[Refactor] Move MXFP8 GEMM management into MxFp8LinearKernel (#39205)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2026-04-10 14:02:03 -07:00
committed by GitHub
parent fc645f1acc
commit 11e2375fe2
10 changed files with 365 additions and 258 deletions
@@ -58,6 +58,19 @@ from vllm.model_executor.kernels.linear.mixed_precision.xpu import (
XPUW4A8IntLinearKernel,
XPUwNa16LinearKernel,
)
from vllm.model_executor.kernels.linear.mxfp8 import (
Mxfp8LinearKernel,
Mxfp8LinearLayerConfig,
)
from vllm.model_executor.kernels.linear.mxfp8.emulation import (
EmulationMxfp8LinearKernel,
)
from vllm.model_executor.kernels.linear.mxfp8.flashinfer import (
FlashInferCutlassMxfp8LinearKernel,
)
from vllm.model_executor.kernels.linear.mxfp8.marlin import (
MarlinMxfp8LinearKernel,
)
from vllm.model_executor.kernels.linear.nvfp4 import (
NvFp4LinearKernel,
NvFp4LinearLayerConfig,
@@ -221,6 +234,17 @@ _POSSIBLE_KERNELS: dict[PlatformEnum, list[type[MPLinearKernel]]] = {
}
# in priority/performance order (when available)
_POSSIBLE_MXFP8_KERNELS: dict[PlatformEnum, list[type[Mxfp8LinearKernel]]] = {
PlatformEnum.CUDA: [
FlashInferCutlassMxfp8LinearKernel,
MarlinMxfp8LinearKernel,
EmulationMxfp8LinearKernel,
],
PlatformEnum.ROCM: [
EmulationMxfp8LinearKernel,
],
}
_POSSIBLE_NVFP4_KERNELS: dict[PlatformEnum, list[type[NvFp4LinearKernel]]] = {
PlatformEnum.CUDA: [
FlashInferCutlassNvFp4LinearKernel,
@@ -482,6 +506,41 @@ def choose_mp_linear_kernel(
)
def init_mxfp8_linear_kernel() -> Mxfp8LinearKernel:
"""Select and instantiate the best MXFP8 linear kernel for the
current platform."""
config = Mxfp8LinearLayerConfig()
platform = current_platform._enum
possible = _POSSIBLE_MXFP8_KERNELS.get(platform, [])
failure_reasons = []
for kernel_cls in possible:
if kernel_cls.__name__ in envs.VLLM_DISABLED_KERNELS:
failure_reasons.append(
f" {kernel_cls.__name__} disabled by environment variable"
)
continue
is_supported, reason = kernel_cls.is_supported()
if not is_supported:
failure_reasons.append(f"{kernel_cls.__name__}: {reason}")
continue
can_implement, reason = kernel_cls.can_implement(config)
if not can_implement:
failure_reasons.append(f"{kernel_cls.__name__}: {reason}")
continue
logger.info_once("Using %s for MXFP8 GEMM", kernel_cls.__name__)
return kernel_cls(config)
raise ValueError(
"Failed to find a kernel that can implement the "
"MXFP8 linear layer. Reasons: \n" + "\n".join(failure_reasons)
)
def init_wfp8_a16_linear_kernel(
weight_quant_key: QuantKey,
activation_quant_key: QuantKey,
@@ -628,6 +687,10 @@ def register_linear_kernel(
if platform not in _POSSIBLE_FP8_KERNELS:
_POSSIBLE_FP8_KERNELS[platform] = []
_POSSIBLE_FP8_KERNELS[platform].append(kernel_class)
elif kernel_type == "mxfp8":
if platform not in _POSSIBLE_MXFP8_KERNELS:
_POSSIBLE_MXFP8_KERNELS[platform] = []
_POSSIBLE_MXFP8_KERNELS[platform].append(kernel_class)
elif kernel_type == "nvfp4":
if platform not in _POSSIBLE_NVFP4_KERNELS:
_POSSIBLE_NVFP4_KERNELS[platform] = []
@@ -674,6 +737,12 @@ __all__ = [
"TritonW4A16LinearKernel",
"XPUW4A8IntLinearKernel",
"XPUwNa16LinearKernel",
"init_mxfp8_linear_kernel",
"Mxfp8LinearKernel",
"Mxfp8LinearLayerConfig",
"FlashInferCutlassMxfp8LinearKernel",
"MarlinMxfp8LinearKernel",
"EmulationMxfp8LinearKernel",
"CutlassNvFp4LinearKernel",
"EmulationNvFp4LinearKernel",
"FbgemmNvFp4LinearKernel",
@@ -0,0 +1,56 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from abc import ABC, abstractmethod
from dataclasses import dataclass
import torch
@dataclass
class Mxfp8LinearLayerConfig:
"""Configuration for an MXFP8 linear layer.
All MXFP8 layers share the same structure: FP8-E4M3 weights with
uint8 (E8M0) per-block scales at block size 32.
"""
pass
class Mxfp8LinearKernel(ABC):
"""Base class for MXFP8 quantized linear kernels.
Each subclass implements a specific GEMM backend (FlashInfer CUTLASS,
Marlin, emulation).
