[Feature] Support batch invariance on ROCm (#52231)

Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Matt
2026-10-12 10:53:26 -05:00
committed by GitHub
co-authored by Claude Opus 5.5 Andreas Karatzas
parent 50b21a2667
commit e0317b28ce
31 changed files with 1106 additions and 122 deletions
@@ -39,6 +39,7 @@ from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import (
ScaledMMLinearKernel,
)
from vllm.model_executor.kernels.linear.scaled_mm.triton import (
TritonFP8ScaledMMLinearKernel,
TritonInt8ScaledMMLinearKernel,
)
from vllm.model_executor.kernels.linear.scaled_mm.xpu import (
@@ -65,6 +66,7 @@ __all__ = [
"PerTensorTorchFP8ScaledMMLinearKernel",
"RowWiseTorchFP8ScaledMMLinearKernel",
"ROCmFP8ScaledMMLinearKernel",
"TritonFP8ScaledMMLinearKernel",
"TritonInt8ScaledMMLinearKernel",
"ZentorchInt8ScaledMMLinearKernel",
"Fp8BlockScaledMMLinearKernel",
@@ -20,6 +20,8 @@ from .BlockScaledMMLinearKernel import (
)
from .cutlass import CutlassInt8ScaledMMLinearKernel
from .ScaledMMLinearKernel import (
FP8ScaledMMLinearKernel,
FP8ScaledMMLinearLayerConfig,
Int8ScaledMMLinearLayerConfig,
)
@@ -156,6 +158,46 @@ class TritonInt8ScaledMMLinearKernel(CutlassInt8ScaledMMLinearKernel):
return out
class TritonFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
if not current_platform.is_cuda_alike():
return False, "requires ROCm or CUDA."
return True, None
@classmethod
def can_implement(cls, c: FP8ScaledMMLinearLayerConfig) -> tuple[bool, str | None]:
act_group = c.activation_quant_key.scale.group_shape
weight_group = c.weight_quant_key.scale.group_shape
if not (act_group.is_per_tensor() or act_group.is_per_token()):
return False, "requires per-tensor or per-token activation scales."
if not (
weight_group.is_per_tensor()
or weight_group.is_per_channel()
or weight_group.is_per_token()
):
return False, "requires per-tensor, per-channel or per-token weight scales."
return True, None
def apply_scaled_mm(
self,
*,
A: torch.Tensor,
B: torch.Tensor,
out_dtype: torch.dtype,
As: torch.Tensor,
Bs: torch.Tensor,
bias: torch.Tensor | None,
output_shape: list,
) -> torch.Tensor:
out = triton_scaled_mm(
A, B, scale_a=As, scale_b=Bs, out_dtype=out_dtype, bias=bias
)
return out.view(*output_shape)
class TritonFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel):
@classmethod
def is_supported(cls, compute_capability=None):