From 429f4057487076d569f4a80413738462513dbe5a Mon Sep 17 00:00:00 2001 From: Tsvika Shapira Date: Thu, 9 Jul 2026 17:10:53 +0300 Subject: [PATCH] [Bugfix] Guard CUDA-only rms_norm_per_block_quant in FUSED_OPS for non-CUDA builds (#47296) Signed-off-by: Tsvika Shapira Co-authored-by: Claude Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- .../passes/fusion/rms_quant_fusion.py | 19 +++++++------------ 1 file changed, 7 insertions(+), 12 deletions(-) diff --git a/vllm/compilation/passes/fusion/rms_quant_fusion.py b/vllm/compilation/passes/fusion/rms_quant_fusion.py index 670349a08b..6c55261e87 100644 --- a/vllm/compilation/passes/fusion/rms_quant_fusion.py +++ b/vllm/compilation/passes/fusion/rms_quant_fusion.py @@ -128,19 +128,14 @@ FUSED_OPS: dict[FusedRMSQuantKey, OpOverload] = { FusedRMSQuantKey( kFp8DynamicTokenSym, True ): torch.ops._C.rms_norm_dynamic_per_token_quant.default, # noqa: E501 - FusedRMSQuantKey( - kFp8Dynamic128Sym, False - ): torch.ops._C.rms_norm_per_block_quant.default, # noqa: E501 - FusedRMSQuantKey( - kFp8Dynamic128Sym, True - ): torch.ops._C.rms_norm_per_block_quant.default, # noqa: E501 - FusedRMSQuantKey( - kFp8Dynamic64Sym, False - ): torch.ops._C.rms_norm_per_block_quant.default, # noqa: E501 - FusedRMSQuantKey( - kFp8Dynamic64Sym, True - ): torch.ops._C.rms_norm_per_block_quant.default, # noqa: E501 } +# rms_norm_per_block_quant is CUDA-only; guard it like per_token_group_fp8_quant above. +if hasattr(torch.ops._C, "rms_norm_per_block_quant"): + _rms_norm_per_block_quant = torch.ops._C.rms_norm_per_block_quant.default + FUSED_OPS[FusedRMSQuantKey(kFp8Dynamic128Sym, False)] = _rms_norm_per_block_quant # noqa: E501 + FUSED_OPS[FusedRMSQuantKey(kFp8Dynamic128Sym, True)] = _rms_norm_per_block_quant # noqa: E501 + FUSED_OPS[FusedRMSQuantKey(kFp8Dynamic64Sym, False)] = _rms_norm_per_block_quant # noqa: E501 + FUSED_OPS[FusedRMSQuantKey(kFp8Dynamic64Sym, True)] = _rms_norm_per_block_quant # noqa: E501 class RMSNormQuantPattern: