| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | |
| 6 | import vllm.envs as envs |
| 7 | from vllm.config import get_current_vllm_config |
| 8 | from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8 |
| 9 | from vllm.model_executor.layers.quantization.utils.fp8_utils import ( |
| 10 | deepgemm_post_process_fp8_weight_block, |
| 11 | ) |
| 12 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 13 | GroupShape, |
| 14 | ) |
| 15 | from vllm.model_executor.utils import replace_parameter |
| 16 | from vllm.platforms import current_platform |
| 17 | from vllm.utils.deep_gemm import ( |
| 18 | fp8_gemm_nt, |
| 19 | is_deep_gemm_e8m0_used, |
| 20 | is_deep_gemm_supported, |
| 21 | should_auto_disable_deep_gemm, |
| 22 | should_use_deepgemm_for_fp8_linear, |
| 23 | ) |
| 24 | from vllm.utils.torch_utils import direct_register_custom_op |
| 25 | |
| 26 | from .BlockScaledMMLinearKernel import ( |
| 27 | Fp8BlockScaledMMLinearKernel, |
| 28 | FP8ScaledMMLinearLayerConfig, |
| 29 | ) |
| 30 | |
| 31 | |
| 32 | class DeepGemmFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel): |
| 33 | def __init__(self, config: FP8ScaledMMLinearLayerConfig): |
| 34 | super().__init__(config) |
| 35 | self.use_deep_gemm_e8m0 = is_deep_gemm_e8m0_used() |
| 36 | act_scale_descriptor = config.activation_quant_key.scale |
| 37 | self.is_deep_gemm_supported = is_deep_gemm_supported() |
| 38 | self.quant_fp8 = QuantFP8( |
| 39 | static=False, |
| 40 | group_shape=act_scale_descriptor.group_shape, |
| 41 | use_ue8m0=self.use_deep_gemm_e8m0, |
| 42 | tma_aligned_scales=envs.VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES, |
| 43 | column_major_scales=True, |
| 44 | ) |
| 45 | |
| 46 | @classmethod |
| 47 | def is_supported(cls, compute_capability=None): |
| 48 | if not current_platform.is_cuda(): |
| 49 | return False, "DeepGEMM is only supported on cuda platform" |
| 50 | if not is_deep_gemm_supported(): |
| 51 | return False, "Currently, only Hopper and Blackwell GPUs are supported." |
| 52 | return True, None |
| 53 | |
| 54 | @classmethod |
| 55 | def can_implement(cls, config): |
| 56 | can_implement_base, reason = super().can_implement(config) |
| 57 | if not can_implement_base: |
| 58 | return can_implement_base, reason |
| 59 | if config.out_dtype != torch.bfloat16: |
| 60 | return (False, "Supports only output dtype of bfloat16") |
| 61 | |
| 62 | act_quant_desc = config.activation_quant_key.scale |
| 63 | if act_quant_desc.group_shape != GroupShape(1, 128): |
| 64 | return ( |
| 65 | False, |
| 66 | "Supports only dynamic per token group activation " |
| 67 | "quantization with group_shape=(1,128).", |
| 68 | ) |
| 69 | model_config = get_current_vllm_config().model_config |
| 70 | |
| 71 | if model_config is None: |
| 72 | return False, "Model configuration is required." |
| 73 | |
| 74 | model_type = getattr(model_config.hf_text_config, "model_type", None) |
| 75 | if should_auto_disable_deep_gemm(model_type): |
| 76 | return False, f"Should not use deepgemm for model {model_type}" |
| 77 | |
| 78 | if not should_use_deepgemm_for_fp8_linear( |
| 79 | config.out_dtype, config.weight_shape |
| 80 | ): |
| 81 | return False, "The provided metadata is not supported." |
| 82 | return True, None |
| 83 | |
| 84 | def process_weights_after_loading(self, layer): |
| 85 | super().process_weights_after_loading(layer) |
| 86 | params = self._get_layer_params(layer) |
| 87 | assert layer.weight_block_size is not None |
| 88 | |
| 89 | is_bmm = getattr(layer, "is_bmm", False) |
| 90 | |
| 91 | if self.is_deep_gemm_supported: |
| 92 | dg_weight, dg_weight_scale = deepgemm_post_process_fp8_weight_block( |
| 93 | wq=params.weight, |
| 94 | ws=params.block_scale, |
| 95 | quant_block_shape=tuple(layer.weight_block_size), |
| 96 | use_e8m0=self.use_deep_gemm_e8m0, |
| 97 | is_bmm=is_bmm, |
| 98 | bmm_batch_size=getattr(layer, "bmm_batch_size", 0), |
| 99 | ) |
| 100 | replace_parameter(layer, params.WEIGHT, dg_weight) |
| 101 | replace_parameter(layer, params.block_scale_attr, dg_weight_scale) |
| 102 | # bmm layers bypass this kernel and run through fp8_einsum. |
| 103 | if not is_bmm: |
| 104 | layer.deep_gemm_warmup_provider = self |
| 105 | |
| 106 | def get_deep_gemm_warmup_weights( |
| 107 | self, layer: torch.nn.Module |
| 108 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 109 | params = self._get_layer_params(layer) |
| 110 | return params.weight, params.block_scale |
| 111 | |
| 112 | def apply_block_scaled_mm( |
| 113 | self, |
| 114 | A: torch.Tensor, |
| 115 | B: torch.Tensor, |
| 116 | As: torch.Tensor, |
| 117 | Bs: torch.Tensor, |
| 118 | ) -> torch.Tensor: |
| 119 | out_dtype = self.config.out_dtype |
| 120 | output = torch.empty( |
| 121 | (A.shape[0], B.shape[0]), |
| 122 | dtype=out_dtype, |
| 123 | device=A.device, |
| 124 | ) |
| 125 | torch.ops.vllm.fp8_gemm_nt_op(A, As, B, Bs, output, self.use_deep_gemm_e8m0) |
| 126 | return output |
| 127 | |
| 128 | |
| 129 | def _fp8_gemm_nt_op( |
| 130 | q_input: torch.Tensor, |
| 131 | input_scale: torch.Tensor, |
| 132 | weight: torch.Tensor, |
| 133 | weight_scale: torch.Tensor, |
| 134 | output: torch.Tensor, |
| 135 | use_deep_gemm_e8m0: bool, |
| 136 | ) -> None: |
| 137 | fp8_gemm_nt( |
| 138 | (q_input, input_scale), |
| 139 | (weight, weight_scale), |
| 140 | output, |
| 141 | is_deep_gemm_e8m0_used=use_deep_gemm_e8m0, |
| 142 | ) |
| 143 | |
| 144 | |
| 145 | direct_register_custom_op( |
| 146 | "fp8_gemm_nt_op", |
| 147 | _fp8_gemm_nt_op, |
| 148 | mutates_args=["output"], |
| 149 | ) |