| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Utility helpers for NVFP4 + FlashInfer fused-MoE path.""" |
| 4 | |
| 5 | from typing import TYPE_CHECKING |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | from vllm.logger import init_logger |
| 10 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 11 | from vllm.model_executor.layers.quantization.utils.flashinfer_utils import ( |
| 12 | align_fp4_moe_hidden_dim_for_fi, |
| 13 | align_fp4_moe_weights_for_fi, |
| 14 | ) |
| 15 | from vllm.model_executor.layers.quantization.utils.nvfp4_utils import ( |
| 16 | swizzle_blockscale, |
| 17 | ) |
| 18 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 19 | amax_for_moe_activation_quant, |
| 20 | ) |
| 21 | |
| 22 | if TYPE_CHECKING: |
| 23 | from vllm.model_executor.layers.fused_moe import RoutedExperts |
| 24 | from vllm.model_executor.layers.fused_moe.oracle.nvfp4 import ( |
| 25 | NvFp4MoeBackend, |
| 26 | ) |
| 27 | |
| 28 | logger = init_logger(__name__) |
| 29 | |
| 30 | __all__ = [ |
| 31 | "reorder_w1w3_to_w3w1", |
| 32 | ] |
| 33 | |
| 34 | |
| 35 | def reorder_w1w3_to_w3w1( |
| 36 | weight: torch.Tensor, scale: torch.Tensor, dim: int = -2 |
| 37 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 38 | """Re-order concatenated `[w1, w3]` tensors to `[w3, w1]` in-place. |
| 39 | |
| 40 | `weight` and `scale` must be contiguous; they remain contiguous on return. |
| 41 | """ |
| 42 | assert weight.is_contiguous(), "weight must be contiguous" |
| 43 | assert scale.is_contiguous(), "scale must be contiguous" |
| 44 | size = weight.size(dim) |
| 45 | assert size % 2 == 0, f"Expected even size in dim {dim}, got {size}" |
| 46 | half = size // 2 |
| 47 | d = dim % weight.dim() |
| 48 | |
| 49 | # 64 MB transient cap |
| 50 | bytes_per_row = max( |
| 51 | weight.numel() // size * weight.element_size(), |
| 52 | scale.numel() // size * scale.element_size(), |
| 53 | ) |
| 54 | chunk = max(1, min(half, (64 << 20) // max(bytes_per_row, 1))) |
| 55 | |
| 56 | fa, fb = [slice(None)] * weight.dim(), [slice(None)] * weight.dim() |
| 57 | for off in range(0, half, chunk): |
| 58 | end = min(off + chunk, half) |
| 59 | fa[d], fb[d] = slice(off, end), slice(half + off, half + end) |
| 60 | a, b = tuple(fa), tuple(fb) |
| 61 | for t in (weight, scale): |
| 62 | tmp = t[b].clone() |
| 63 | t[b] = t[a] |
| 64 | t[a] = tmp |
| 65 | |
| 66 | return weight, scale |
| 67 | |
| 68 | |
| 69 | def interleave_linear_and_gate( |
| 70 | x: torch.Tensor, |
| 71 | group_size: int = 64, |
| 72 | dim: int = -1, |
| 73 | ) -> torch.Tensor: |
| 74 | """Interleave gate and linear weight rows for CuteDSL wrapper.""" |
| 75 | sizes = x.size() |
| 76 | dim = dim % x.dim() |
| 77 | assert sizes[dim] % (group_size * 2) == 0, ( |
| 78 | f"dim {dim} size {sizes[dim]} must be divisible by {group_size * 2}" |
| 79 | ) |
| 80 | prev_sizes = sizes[:dim] |
| 81 | post_sizes = sizes[dim + 1 :] |
| 82 | x = x.view(*prev_sizes, 2, sizes[dim] // (group_size * 2), group_size, *post_sizes) |
| 83 | x = x.transpose(dim, dim + 1).contiguous().view(*sizes) |
| 84 | return x |
| 85 | |
| 86 | |
| 87 | def reorder_w13_to_w31_for_flashinfer_cutedsl( |
| 88 | activation: MoEActivation, |
| 89 | w13: torch.Tensor, |
| 90 | w13_scale: torch.Tensor, |
| 91 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 92 | """Normalize gated w13 rows to the [up; gate] order used by FlashInfer.""" |
| 93 | if activation == MoEActivation.SWIGLUOAI: |
| 94 | # gpt-oss checkpoints store w13 interleaved as [gate0, up0, gate1, ...]. |
| 95 | gate, up = w13[:, 0::2], w13[:, 1::2] |
| 96 | gate_scale, up_scale = w13_scale[:, 0::2], w13_scale[:, 1::2] |
| 97 | return ( |
| 98 | torch.cat([up, gate], dim=1).contiguous(), |
| 99 | torch.cat([up_scale, gate_scale], dim=1).contiguous(), |
| 100 | ) |
| 101 | |
| 102 | half = w13.shape[1] // 2 |
| 103 | return ( |
| 104 | torch.cat([w13[:, half:], w13[:, :half]], dim=1).contiguous(), |
| 105 | torch.cat([w13_scale[:, half:], w13_scale[:, :half]], dim=1).contiguous(), |
| 106 | ) |
| 107 | |
| 108 | |
| 109 | def prepare_nvfp4_moe_layer_for_flashinfer_cutedsl( |
| 110 | layer: "RoutedExperts", |
| 111 | w13: torch.Tensor, |
| 112 | w13_scale: torch.Tensor, |
| 113 | w13_scale_2: torch.Tensor, |
| 114 | a13_scale: torch.Tensor, |
| 115 | w2: torch.Tensor, |
| 116 | w2_scale: torch.Tensor, |
| 117 | w2_scale_2: torch.Tensor, |
| 118 | a2_scale: torch.Tensor, |
| 119 | ) -> tuple[ |
| 120 | torch.Tensor, |
| 121 | torch.Tensor, |
| 122 | torch.Tensor, |
| 123 | torch.Tensor, |
| 124 | torch.Tensor, |
| 125 | torch.Tensor, |
| 126 | torch.Tensor, |
| 127 | torch.Tensor, |
| 128 | ]: |
| 129 | """Prepare weights for the CuteDSL wrapper-based NvFP4 MoE backend. |
| 130 | |
| 131 | Pads the runtime expert tensors to the kernel's GEMM alignment, converts |
| 132 | weight scale factors to the MMA layout expected by CuteDslMoEWrapper, and |
| 133 | interleaves w13 gate/linear rows for gated activations. |
| 134 | """ |
| 135 | # Global scaling factors (same as other FlashInfer backends). |
| 136 | num_experts = w13.shape[0] |
| 137 | enable_eplb = layer.moe_config.moe_parallel_config.enable_eplb |
| 138 | a13_scale = amax_for_moe_activation_quant(a13_scale, enable_eplb).repeat( |
| 139 | num_experts |
| 140 | ) |
| 141 | a2_scale = amax_for_moe_activation_quant(a2_scale, enable_eplb).repeat(num_experts) |
| 142 | |
| 143 | gated = layer.activation.is_gated |
| 144 | if gated: |
| 145 | w13, w13_scale = reorder_w13_to_w31_for_flashinfer_cutedsl( |
| 146 | layer.activation, w13, w13_scale |
| 147 | ) |
| 148 | |
| 149 | # GEMM1's output dimension must be a multiple of 128: 2I for gated |
| 150 | # activations (also required by interleaving), but only I for non-gated. |
| 151 | # Keep the checkpoint tensors unchanged and pad only the kernel's runtime |
| 152 | # representation. Zero rows also make the padded GEMM2 contraction a no-op. |
| 153 | w13, w13_scale, w2, w2_scale, padded_intermediate = align_fp4_moe_weights_for_fi( |
| 154 | w13, |
