| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # ruff: noqa: E501 |
| 4 | |
| 5 | from dataclasses import dataclass |
| 6 | from typing import Any |
| 7 | |
| 8 | import cutlass |
| 9 | import cutlass.cute as cute |
| 10 | import torch |
| 11 | from cuda.bindings.driver import CUstream |
| 12 | from cutlass import BFloat16, Float32, Int64, Uint8, Uint32, const_expr |
| 13 | from quack.compile_utils import make_fake_tensor |
| 14 | |
| 15 | from vllm.cute_utils import ( |
| 16 | _bf16x2_abs, |
| 17 | _bf16x2_max, |
| 18 | cvt, |
| 19 | recast_val, |
| 20 | torch_to_cute_dtype, |
| 21 | ) |
| 22 | from vllm.model_executor.warmup.jit_warmup import kernel_launcher |
| 23 | from vllm.model_executor.warmup.jit_warmup_cutedsl_helper import ( |
| 24 | CuTeDSLLaunchSpec, |
| 25 | VllmCuTeDSLJitKernel, |
| 26 | ) |
| 27 | from vllm.vllm_flash_attn.cute import utils as cute_utils |
| 28 | |
| 29 | # MXFP4: 32 elements per block, packed 2 nibbles per byte, ue8m0 block scale. |
| 30 | MXFP4_BLOCK_SIZE = 32 |
| 31 | |
| 32 | |
| 33 | @cute.jit |
| 34 | def _load_q_and_rope( |
| 35 | positions: cute.Tensor, |
| 36 | q: cute.Tensor, |
| 37 | cos_sin_cache: cute.Tensor, |
| 38 | num_heads: cutlass.Constexpr, |
| 39 | coarsen: cutlass.Constexpr, |
| 40 | tb_size: cutlass.Constexpr, |
| 41 | subwarp_size: cutlass.Constexpr, |
| 42 | threads_per_token: cutlass.Constexpr, |
| 43 | nope_dim: cutlass.Constexpr, |
| 44 | rope_dim: cutlass.Constexpr, |
| 45 | cos_sin_dtype: cutlass.Constexpr, |
| 46 | ): |
| 47 | """Compute thread indices, load Q (BF16), and apply interleaved RoPE. |
| 48 | |
| 49 | Returns a tuple |
| 50 | (q_bf16x2, tid, global_tid, sublane, token_id, head_tile_id, |
| 51 | head_start, in_bounds, num_token_heads) |
| 52 | where ``q_bf16x2`` is a (coarsen, 8) rmem tile of Uint32 packed |
| 53 | bf16x2 pairs covering the 16 BF16 lanes owned by this thread for |
| 54 | each of ``coarsen`` heads. RoPE is applied in place to the |
| 55 | trailing ``rope_dim`` lanes; the leading nope lanes pass through. |
| 56 | """ |
| 57 | block_id, _, _ = cute.arch.block_idx() |
| 58 | tid, _, _ = cute.arch.thread_idx() |
| 59 | |
| 60 | num_tokens = q.shape[0] |
| 61 | num_token_heads = num_tokens * num_heads |
| 62 | global_tid = block_id * tb_size + tid |
| 63 | |
| 64 | global_subwarp_id = global_tid // subwarp_size |
| 65 | sublane = tid % subwarp_size |
| 66 | |
| 67 | token_id = global_subwarp_id // (num_heads // coarsen) |
| 68 | head_tile_id = global_subwarp_id % (num_heads // coarsen) |
| 69 | head_start = head_tile_id * coarsen |
| 70 | |
| 71 | # NOTE: token_id may exceed bounds, hence we need to add load/store guards |
| 72 | # we can't do early exit because CuteDSL doesn't support it. and we also need |
| 73 | # all threads in a warp to be active since we utilize warp shuffle later. |
| 74 | # must_in_bounds is constexpr, True when a token spans a whole number of |
| 75 | # threadblocks. Then num_tokens * threads_per_token is always a multiple of |
| 76 | # tb_size, so the trailing block never straddles num_tokens and the compiler |
| 77 | # removes the bounds check. Otherwise the guard has to stay: e.g. 32 heads |
| 78 | # with coarsen=4 needs 64 threads per token, so an odd num_tokens leaves half |
| 79 | # of the last block addressing token num_tokens. |
| 80 | must_in_bounds = cutlass.const_expr(threads_per_token % tb_size == 0) |
| 81 | in_bounds = must_in_bounds or (token_id < num_tokens) |
| 82 | |
| 83 | cp_op = cute.nvgpu.CopyUniversalOp() |
| 84 | |
| 85 | _layout = cute.make_layout((coarsen, 8), stride=(8, 1)) |
| 86 | q_bf16x2 = cute.make_rmem_tensor(_layout, Uint32) |
| 87 | |
| 88 | if in_bounds: |
| 89 | # we can't do cute.copy() on the whole 2D tile directly because |
| 90 | # cute.copy() wants the 1st mode to be covered by the copy atom, |
| 91 | # and other modes as for loop. there is no fast way to |
| 92 | # "transpose" the tensor view. |
| 93 | q_tile = cute.local_tile( |
| 94 | q[token_id, None, None], |
| 95 | tiler=(coarsen, 16), |
| 96 | coord=(head_tile_id, sublane), |
| 97 | ) |
| 98 | cp_u32x8 = cute.make_copy_atom(cp_op, Uint32, num_bits_per_copy=256) |
| 99 | for i in cutlass.range_constexpr(coarsen): |
| 100 | src = cute.recast_tensor(q_tile[i, None], Uint32) |
| 101 | cute.copy(cp_u32x8, src, q_bf16x2[i, None]) |
| 102 | |
| 103 | # RoPE applies only to the trailing rope_dim values. We keep the rounded |
| 104 | # BF16 result in q_bits so the later amax and quantization see BF16. |
| 105 | # cos_sin_cache layout: [max_pos, rope_dim] |
| 106 | if in_bounds and sublane * 16 >= nope_dim: |
| 107 | cos_vals = cute.make_rmem_tensor((8,), Float32) |
| 108 | sin_vals = cute.make_rmem_tensor((8,), Float32) |
| 109 | |
| 110 | pos = positions[token_id] |
| 111 | |
| 112 | # select 8 elems from cos and sin |
| 113 | cos_id = sublane - nope_dim // 16 |
| 114 | sin_id = cos_id + rope_dim // 16 |
| 115 | cos_src = cute.local_tile(cos_sin_cache[pos, None], tiler=(8,), coord=(cos_id,)) |
| 116 | sin_src = cute.local_tile(cos_sin_cache[pos, None], tiler=(8,), coord=(sin_id,)) |
| 117 | |
| 118 | cp_f32x8 = cute.make_copy_atom(cp_op, Float32, num_bits_per_copy=256) |
| 119 | cp_u32x4 = cute.make_copy_atom(cp_op, Uint32, num_bits_per_copy=128) |
| 120 | |
| 121 | if const_expr(cos_sin_dtype is Float32): |
| 122 | cute.copy(cp_f32x8, cos_src, cos_vals) |
| 123 | cute.copy(cp_f32x8, sin_src, sin_vals) |
| 124 | else: |
| 125 | cos_bf16x2 = cute.make_rmem_tensor((4,), Uint32) |
| 126 | sin_bf16x2 = cute.make_rmem_tensor((4,), Uint32) |
| 127 | cute.copy(cp_u32x4, cute.recast_tensor(cos_src, Uint32), cos_bf16x2) |
| 128 | cute.copy(cp_u32x4, cute.recast_tensor(sin_src, Uint32), sin_bf16x2) |
| 129 | |
| 130 | for i in cutlass.range_constexpr(4): |
| 131 | cos0, cos1 = cvt.bf16x2_to_fp32x2(cos_bf16x2[i]) |
| 132 | sin0, sin1 = cvt.bf16x2_to_fp32x2(sin_bf16x2[i]) |
| 133 | cos_vals[i * 2] = cos0 |
| 134 | cos_vals[i * 2 + 1] = cos1 |
| 135 | sin_vals[i * 2] = sin0 |
| 136 | sin_vals[i * 2 + 1] = sin1 |
| 137 | |
| 138 | for i in cutlass.range_constexpr(coarsen): |
| 139 | for j in cutlass.range_constexpr(8): |
| 140 | q0, q1 = cvt.bf16x2_to_fp32x2(q_bf16x2[i, j]) |
| 141 | rot0 = q0 * cos_vals[j] - q1 * sin_vals[j] |
| 142 | rot1 = q0 * sin_vals[j] + q1 * cos_vals[j] |
| 143 | # convert back to BF16 to match numerics |
| 144 | q_bf16x2[i, j] = cvt.fp32x2_to_bf16x2(rot0, rot1) |
| 145 | |
| 146 | return ( |
| 147 | q_bf16x2, |
| 148 | tid, |
| 149 | global_tid, |
| 150 | sublane, |
| 151 | token_id, |
| 152 | head_tile_id, |
| 153 | head_start, |
| 154 | in_bounds, |
| 155 | num_token_heads, |
| 156 | ) |
| 157 | |
| 158 | |
| 159 | class IndexerQMxFp4Kernel(VllmCuTeDSLJitKernel["IndexerQMxFp4Kernel.CompileKey"]): |
| 160 | tb_size = 128 |
| 161 | |
| 162 | @dataclass(frozen=True) |
| 163 | class CompileKey: |
| 164 | head_dim: int |
| 165 | rope_dim: int |
| 166 | num_heads: int |
| 167 | cos_sin_dtype: type[cutlass.Numeric] |
| 168 | weights_out_dtype: type[cutlass.Numeric] |
| 169 | coarsen: int |
| 170 | |
| 171 | @staticmethod |
| 172 | def kernel(compile_key: CompileKey) -> Any: |
| 173 | head_dim = compile_key.head_dim |
| 174 | rope_dim = compile_key.rope_dim |
| 175 | nope_dim = head_dim - rope_dim |
| 176 | num_heads = compile_key.num_heads |
| 177 | cos_sin_dtype = compile_key.cos_sin_dtype |
| 178 | # process multiple heads at the same time to armotize RoPE load costs |
| 179 | coarsen = compile_key.coarsen |
| 180 | tb_size = IndexerQMxFp4Kernel.tb_size |
| 181 | weights_out_dtype = compile_key.weights_out_dtype |
| 182 | assert num_heads % coarsen == 0 |
| 183 | |
| 184 | # later we will use 32B load = 16 BF16 elems |
| 185 | # thus, head_dim=128 requires 8 threads to handle. |
| 186 | # let's call subwarp = 8 threads. |
| 187 | subwarp_size = head_dim // 16 |
| 188 | threads_per_token = (num_heads // coarsen) * subwarp_size |
| 189 | |
| 190 | @cute.kernel |
| 191 | def device_kernel( |
| 192 | positions: cute.Tensor, |
| 193 | q: cute.Tensor, |
| 194 | cos_sin_cache: cute.Tensor, |
| 195 | weights: cute.Tensor, |
| 196 | q_quant: cute.Tensor, |
| 197 | q_scale: cute.Tensor, |
| 198 | weights_out: cute.Tensor, |
| 199 | scale: Float32, |
| 200 | ): |
| 201 | ( |
| 202 | q_bf16x2, |
| 203 | tid, |
| 204 | global_tid, |
| 205 | sublane, |
| 206 | token_id, |
| 207 | head_tile_id, |
| 208 | head_start, |
| 209 | in_bounds, |
| 210 | num_token_heads, |
| 211 | ) = _load_q_and_rope( |
| 212 | positions, |
| 213 | q, |
| 214 | cos_sin_cache, |
| 215 | num_heads, |
| 216 | coarsen, |
| 217 | tb_size, |
| 218 | subwarp_size, |
| 219 | threads_per_token, |
| 220 | nope_dim, |
| 221 | rope_dim, |
| 222 | cos_sin_dtype, |
| 223 | ) |
| 224 | |
| 225 | cp_op = cute.nvgpu.CopyUniversalOp() |
| 226 | |
| 227 | # layout: [coarsen, 8] |
| 228 | q_fp4_tile = cute.local_tile( |
| 229 | q_quant[token_id, None, None], |
| 230 | tiler=(coarsen, 8), |
| 231 | coord=(head_tile_id, sublane), |
| 232 | ) |
| 233 | |
| 234 | for i in cutlass.range_constexpr(coarsen): |
| 235 | # compute amax in packed bf16x2 to save instructions |
| 236 | # Each thread holds 16 elems. Two adjacent threads form one 32-elem |
| 237 | # MXFP4 block, so a width-2 shuffle gives the block amax. |
| 238 | amax_bf16x2 = _bf16x2_abs(q_bf16x2[i, 0]) |
| 239 | for j in cutlass.range_constexpr(1, 8): |
| 240 | amax_bf16x2 = _bf16x2_max(amax_bf16x2, _bf16x2_abs(q_bf16x2[i, j])) |
| 241 | amax_bf16x2 = cute_utils.warp_reduce( |
| 242 | amax_bf16x2, |
| 243 | _bf16x2_max, |
| 244 | width=MXFP4_BLOCK_SIZE // 16, |
| 245 | ) |
| 246 | amax_pair = cvt.bf16x2_to_fp32x2(amax_bf16x2) |
| 247 | amax = cute_utils.fmax(amax_pair[0], amax_pair[1]) |
| 248 | |
| 249 | if in_bounds: |
| 250 | # compute block scale with bit manipulation |
| 251 | # UE8M0 stores ceil(log2(fp4_scale)) + 127. Adding the mantissa mask |
| 252 | # increments the exponent whenever fp4_scale is not exactly a power of 2 |
| 253 | eps = cutlass.const_expr(float.fromhex("0x6p-126")) |
| 254 | fp4_scale = cute_utils.fmax(amax, eps) * Float32(1.0 / 6.0) |
| 255 | bits = recast_val(fp4_scale, Uint32) |
| 256 | ue8m0 = cute_utils.shr_u32( |
| 257 | bits + Uint32(0x7FFFFF), Uint32(23) |
| 258 | ) & Uint32(0xFF) |
| 259 | |
| 260 | # Only one of the two threads in an MXFP4 block writes the shared scale. |
| 261 | if tid % 2 == 0: |
| 262 | mx_block = sublane // 2 |
| 263 | q_scale[token_id, head_start + i, mx_block] = Uint8(ue8m0) |
| 264 | |
| 265 | # If scale = 2^A and ue8m0 = A + 127, then inverse scale has exponent |
| 266 | # -A + 127 = 254 - ue8m0. |
| 267 | inv_scale_bits = (Uint32(254) - ue8m0) << Uint32(23) |
| 268 | inv_fp4_scale = recast_val(inv_scale_bits, Float32) |
| 269 | |
| 270 | vals = cute.make_rmem_tensor(16, Float32) |
| 271 | for j in cutlass.range_constexpr(8): |
| 272 | q0, q1 = cvt.bf16x2_to_fp32x2(q_bf16x2[i, j]) |
| 273 | vals[j * 2] = q0 * inv_fp4_scale |
| 274 | vals[j * 2 + 1] = q1 * inv_fp4_scale |
| 275 | |
| 276 | # pack to FP4 |
| 277 | packed = cute.make_rmem_tensor((2,), Uint32) |
| 278 | packed[0] = cvt.fp32x8_to_fp4x8(vals, 0) |
| 279 | packed[1] = cvt.fp32x8_to_fp4x8(vals, 8) |
| 280 | |
| 281 | dst = q_fp4_tile[i, None] |
| 282 | cp_u32x2 = cute.make_copy_atom(cp_op, Uint32, num_bits_per_copy=64) |
| 283 | cute.copy(cp_u32x2, packed, cute.recast_tensor(dst, Uint32)) |
| 284 | |
| 285 | # Weight scaling is independent of the Q subwarp work. The first |
| 286 | # num_tokens * num_heads logical threads cover one weight each. |
| 287 | if global_tid < num_token_heads: |
| 288 | weight_token_id = global_tid // num_heads |
| 289 | weight_head_id = global_tid % num_heads |
| 290 | weights_out[weight_token_id, weight_head_id] = ( |
| 291 | weights[weight_token_id, weight_head_id].to(Float32) * scale |
| 292 | ).to(weights_out_dtype) |
| 293 | |
| 294 | @cute.jit |
| 295 | def host_entrypoint( |
| 296 | positions: cute.Tensor, |
| 297 | q: cute.Tensor, |
| 298 | cos_sin_cache: cute.Tensor, |
| 299 | weights: cute.Tensor, |
| 300 | q_quant: cute.Tensor, |
| 301 | q_scale: cute.Tensor, |
| 302 | weights_out: cute.Tensor, |
| 303 | scale: Float32, |
| 304 | stream: CUstream, |
| 305 | ): |
| 306 | total_threads = q.shape[0] * threads_per_token |
| 307 | grid = (cute.ceil_div(total_threads, tb_size), 1, 1) |
| 308 | device_kernel( |
| 309 | positions, |
| 310 | q, |
| 311 | cos_sin_cache, |
| 312 | weights, |
| 313 | q_quant, |
| 314 | q_scale, |
| 315 | weights_out, |
| 316 | scale, |
| 317 | ).launch(grid=grid, block=(tb_size, 1, 1), stream=stream) |
| 318 | |
| 319 | return host_entrypoint |
| 320 | |
| 321 | def dispatch( # type: ignore[override] |
| 322 | self, |
| 323 | *, |
| 324 | cos_sin_dtype: type[cutlass.Numeric], |
| 325 | weights_out_dtype: type[cutlass.Numeric], |
| 326 | **compile_key_fields: int, |
| 327 | ) -> CompileKey: |
| 328 | return self.CompileKey( |
| 329 | **compile_key_fields, |
| 330 | cos_sin_dtype=cos_sin_dtype, |
| 331 | weights_out_dtype=weights_out_dtype, |
| 332 | ) |
| 333 | |
| 334 | def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]: |
| 335 | from vllm.v1.attention.backends.mla.indexer import dsa_indexer_uses_fp4 |
| 336 | |
