[Bugfix] Guard CUDA-only rms_norm_per_block_quant in FUSED_OPS for non-CUDA builds (#47296)

Signed-off-by: Tsvika Shapira <tsvika@moonmath.ai>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Tsvika Shapira
2026-07-09 10:10:53 -04:00
committed by GitHub
co-authored by Claude mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 753c5039f0
commit 429f405748
@@ -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: