| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from abc import ABC, abstractmethod |
| 5 | from dataclasses import dataclass |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | from vllm.model_executor.layers.quantization.utils.quant_utils import QuantKey |
| 10 | |
| 11 | |
| 12 | @dataclass |
| 13 | class Mxfp8LinearLayerConfig: |
| 14 | """Configuration for an MXFP8 linear layer. |
| 15 | |
| 16 | All MXFP8 layers share the same structure: FP8-E4M3 weights with |
| 17 | uint8 (E8M0) per-block scales at block size 32. |
| 18 | |
| 19 | Attributes: |
| 20 | weight_shape: The layer's `(out_features, in_features)`, i.e. `(N, K)`. |
| 21 | |
| 22 | """ |
| 23 | |
| 24 | weight_shape: tuple[int, int] |
| 25 | bmm_batch_size: int | None = None |
| 26 | |
| 27 | |
| 28 | class Mxfp8LinearKernel(ABC): |
| 29 | """Base class for MXFP8 quantized linear kernels. |
| 30 | |
| 31 | Each subclass implements a specific GEMM backend (FlashInfer CUTLASS, |
| 32 | Marlin, emulation). |
| 33 | """ |
| 34 | |
| 35 | supports_pre_processed_weights: bool = False |
| 36 | """True if ``process_weights_after_loading`` only rewrites parameters, so |
| 37 | weights exported by the weight cache daemon can be used as-is.""" |
| 38 | |
| 39 | def __init__(self, c: Mxfp8LinearLayerConfig) -> None: |
| 40 | assert self.can_implement(c)[0] |
| 41 | assert self.is_supported()[0] |
| 42 | self.config = c |
| 43 | |
| 44 | def input_quant_key(self) -> QuantKey | None: |
| 45 | """Return the input quantization key supported by this kernel. If the kernel |
| 46 | does not support input quantization outside of the kernel, return None. |
| 47 | """ |
| 48 | return None |
| 49 | |
| 50 | @classmethod |
| 51 | @abstractmethod |
| 52 | def is_supported( |
| 53 | cls, compute_capability: int | None = None |
| 54 | ) -> tuple[bool, str | None]: |
| 55 | raise NotImplementedError |
| 56 | |
| 57 | @classmethod |
| 58 | @abstractmethod |
| 59 | def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]: |
| 60 | raise NotImplementedError |
| 61 | |
| 62 | @abstractmethod |
| 63 | def process_weights_after_loading(self, layer: torch.nn.Module) -> None: |
| 64 | raise NotImplementedError |
| 65 | |
| 66 | @abstractmethod |
| 67 | def apply_weights( |
| 68 | self, |
| 69 | layer: torch.nn.Module, |
| 70 | x: torch.Tensor, |
| 71 | bias: torch.Tensor | None = None, |
| 72 | ) -> torch.Tensor: |
| 73 | raise NotImplementedError |