| 337 | hf_config = vllm_config.model_config.hf_config |
| 338 | if hf_config is None or not dsa_indexer_uses_fp4(vllm_config): |
| 339 | return [] |
| 340 | |
| 341 | num_heads = int(getattr(hf_config, "index_n_heads", 0) or 0) |
| 342 | head_dim = int(getattr(hf_config, "index_head_dim", 0) or 0) |
| 343 | rope_dim = int(getattr(hf_config, "qk_rope_head_dim", 0) or 0) |
| 344 | if num_heads <= 0 or head_dim <= 0 or rope_dim <= 0: |
| 345 | return [] |
| 346 | |
| 347 | # The DeepSeek V4.1 sparse-logits indexer takes bf16 weights; the |
| 348 | # dense scoring kernels take fp32. |
| 349 | weights_out_dtypes: tuple[type[cutlass.Numeric], ...] = (Float32,) |
| 350 | if vllm_config.attention_config.indexer_sparse_logits: |
| 351 | weights_out_dtypes = (Float32, BFloat16) |
| 352 | return self._trace_dispatch(self.dispatch)( |
| 353 | head_dim=head_dim, |
| 354 | rope_dim=rope_dim, |
| 355 | num_heads=num_heads, |
| 356 | cos_sin_dtype=Float32, |
| 357 | weights_out_dtype=weights_out_dtypes, |
| 358 | coarsen=(1, 4), |
| 359 | ) |
| 360 | |
| 361 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 362 | num_tokens = cute.sym_int() |
| 363 | max_pos = cute.sym_int() |
| 364 | q = make_fake_tensor( |
| 365 | BFloat16, |
| 366 | (num_tokens, compile_key.num_heads, compile_key.head_dim), |
| 367 | divisibility=16, |
| 368 | ) |
| 369 | positions = make_fake_tensor(Int64, (num_tokens,), divisibility=1) |
| 370 | cos_sin_cache = make_fake_tensor( |
| 371 | compile_key.cos_sin_dtype, |
| 372 | (max_pos, compile_key.rope_dim), |
| 373 | divisibility=8, |
| 374 | ) |
| 375 | weights = make_fake_tensor( |
| 376 | BFloat16, (num_tokens, compile_key.num_heads), divisibility=8 |
| 377 | ) |
| 378 | q_packed = make_fake_tensor( |
| 379 | Uint8, |
| 380 | (num_tokens, compile_key.num_heads, compile_key.head_dim // 2), |
| 381 | divisibility=16, |
| 382 | ) |
| 383 | q_scale = make_fake_tensor( |
| 384 | Uint8, |
| 385 | ( |
| 386 | num_tokens, |
| 387 | compile_key.num_heads, |
| 388 | compile_key.head_dim // MXFP4_BLOCK_SIZE, |
| 389 | ), |
| 390 | divisibility=4, |
| 391 | ) |
| 392 | weights_out = make_fake_tensor( |
| 393 | compile_key.weights_out_dtype, |
| 394 | (num_tokens, compile_key.num_heads), |
| 395 | divisibility=4, |
| 396 | ) |
| 397 | return ( |
| 398 | positions, |
| 399 | q, |
| 400 | cos_sin_cache, |
| 401 | weights, |
| 402 | q_packed, |
| 403 | q_scale, |
| 404 | weights_out, |
| 405 | Float32(0.0), |
| 406 | ) |
| 407 | |
| 408 | @kernel_launcher |
| 409 | def __call__( |
| 410 | self, |
| 411 | *, |
| 412 | positions: torch.Tensor, |
| 413 | q: torch.Tensor, |
| 414 | cos_sin_cache: torch.Tensor, |
| 415 | weights: torch.Tensor, |
| 416 | weights_softmax_scale: float, |
| 417 | weights_head_scale: float, |
| 418 | q_packed: torch.Tensor, |
| 419 | q_scale: torch.Tensor, |
| 420 | weights_out: torch.Tensor, |
| 421 | ) -> CuTeDSLLaunchSpec["IndexerQMxFp4Kernel.CompileKey"]: |
| 422 | num_tokens, num_heads, head_dim = q.shape |
| 423 | compile_key = self.dispatch( |
| 424 | head_dim=head_dim, |
| 425 | rope_dim=cos_sin_cache.shape[-1], |
| 426 | num_heads=num_heads, |
| 427 | cos_sin_dtype=torch_to_cute_dtype(cos_sin_cache.dtype), |
| 428 | weights_out_dtype=torch_to_cute_dtype(weights_out.dtype), |
| 429 | coarsen=1 if num_tokens < 512 else 4, |
| 430 | ) |
| 431 | launch_args = ( |
| 432 | positions, |
| 433 | q, |
| 434 | cos_sin_cache, |
| 435 | weights, |
| 436 | q_packed, |
| 437 | q_scale, |
| 438 | weights_out, |
| 439 | float(weights_softmax_scale * weights_head_scale), |
| 440 | ) |
| 441 | return compile_key, launch_args |
| 442 | |
| 443 | |
| 444 | class IndexerQFp8Kernel(VllmCuTeDSLJitKernel["IndexerQFp8Kernel.CompileKey"]): |
| 445 | tb_size = 128 |
| 446 | |
| 447 | @dataclass(frozen=True) |
| 448 | class CompileKey: |
| 449 | head_dim: int |
| 450 | rope_dim: int |
| 451 | num_heads: int |
| 452 | cos_sin_dtype: type[cutlass.Numeric] |
| 453 | coarsen: int |
| 454 | |
| 455 | @staticmethod |
| 456 | def kernel(compile_key: CompileKey) -> Any: |
| 457 | head_dim = compile_key.head_dim |
| 458 | rope_dim = compile_key.rope_dim |
| 459 | nope_dim = head_dim - rope_dim |
| 460 | num_heads = compile_key.num_heads |
| 461 | cos_sin_dtype = compile_key.cos_sin_dtype |
| 462 | # process multiple heads at the same time to armotize RoPE load costs |
| 463 | coarsen = compile_key.coarsen |
| 464 | tb_size = IndexerQFp8Kernel.tb_size |
| 465 | assert num_heads % coarsen == 0 |
| 466 | |
| 467 | # later we will use 32B load = 16 BF16 elems |
| 468 | # thus, head_dim=128 requires 8 threads to handle. |
| 469 | # let's call subwarp = 8 threads. |
| 470 | subwarp_size = head_dim // 16 |
| 471 | threads_per_token = (num_heads // coarsen) * subwarp_size |
| 472 | # Each subwarp owns `coarsen` heads; we use the first `coarsen` |
| 473 | # threads of the subwarp to write the per-head weights using the |
| 474 | # fp8 scale computed in the matching loop iteration. |
| 475 | assert coarsen <= subwarp_size, ( |
| 476 | f"FP8 kernel requires coarsen ({coarsen}) <= " |
| 477 | f"subwarp_size ({subwarp_size}) for the weight-fold step" |
| 478 | ) |
| 479 | |
| 480 | @cute.kernel |
| 481 | def device_kernel( |
| 482 | positions: cute.Tensor, |
| 483 | q: cute.Tensor, |
| 484 | cos_sin_cache: cute.Tensor, |
| 485 | weights: cute.Tensor, |
| 486 | q_fp8: cute.Tensor, |
| 487 | weights_out: cute.Tensor, |
| 488 | scale: Float32, |
| 489 | ): |
| 490 | ( |
| 491 | q_bf16x2, |
| 492 | _tid, |
| 493 | _global_tid, |
| 494 | sublane, |
| 495 | token_id, |
| 496 | head_tile_id, |
| 497 | head_start, |
| 498 | in_bounds, |
| 499 | _num_token_heads, |
| 500 | ) = _load_q_and_rope( |
| 501 | positions, |
| 502 | q, |
| 503 | cos_sin_cache, |
| 504 | num_heads, |
| 505 | coarsen, |
| 506 | tb_size, |
| 507 | subwarp_size, |
| 508 | threads_per_token, |
| 509 | nope_dim, |
| 510 | rope_dim, |
| 511 | cos_sin_dtype, |
| 512 | ) |
| 513 | |
| 514 | cp_op = cute.nvgpu.CopyUniversalOp() |
| 515 | |
| 516 | # layout: [coarsen, 16] bytes (one e4m3fn per element). |
| 517 | q_fp8_tile = cute.local_tile( |
| 518 | q_fp8[token_id, None, None], |
| 519 | tiler=(coarsen, 16), |
| 520 | coord=(head_tile_id, sublane), |
| 521 | ) |
| 522 | |
| 523 | for i in cutlass.range_constexpr(coarsen): |
| 524 | # Reduce amax across the full head_dim: each thread already holds |
| 525 | # the max over its 16 lanes; a width=subwarp_size warp shuffle |
| 526 | # spreads the head-wide max to every lane in the subwarp. |
| 527 | amax_bf16x2 = _bf16x2_abs(q_bf16x2[i, 0]) |
| 528 | for j in cutlass.range_constexpr(1, 8): |
| 529 | amax_bf16x2 = _bf16x2_max(amax_bf16x2, _bf16x2_abs(q_bf16x2[i, j])) |
| 530 | amax_bf16x2 = cute_utils.warp_reduce( |
| 531 | amax_bf16x2, |
| 532 | _bf16x2_max, |
| 533 | width=subwarp_size, |
| 534 | ) |
| 535 | amax_pair = cvt.bf16x2_to_fp32x2(amax_bf16x2) |
| 536 | amax = cute_utils.fmax(amax_pair[0], amax_pair[1]) |
| 537 | |
| 538 | # scale = max(amax, eps) / fp8_max, then rounded UP to the next |
| 539 | # power of two. Adding the mantissa mask before shifting out the |
| 540 | # mantissa bumps the exponent whenever s isn't a pure pow2. |
| 541 | fp32_scale = cute_utils.fmax(amax, Float32(1e-4)) * Float32(1.0 / 448.0) |
| 542 | bits = recast_val(fp32_scale, Uint32) |
| 543 | scale_exp = cute_utils.shr_u32( |
| 544 | bits + Uint32(0x7FFFFF), Uint32(23) |
| 545 | ) & Uint32(0xFF) |
| 546 | |
| 547 | # rounded scale = 2^(scale_exp - 127); bit pattern is scale_exp << 23 |
| 548 | fp8_scale_bits = scale_exp << Uint32(23) |
| 549 | fp8_scale = recast_val(fp8_scale_bits, Float32) |
| 550 | # inverse = 2^-(scale_exp - 127); bit pattern is (254 - scale_exp) << 23 |
| 551 | inv_scale_bits = (Uint32(254) - scale_exp) << Uint32(23) |
| 552 | inv_fp8_scale = recast_val(inv_scale_bits, Float32) |
| 553 | |
| 554 | # Weight fold: weights_out = weights * q_scale * scale_combined. |
| 555 | # All threads in the subwarp share the same fp8_scale after the |
| 556 | # warp_reduce above, so we let thread `sublane == i` write the |
| 557 | # weight for head `head_start + i`. |
| 558 | if in_bounds and sublane == i: |
| 559 | head_id = head_start + i |
| 560 | weights_out[token_id, head_id] = ( |
| 561 | weights[token_id, head_id].to(Float32) * scale * fp8_scale |
| 562 | ) |
| 563 | |
| 564 | if in_bounds: |
| 565 | # 16 BF16 → 16 e4m3 bytes per thread, packed into 4 b32s |
| 566 | # (one cp.async-shaped 128-bit store per row). |
| 567 | packed = cute.make_rmem_tensor((4,), Uint32) |
| 568 | for j in cutlass.range_constexpr(4): |
| 569 | q0, q1 = cvt.bf16x2_to_fp32x2(q_bf16x2[i, j * 2]) |
| 570 | q2, q3 = cvt.bf16x2_to_fp32x2(q_bf16x2[i, j * 2 + 1]) |
| 571 | packed[j] = cvt.fp32x4_to_fp8x4( |
| 572 | q0 * inv_fp8_scale, |
| 573 | q1 * inv_fp8_scale, |
| 574 | q2 * inv_fp8_scale, |
| 575 | q3 * inv_fp8_scale, |
| 576 | ) |
| 577 | |
| 578 | dst = q_fp8_tile[i, None] |
| 579 | cp_u32x4 = cute.make_copy_atom(cp_op, Uint32, num_bits_per_copy=128) |
| 580 | cute.copy(cp_u32x4, packed, cute.recast_tensor(dst, Uint32)) |
| 581 | |
| 582 | @cute.jit |
| 583 | def host_entrypoint( |
| 584 | positions: cute.Tensor, |
| 585 | q: cute.Tensor, |
| 586 | cos_sin_cache: cute.Tensor, |
| 587 | weights: cute.Tensor, |
| 588 | q_fp8: cute.Tensor, |
| 589 | weights_out: cute.Tensor, |
| 590 | scale: Float32, |
| 591 | stream: CUstream, |
| 592 | ): |
| 593 | total_threads = q.shape[0] * threads_per_token |
| 594 | grid = (cute.ceil_div(total_threads, tb_size), 1, 1) |
| 595 | device_kernel( |
| 596 | positions, |
| 597 | q, |
| 598 | cos_sin_cache, |
| 599 | weights, |
| 600 | q_fp8, |
| 601 | weights_out, |
| 602 | scale, |
| 603 | ).launch(grid=grid, block=(tb_size, 1, 1), stream=stream) |
| 604 | |
| 605 | return host_entrypoint |
| 606 | |
| 607 | def dispatch( # type: ignore[override] |
| 608 | self, |
| 609 | *, |
| 610 | cos_sin_dtype: type[cutlass.Numeric], |
