main
py 57 lines 1.82 KB
Raw
1 # SPDX-License-Identifier: Apache-2.0
2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4 import torch
5
6 from vllm.logger import init_logger
7 from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
8 apply_fp4_marlin_linear,
9 is_fp4_marlin_supported,
10 prepare_fp4_layer_for_marlin,
11 )
12
13 from .base import NvFp4LinearKernel, NvFp4LinearLayerConfig
14
15 logger = init_logger(__name__)
16
17
18 class MarlinNvFp4LinearKernel(NvFp4LinearKernel):
19 """NVFP4 weight-only GEMM via Marlin (W4A16)."""
20
21 @classmethod
22 def is_supported(
23 cls, compute_capability: int | None = None
24 ) -> tuple[bool, str | None]:
25 if is_fp4_marlin_supported():
26 return True, None
27 return False, "Marlin FP4 not available"
28
29 @classmethod
30 def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
31 return True, None
32
33 def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
34 logger.warning_once(
35 "Your GPU does not have native support for FP4 computation but "
36 "FP4 quantization is being used. Weight-only FP4 compression "
37 "will be used leveraging the Marlin kernel. This may degrade "
38 "performance for compute-heavy workloads."
39 )
40 prepare_fp4_layer_for_marlin(layer)
41
42 def apply_weights(
43 self,
44 layer: torch.nn.Module,
45 x: torch.Tensor,
46 bias: torch.Tensor | None = None,
47 ) -> torch.Tensor:
48 return apply_fp4_marlin_linear(
49 input=x,
50 weight=layer.weight,
51 weight_scale=layer.weight_scale,
52 weight_global_scale=layer.weight_global_scale,
53 workspace=None,
54 size_n=layer.output_size_per_partition,
55 size_k=layer.input_size_per_partition,
56 bias=bias,
57 )