[MXFP4] Support for linear layers + compressed-tensors integration (#41664)
This commit is contained in:
@@ -58,6 +58,16 @@ from vllm.model_executor.kernels.linear.mixed_precision.xpu import (
|
||||
XPUW4A8IntLinearKernel,
|
||||
XPUwNa16LinearKernel,
|
||||
)
|
||||
from vllm.model_executor.kernels.linear.mxfp4 import (
|
||||
MxFp4LinearKernel,
|
||||
MxFp4LinearLayerConfig,
|
||||
)
|
||||
from vllm.model_executor.kernels.linear.mxfp4.flashinfer import (
|
||||
FlashInferMxFp4LinearKernel,
|
||||
)
|
||||
from vllm.model_executor.kernels.linear.mxfp4.marlin import (
|
||||
MarlinMxFp4LinearKernel,
|
||||
)
|
||||
from vllm.model_executor.kernels.linear.mxfp8 import (
|
||||
Mxfp8LinearKernel,
|
||||
Mxfp8LinearLayerConfig,
|
||||
@@ -276,6 +286,13 @@ _POSSIBLE_NVFP4_KERNELS: dict[PlatformEnum, list[type[NvFp4LinearKernel]]] = {
|
||||
],
|
||||
}
|
||||
|
||||
_POSSIBLE_MXFP4_KERNELS: dict[PlatformEnum, list[type[MxFp4LinearKernel]]] = {
|
||||
PlatformEnum.CUDA: [
|
||||
FlashInferMxFp4LinearKernel,
|
||||
MarlinMxFp4LinearKernel,
|
||||
],
|
||||
}
|
||||
|
||||
# TODO make all kernels inherit from MMLinearKernel
|
||||
# then bound _KernelT only to MMLinearKernel
|
||||
_KernelT = TypeVar("_KernelT", bound=ScaledMMLinearKernel | MMLinearKernel)
|
||||
@@ -570,6 +587,48 @@ def init_mxfp8_linear_kernel() -> Mxfp8LinearKernel:
|
||||
)
|
||||
|
||||
|
||||
def init_mxfp4_linear_kernel() -> MxFp4LinearKernel:
|
||||
"""Select and instantiate the best MXFP4 linear kernel for the
|
||||
current platform."""
|
||||
force_kernel: type[MxFp4LinearKernel] | None = None
|
||||
if envs.VLLM_MXFP4_USE_MARLIN:
|
||||
force_kernel = MarlinMxFp4LinearKernel
|
||||
|
||||
if force_kernel is not None:
|
||||
is_supported, reason = force_kernel.is_supported()
|
||||
if not is_supported:
|
||||
raise ValueError(
|
||||
f"Forced MXFP4 kernel {force_kernel.__name__} is not "
|
||||
f"supported: {reason}"
|
||||
)
|
||||
logger.info_once("Using %s for MXFP4 GEMM", force_kernel.__name__)
|
||||
return force_kernel(MxFp4LinearLayerConfig())
|
||||
|
||||
platform = current_platform._enum
|
||||
possible = _POSSIBLE_MXFP4_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
|
||||
|
||||
logger.info_once("Using %s for MXFP4 GEMM", kernel_cls.__name__)
|
||||
return kernel_cls(MxFp4LinearLayerConfig())
|
||||
|
||||
raise ValueError(
|
||||
"Failed to find a kernel that can implement the "
|
||||
"MXFP4 linear layer. Reasons: \n" + "\n".join(failure_reasons)
|
||||
)
|
||||
|
||||
|
||||
def init_wfp8_a16_linear_kernel(
|
||||
weight_quant_key: QuantKey,
|
||||
activation_quant_key: QuantKey,
|
||||
@@ -730,6 +789,10 @@ def register_linear_kernel(
|
||||
if platform not in _POSSIBLE_NVFP4_KERNELS:
|
||||
_POSSIBLE_NVFP4_KERNELS[platform] = []
|
||||
_POSSIBLE_NVFP4_KERNELS[platform].append(kernel_class)
|
||||
elif kernel_type == "mxfp4":
|
||||
if platform not in _POSSIBLE_MXFP4_KERNELS:
|
||||
_POSSIBLE_MXFP4_KERNELS[platform] = []
|
||||
_POSSIBLE_MXFP4_KERNELS[platform].append(kernel_class)
|
||||
else:
|
||||
raise ValueError(f"Unrecognized kernel type: {kernel_type}")
|
||||
|
||||
@@ -777,6 +840,11 @@ __all__ = [
|
||||
"init_mxfp8_linear_kernel",
|
||||
"Mxfp8LinearKernel",
|
||||
"Mxfp8LinearLayerConfig",
|
||||
"init_mxfp4_linear_kernel",
|
||||
"MxFp4LinearKernel",
|
||||
"MxFp4LinearLayerConfig",
|
||||
"FlashInferMxFp4LinearKernel",
|
||||
"MarlinMxFp4LinearKernel",
|
||||
"FlashInferCutlassMxfp8LinearKernel",
|
||||
"MarlinMxfp8LinearKernel",
|
||||
"XPUMxFp8LinearKernel",
|
||||
|
||||
@@ -0,0 +1,12 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from vllm.model_executor.kernels.linear.mxfp4.base import (
|
||||
MxFp4LinearKernel,
|
||||
MxFp4LinearLayerConfig,
|
||||
)
|
||||
|
||||
__all__ = [
|
||||
"MxFp4LinearKernel",
|
||||
"MxFp4LinearLayerConfig",
|
||||
]
|
||||
@@ -0,0 +1,67 @@
|
||||
# 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 MxFp4LinearLayerConfig:
|
||||
"""Configuration for an MXFP4 linear layer.