| 611 | **compile_key_fields: int, |
| 612 | ) -> CompileKey: |
| 613 | return self.CompileKey( |
| 614 | **compile_key_fields, |
| 615 | cos_sin_dtype=cos_sin_dtype, |
| 616 | ) |
| 617 | |
| 618 | def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]: |
| 619 | from vllm.v1.attention.backends.mla.indexer import dsa_indexer_uses_fp4 |
| 620 | |
| 621 | hf_config = vllm_config.model_config.hf_config |
| 622 | if hf_config is None or dsa_indexer_uses_fp4(vllm_config): |
| 623 | return [] |
| 624 | |
| 625 | num_heads = int(getattr(hf_config, "index_n_heads", 0) or 0) |
| 626 | head_dim = int(getattr(hf_config, "index_head_dim", 0) or 0) |
| 627 | rope_dim = int(getattr(hf_config, "qk_rope_head_dim", 0) or 0) |
| 628 | if num_heads <= 0 or head_dim <= 0 or rope_dim <= 0: |
| 629 | return [] |
| 630 | |
| 631 | return self._trace_dispatch(self.dispatch)( |
| 632 | head_dim=head_dim, |
| 633 | rope_dim=rope_dim, |
| 634 | num_heads=num_heads, |
| 635 | cos_sin_dtype=Float32, |
| 636 | coarsen=(1, 4), |
| 637 | ) |
| 638 | |
| 639 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 640 | num_tokens = cute.sym_int() |
| 641 | max_pos = cute.sym_int() |
| 642 | q = make_fake_tensor( |
| 643 | BFloat16, |
| 644 | (num_tokens, compile_key.num_heads, compile_key.head_dim), |
| 645 | divisibility=16, |
| 646 | ) |
| 647 | positions = make_fake_tensor(Int64, (num_tokens,), divisibility=1) |
| 648 | cos_sin_cache = make_fake_tensor( |
| 649 | compile_key.cos_sin_dtype, |
| 650 | (max_pos, compile_key.rope_dim), |
| 651 | divisibility=8, |
| 652 | ) |
| 653 | weights = make_fake_tensor( |
| 654 | BFloat16, (num_tokens, compile_key.num_heads), divisibility=8 |
| 655 | ) |
| 656 | q_fp8 = make_fake_tensor( |
| 657 | Uint8, |
| 658 | (num_tokens, compile_key.num_heads, compile_key.head_dim), |
| 659 | divisibility=16, |
| 660 | ) |
| 661 | weights_out = make_fake_tensor( |
| 662 | Float32, (num_tokens, compile_key.num_heads), divisibility=4 |
| 663 | ) |
| 664 | return ( |
| 665 | positions, |
| 666 | q, |
| 667 | cos_sin_cache, |
| 668 | weights, |
| 669 | q_fp8, |
| 670 | weights_out, |
| 671 | Float32(0.0), |
| 672 | ) |
| 673 | |
| 674 | @kernel_launcher |
| 675 | def __call__( |
| 676 | self, |
| 677 | *, |
| 678 | positions: torch.Tensor, |
| 679 | q: torch.Tensor, |
| 680 | cos_sin_cache: torch.Tensor, |
| 681 | weights: torch.Tensor, |
| 682 | weights_softmax_scale: float, |
| 683 | weights_head_scale: float, |
| 684 | q_fp8: torch.Tensor, |
| 685 | weights_out: torch.Tensor, |
| 686 | ) -> CuTeDSLLaunchSpec["IndexerQFp8Kernel.CompileKey"]: |
| 687 | num_tokens, num_heads, head_dim = q.shape |
| 688 | compile_key = self.dispatch( |
| 689 | head_dim=head_dim, |
| 690 | rope_dim=cos_sin_cache.shape[-1], |
| 691 | num_heads=num_heads, |
| 692 | cos_sin_dtype=torch_to_cute_dtype(cos_sin_cache.dtype), |
| 693 | coarsen=1 if num_tokens < 512 else 4, |
| 694 | ) |
| 695 | launch_args = ( |
| 696 | positions, |
| 697 | q, |
| 698 | cos_sin_cache, |
| 699 | weights, |
| 700 | q_fp8.view(torch.uint8), |
| 701 | weights_out, |
| 702 | float(weights_softmax_scale * weights_head_scale), |
| 703 | ) |
| 704 | return compile_key, launch_args |
| 705 | |
| 706 | |
| 707 | _INDEXER_Q_MXFP4_KERNEL = IndexerQMxFp4Kernel() |
| 708 | _INDEXER_Q_FP8_KERNEL = IndexerQFp8Kernel() |