| 155 | w13_scale, |
| 156 | w2, |
| 157 | w2_scale, |
| 158 | is_act_and_mul=gated, |
| 159 | min_alignment=64 if gated else 128, |
| 160 | ) |
| 161 | layer.moe_config.intermediate_size_per_partition = padded_intermediate |
| 162 | |
| 163 | # GEMM1 gathers full 256-element K tiles of activations and their scales |
| 164 | # without a K-tail predicate. Pad runtime weights to match; the MoE runner |
| 165 | # pads activations and trims the output using the original hidden size. |
| 166 | w13, w13_scale, w2, w2_scale, padded_hidden = align_fp4_moe_hidden_dim_for_fi( |
| 167 | w13, w13_scale, w2, w2_scale |
| 168 | ) |
| 169 | if layer.moe_config.hidden_dim_unpadded is None: |
| 170 | layer.moe_config.hidden_dim_unpadded = layer.moe_config.hidden_dim |
| 171 | layer.moe_config.hidden_dim = padded_hidden |
| 172 | |
| 173 | if gated: |
| 174 | # Interleave up/gate rows for the fused gated activation. |
| 175 | w13 = interleave_linear_and_gate(w13, group_size=64, dim=1) |
| 176 | w13_scale = interleave_linear_and_gate(w13_scale, group_size=64, dim=1) |
| 177 | |
| 178 | w13_scale = swizzle_blockscale(w13_scale) |
| 179 | w2_scale = swizzle_blockscale(w2_scale) |
| 180 | |
| 181 | return ( |
| 182 | w13, |
| 183 | w13_scale, |
| 184 | w13_scale_2, |
| 185 | a13_scale, |
| 186 | w2, |
| 187 | w2_scale, |
| 188 | w2_scale_2, |
| 189 | a2_scale, |
| 190 | ) |
| 191 | |
| 192 | |
| 193 | def nvfp4_swizzled_scale_to_cutedsl_mma_view(scale: torch.Tensor) -> torch.Tensor: |
| 194 | """View a swizzled (E, M_padded, K_sf_padded) block-scale tensor in the |
| 195 | MMA layout expected by the CuteDSL MoE kernel. |
| 196 | |
| 197 | The returned tensor aliases `scale`'s storage, so in-place updates of the |
| 198 | registered Parameter (weight reloads, EPLB rearrangement) are visible to |
| 199 | the kernel with no extra bookkeeping. |
| 200 | """ |
| 201 | from flashinfer.cute_dsl.utils import convert_sf_to_mma_layout |
| 202 | |
| 203 | num_experts, m_padded, k_sf_padded = scale.shape |
| 204 | mma_view = convert_sf_to_mma_layout( |
| 205 | scale.reshape(num_experts * m_padded, k_sf_padded), |
| 206 | m=m_padded, |
| 207 | k=k_sf_padded * 16, |
| 208 | num_groups=num_experts, |
| 209 | sf_vec_size=16, |
| 210 | ) |
| 211 | assert mma_view.data_ptr() == scale.data_ptr(), ( |
| 212 | "convert_sf_to_mma_layout no longer returns a view of its input; " |
| 213 | "the quant config would go stale after weight updates." |
| 214 | ) |
| 215 | return mma_view |
| 216 | |
| 217 | |
| 218 | def prepare_static_weights_for_trtllm_fp4_moe( |
| 219 | gemm1_weights: torch.Tensor, |
| 220 | gemm2_weights: torch.Tensor, |
| 221 | gemm1_scales_linear_fp4_bytes: torch.Tensor, |
| 222 | gemm2_scales_linear_fp4_bytes: torch.Tensor, |
| 223 | hidden_size: int, |
| 224 | intermediate_size: int, |
| 225 | num_experts: int, |
| 226 | is_gated_activation: bool, |
| 227 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: |
| 228 | """Shuffle NVFP4 weights into the FlashInfer TRT-LLM layout.""" |
