[ROCm][FEAT] Integrate aiter gemm w8a8 ptpc (#33773)

Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
This commit is contained in:
vllmellm
2026-04-16 09:55:29 +08:00
committed by GitHub
parent 343f65234b
commit 5f7fab881a
9 changed files with 279 additions and 23 deletions
@@ -106,6 +106,8 @@ from vllm.model_executor.kernels.linear.scaled_mm import (
from vllm.model_executor.kernels.linear.scaled_mm.aiter import (
AiterFp8BlockScaledMMKernel,
AiterInt8ScaledMMLinearKernel,
AiterPerTokenFp8ScaledMMLinearKernel,
AiterPreshuffledPerTokenFp8ScaledMMLinearKernel,
)
from vllm.model_executor.kernels.linear.scaled_mm.cpu import (
CPUInt8ScaledMMLinearKernel,
@@ -165,6 +167,8 @@ _POSSIBLE_FP8_KERNELS: dict[PlatformEnum, list[type[FP8ScaledMMLinearKernel]]] =
ChannelWiseTorchFP8ScaledMMLinearKernel,
],
PlatformEnum.ROCM: [
AiterPreshuffledPerTokenFp8ScaledMMLinearKernel,
AiterPerTokenFp8ScaledMMLinearKernel,
ROCmFP8ScaledMMLinearKernel,
PerTensorTorchFP8ScaledMMLinearKernel,
RowWiseTorchFP8ScaledMMLinearKernel,
@@ -360,18 +364,18 @@ def choose_scaled_mm_linear_kernel(
def init_fp8_linear_kernel(
activation_quant_key: QuantKey,
weight_quant_key: QuantKey,
weight_shape: tuple[int, int],
input_dtype: torch.dtype,
out_dtype: torch.dtype,
force_kernel: type[_KernelT] | None = None,
weight_shape: tuple[int, int],
force_kernel: type[FP8ScaledMMLinearKernel] | None = None,
module_name: str | None = None,
) -> FP8ScaledMMLinearKernel | Fp8BlockScaledMMLinearKernel:
scaled_mm_linear_kernel_config = FP8ScaledMMLinearLayerConfig(
weight_quant_key=weight_quant_key,
activation_quant_key=activation_quant_key,
weight_shape=weight_shape,
input_dtype=input_dtype,
out_dtype=out_dtype,
weight_shape=weight_shape,
)
if activation_quant_key.scale.group_shape.is_per_group():
@@ -725,6 +729,8 @@ __all__ = [
"FP8ScaledMMLinearLayerConfig",
"Int8ScaledMMLinearLayerConfig",
"ScaledMMLinearLayerConfig",
"AiterPreshuffledPerTokenFp8ScaledMMLinearKernel",
"AiterPerTokenFp8ScaledMMLinearKernel",
"NvFp4LinearKernel",
"NvFp4LinearLayerConfig",
"AiterInt8ScaledMMLinearKernel",
@@ -5,18 +5,27 @@
import torch
from vllm import _custom_ops as ops
from vllm._aiter_ops import rocm_aiter_ops
from vllm._aiter_ops import (
rocm_aiter_ops,
)
from vllm.logger import init_logger
from vllm.model_executor.layers.quantization.utils.quant_utils import (
GroupShape,
)
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from .BlockScaledMMLinearKernel import (
Fp8BlockScaledMMLinearKernel,
FP8ScaledMMLinearLayerConfig,
)
from .cutlass import CutlassInt8ScaledMMLinearKernel
from .ScaledMMLinearKernel import Int8ScaledMMLinearLayerConfig
from .ScaledMMLinearKernel import (
FP8ScaledMMLinearKernel,
FP8ScaledMMLinearLayerConfig,
Int8ScaledMMLinearLayerConfig,
)
logger = init_logger(__name__)
class AiterInt8ScaledMMLinearKernel(CutlassInt8ScaledMMLinearKernel):
@@ -113,7 +122,154 @@ class AiterInt8ScaledMMLinearKernel(CutlassInt8ScaledMMLinearKernel):
# a to be [M, K]
# b to be [N, K]
# CutlassInt8ScaledMMLinearKernel prepare weight `w_q` in [K, N] format
return rocm_aiter_ops.gemm_a8w8(x_q, w_q.t(), x_s, w_s, bias, out_dtype)
return rocm_aiter_ops.w8a8_gemm(x_q, w_q.t(), x_s, w_s, bias, out_dtype)
class AiterPreshuffledPerTokenFp8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
if not current_platform.is_rocm():
return False, "requires ROCm."
if not rocm_aiter_ops.is_linear_fp8_enabled():
return (
False,
"requires setting `VLLM_ROCM_USE_AITER=1` "
"and `VLLM_ROCM_USE_AITER_LINEAR=1`. "
"`VLLM_ROCM_USE_AITER_LINEAR` default is True.",
)
try:
import aiter # noqa: F401
except Exception:
return False, "requires aiter library to be installed."
return True, None
@classmethod
def can_implement(cls, c: FP8ScaledMMLinearLayerConfig) -> tuple[bool, str | None]:
is_ptpc = (
c.activation_quant_key.scale.group_shape.is_per_token()
and c.weight_quant_key.scale.group_shape.is_per_channel()
)
if c.weight_shape is None:
return False, "weight_shape is required for Aiter kernels"
N, K = c.weight_shape
fp8_dtype = current_platform.fp8_dtype()
if c.out_dtype is not torch.bfloat16:
return False, "requires bfloat16 output dtype."
if not is_ptpc:
return (
False,
"requires per token activation scales and per channel weight scales.",
)
if not (N % 16 == 0 and K % 16 == 0):
return (
False,
f"requires N and K dimensions divisible by 16, received "
f"N={N} and K={K}.",
)
# Aiter's shuffled per-token Gemm performs better than torch only when its
# tuned.
if not rocm_aiter_ops.is_shuffled_per_token_w8a8_gemm_tuned(N, K, fp8_dtype):
return (
False,
f"requires a tuned configuration for N: {N} and K: {K} "
f"and fp8 dtype {fp8_dtype}.",
)
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
w_name, *_ = self.layer_param_names
w, *_ = self._get_layer_params(layer)
replace_parameter(
layer,
w_name,
torch.nn.Parameter(
rocm_aiter_ops.shuffle_weight(w.t().contiguous()).data,
requires_grad=False,
),
)
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:
return rocm_aiter_ops.preshuffled_per_token_w8a8_gemm(
A, B, As, Bs, bias, out_dtype
)
class AiterPerTokenFp8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
return AiterPreshuffledPerTokenFp8ScaledMMLinearKernel.is_supported(
compute_capability
)
@classmethod
def can_implement(cls, c: FP8ScaledMMLinearLayerConfig) -> tuple[bool, str | None]:
is_ptpc = (
c.activation_quant_key.scale.group_shape.is_per_token()
and c.weight_quant_key.scale.group_shape.is_per_channel()
)
if c.weight_shape is None:
return False, "weight_shape is required for Aiter kernels"
N, K = c.weight_shape
fp8_dtype = current_platform.fp8_dtype()
if not is_ptpc:
return (
False,
"requires per token activation scales and per channel weight scales.",
)
# Aiter's per-token Gemm performs better than torch oonly when its
# tuned.
if not rocm_aiter_ops.is_per_token_w8a8_gemm_tuned(N, K, fp8_dtype):
return (
False,
f"requires a tuned configuration for N: {N} and K: {K} "
f"and fp8 dtype {fp8_dtype}.",
)
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
w_name, *_ = self.layer_param_names
w, *_ = self._get_layer_params(layer)
replace_parameter(
layer,
w_name,
torch.nn.Parameter(w.t(), requires_grad=False),
)
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:
return rocm_aiter_ops.w8a8_gemm(A, B, As, Bs, bias, out_dtype)
class AiterFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel):