Signed-off-by: samuelkim7 <samuelmwkim@gmail.com> Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: mgoin <mgoin64@gmail.com>
68 lines
2.1 KiB
Python
68 lines
2.1 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import torch
|
|
|
|
from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
|
|
MXFP8_BLOCK_SIZE,
|
|
)
|
|
|
|
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]:
|
|
N, K = c.weight_shape
|
|
if N <= 0 or K <= 0:
|
|
return False, f"Marlin MXFP8 requires positive N and K, got N={N}, K={K}."
|
|
if K % MXFP8_BLOCK_SIZE != 0:
|
|
return (
|
|
False,
|
|
f"Marlin MXFP8 requires K to be divisible by {MXFP8_BLOCK_SIZE}, "
|
|
f"got K={K}.",
|
|
)
|
|
# Tile misalignment is fixed by zero-padding at weight prep.
|
|
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=None,
|
|
size_n=layer.output_size_per_partition,
|
|
size_k=layer.input_size_per_partition,
|
|
bias=bias,
|
|
)
|