| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | |
| 6 | from vllm.model_executor.layers.quantization.utils.mxfp4_utils import ( |
| 7 | xpu_mxfp4_quantize as quant_mxfp4, |
| 8 | ) |
| 9 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 10 | kMxfp4Dynamic, |
| 11 | ) |
| 12 | from vllm.model_executor.utils import replace_parameter |
| 13 | from vllm.platforms import current_platform |
| 14 | |
| 15 | from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig |
| 16 | |
| 17 | |
| 18 | class XPUMxFp4LinearKernel(MxFp4LinearKernel): |
| 19 | """MXFP4 W4A4 GEMM on XPU.""" |
| 20 | |
| 21 | @classmethod |
| 22 | def is_supported( |
| 23 | cls, compute_capability: int | None = None |
| 24 | ) -> tuple[bool, str | None]: |
| 25 | if not current_platform.is_xpu(): |
| 26 | return False, "XPUMxFp4 only support on XPU" |
| 27 | return True, None |
| 28 | |
| 29 | @classmethod |
| 30 | def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]: |
| 31 | if config.activation_quant_key != kMxfp4Dynamic: |
| 32 | return False, "only supports MXFP4 dynamic activation" |
| 33 | return True, None |
| 34 | |
| 35 | def process_weights_after_loading(self, layer: torch.nn.Module) -> None: |
| 36 | weight = layer.weight.view(torch.float4_e2m1fn_x2) |
| 37 | replace_parameter(layer, "weight", weight.data.t()) |
| 38 | |
| 39 | weight_scale = layer.weight_scale.view(torch.float8_e8m0fnu) |
| 40 | weight_scale = weight_scale.t().contiguous() |
| 41 | replace_parameter(layer, "weight_scale", weight_scale.data) |
| 42 | |
| 43 | def apply_weights( |
| 44 | self, |
| 45 | layer: torch.nn.Module, |
| 46 | x: torch.Tensor, |
| 47 | bias: torch.Tensor | None = None, |
| 48 | ) -> torch.Tensor: |
| 49 | out_dtype = x.dtype |
| 50 | x_fp4, x_blockscale = quant_mxfp4(x) |
| 51 | return torch.ops._xpu_C.fp4_gemm( |
| 52 | x_fp4, |
| 53 | layer.weight, |
| 54 | x_blockscale, |
| 55 | layer.weight_scale, |
| 56 | out_dtype, |
| 57 | bias, |
| 58 | ) |