| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from dataclasses import dataclass |
| 5 | from typing import Any |
| 6 | |
| 7 | import cutlass |
| 8 | import cutlass.cute as cute |
| 9 | import torch |
| 10 | from cuda.bindings.driver import CUstream |
| 11 | from cutlass import Float32, Int32, Uint32, Uint64 |
| 12 | from quack.compile_utils import make_fake_tensor |
| 13 | |
| 14 | from vllm.cute_utils import recast_val |
| 15 | from vllm.model_executor.warmup.jit_warmup import kernel_launcher |
| 16 | from vllm.model_executor.warmup.jit_warmup_cutedsl_helper import ( |
| 17 | CuTeDSLLaunchSpec, |
| 18 | VllmCuTeDSLJitKernel, |
| 19 | ) |
| 20 | from vllm.model_executor.warmup.jit_warmup_triton_helper import ( |
| 21 | LaunchSpec, |
| 22 | TritonWarmupTensor, |
| 23 | VllmTritonJitKernel, |
| 24 | ) |
| 25 | from vllm.triton_utils import tl, triton |
| 26 | |
| 27 | |
| 28 | def stable_topk_from_gathered_candidates_cutedsl( |
| 29 | gathered: torch.Tensor, |
| 30 | topk: int, |
| 31 | out: torch.Tensor | None = None, |
| 32 | ) -> torch.Tensor: |
| 33 | if out is None: |
| 34 | out = torch.empty( |
| 35 | (gathered.shape[0], topk), |
| 36 | dtype=torch.int32, |
| 37 | device=gathered.device, |
| 38 | ) |
| 39 | _STABLE_TOPK_FROM_GATHERED_CANDIDATES_KERNEL(gathered, out, topk=topk) |
| 40 | return out |
| 41 | |
| 42 | |
| 43 | def pack_dcp_topk_candidates_cutedsl( |
| 44 | logits: torch.Tensor, |
| 45 | topk_indices: torch.Tensor, |
| 46 | packed: torch.Tensor, |
| 47 | dcp_rank: int, |
| 48 | dcp_world_size: int, |
| 49 | cp_interleave: int, |
| 50 | row_starts: torch.Tensor | None, |
| 51 | ) -> None: |
| 52 | topk = topk_indices.shape[1] |
| 53 | row_starts_arg = row_starts if row_starts is not None else topk_indices |
| 54 | _PACK_DCP_TOPK_CANDIDATES_KERNEL( |
| 55 | logits, |
| 56 | topk_indices, |
| 57 | packed, |
| 58 | row_starts_arg, |
| 59 | logits_stride0=logits.stride(0), |
| 60 | logits_stride1=logits.stride(1), |
| 61 | topk_stride0=topk_indices.stride(0), |
| 62 | topk_stride1=topk_indices.stride(1), |
| 63 | packed_stride0=packed.stride(0), |
| 64 | packed_stride1=packed.stride(1), |
| 65 | packed_stride2=packed.stride(2), |
| 66 | num_cols=logits.shape[1], |
| 67 | dcp_rank=dcp_rank, |
| 68 | dcp_world_size=dcp_world_size, |
| 69 | cp_interleave=cp_interleave, |
| 70 | has_row_starts=row_starts is not None, |
| 71 | topk=topk, |
| 72 | block_size=512, |
| 73 | ) |
| 74 | |
| 75 | |
| 76 | class PackDCPTopkCandidatesKernel( |
| 77 | VllmTritonJitKernel["PackDCPTopkCandidatesKernel.CompileKey"] |
| 78 | ): |
| 79 | @dataclass(frozen=True) |
| 80 | class CompileKey: |
| 81 | dcp_rank: int |
| 82 | dcp_world_size: int |
| 83 | cp_interleave: int |
| 84 | has_row_starts: bool |
| 85 | topk: int |
| 86 | block_size: int |
| 87 | |
| 88 | # These scalars only describe runtime layouts and bounds; the constexpr |
| 89 | # fields below own the launch geometry and algorithmic specialization. |
| 90 | @staticmethod |
| 91 | @triton.jit( |
| 92 | do_not_specialize=[ |
| 93 | "logits_stride0", |
| 94 | "logits_stride1", |
| 95 | "topk_stride0", |
| 96 | "topk_stride1", |
| 97 | "packed_stride0", |
| 98 | "packed_stride1", |
| 99 | "packed_stride2", |
| 100 | "num_cols", |
| 101 | ] |
| 102 | ) |
| 103 | def kernel( |
| 104 | logits, |
| 105 | topk_indices, |
| 106 | packed, |
| 107 | row_starts, |
| 108 | logits_stride0, |
| 109 | logits_stride1, |
| 110 | topk_stride0, |
| 111 | topk_stride1, |
| 112 | packed_stride0, |
| 113 | packed_stride1, |
| 114 | packed_stride2, |
| 115 | num_cols, |
| 116 | DCP_RANK: tl.constexpr, |
| 117 | DCP_WORLD_SIZE: tl.constexpr, |
| 118 | CP_INTERLEAVE: tl.constexpr, |
| 119 | HAS_ROW_STARTS: tl.constexpr, |
| 120 | TOPK: tl.constexpr, |
| 121 | BLOCK_SIZE: tl.constexpr, |
| 122 | ): |
| 123 | row = tl.program_id(0) |
| 124 | tile = tl.program_id(1) |
| 125 | cols = tile * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) |
| 126 | mask = cols < TOPK |
| 127 | |
| 128 | local_idx = tl.load( |
| 129 | topk_indices + row * topk_stride0 + cols * topk_stride1, |
| 130 | mask=mask, |
| 131 | other=-1, |
| 132 | ) |
| 133 | valid = local_idx >= 0 |
| 134 | safe_local_idx = tl.maximum(local_idx, 0) |
| 135 | |
| 136 | row_start = 0 |
| 137 | if HAS_ROW_STARTS: |
| 138 | row_start = tl.load(row_starts + row) |
| 139 | |
| 140 | score_col = safe_local_idx + row_start |
| 141 | score_col = tl.minimum(score_col, tl.maximum(num_cols - 1, 0)) |
| 142 | score = tl.load( |
| 143 | logits + row * logits_stride0 + score_col * logits_stride1, |
| 144 | mask=mask & valid, |
| 145 | other=-float("inf"), |
| 146 | ) |
| 147 | |
| 148 | global_id = ( |
| 149 | (safe_local_idx // CP_INTERLEAVE) * (DCP_WORLD_SIZE * CP_INTERLEAVE) |
| 150 | + DCP_RANK * CP_INTERLEAVE |
| 151 | + safe_local_idx % CP_INTERLEAVE |
| 152 | ) |
| 153 | global_id = tl.where(valid, global_id, -1) |
| 154 | |
| 155 | packed_base = packed + row * packed_stride0 + cols * packed_stride1 |
| 156 | tl.store(packed_base, score, mask=mask) |
| 157 | tl.store(packed_base + packed_stride2, global_id.to(tl.float32), mask=mask) |
| 158 | |
| 159 | def dispatch( # type: ignore[override] |
| 160 | self, |
| 161 | *, |
| 162 | has_row_starts: bool, |
| 163 | **compile_key_fields: int, |
| 164 | ) -> CompileKey: |
| 165 | return self.CompileKey( |
| 166 | **compile_key_fields, |
| 167 | has_row_starts=has_row_starts, |
| 168 | ) |
| 169 | |
| 170 | def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]: |
| 171 | dcp_world_size = vllm_config.parallel_config.decode_context_parallel_size |
| 172 | if dcp_world_size <= 1: |
| 173 | return [] |
| 174 | cp_interleave = vllm_config.parallel_config.cp_kv_cache_interleave_size |
| 175 | topk = vllm_config.model_config.hf_text_config.index_topk |
| 176 | if topk <= 0: |
| 177 | return [] |
| 178 | |
| 179 | try: |
| 180 | from vllm.distributed.parallel_state import get_dcp_group |
| 181 | |
| 182 | dcp_rank = get_dcp_group().rank_in_group |
| 183 | except Exception: |
| 184 | dcp_rank = 0 |
| 185 | |
| 186 | return self._trace_dispatch(self.dispatch)( |
| 187 | dcp_rank=dcp_rank, |
| 188 | dcp_world_size=dcp_world_size, |
| 189 | cp_interleave=cp_interleave, |
| 190 | has_row_starts=(False, True), |
| 191 | topk=topk, |
| 192 | block_size=512, |
| 193 | ) |
| 194 | |
| 195 | def warmup_inputs(self, compile_key: CompileKey) -> dict[str, Any]: |
| 196 | fp32_ptr = TritonWarmupTensor(torch.float32) |
| 197 | int32_ptr = TritonWarmupTensor(torch.int32) |
| 198 | return dict( |
| 199 | logits=fp32_ptr, |
| 200 | topk_indices=TritonWarmupTensor(torch.int32, shape=(1, compile_key.topk)), |
| 201 | packed=fp32_ptr, |
| 202 | row_starts_arg=int32_ptr, |
| 203 | logits_stride0=1, |
| 204 | logits_stride1=1, |
| 205 | topk_stride0=1, |
| 206 | topk_stride1=1, |
| 207 | packed_stride0=1, |
| 208 | packed_stride1=1, |
| 209 | packed_stride2=1, |
| 210 | num_cols=1, |
| 211 | dcp_rank=compile_key.dcp_rank, |
| 212 | dcp_world_size=compile_key.dcp_world_size, |
| 213 | cp_interleave=compile_key.cp_interleave, |
| 214 | has_row_starts=compile_key.has_row_starts, |
| 215 | topk=compile_key.topk, |
| 216 | block_size=compile_key.block_size, |
| 217 | ) |
| 218 | |
| 219 | @kernel_launcher |
| 220 | def __call__( |
| 221 | self, |
| 222 | logits: torch.Tensor, |
| 223 | topk_indices: torch.Tensor, |
| 224 | packed: torch.Tensor, |
| 225 | row_starts_arg: torch.Tensor, |
| 226 | *, |
| 227 | logits_stride0: int, |
| 228 | logits_stride1: int, |
| 229 | topk_stride0: int, |
| 230 | topk_stride1: int, |
| 231 | packed_stride0: int, |
| 232 | packed_stride1: int, |
| 233 | packed_stride2: int, |
| 234 | num_cols: int, |
| 235 | dcp_rank: int, |
| 236 | dcp_world_size: int, |
| 237 | cp_interleave: int, |
| 238 | has_row_starts: bool, |
| 239 | topk: int, |
| 240 | block_size: int, |
| 241 | ) -> LaunchSpec: |
| 242 | grid = (topk_indices.shape[0], triton.cdiv(topk, block_size)) |
| 243 | return grid, dict( |
| 244 | row_starts=row_starts_arg, |
| 245 | DCP_RANK=dcp_rank, |
| 246 | DCP_WORLD_SIZE=dcp_world_size, |
| 247 | CP_INTERLEAVE=cp_interleave, |
| 248 | HAS_ROW_STARTS=has_row_starts, |
| 249 | TOPK=topk, |
| 250 | BLOCK_SIZE=block_size, |
| 251 | num_warps=8, |
| 252 | ) |
| 253 | |
| 254 | |
| 255 | class StableTopKFromGatheredCandidatesKernel( |
| 256 | VllmCuTeDSLJitKernel["StableTopKFromGatheredCandidatesKernel.CompileKey"] |
| 257 | ): |
| 258 | tb_size = 512 |
| 259 | hist_bins = 2048 |
| 260 | radix_bits = (hist_bins - 1).bit_length() |
| 261 | key_bits = Uint64.width |
| 262 | radix_passes = (key_bits + radix_bits - 1) // radix_bits |
| 263 | final_radix_bits = key_bits - radix_bits * (radix_passes - 1) |
| 264 | hist_chunks = (hist_bins + tb_size - 1) // tb_size |
| 265 | warps_per_block = tb_size // cute.arch.WARP_SIZE |
| 266 | |
| 267 | @dataclass(frozen=True) |
| 268 | class CompileKey: |
| 269 | topk: int |
| 270 | num_candidates: int |
| 271 | |
| 272 | @staticmethod |
| 273 | def kernel(compile_key: CompileKey) -> Any: |
| 274 | tb_size = StableTopKFromGatheredCandidatesKernel.tb_size |
| 275 | hist_bins = StableTopKFromGatheredCandidatesKernel.hist_bins |
| 276 | radix_bits = StableTopKFromGatheredCandidatesKernel.radix_bits |
| 277 | key_bits = StableTopKFromGatheredCandidatesKernel.key_bits |
| 278 | radix_passes = StableTopKFromGatheredCandidatesKernel.radix_passes |
| 279 | final_radix_bits = StableTopKFromGatheredCandidatesKernel.final_radix_bits |
| 280 | hist_chunks = StableTopKFromGatheredCandidatesKernel.hist_chunks |
| 281 | warps_per_block = StableTopKFromGatheredCandidatesKernel.warps_per_block |
| 282 | topk = compile_key.topk |
| 283 | assert hist_bins == 1 << radix_bits |
| 284 | assert compile_key.num_candidates % tb_size == 0, ( |
| 285 | "StableTopKFromGatheredCandidatesKernel requires candidate count " |
| 286 | f"to be a multiple of {tb_size}, got {compile_key.num_candidates}" |
| 287 | ) |
| 288 | keys_per_thread = compile_key.num_candidates // tb_size |
| 289 | |
| 290 | @cute.struct |
| 291 | class SharedStorage: |
| 292 | hist: cute.struct.MemRange[Int32, hist_bins] |
| 293 | committed_count: cute.struct.MemRange[Int32, 1] |
| 294 | running_count: cute.struct.MemRange[Int32, 1] |
| 295 | threshold_bin: cute.struct.MemRange[Int32, 1] |
| 296 | threshold_found: cute.struct.MemRange[Int32, 1] |
| 297 | include_threshold_bin: cute.struct.MemRange[Int32, 1] |
| 298 | prefix_s: cute.struct.Align[cute.struct.MemRange[Uint64, 1], 8] |
| 299 | warp_totals: cute.struct.MemRange[Int32, warps_per_block] |
| 300 | |
| 301 | shared_storage = SharedStorage |
| 302 | |
| 303 | @cute.jit |
| 304 | def warp_scan_inclusive_i32(val: Int32, lane: Int32) -> Int32: |
| 305 | for i in cutlass.range_constexpr(cute.arch.WARP_SIZE.bit_length() - 1): |
| 306 | offset = 1 << i |
| 307 | partial = cute.arch.shuffle_sync_up( |
| 308 | val, offset=offset, mask_and_clamp=0 |
| 309 | ) |
| 310 | if lane >= offset: |
| 311 | val += partial |
| 312 | return val |
| 313 | |
| 314 | @cute.jit |
| 315 | def block_scan_inclusive_i32( |
| 316 | val: Int32, |
| 317 | lane: Int32, |
| 318 | warp_id: Int32, |
| 319 | warp_scratch: cute.Tensor, |
| 320 | warps_per_block: int, |
| 321 | ) -> Int32: |
| 322 | prefix = warp_scan_inclusive_i32(val, lane) |
| 323 | if lane == Int32(cute.arch.WARP_SIZE - 1): |
| 324 | warp_scratch[0, warp_id] = prefix |
| 325 | cute.arch.sync_threads() |
| 326 | |
| 327 | if warp_id == Int32(0): |
| 328 | warp_total = Int32(0) |
| 329 | if lane < Int32(warps_per_block): |
| 330 | warp_total = warp_scratch[0, lane] |
| 331 | warp_prefix = warp_scan_inclusive_i32(warp_total, lane) |
| 332 | if lane < Int32(warps_per_block): |
| 333 | warp_scratch[0, lane] = warp_prefix - warp_total |
| 334 | cute.arch.sync_threads() |
| 335 | |
| 336 | return prefix + warp_scratch[0, warp_id] |
| 337 | |
| 338 | @cute.jit |
| 339 | def stable_key(score: Float32, token_id: Int32) -> Uint64: |
| 340 | bits = recast_val(score, Uint32) |
| 341 | mask = Uint32(0x80000000) |
| 342 | if (bits & Uint32(0x80000000)) != Uint32(0): |
| 343 | mask = Uint32(0xFFFFFFFF) |
| 344 | score_key = Uint64(bits ^ mask) << Uint64(32) |
| 345 | id_key = Uint64(~Uint32(token_id)) |
| 346 | key = score_key | id_key |
| 347 | if token_id < Int32(0): |
| 348 | key = Uint64(0) |
| 349 | return key |
| 350 | |
| 351 | @cute.jit |
| 352 | def prefix_matches( |
| 353 | key: Uint64, |
| 354 | prefix: Uint64, |
| 355 | prefix_bits: Int32, |
| 356 | ): |
| 357 | matches = prefix_bits == Int32(0) |
| 358 | if prefix_bits != Int32(0): |
| 359 | shift = Int32(key_bits) - prefix_bits |
| 360 | matches = (key >> Uint64(shift)) == (prefix >> Uint64(shift)) |
| 361 | return matches |
| 362 | |
| 363 | @cute.jit |
| 364 | def radix_pass( |
| 365 | keys: cute.Tensor, |
| 366 | output: cute.Tensor, |
| 367 | storage, |
| 368 | tid: Int32, |
| 369 | step: Int32, |
| 370 | bits: int, |
| 371 | is_final_pass: bool, |
| 372 | ): |
| 373 | hist_smem = storage.hist.get_tensor(cute.make_layout((hist_bins,))) |
| 374 | committed_count_smem = storage.committed_count.data_ptr() |
| 375 | running_count_smem = storage.running_count.data_ptr() |
| 376 | threshold_bin_smem = storage.threshold_bin.data_ptr() |
| 377 | threshold_found_smem = storage.threshold_found.data_ptr() |
| 378 | include_threshold_bin_smem = storage.include_threshold_bin.data_ptr() |
| 379 | prefix_smem = storage.prefix_s.data_ptr() |
| 380 | warp_totals_smem = storage.warp_totals.get_tensor( |
| 381 | cute.make_layout((1, warps_per_block)) |
| 382 | ) |
| 383 | |
| 384 | prefix_bits = step * Int32(radix_bits) |
| 385 | num_bins = 1 << bits |
| 386 | block_scan_iterations = (num_bins + tb_size - 1) // tb_size |
| 387 | shift = Int32(key_bits) - prefix_bits - Int32(bits) |
| 388 | bin_mask = Uint64(num_bins - 1) |
| 389 | prefix = prefix_smem.load() |
| 390 | |
| 391 | for chunk in cutlass.range_constexpr(hist_chunks): |
| 392 | hist_smem[tid + Int32(chunk * tb_size)] = Int32(0) |
| 393 | if tid == Int32(0): |
| 394 | running_count_smem.store(committed_count_smem.load()) |
| 395 | include_threshold_bin_smem.store(Int32(0)) |
| 396 | threshold_found_smem.store(Int32(0)) |
| 397 | cute.arch.sync_threads() |
| 398 | |
| 399 | for key_idx in cutlass.range_constexpr(keys_per_thread): |
| 400 | key = keys[key_idx] |
| 401 | if prefix_matches(key, prefix, prefix_bits): |
| 402 | bin_idx = Int32((key >> Uint64(shift)) & bin_mask) |
| 403 | cute.arch.atomic_add( |
| 404 | hist_smem.iterator + bin_idx, |
| 405 | Int32(1), |
| 406 | sem="relaxed", |
| 407 | scope="cta", |
| 408 | ) |
| 409 | cute.arch.sync_threads() |
| 410 | |
| 411 | lane = cute.arch.lane_idx() |
| 412 | warp_id = cute.arch.warp_idx() |
| 413 | # Each iteration scans one tb_size-wide slice of bins, high to low. |
| 414 | iter = Int32(0) |
| 415 | threshold_found = threshold_found_smem.load() |
| 416 | while threshold_found == Int32(0) and iter < Int32(block_scan_iterations): |
| 417 | bin_idx = Int32(num_bins - 1) - (iter * Int32(tb_size) + tid) |
| 418 | count = hist_smem[bin_idx] |
