| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Utility helpers for MxInt4 + FlashInfer fused-MoE path.""" |
| 4 | |
| 5 | import functools |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | import vllm.envs as envs |
| 10 | from vllm.logger import init_logger |
| 11 | from vllm.platforms import current_platform |
| 12 | from vllm.utils.flashinfer import has_flashinfer_trtllm_fused_moe |
| 13 | |
| 14 | __all__ = [ |
| 15 | "prepare_static_weights_for_trtllm_mxint4_moe", |
| 16 | "flashinfer_trtllm_mxint4_moe", |
| 17 | "is_flashinfer_mxint4_moe_available", |
| 18 | ] |
| 19 | |
| 20 | logger = init_logger(__name__) |
| 21 | |
| 22 | |
| 23 | @functools.cache |
| 24 | def is_flashinfer_mxint4_moe_available() -> bool: |
| 25 | """Return `True` when FlashInfer MxInt4 kernels can be used.""" |
| 26 | return ( |
| 27 | envs.VLLM_USE_FLASHINFER_MOE_INT4 |
| 28 | and has_flashinfer_trtllm_fused_moe() |
| 29 | and current_platform.is_cuda() |
| 30 | and current_platform.is_device_capability_family(100) |
| 31 | ) |
| 32 | |
| 33 | |
| 34 | def prepare_static_weights_for_trtllm_mxint4_moe( |
| 35 | gemm1_weights: torch.Tensor, |
| 36 | gemm1_scales: torch.Tensor, |
| 37 | gemm2_weights: torch.Tensor, |
| 38 | gemm2_scales: torch.Tensor, |
| 39 | ) -> dict[str, torch.Tensor]: |
| 40 | """Prepare MxInt4 weights for TRT-LLM kernel. |
| 41 | |
| 42 | Input: |
| 43 | gemm1_weights: [num_experts, 2*intermediate_size, hidden_size//8] int32 |
| 44 | (checkpoint uint4b8 packed) or uint8 (already packed signed int4) |
| 45 | gemm1_scales: [num_experts, 2*intermediate_size, hidden_size//32] bf16 |
| 46 | gemm2_weights: [num_experts, hidden_size, intermediate_size//8] int32 |
| 47 | (checkpoint uint4b8 packed) or uint8 (already packed signed int4) |
| 48 | gemm2_scales: [num_experts, hidden_size, intermediate_size//32] bf16 |
| 49 | |
| 50 | Returns: |
| 51 | Dict with keys 'gemm1_weights', 'gemm1_scales', 'gemm2_weights', |
| 52 | 'gemm2_scales' containing shuffled/packed tensors ready for kernel |
| 53 | |
| 54 | """ |
| 55 | from flashinfer import block_scale_interleave |
| 56 | from flashinfer.fused_moe import ( |
| 57 | convert_to_block_layout, |
| 58 | ) |
| 59 | from flashinfer.fused_moe.core import ( |
| 60 | _maybe_get_cached_w3_w1_permute_indices, |
| 61 | get_w2_permute_indices_with_cache, |
| 62 | ) |
| 63 | |
| 64 | from vllm.model_executor.layers.quantization.utils.flashinfer_fp4_moe import ( |
| 65 | reorder_w1w3_to_w3w1, |
| 66 | ) |
| 67 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 68 | convert_packed_uint4b8_to_signed_int4_inplace, |
| 69 | ) |
| 70 | |
| 71 | device = gemm1_weights.device |
| 72 | assert gemm1_weights.ndim == 3, ( |
| 73 | f"Expected a 3D gemm1_weights tensor, got {gemm1_weights.shape}" |
| 74 | ) |
| 75 | assert gemm1_scales.ndim == 3, ( |
| 76 | f"Expected a 3D gemm1_scales tensor, got {gemm1_scales.shape}" |
| 77 | ) |
| 78 | assert gemm2_weights.ndim == 3, ( |
| 79 | f"Expected a 3D gemm2_weights tensor, got {gemm2_weights.shape}" |
| 80 | ) |
| 81 | assert gemm2_scales.ndim == 3, ( |
| 82 | f"Expected a 3D gemm2_scales tensor, got {gemm2_scales.shape}" |
| 83 | ) |
| 84 | |
| 85 | # Convert checkpoint format (uint4b8 in int32) to signed int4 |
| 86 | # Checkpoint stores INT4 as unsigned [0, 15], kernel expects signed [-8, 7] |
| 87 | if gemm1_weights.dtype == torch.int32 and gemm2_weights.dtype == torch.int32: |
| 88 | convert_packed_uint4b8_to_signed_int4_inplace(gemm1_weights) |
| 89 | convert_packed_uint4b8_to_signed_int4_inplace(gemm2_weights) |
| 90 | |
| 91 | gemm1_weights, gemm1_scales = reorder_w1w3_to_w3w1( |
| 92 | gemm1_weights, gemm1_scales, dim=-2 |
| 93 | ) |
| 94 | |
| 95 | _cache_permute_indices: dict[torch.Size, torch.Tensor] = {} |
| 96 | num_experts = gemm1_weights.shape[0] |
| 97 | |
| 98 | # Convert quantized weights to proper formats - |
| 99 | gemm1_weights_mxint4 = gemm1_weights.view(torch.uint8) |
| 100 | assert gemm1_scales.dtype == torch.bfloat16 |
| 101 | gemm2_weights_mxint4 = gemm2_weights.view(torch.uint8) |
| 102 | assert gemm2_scales.dtype == torch.bfloat16 |
| 103 | |
| 104 | epilogue_tile_m = 128 |
| 105 | gemm1_weights_mxint4_shuffled = [] |
| 106 | gemm1_scales_shuffled = [] |
| 107 | gemm2_weights_mxint4_shuffled = [] |
| 108 | gemm2_scales_shuffled = [] |
| 109 | |
| 110 | for i in range(num_experts): |
| 111 | # Calculate the permute indices for the following: |
| 112 | # 1. Reorder rows of W1 and scales for fused gated activation |
| 113 | # 2. Shuffle weights and scaling factors for transposed mma output |
| 114 | # for both w3_w1 and w2 weights and scale factors |
| 115 | permute_indices = _maybe_get_cached_w3_w1_permute_indices( |
| 116 | _cache_permute_indices, |
| 117 | gemm1_weights_mxint4[i], |
| 118 | epilogue_tile_m, |
| 119 | ) |
| 120 | gemm1_weights_shuffled = gemm1_weights_mxint4[i][ |
| 121 | permute_indices.to(gemm1_weights.device) |
| 122 | ].contiguous() |
| 123 | permute_sf_indices = _maybe_get_cached_w3_w1_permute_indices( |
| 124 | _cache_permute_indices, |
| 125 | gemm1_scales[i], |
| 126 | epilogue_tile_m, |
| 127 | num_elts_per_sf=32, |
| 128 | ).to(device) |
| 129 | gemm1_scales_shuffled.append( |
| 130 | block_scale_interleave(gemm1_scales[i][permute_sf_indices].contiguous()) |
| 131 | ) |
| 132 | |
| 133 | permute_indices = get_w2_permute_indices_with_cache( |
| 134 | _cache_permute_indices, |
| 135 | gemm2_weights_mxint4[i], |
| 136 | epilogue_tile_m, |
| 137 | ) |
| 138 | gemm2_weights_shuffled = gemm2_weights_mxint4[i][ |
| 139 | permute_indices.to(gemm2_weights.device) |
| 140 | ].contiguous() |
| 141 | |
| 142 | permute_sf_indices = get_w2_permute_indices_with_cache( |
| 143 | _cache_permute_indices, |
| 144 | gemm2_scales[i], |
| 145 | epilogue_tile_m, |
| 146 | num_elts_per_sf=16, |
| 147 | ) |
| 148 | gemm2_scales_shuffled.append( |
| 149 | block_scale_interleave( |
| 150 | gemm2_scales[i][permute_sf_indices.to(gemm2_scales.device)].contiguous() |
| 151 | ) |
| 152 | ) |
| 153 | |
| 154 | block_k = 128 |
| 155 | gemm1_weights_shuffled = convert_to_block_layout( |
| 156 | gemm1_weights_shuffled.view(torch.uint8), block_k |
| 157 | ) |
| 158 | gemm2_weights_shuffled = convert_to_block_layout( |
| 159 | gemm2_weights_shuffled.view(torch.uint8), block_k |
| 160 | ) |
| 161 | |
| 162 | gemm1_weights_mxint4_shuffled.append(gemm1_weights_shuffled) |
| 163 | gemm2_weights_mxint4_shuffled.append(gemm2_weights_shuffled) |
| 164 | |
| 165 | gemm1_weights_mxint4_shuffled = torch.stack(gemm1_weights_mxint4_shuffled) |
| 166 | gemm2_weights_mxint4_shuffled = torch.stack(gemm2_weights_mxint4_shuffled) |
| 167 | gemm1_scales_shuffled = torch.stack(gemm1_scales_shuffled).view(torch.bfloat16) |
| 168 | gemm2_scales_shuffled = torch.stack(gemm2_scales_shuffled).view(torch.bfloat16) |
| 169 | return { |
| 170 | "gemm1_weights": gemm1_weights_mxint4_shuffled, |
| 171 | "gemm1_scales": gemm1_scales_shuffled, |
| 172 | "gemm2_weights": gemm2_weights_mxint4_shuffled, |
| 173 | "gemm2_scales": gemm2_scales_shuffled, |
| 174 | } |
| 175 | |
| 176 | |
| 177 | def flashinfer_trtllm_mxint4_moe( |
| 178 | x: torch.Tensor, |
| 179 | router_logits: torch.Tensor, |
| 180 | w13_weight_packed: torch.Tensor, |
| 181 | w13_weight_scale: torch.Tensor, |
| 182 | w2_weight_packed: torch.Tensor, |
| 183 | w2_weight_scale: torch.Tensor, |
| 184 | global_num_experts: int, |
| 185 | top_k: int, |
| 186 | intermediate_size_per_partition: int, |
| 187 | local_num_experts: int, |
| 188 | ep_rank: int = 0, |
| 189 | num_expert_group: int | None = None, |
| 190 | topk_group: int | None = None, |
| 191 | e_score_correction_bias: torch.Tensor | None = None, |