| 229 | from flashinfer import nvfp4_block_scale_interleave |
| 230 | from flashinfer.fused_moe.core import ( |
| 231 | _maybe_get_cached_w3_w1_permute_indices, |
| 232 | get_w2_permute_indices_with_cache, |
| 233 | ) |
| 234 | |
| 235 | permute_indices_cache: dict[tuple[object, ...], torch.Tensor] = {} |
| 236 | epilogue_tile_m = 128 # FIXME: this depends on the kernel internals |
| 237 | gemm1_intermediate_size = ( |
| 238 | 2 * intermediate_size if is_gated_activation else intermediate_size |
| 239 | ) |
| 240 | |
| 241 | # Convert quantized weights to proper formats |
| 242 | gemm1_weights_fp4 = gemm1_weights.view(torch.float8_e4m3fn).reshape( |
| 243 | num_experts, gemm1_intermediate_size, hidden_size // 2 |
| 244 | ) # packed fp4 |
| 245 | gemm1_scales_linear_fp4 = gemm1_scales_linear_fp4_bytes.view( |
| 246 | torch.float8_e4m3fn |
| 247 | ).reshape( |
| 248 | num_experts, gemm1_intermediate_size, hidden_size // 16 |
| 249 | ) # fp8 scaling factors |
| 250 | |
| 251 | gemm2_weights_fp4 = gemm2_weights.view(torch.float8_e4m3fn).reshape( |
| 252 | num_experts, hidden_size, intermediate_size // 2 |
| 253 | ) # packed fp4 |
| 254 | gemm2_scales_linear_fp4 = gemm2_scales_linear_fp4_bytes.view( |
| 255 | torch.float8_e4m3fn |
| 256 | ).reshape(num_experts, hidden_size, intermediate_size // 16) # fp8 scaling factors |
| 257 | |
| 258 | gemm1_weight_indices = _maybe_get_cached_w3_w1_permute_indices( |
| 259 | permute_indices_cache, |
| 260 | gemm1_weights_fp4[0].view(torch.uint8), |
| 261 | epilogue_tile_m, |
| 262 | is_gated_act_gemm=is_gated_activation, |
| 263 | ) |
| 264 | gemm1_scale_indices = _maybe_get_cached_w3_w1_permute_indices( |
| 265 | permute_indices_cache, |
| 266 | gemm1_scales_linear_fp4[0].view(torch.uint8), |
| 267 | epilogue_tile_m, |
| 268 | num_elts_per_sf=16, |
| 269 | is_gated_act_gemm=is_gated_activation, |
| 270 | ) |
| 271 | gemm2_weight_indices = get_w2_permute_indices_with_cache( |
| 272 | permute_indices_cache, |
| 273 | gemm2_weights_fp4[0].view(torch.uint8), |
| 274 | epilogue_tile_m, |
| 275 | ) |
| 276 | gemm2_scale_indices = get_w2_permute_indices_with_cache( |
| 277 | permute_indices_cache, |
| 278 | gemm2_scales_linear_fp4[0].view(torch.uint8), |
| 279 | epilogue_tile_m, |
| 280 | num_elts_per_sf=16, |
| 281 | ) |
| 282 | |
| 283 | gemm1_weights_fp4_shuffled = torch.empty_like(gemm1_weights_fp4, dtype=torch.uint8) |
| 284 | gemm1_scales_fp4_shuffled = torch.empty_like(gemm1_scales_linear_fp4) |
| 285 | gemm2_weights_fp4_shuffled = torch.empty_like(gemm2_weights_fp4, dtype=torch.uint8) |
| 286 | gemm2_scales_fp4_shuffled = torch.empty_like(gemm2_scales_linear_fp4) |
| 287 | gemm1_scale_scratch = torch.empty_like( |
| 288 | gemm1_scales_linear_fp4[0], dtype=torch.uint8 |
| 289 | ) |
| 290 | gemm2_scale_scratch = torch.empty_like( |
| 291 | gemm2_scales_linear_fp4[0], dtype=torch.uint8 |
| 292 | ) |
| 293 | |
| 294 | for expert_id in range(num_experts): |
| 295 | torch.index_select( |
| 296 | gemm1_weights_fp4[expert_id].view(torch.uint8), |
| 297 | 0, |
| 298 | gemm1_weight_indices, |
| 299 | out=gemm1_weights_fp4_shuffled[expert_id], |
| 300 | ) |
| 301 | torch.index_select( |
| 302 | gemm1_scales_linear_fp4[expert_id].view(torch.uint8), |
| 303 | 0, |
| 304 | gemm1_scale_indices, |
| 305 | out=gemm1_scale_scratch, |
| 306 | ) |
| 307 | gemm1_scales_fp4_shuffled[expert_id].view(torch.uint8).reshape(-1).copy_( |
| 308 | nvfp4_block_scale_interleave(gemm1_scale_scratch) |
| 309 | ) |
| 310 | |
| 311 | torch.index_select( |
| 312 | gemm2_weights_fp4[expert_id].view(torch.uint8), |
| 313 | 0, |
| 314 | gemm2_weight_indices, |
| 315 | out=gemm2_weights_fp4_shuffled[expert_id], |
| 316 | ) |
| 317 | torch.index_select( |
| 318 | gemm2_scales_linear_fp4[expert_id].view(torch.uint8), |
| 319 | 0, |
| 320 | gemm2_scale_indices, |
| 321 | out=gemm2_scale_scratch, |
| 322 | ) |
| 323 | gemm2_scales_fp4_shuffled[expert_id].view(torch.uint8).reshape(-1).copy_( |
| 324 | nvfp4_block_scale_interleave(gemm2_scale_scratch) |
| 325 | ) |
| 326 | |
| 327 | return ( |
| 328 | gemm1_weights_fp4_shuffled, |
| 329 | gemm1_scales_fp4_shuffled, |
| 330 | gemm2_weights_fp4_shuffled, |
| 331 | gemm2_scales_fp4_shuffled, |
| 332 | ) |
| 333 | |
| 334 | |
| 335 | def prepare_nvfp4_moe_layer_for_fi_or_cutlass( |
| 336 | backend: "NvFp4MoeBackend", |
| 337 | layer: "RoutedExperts", |
| 338 | w13: torch.Tensor, |
| 339 | w13_scale: torch.Tensor, |
| 340 | w13_scale_2: torch.Tensor, |
| 341 | a13_scale: torch.Tensor, |
| 342 | w2: torch.Tensor, |
| 343 | w2_scale: torch.Tensor, |
| 344 | w2_scale_2: torch.Tensor, |
| 345 | a2_scale: torch.Tensor, |
| 346 | is_act_and_mul: bool, |
| 347 | ) -> tuple[ |
| 348 | torch.Tensor, |
| 349 | torch.Tensor, |
| 350 | torch.Tensor, |
| 351 | torch.Tensor, |
| 352 | torch.Tensor, |
| 353 | torch.Tensor, |
| 354 | torch.Tensor, |
| 355 | torch.Tensor, |
| 356 | ]: |
| 357 | # Delayed import for circular dependency avoidance. |
| 358 | from vllm.model_executor.layers.fused_moe.oracle.nvfp4 import ( |
| 359 | NvFp4MoeBackend, |
| 360 | is_global_sf_supported_for_nvfp4_backend, |
| 361 | ) |
| 362 | |
| 363 | assert backend in [ |
| 364 | NvFp4MoeBackend.VLLM_CUTLASS, |
| 365 | NvFp4MoeBackend.FLASHINFER_CUTLASS, |
| 366 | NvFp4MoeBackend.FLASHINFER_TRTLLM, |
| 367 | NvFp4MoeBackend.FLASHINFER_CUTEDSL_BATCHED, |
| 368 | NvFp4MoeBackend.FLASHINFER_B12X, |
| 369 | ] |
| 370 | |
| 371 | # Reorder [w1, w3] to [w3, w1] for FI NVFP4 MoE kernels. |
| 372 | is_gated = layer.activation.is_gated |
| 373 | if ( |
| 374 | is_gated |
| 375 | and is_act_and_mul |
| 376 | and backend |
| 377 | in [ |
| 378 | NvFp4MoeBackend.FLASHINFER_CUTLASS, |
| 379 | NvFp4MoeBackend.FLASHINFER_TRTLLM, |
| 380 | NvFp4MoeBackend.FLASHINFER_B12X, |
| 381 | ] |
| 382 | ): |
| 383 | w13, w13_scale = reorder_w1w3_to_w3w1(w13, w13_scale) |
| 384 | |
| 385 | # For some FI kernels, the input scales are shared by all experts. |
| 386 | if is_global_sf_supported_for_nvfp4_backend(backend): |
| 387 | num_experts = w13.shape[0] |
| 388 | enable_eplb = layer.moe_config.moe_parallel_config.enable_eplb |
| 389 | a13_scale = amax_for_moe_activation_quant(a13_scale, enable_eplb).repeat( |
| 390 | num_experts |
| 391 | ) |
| 392 | a2_scale = amax_for_moe_activation_quant(a2_scale, enable_eplb).repeat( |
| 393 | num_experts |
| 394 | ) |
| 395 | else: |
| 396 | a13_scale = a13_scale.max(dim=1).values.to(torch.float32) |
| 397 | |
| 398 | # Shuffle weights and scales for FI TRTLLM NVFP4 MoE kernels. |
| 399 | if backend == NvFp4MoeBackend.FLASHINFER_TRTLLM: |
| 400 | w13, w13_scale, w2, w2_scale, padded_hidden = align_fp4_moe_hidden_dim_for_fi( |
| 401 | w13, w13_scale, w2, w2_scale |
| 402 | ) |
| 403 | if layer.moe_config.hidden_dim_unpadded is None: |
| 404 | layer.moe_config.hidden_dim_unpadded = layer.moe_config.hidden_dim |
| 405 | layer.moe_config.hidden_dim = padded_hidden |
| 406 | |
| 407 | # Align weights for FI NVFP4 MoE kernels. |
| 408 | # FlashInfer's TRT-LLM block-scale shuffle asserts the gate/up row dim |
| 409 | # (= up_mult * padded_intermediate, up_mult=2 when gated) is a multiple of |
| 410 | # 128. So gated needs padded_intermediate % 64 (2*64=128); the old value 16 |
| 411 | # left 2*intermediate a multiple of only 32, so an NVFP4 MoE whose rank-local |
| 412 | # intermediate is not 128-aligned at TP>1 (e.g. Gemma-4-26B-A4B at tp4) hit |
| 413 | # `assert M % 128 == 0`. Padded rows are zero -> outputs unchanged. |
| 414 | min_alignment = 64 if is_gated else 128 |
| 415 | w13, w13_scale, w2, w2_scale, padded_intermediate = ( |
| 416 | align_fp4_moe_weights_for_fi( |
| 417 | w13, w13_scale, w2, w2_scale, is_act_and_mul, min_alignment |
| 418 | ) |
| 419 | ) |
| 420 | layer.moe_config.intermediate_size_per_partition = padded_intermediate |
| 421 | |
| 422 | w13, w13_scale, w2, w2_scale = prepare_static_weights_for_trtllm_fp4_moe( |
| 423 | w13, |
| 424 | w2, |
| 425 | w13_scale, |
| 426 | w2_scale, |
| 427 | hidden_size=w2.size(-2), |
| 428 | intermediate_size=w13.size(-2) // 2 if is_gated else w13.size(-2), |
| 429 | num_experts=w13.size(0), |
| 430 | is_gated_activation=is_gated, |
| 431 | ) |
| 432 | else: |
| 433 | # Swizzle the block scales for other FI NVFP4 MoE kernels. |
| 434 | w13_scale = swizzle_blockscale(w13_scale) |
| 435 | |
| 436 | # Apply padding if needed. |
| 437 | pad_size = w13_scale.size(1) - w13.size(1) |
| 438 | if pad_size > 0: |
| 439 | if is_act_and_mul: |
| 440 | raise NotImplementedError( |
| 441 | "Intermediate size padding for w1 and w3, for %s " |
| 442 | "NvFp4 backend, but this is not currently supported", |
| 443 | backend.value, |
| 444 | ) |
| 445 | w13 = torch.nn.functional.pad(w13, (0, 0, 0, pad_size)) |
| 446 | w2 = torch.nn.functional.pad(w2, (0, pad_size // 2, 0, 0)) |
| 447 | w2_scale = torch.nn.functional.pad(w2_scale, (0, pad_size // 16)) |
| 448 | |
| 449 | w2_scale = swizzle_blockscale(w2_scale) |
| 450 | |
| 451 | return w13, w13_scale, w13_scale_2, a13_scale, w2, w2_scale, w2_scale_2, a2_scale |