[Bugfix][Compile] Guard per_token_group_fp8_quant lookup on non-CUDA platforms (#44476)

This commit is contained in:
QiliangCui2023
2026-06-04 09:31:50 -04:00
committed by GitHub
parent f35b557239
commit 9354fb1ba5
2 changed files with 7 additions and 4 deletions
@@ -36,10 +36,12 @@ QUANT_OPS: dict[QuantKey, OpOverload] = {
kFp8StaticTensorSym: torch.ops._C.static_scaled_fp8_quant.default, # noqa: E501
kFp8DynamicTensorSym: torch.ops._C.dynamic_scaled_fp8_quant.default, # noqa: E501
kFp8DynamicTokenSym: torch.ops._C.dynamic_per_token_scaled_fp8_quant.default, # noqa: E501
kFp8Dynamic128Sym: torch.ops._C.per_token_group_fp8_quant.default, # noqa: E501
kFp8Dynamic64Sym: torch.ops._C.per_token_group_fp8_quant.default, # noqa: E501
}
if hasattr(torch.ops._C, "per_token_group_fp8_quant"):
QUANT_OPS[kFp8Dynamic128Sym] = torch.ops._C.per_token_group_fp8_quant.default # noqa: E501
QUANT_OPS[kFp8Dynamic64Sym] = torch.ops._C.per_token_group_fp8_quant.default # noqa: E501
if current_platform.is_cuda() and hasattr(torch.ops._C, "scaled_fp4_quant"):
QUANT_OPS[kNvfp4Dynamic] = torch.ops._C.scaled_fp4_quant.out # noqa: E501
@@ -84,9 +84,10 @@ QUANT_OPS: dict[QuantKey, OpOverload] = {
kFp8StaticTensorSym: torch.ops._C.static_scaled_fp8_quant.default, # noqa: E501
kFp8DynamicTensorSym: torch.ops._C.dynamic_scaled_fp8_quant.default, # noqa: E501
kFp8DynamicTokenSym: torch.ops._C.dynamic_per_token_scaled_fp8_quant.default, # noqa: E501
kFp8Dynamic128Sym: torch.ops._C.per_token_group_fp8_quant.default, # noqa: E501
kFp8Dynamic64Sym: torch.ops._C.per_token_group_fp8_quant.default, # noqa: E501
}
if hasattr(torch.ops._C, "per_token_group_fp8_quant"):
QUANT_OPS[kFp8Dynamic128Sym] = torch.ops._C.per_token_group_fp8_quant.default # noqa: E501
QUANT_OPS[kFp8Dynamic64Sym] = torch.ops._C.per_token_group_fp8_quant.default # noqa: E501
if current_platform.is_cuda() and hasattr(torch.ops._C, "scaled_fp4_quant"):
QUANT_OPS[kNvfp4Dynamic] = torch.ops._C.scaled_fp4_quant.out