| 419 | chunk_inclusive = block_scan_inclusive_i32( |
| 420 | count, |
| 421 | lane, |
| 422 | warp_id, |
| 423 | warp_totals_smem, |
| 424 | warps_per_block, |
| 425 | ) |
| 426 | running_count = running_count_smem.load() |
| 427 | prior_in_scan_slice = chunk_inclusive - count |
| 428 | remaining = Int32(topk) - running_count - prior_in_scan_slice |
| 429 | if count > Int32(0) and remaining > Int32(0) and remaining <= count: |
| 430 | threshold_bin_smem.store(bin_idx) |
| 431 | if count <= remaining or cutlass.const_expr(is_final_pass): |
| 432 | include_threshold_bin_smem.store(Int32(1)) |
| 433 | threshold_found_smem.store(Int32(1)) |
| 434 | # Barrier: every thread must finish reading running_count for this |
| 435 | # slice before tb_size-1 advances it, else a warp racing ahead to |
| 436 | # the store makes a lagging thread double-count the slice total |
| 437 | # (-> remaining too small -> threshold too high -> under-fill). |
| 438 | cute.arch.sync_threads() |
| 439 | if tid == Int32(tb_size - 1): |
| 440 | running_count_smem.store(running_count + chunk_inclusive) |
| 441 | cute.arch.sync_threads() |
| 442 | |
| 443 | threshold_found = threshold_found_smem.load() |
| 444 | iter += Int32(1) |
| 445 | |
| 446 | threshold = threshold_bin_smem.load() |
| 447 | should_include_threshold = include_threshold_bin_smem.load() != Int32(0) |
| 448 | for key_idx in cutlass.range_constexpr(keys_per_thread): |
| 449 | key = keys[key_idx] |
| 450 | if prefix_matches(key, prefix, prefix_bits): |
| 451 | bin_idx = Int32((key >> Uint64(shift)) & bin_mask) |
| 452 | selected = bin_idx > threshold |
| 453 | if should_include_threshold: |
| 454 | selected = selected or bin_idx == threshold |
| 455 | if selected: |
| 456 | dst = cute.arch.atomic_add( |
| 457 | committed_count_smem, |
| 458 | Int32(1), |
| 459 | sem="relaxed", |
| 460 | scope="cta", |
| 461 | ) |
| 462 | if dst < Int32(topk): |
| 463 | output[dst] = recast_val(~Uint32(key), Int32) |
| 464 | cute.arch.sync_threads() |
| 465 | |
| 466 | pass_finished = include_threshold_bin_smem.load() |
| 467 | if tid == Int32(0) and pass_finished == Int32(0): |
| 468 | prefix_smem.store(prefix | (Uint64(threshold) << Uint64(shift))) |
| 469 | cute.arch.sync_threads() |
| 470 | return pass_finished |
| 471 | |
| 472 | @cute.kernel |
| 473 | def device_kernel(input: cute.Tensor, out: cute.Tensor): |
| 474 | row, _, _ = cute.arch.block_idx() |
| 475 | tid, _, _ = cute.arch.thread_idx() |
| 476 | input_row = input[row, None, None] |
| 477 | output_row = out[row, None] |
| 478 | keys = cute.make_rmem_tensor((keys_per_thread,), Uint64) |
| 479 | |
| 480 | smem = cutlass.utils.SmemAllocator() |
| 481 | storage = smem.allocate(shared_storage, 8) |
| 482 | committed_count_smem = storage.committed_count.data_ptr() |
| 483 | prefix_smem = storage.prefix_s.data_ptr() |
| 484 | for i in range(tid, topk, tb_size): |
| 485 | output_row[i] = Int32(-1) |
| 486 | |
| 487 | for key_idx in cutlass.range_constexpr(keys_per_thread): |
