| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Warmup kernels used during model execution. |
| 4 | This is useful specifically for JIT'ed kernels as we don't want JIT'ing to |
| 5 | happen during model execution. |
| 6 | """ |
| 7 | |
| 8 | import sys |
| 9 | import time |
| 10 | from typing import TYPE_CHECKING |
| 11 | |
| 12 | import torch |
| 13 | |
| 14 | import vllm.envs as envs |
| 15 | from vllm.logger import init_logger |
| 16 | from vllm.model_executor.warmup.b12x_warmup import b12x_warmup |
| 17 | from vllm.model_executor.warmup.cutedsl_warmup import cutedsl_warmup |
| 18 | from vllm.model_executor.warmup.deep_gemm_warmup import deep_gemm_warmup |
| 19 | from vllm.model_executor.warmup.flashinfer_autotune_cache import ( |
| 20 | resolve_flashinfer_autotune_file, |
| 21 | write_flashinfer_autotune_cache, |
| 22 | ) |
| 23 | from vllm.model_executor.warmup.flashinfer_sparse_mla_warmup import ( |
| 24 | autotune_hisparse_flashinfer_attention, |
| 25 | deepseek_v4_sparse_mla_attention_warmup, |
| 26 | flashinfer_sparse_mla_decode_autotune_warmup, |
| 27 | ) |
| 28 | from vllm.model_executor.warmup.kimi_k3_triton_warmup import ( |
| 29 | kimi_k3_triton_warmup, |
| 30 | ) |
| 31 | from vllm.model_executor.warmup.mamba_triton_warmup import mamba_triton_warmup |
| 32 | from vllm.model_executor.warmup.qwen4_exp_qsa_warmup import ( |
| 33 | qwen4_exp_qsa_triton_warmup, |
| 34 | ) |
| 35 | from vllm.model_executor.warmup.qwen_triton_warmup import qwen_triton_warmup |
| 36 | from vllm.model_executor.warmup.qwen_vl_triton_warmup import qwen_vl_triton_warmup |
| 37 | from vllm.model_executor.warmup.replayssm_warmup import ( |
| 38 | replayssm_autotune_warmup, |
| 39 | ) |
| 40 | from vllm.model_executor.warmup.watermark_sample_warmup import ( |
| 41 | watermark_sample_warmup, |
| 42 | ) |
| 43 | from vllm.platforms import current_platform |
| 44 | from vllm.utils.deep_gemm import is_deep_gemm_supported |
| 45 | from vllm.utils.flashinfer import has_flashinfer |
| 46 | |
| 47 | if TYPE_CHECKING: |
| 48 | from vllm.v1.worker.gpu_model_runner import GPUModelRunner |
| 49 | from vllm.v1.worker.gpu_worker import Worker |
| 50 | |
| 51 | logger = init_logger(__name__) |
| 52 | |
| 53 | _LL_BF16_WARMUP_M_RANGE = range(1, 17) |
| 54 | |
| 55 | |
| 56 | def _ll_bf16_router_shapes_from_model( |
| 57 | model: torch.nn.Module, |
| 58 | ) -> tuple[tuple[int, int], ...]: |
| 59 | from vllm.model_executor.layers.fused_moe.router.gate_linear import GateLinear |
| 60 | |
| 61 | shapes: set[tuple[int, int]] = set() |
| 62 | for module in model.modules(): |
| 63 | if not isinstance(module, GateLinear): |
| 64 | continue |
| 65 | weight = getattr(module, "weight", None) |
| 66 | if not isinstance(weight, torch.Tensor): |
| 67 | continue |
| 68 | if weight.dim() != 2 or weight.dtype != torch.bfloat16: |
| 69 | continue |
| 70 | n, k = weight.shape |
| 71 | if k % 8 == 0: |
| 72 | shapes.add((int(k), int(n))) |
| 73 | return tuple(sorted(shapes)) |
| 74 | |
| 75 | |
| 76 | def _warmup_ll_bf16_router_gemm(model: torch.nn.Module) -> None: |
| 77 | from vllm.model_executor.kernels.linear.cute_dsl.ll_bf16 import ( |
| 78 | is_available as is_ll_bf16_gemm_available, |
| 79 | ) |