|
||||
|
||||
All MXFP4 layers share the same structure: packed uint8 weights (2 FP4 values per
|
||||
byte) and per-block weight scales (group size 32).
|
||||
"""
|
||||
|
||||
pass
|
||||
|
||||
|
||||
class MxFp4LinearKernel(ABC):
|
||||
"""Base class for MXFP4 quantized linear kernels.
|
||||
|
||||
Each subclass implements a specific GEMM backend (CUTLASS, Marlin, etc).
|
||||
The kernel selection mechanism iterates over registered subclasses in
|
||||
priority order,calling ``is_supported`` and ``can_implement`` to find the best
|
||||
match for the current hardware.
|
||||
"""
|
||||
|
||||
def __init__(self, config: MxFp4LinearLayerConfig) -> None:
|
||||
assert self.can_implement(config)[0]
|
||||
assert self.is_supported()[0]
|
||||
self.config = config
|
||||
|
||||
@classmethod
|
||||
@abstractmethod
|
||||
def is_supported(
|
||||
cls, compute_capability: int | None = None
|
||||
) -> tuple[bool, str | None]:
|
||||
"""Return whether this kernel can run on the current platform."""
|
||||
raise NotImplementedError
|
||||
|
||||
@classmethod
|
||||
@abstractmethod
|
||||
def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
|
||||
"""Return whether this kernel can handle *config*."""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
"""Transform weights into the format required by this kernel.
|
||||
|
||||
Called once after checkpoint weights have been loaded onto the
|
||||
device. Implementations should repack / swizzle / pad weights
|
||||
and scales in-place on *layer*.
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def apply_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: torch.Tensor | None = None,
|
||||
) -> torch.Tensor:
|
||||
"""Run the quantized GEMM."""
|
||||
raise NotImplementedError
|
||||
@@ -0,0 +1,74 @@
|
||||
# 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.fused_moe.experts.cutlass_moe import (
|
||||
swizzle_mxfp4_scales,
|
||||
)
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.flashinfer import has_flashinfer_cutedsl
|
||||
|
||||
from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig
|
||||
|
||||
_MXFP4_GROUP_SIZE = 32
|
||||
|
||||
|
||||
class FlashInferMxFp4LinearKernel(MxFp4LinearKernel):
|
||||
"""MXFP4 W4A4 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) and has_flashinfer_cutedsl():
|
||||
return True, None
|
||||
return False, "FlashInfer + >=sm_100 (Blackwell) required"
|
||||
|
||||
@classmethod
|
||||
def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
|
||||
return True, None
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
N, scale_K = layer.weight_scale.shape
|
||||
K = scale_K * _MXFP4_GROUP_SIZE
|
||||
|
||||
# swizzle pads N to the next multiple of 128 for CUTLASS tiling
|
||||
padded_N = ((N + 127) // 128) * 128
|
||||
layer.weight_scale = Parameter(
|
||||
swizzle_mxfp4_scales(layer.weight_scale.data, N, K).reshape(padded_N, -1),
|
||||
requires_grad=False,
|
||||
)
|
||||
|
||||
def apply_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: torch.Tensor | None = None,
|
||||
) -> torch.Tensor:
|
||||
from vllm.utils.flashinfer import (
|
||||
flashinfer_mxfp4_quantize,
|
||||
flashinfer_scaled_fp4_mm,
|
||||
)
|
||||
|
||||
weight = layer.weight
|
||||
out_shape = x.shape[:-1] + (layer.output_size_per_partition,)
|
||||
x_2d = x.reshape(-1, x.shape[-1])
|
||||
|
||||
x_fp4, x_scale = flashinfer_mxfp4_quantize(x_2d)
|
||||
out = flashinfer_scaled_fp4_mm(
|
||||
x_fp4,
|
||||
weight,
|
||||
x_scale,
|
||||
layer.weight_scale,
|
||||
alpha=None,
|
||||
out_dtype=x.dtype,
|
||||
backend="cute-dsl",
|
||||
block_size=_MXFP4_GROUP_SIZE,
|
||||
use_nvfp4=False,
|
||||
)
|
||||
|
||||
if bias is not None:
|
||||
out = out + bias
|
||||
return out.view(out_shape)
|
||||
@@ -0,0 +1,52 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import torch
|
||||
|
||||
from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig
|
||||
|
||||
|
||||
class MarlinMxFp4LinearKernel(MxFp4LinearKernel):
|
||||
@classmethod
|
||||
def is_supported(
|
||||
cls, compute_capability: int | None = None
|
||||
) -> tuple[bool, str | None]:
|
||||
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
|
||||
is_fp4_marlin_supported,
|
||||
)
|
||||
|
||||
if is_fp4_marlin_supported():
|
||||
return True, None
|
||||
return False, "Marlin FP4 not available"
|
||||
|
||||
@classmethod
|
||||
def can_implement(cls, c: MxFp4LinearLayerConfig) -> 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_fp4 import (
|
||||
prepare_fp4_layer_for_marlin,
|
||||
)
|
||||
|
||||
prepare_fp4_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_fp4 import (
|
||||
apply_fp4_marlin_linear,
|
||||
)
|
||||
|
||||
return apply_fp4_marlin_linear(
|
||||
input=x,
|
||||
weight=layer.weight,
|
||||
weight_scale=layer.weight_scale,
|
||||
weight_global_scale=None,
|
||||
workspace=layer.workspace,
|
||||
size_n=layer.output_size_per_partition,
|
||||
size_k=layer.input_size_per_partition,
|
||||
bias=bias,
|
||||
)
|
||||
Reference in New Issue
Block a user