acc 100644 --- a/tests/model_executor/layers/test_mm_input_norm.py +++ b/tests/model_executor/layers/test_mm_input_norm.py @@ -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``.""" diff --git a/vllm/model_executor/layers/fusion/mm_input_norm.py b/vllm/model_executor/layers/fusion/mm_input_norm.py index 5127f1007e..d125f8d2d2 100644 --- a/vllm/model_executor/layers/fusion/mm_input_norm.py +++ b/vllm/model_executor/layers/fusion/mm_input_norm.py @@ -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