| 80 | from vllm.model_executor.kernels.linear.cute_dsl.ll_bf16 import ( |
| 81 | ll_bf16_gemm_kernel, |
| 82 | ) |
| 83 | |
| 84 | if not is_ll_bf16_gemm_available(): |
| 85 | return |
| 86 | |
| 87 | shapes = _ll_bf16_router_shapes_from_model(model) |
| 88 | if not shapes: |
| 89 | logger.debug_once( |
| 90 | "Skipping ll_bf16 router GEMM warmup: no bf16 GateLinear shapes found." |
| 91 | ) |
| 92 | return |
| 93 | |
| 94 | logger.info_once("Warming up ll_bf16 router GEMM kernels for shapes: %s.", shapes) |
| 95 | ll_bf16_gemm_kernel.warmup( |
| 96 | shapes=shapes, |
| 97 | m_values=_LL_BF16_WARMUP_M_RANGE, |
| 98 | ) |
| 99 | |
| 100 | |
| 101 | def _warmup_bf16x3_router_gemm( |
| 102 | model: torch.nn.Module, |
| 103 | max_num_tokens: int, |
| 104 | ) -> None: |
| 105 | from vllm.model_executor.layers.fused_moe.router.bf16x3_router_gemm_cutedsl import ( # noqa: E501 |
| 106 | warmup_bf16x3_router_gemm, |
| 107 | ) |
| 108 | from vllm.model_executor.layers.fused_moe.router.gate_linear import GateLinear |
| 109 | |
| 110 | gate = next( |
| 111 | ( |
| 112 | module |
| 113 | for module in model.modules() |
| 114 | if isinstance(module, GateLinear) and module.allow_bf16x3_router_gemm |
| 115 | ), |
| 116 | None, |
| 117 | ) |
| 118 | if gate is None: |
| 119 | logger.debug_once( |
| 120 | "Skipping BF16x3 router GEMM warmup: no eligible GateLinear found." |
| 121 | ) |
| 122 | return |
| 123 | |
| 124 | min_num_tokens = gate.FP32_MAX_TOKENS + 1 if gate.allow_fp32_router_gemm else 1 |
| 125 | logger.info_once( |
| 126 | "Warming up BF16x3 router GEMM for K=%d, M=%d.", |
| 127 | gate.input_size, |
| 128 | gate.output_size, |
| 129 | ) |
| 130 | configs = warmup_bf16x3_router_gemm( |
| 131 | gate.input_size, |
| 132 | gate.output_size, |
| 133 | min_num_tokens, |
| 134 | max_num_tokens, |
| 135 | ) |
| 136 | logger.info_once("Warmed up BF16x3 router GEMM configs: %s.", configs) |
| 137 | |
| 138 | |
| 139 | def _warmup_gemm_rs_ar() -> None: |
| 140 | # Model construction (Kimi-K3, DeepSeek-V4.1) imports this module only |
| 141 | # when GEMM-RS/AR is enabled and initializes its singleton before |
| 142 | # kernel_warmup runs. Avoid importing it here so other models do not |
| 143 | # compile the RS/AR variants. |
| 144 | module = sys.modules.get("vllm.model_executor.kernels.linear.cute_dsl.gemm_rs_ar") |
| 145 | if module is None: |
| 146 | return |
| 147 | compiled = module.warmup_gemm_rs_ar() |
| 148 | if compiled: |
| 149 | logger.info_once("Warmed up %d GEMM-RS/AR variants.", compiled) |
| 150 | |
| 151 | |
| 152 | def _autotune_kimi_k3_kda_qkvg(model: torch.nn.Module) -> None: |
| 153 | module = sys.modules.get("vllm.models.kimi_k3.nvidia.low_latency_gemm") |
| 154 | if module is not None: |
| 155 | module.autotune_kda_qkvg(model) |
| 156 | |
| 157 | |
| 158 | def kernel_warmup(worker: "Worker", *, process_local_only: bool = False): |
| 159 | from vllm.model_executor.warmup.minimax_m3_msa_warmup import ( |
| 160 | minimax_m3_msa_warmup, |
| 161 | ) |
| 162 | |
| 163 | if not worker.use_v2_model_runner: |
| 164 | # The KV-block zeroing kernel is driven by the scheduler's |
| 165 | # `new_block_ids_to_zero`, so no dummy run ever reaches it. |
| 166 | zeroer = getattr(worker.model_runner, "_kv_block_zeroer", None) |