"""
def __init__(self, c: Mxfp8LinearLayerConfig) -> None:
assert self.can_implement(c)[0]
assert self.is_supported()[0]
self.config = c
@classmethod
@abstractmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
raise NotImplementedError
@classmethod
@abstractmethod
def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
raise NotImplementedError
@abstractmethod
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
raise NotImplementedError
@abstractmethod
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
raise NotImplementedError
@@ -0,0 +1,12 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.model_executor.kernels.linear.mxfp8.Mxfp8LinearKernel import (
Mxfp8LinearKernel,
Mxfp8LinearLayerConfig,
)
__all__ = [
"Mxfp8LinearKernel",
"Mxfp8LinearLayerConfig",
]
@@ -0,0 +1,60 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from torch.nn.parameter import Parameter
from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
MXFP8_BLOCK_SIZE,
MXFP8_SCALE_DTYPE,
dequant_mxfp8_to_bf16,
)
from .Mxfp8LinearKernel import Mxfp8LinearKernel, Mxfp8LinearLayerConfig
class EmulationMxfp8LinearKernel(Mxfp8LinearKernel):
"""Software emulation fallback for MXFP8 (dequant to BF16)."""
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
return True, None
@classmethod
def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
weight = layer.weight.data # [N, K]
N, K = weight.shape
scale_k = K // MXFP8_BLOCK_SIZE
weight_scale = layer.weight_scale.data[:N, :scale_k].contiguous()
layer.weight = Parameter(weight.contiguous(), requires_grad=False)
layer.weight_scale = Parameter(weight_scale, requires_grad=False)
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
weight_scale = layer.weight_scale
if weight_scale.dtype != MXFP8_SCALE_DTYPE:
raise ValueError(
f"Emulation backend requires {MXFP8_SCALE_DTYPE} "
f"weight_scale dtype, got {weight_scale.dtype}."
)
if weight_scale.ndim != 2:
raise ValueError(
f"Emulation backend requires 2D weight_scale, "
f"got {weight_scale.ndim}D. "
f"Ensure process_weights_after_loading was called."
)
weight_bf16 = dequant_mxfp8_to_bf16(layer.weight, weight_scale)
output = torch.nn.functional.linear(x, weight_bf16, bias)
return output.to(x.dtype)
@@ -0,0 +1,103 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from torch.nn.parameter import Parameter
from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
MXFP8_BLOCK_SIZE,
mxfp8_e4m3_quantize,
swizzle_mxfp8_scale,
)
from vllm.platforms import current_platform
from vllm.utils import flashinfer as vllm_flashinfer
from .Mxfp8LinearKernel import Mxfp8LinearKernel, Mxfp8LinearLayerConfig
class FlashInferCutlassMxfp8LinearKernel(Mxfp8LinearKernel):
"""MXFP8 W8A8 GEMM via FlashInfer CUTLASS (SM100+)."""
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
if current_platform.has_device_capability(100):
return True, None
return False, "requires >=sm_100 (Blackwell)"
@classmethod
def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
weight = layer.weight.data # [N, K]
N, K = weight.shape
scale_k = K // MXFP8_BLOCK_SIZE
weight_scale_2d = layer.weight_scale.data[:N, :scale_k].contiguous()
weight_scale_swizzled = swizzle_mxfp8_scale(weight_scale_2d, M=N, K=K)
layer.weight = Parameter(weight.contiguous(), requires_grad=False)
layer.weight_scale = Parameter(
weight_scale_swizzled.contiguous(), requires_grad=False
)
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
weight = layer.weight
weight_scale = layer.weight_scale
out_dtype = x.dtype
N, K = weight.shape
input_shape = x.shape
input_2d = x.view(-1, K)
M_orig = input_2d.shape[0]
min_dim = 128
assert min_dim <= K, (
f"mm_mxfp8 requires K >= {min_dim}, got K={K}. "
f"in_features is too small for mm_mxfp8."
)
assert K % MXFP8_BLOCK_SIZE == 0, (
f"mm_mxfp8 requires K to be divisible by {MXFP8_BLOCK_SIZE}, got K={K}."
)
assert min_dim <= N, (
f"mm_mxfp8 requires N >= {min_dim}, got N={N}. "
f"out_features is too small for mm_mxfp8."
)
M_padded = ((M_orig + min_dim - 1) // min_dim) * min_dim
if M_padded != M_orig:
pad_rows = M_padded - M_orig
input_2d = torch.nn.functional.pad(input_2d, (0, 0, 0, pad_rows))
input_mxfp8, input_scale = mxfp8_e4m3_quantize(
input_2d, is_sf_swizzled_layout=True
)
if not weight.is_contiguous():
weight = weight.contiguous()
output = vllm_flashinfer.mm_mxfp8(
input_mxfp8,
weight.t(),
input_scale,
weight_scale,
out_dtype=out_dtype,
backend="cutlass",
)
if M_padded != M_orig:
output = output[:M_orig, :]
if bias is not None:
output = output + bias
output_shape = (*input_shape[:-1], N)
return output.view(output_shape)
@@ -0,0 +1,53 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from .Mxfp8LinearKernel import Mxfp8LinearKernel, Mxfp8LinearLayerConfig
class MarlinMxfp8LinearKernel(Mxfp8LinearKernel):
"""MXFP8 W8A16 GEMM via Marlin (SM80+)."""
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
is_fp8_marlin_supported,
)
if is_fp8_marlin_supported():
return True, None
return False, "Marlin FP8 not available"
@classmethod
def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
prepare_mxfp8_layer_for_marlin,
)
prepare_mxfp8_layer_for_marlin(layer)
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
apply_mxfp8_marlin_linear,
)
return apply_mxfp8_marlin_linear(
input=x,
weight=layer.weight,
weight_scale=layer.weight_scale,
workspace=layer.workspace,
size_n=layer.output_size_per_partition,
size_k=layer.input_size_per_partition,
bias=bias,
)