| 192 | routing_method_type: int | None = None, |
| 193 | routing_replay_out: torch.Tensor | None = None, |
| 194 | ) -> torch.Tensor: |
| 195 | """Apply FlashInfer TensorRT-LLM MxInt4 MoE kernel. |
| 196 | |
| 197 | Args: |
| 198 | x: Input hidden states. dtype: bfloat16 |
| 199 | router_logits: Router logits for expert selection. dtype: bfloat16/float32 |
| 200 | w13_weight_packed: Packed gate+up weights. dtype: uint8 |
| 201 | w13_weight_scale: Scales for gate+up weights. dtype: bfloat16 |
| 202 | w2_weight_packed: Packed down weights. dtype: uint8 |
| 203 | w2_weight_scale: Scales for down weights. dtype: bfloat16 |
| 204 | global_num_experts: Total number of experts across all ranks |
| 205 | top_k: Number of experts to select per token |
| 206 | intermediate_size_per_partition: Intermediate size per partition |
| 207 | local_num_experts: Number of experts on this rank |
| 208 | ep_rank: Expert parallelism rank (default: 0) |
| 209 | num_expert_group: Number of expert groups (default: None -> 0) |
| 210 | topk_group: Top-k within groups (default: None -> 0) |
| 211 | routing_replay_out: Optional buffer receiving the routing decisions |
| 212 | so they can be replayed later |
| 213 | e_score_correction_bias: Optional routing bias. dtype: bfloat16 |
| 214 | routing_method_type: FlashInfer RoutingMethodType enum value |
| 215 | |
| 216 | Returns: |
| 217 | Output tensor from MoE layer. dtype: same as x (bfloat16) |
| 218 | |
| 219 | """ |
| 220 | from flashinfer import RoutingMethodType |
| 221 | from flashinfer.fused_moe import trtllm_mxint4_block_scale_moe |
| 222 | |
| 223 | assert x.dtype == torch.bfloat16, f"x dtype must be bfloat16, got {x.dtype}" |
| 224 | assert w13_weight_packed.dtype == torch.uint8, ( |
| 225 | f"w13_weight_packed dtype must be uint8, got {w13_weight_packed.dtype}" |
| 226 | ) |
| 227 | assert w13_weight_scale.dtype == torch.bfloat16, ( |
| 228 | f"w13_weight_scale dtype must be bfloat16, got {w13_weight_scale.dtype}" |
| 229 | ) |
| 230 | assert w2_weight_packed.dtype == torch.uint8, ( |
| 231 | f"w2_weight_packed dtype must be uint8, got {w2_weight_packed.dtype}" |
| 232 | ) |
| 233 | assert w2_weight_scale.dtype == torch.bfloat16, ( |
| 234 | f"w2_weight_scale dtype must be bfloat16, got {w2_weight_scale.dtype}" |
| 235 | ) |
| 236 | |
| 237 | routing_bias = None |
| 238 | if e_score_correction_bias is not None: |
| 239 | routing_bias = e_score_correction_bias.to(torch.bfloat16) |
| 240 | |
| 241 | if routing_method_type == RoutingMethodType.DeepSeekV3: |
| 242 | router_logits = router_logits.to(torch.float32) |
| 243 | |
| 244 | out = trtllm_mxint4_block_scale_moe( |
| 245 | routing_logits=router_logits, |
| 246 | routing_bias=routing_bias, |
| 247 | hidden_states=x, |
| 248 | gemm1_weights=w13_weight_packed.data, |
| 249 | gemm1_weights_scale=w13_weight_scale.data, |
| 250 | gemm1_alpha=None, |
| 251 | gemm1_beta=None, |
| 252 | gemm1_clamp_limit=None, |
| 253 | gemm2_weights=w2_weight_packed.data, |
| 254 | gemm2_weights_scale=w2_weight_scale.data, |
| 255 | num_experts=global_num_experts, |
| 256 | top_k=top_k, |
| 257 | n_group=num_expert_group if num_expert_group is not None else 0, |
| 258 | topk_group=topk_group if topk_group is not None else 0, |
| 259 | intermediate_size=intermediate_size_per_partition, |
| 260 | local_expert_offset=ep_rank * local_num_experts, |
| 261 | local_num_experts=local_num_experts, |
| 262 | routed_scaling_factor=None, |
| 263 | routing_method_type=routing_method_type, |
| 264 | enable_pdl=None, |
| 265 | do_finalize=True, |
| 266 | output=None, |
| 267 | tune_max_num_tokens=8192, |
| 268 | routing_replay_out=routing_replay_out, |
| 269 | ) |
| 270 | if isinstance(out, (tuple, list)): |
| 271 | out = out[0] |
| 272 | out = out.to(x.dtype) |
| 273 | |
| 274 | return out |