| 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.quant_utils import kMxfp4Dynamic |
| 8 | |
| 9 | from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig |
| 10 | |
| 11 | logger = init_logger(__name__) |
| 12 | |
| 13 | |
| 14 | class MarlinMxFp4LinearKernel(MxFp4LinearKernel): |
| 15 | @classmethod |
| 16 | def is_supported( |
| 17 | cls, compute_capability: int | None = None |
| 18 | ) -> tuple[bool, str | None]: |
| 19 | from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import ( |
| 20 | is_fp4_marlin_supported, |
| 21 | ) |
| 22 | |
| 23 | if is_fp4_marlin_supported(): |
| 24 | return True, None |
| 25 | return False, "Marlin FP4 not available" |
| 26 | |
| 27 | @classmethod |
| 28 | def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]: |
| 29 | if config.activation_quant_key not in (None, kMxfp4Dynamic): |
| 30 | return False, "only supports MXFP4 dynamic or unquantized activations" |
| 31 | if config.activation_quant_key is not None: |
| 32 | logger.warning_once( |
| 33 | "MarlinMxFp4LinearKernel is a weight-only (A16) kernel; " |
| 34 | "the requested activation quantization (%s) is ignored.", |
| 35 | config.activation_quant_key, |
| 36 | ) |
| 37 | return True, None |
| 38 | |
| 39 | def process_weights_after_loading(self, layer: torch.nn.Module) -> None: |
| 40 | from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import ( |
| 41 | prepare_fp4_layer_for_marlin, |
| 42 | ) |
| 43 | |
| 44 | prepare_fp4_layer_for_marlin(layer) |
| 45 | |
| 46 | def apply_weights( |
| 47 | self, |
| 48 | layer: torch.nn.Module, |
| 49 | x: torch.Tensor, |
| 50 | bias: torch.Tensor | None = None, |
| 51 | ) -> torch.Tensor: |
| 52 | from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import ( |
| 53 | apply_fp4_marlin_linear, |
| 54 | ) |
| 55 | |
| 56 | return apply_fp4_marlin_linear( |
| 57 | input=x, |
| 58 | weight=layer.weight, |
| 59 | weight_scale=layer.weight_scale, |
| 60 | weight_global_scale=None, |
| 61 | workspace=None, |
| 62 | size_n=layer.output_size_per_partition, |
| 63 | size_k=layer.input_size_per_partition, |
| 64 | bias=bias, |
| 65 | ) |