| 488 | col = tid + Int32(key_idx * tb_size) |
| 489 | score = Float32(input_row[col, 0]) |
| 490 | token_id = Int32(input_row[col, 1]) |
| 491 | keys[key_idx] = stable_key(score, token_id) |
| 492 | |
| 493 | if tid == Int32(0): |
| 494 | committed_count_smem.store(Int32(0)) |
| 495 | prefix_smem.store(Uint64(0)) |
| 496 | cute.arch.sync_threads() |
| 497 | |
| 498 | step = Int32(0) |
| 499 | finished = Int32(0) |
| 500 | while finished == Int32(0) and step < Int32(radix_passes - 1): |
| 501 | finished = radix_pass( |
| 502 | keys, |
| 503 | output_row, |
| 504 | storage, |
| 505 | tid, |
| 506 | step, |
| 507 | radix_bits, |
| 508 | False, |
| 509 | ) |
| 510 | step += Int32(1) |
| 511 | |
| 512 | if finished == Int32(0): |
| 513 | radix_pass( |
| 514 | keys, |
| 515 | output_row, |
| 516 | storage, |
| 517 | tid, |
| 518 | Int32(radix_passes - 1), |
| 519 | final_radix_bits, |
| 520 | True, |
| 521 | ) |
| 522 | |
| 523 | @cute.jit |
| 524 | def host_entrypoint( |
| 525 | gathered: cute.Tensor, |
| 526 | out: cute.Tensor, |
| 527 | stream: CUstream, |
| 528 | ): |
| 529 | grid = (gathered.shape[0], 1, 1) |
| 530 | device_kernel(gathered, out).launch( |
| 531 | grid=grid, |
| 532 | block=(tb_size, 1, 1), |
| 533 | stream=stream, |
| 534 | ) |
| 535 | |
| 536 | return host_entrypoint |
| 537 | |
| 538 | def dispatch( # type: ignore[override] |
| 539 | self, |
| 540 | **compile_key_fields: int, |
| 541 | ) -> CompileKey: |
| 542 | return self.CompileKey(**compile_key_fields) |
| 543 | |
| 544 | def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]: |
| 545 | dcp_world_size = vllm_config.parallel_config.decode_context_parallel_size |
| 546 | if dcp_world_size <= 1: |
| 547 | return [] |
| 548 | topk = vllm_config.model_config.hf_text_config.index_topk |
| 549 | if topk <= 0: |
| 550 | return [] |
| 551 | return self._trace_dispatch(self.dispatch)( |
| 552 | topk=topk, |
| 553 | num_candidates=topk * dcp_world_size, |
| 554 | ) |
| 555 | |
| 556 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 557 | num_rows = cute.sym_int() |
| 558 | gathered = cute.runtime.make_fake_tensor( |
| 559 | Float32, |
| 560 | (num_rows, compile_key.num_candidates, 2), |
| 561 | stride=(cute.sym_int64(divisibility=2), 2, 1), |
| 562 | assumed_align=8, |
| 563 | ) |
| 564 | out = make_fake_tensor( |
| 565 | Int32, |
| 566 | (num_rows, compile_key.topk), |
| 567 | divisibility=1, |
| 568 | ) |
| 569 | return gathered, out |
| 570 | |
| 571 | @kernel_launcher |
| 572 | def __call__( |
| 573 | self, |
| 574 | gathered: torch.Tensor, |
| 575 | out: torch.Tensor, |
| 576 | *, |
| 577 | topk: int, |
| 578 | ) -> CuTeDSLLaunchSpec[CompileKey]: |
| 579 | compile_key = self.dispatch( |
| 580 | topk=topk, |
| 581 | num_candidates=gathered.shape[1], |
| 582 | ) |
| 583 | return compile_key, (gathered, out) |
| 584 | |
| 585 | |
| 586 | _PACK_DCP_TOPK_CANDIDATES_KERNEL = PackDCPTopkCandidatesKernel() |
| 587 | _STABLE_TOPK_FROM_GATHERED_CANDIDATES_KERNEL = StableTopKFromGatheredCandidatesKernel() |