From 22720624714d3299d2df87e0c2c67639202c4b8a Mon Sep 17 00:00:00 2001 From: Luciano Martins <22145370+lucianommartins@users.noreply.github.com> Date: Wed, 27 May 2026 07:36:27 -0300 Subject: [PATCH] [Kernel] Enable TritonW4A16LinearKernel as CUDA fallback for non-Marlin-aligned W4A16 shapes (#43731) Signed-off-by: Luciano Martins Co-authored-by: Luciano Martins --- vllm/model_executor/kernels/linear/__init__.py | 1 + .../kernels/linear/mixed_precision/triton_w4a16.py | 4 ++-- 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/vllm/model_executor/kernels/linear/__init__.py b/vllm/model_executor/kernels/linear/__init__.py index b99fcc7d34..cfeb40fecf 100644 --- a/vllm/model_executor/kernels/linear/__init__.py +++ b/vllm/model_executor/kernels/linear/__init__.py @@ -336,6 +336,7 @@ _POSSIBLE_KERNELS: dict[PlatformEnum, list[type[MPLinearKernel]]] = { MarlinLinearKernel, ConchLinearKernel, ExllamaLinearKernel, + TritonW4A16LinearKernel, ], PlatformEnum.ROCM: [ TritonW4A16LinearKernel, diff --git a/vllm/model_executor/kernels/linear/mixed_precision/triton_w4a16.py b/vllm/model_executor/kernels/linear/mixed_precision/triton_w4a16.py index 5cc100b3e1..5ba1caabae 100644 --- a/vllm/model_executor/kernels/linear/mixed_precision/triton_w4a16.py +++ b/vllm/model_executor/kernels/linear/mixed_precision/triton_w4a16.py @@ -288,8 +288,8 @@ class TritonW4A16LinearKernel(MPLinearKernel): @classmethod def can_implement(cls, c: MPLinearLayerConfig) -> tuple[bool, str | None]: - if not current_platform.is_rocm(): - return False, "TritonW4A16LinearKernel only targets ROCm" + if not (current_platform.is_rocm() or current_platform.is_cuda()): + return False, "TritonW4A16LinearKernel requires CUDA or ROCm" if c.weight_type not in cls.SUPPORTED_QUANT_TYPES: return (