[Quantization] Honor --linear-backend for ModelOpt W4A16 (#50273)

Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Netanel Haber
2026-07-30 13:19:56 -07:00
committed by GitHub
co-authored by OpenAI Codex
parent c27b080d26
commit 70bd10930b
3 changed files with 42 additions and 21 deletions
@@ -891,6 +891,7 @@ def init_nvfp4_linear_kernel(use_a16: bool = False) -> NvFp4LinearKernel:
"""Select and instantiate the best NVFP4 linear kernel for the
current platform."""
config = NvFp4LinearLayerConfig()
a16_kernels = (MarlinNvFp4LinearKernel, HummingNvFp4LinearKernel)
# VLLM_BATCH_INVARIANT forces deterministic execution. Prefer the
# batch-invariant CUTLASS implementation when available, otherwise fall
@@ -931,6 +932,8 @@ def init_nvfp4_linear_kernel(use_a16: bool = False) -> NvFp4LinearKernel:
force_kernel = MarlinNvFp4LinearKernel
if force_kernel is not None:
if use_a16 and force_kernel not in a16_kernels:
raise ValueError(f"{force_kernel.__name__} does not support W4A16")
is_supported, reason = force_kernel.is_supported()
if not is_supported:
raise ValueError(
@@ -943,6 +946,8 @@ def init_nvfp4_linear_kernel(use_a16: bool = False) -> NvFp4LinearKernel:
# Auto-select from registry (or --linear-backend filtered).
platform = current_platform._enum
possible = list(_POSSIBLE_NVFP4_KERNELS.get(platform, []))
if use_a16:
possible = [kernel for kernel in possible if kernel in a16_kernels]
# Apply --linear-backend filtering when set.
if linear_backend != "auto":