[XPU] Enable rms_norm/act quant fusions (#43963)

Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
liuzhenwei
2026-06-02 16:14:41 +00:00
committed by GitHub
co-authored by Kunshang Ji
parent 478b49ddec
commit 3f3e2702c2
4 changed files with 31 additions and 14 deletions
@@ -70,7 +70,11 @@ class ActivationQuantPattern(VllmPatternReplacement):
self.silu_and_mul_matcher = MatcherSiluAndMul()
def empty_quant(self, *args: Any, **kwargs: Any) -> torch.Tensor:
kwargs = {"dtype": self.quant_dtype, "device": "cuda", **kwargs}
kwargs = {
"dtype": self.quant_dtype,
"device": current_platform.device_type,
**kwargs,
}
return torch.empty(*args, **kwargs)
@@ -54,19 +54,27 @@ def _rms_input_weight_dtype_match(match: pm.Match) -> bool:
def empty_bf16(*args: Any, **kwargs: Any) -> torch.Tensor:
return torch.empty(*args, **kwargs, dtype=torch.bfloat16, device="cuda")
return torch.empty(
*args, **kwargs, dtype=torch.bfloat16, device=current_platform.device_type
)
def empty_fp32(*args: Any, **kwargs: Any) -> torch.Tensor:
return torch.empty(*args, **kwargs, dtype=torch.float32, device="cuda")
return torch.empty(
*args, **kwargs, dtype=torch.float32, device=current_platform.device_type
)
def empty_i32(*args: Any, **kwargs: Any) -> torch.Tensor:
return torch.empty(*args, **kwargs, dtype=torch.int32, device="cuda")
return torch.empty(
*args, **kwargs, dtype=torch.int32, device=current_platform.device_type
)
def empty_i64(*args: Any, **kwargs: Any) -> torch.Tensor:
return torch.empty(*args, **kwargs, dtype=torch.int64, device="cuda")
return torch.empty(
*args, **kwargs, dtype=torch.int64, device=current_platform.device_type
)
RMS_ADD_OP = torch.ops._C.fused_add_rms_norm.default
+4
View File
@@ -45,6 +45,10 @@ if current_platform.is_cuda():
from .fusion.allreduce_rms_fusion import AllReduceFusionPass
from .fusion.collective_fusion import AsyncTPPass
if current_platform.is_xpu():
from .fusion.act_quant_fusion import ActivationQuantFusionPass
from .fusion.rms_quant_fusion import RMSNormQuantFusionPass
from .inductor_pass import (
CustomGraphPass,
InductorPass,