| 167 | if zeroer is not None: |
| 168 | zeroer.warmup(worker.model_runner.kv_cache_config.num_blocks) |
| 169 | |
| 170 | enable_jit_warmup = worker.vllm_config.kernel_config.enable_jit_warmup |
| 171 | if enable_jit_warmup: |
| 172 | logger.info("JIT kernel warmup starting.") |
| 173 | jit_warmup_start = time.perf_counter() |
| 174 | try: |
| 175 | registry = ( |
| 176 | worker.model_runner.jit_warmup_registry # type: ignore[attr-defined] |
| 177 | ) |
| 178 | registry.warmup() |
| 179 | except Exception: |
| 180 | logger.exception( |
| 181 | "JIT kernel warmup failed after %.2fs.", |
| 182 | time.perf_counter() - jit_warmup_start, |
| 183 | ) |
| 184 | raise |
| 185 | logger.info( |
| 186 | "JIT kernel warmup finished in %.2fs.", |
| 187 | time.perf_counter() - jit_warmup_start, |
| 188 | ) |
| 189 | |
| 190 | qwen_triton_warmup(worker.model_runner, worker.vllm_config.model_config) |
| 191 | qwen_vl_triton_warmup(worker.model_runner) |
| 192 | mamba_triton_warmup(worker.model_runner) |
| 193 | |
| 194 | compilation_config = worker.vllm_config.compilation_config |
| 195 | cudagraph_capture_sizes = list(compilation_config.cudagraph_capture_sizes or []) |
| 196 | |
| 197 | # Run next so input-prep kernels JIT against pristine runner state. |
| 198 | if enable_jit_warmup: |
| 199 | kimi_k3_triton_warmup(worker) |
| 200 | watermark_sample_warmup(worker) |
| 201 | qwen4_exp_qsa_triton_warmup(worker) |
| 202 | |
| 203 | if enable_jit_warmup and current_platform.is_device_capability_family(100): |
| 204 | _warmup_bf16x3_router_gemm( |
| 205 | worker.get_model(), |
| 206 | worker.scheduler_config.max_num_batched_tokens, |
| 207 | ) |
| 208 | |
| 209 | if current_platform.has_device_capability(90): |
| 210 | _warmup_ll_bf16_router_gemm(worker.get_model()) |
| 211 | |
| 212 | _warmup_gemm_rs_ar() |
| 213 | |
| 214 | if worker.vllm_config.kernel_config.enable_cutedsl_warmup: |
| 215 | # TODO(roberto): Remove after registered CuTeDSL warmups are migrated |
| 216 | # to the shared JIT warmup infrastructure. |
| 217 | # https://github.com/vllm-project/vllm/pull/47451 |
| 218 | cutedsl_warmup() |
| 219 | |
| 220 | if process_local_only: |
| 221 | return |
| 222 | |
| 223 | flashinfer_sparse_mla_decode_autotune_warmup(worker) |
| 224 | deepseek_v4_sparse_mla_attention_warmup(worker) |
| 225 | |
| 226 | # Deep GEMM warmup |
| 227 | do_deep_gemm_warmup = ( |
| 228 | is_deep_gemm_supported() and envs.VLLM_DEEP_GEMM_WARMUP != "skip" |
| 229 | ) |
| 230 | if do_deep_gemm_warmup: |
| 231 | model = worker.get_model() |
| 232 | max_tokens = worker.scheduler_config.max_num_batched_tokens |
| 233 | deep_gemm_warmup(model, max_tokens) |
| 234 | |
| 235 | b12x_warmup(worker, cudagraph_capture_sizes) |
| 236 | |
| 237 | minimax_m3_msa_warmup(worker) |
| 238 | |
| 239 | # Allocate the exact decode-sized workspace, autotune cache misses, and |
| 240 | # resolve every CUDA Graph bucket before capture begins. |
| 241 | # Lazy import: flashinfer_pcie_ipc_all_reduce imports flashinfer.comm, |
| 242 | # which initializes CUDA at import time and must not run at module scope. |
| 243 | if envs.VLLM_ALLREDUCE_USE_FLASHINFER_PCIE_IPC: |
| 244 | from vllm.distributed.device_communicators import ( |
| 245 | flashinfer_pcie_ipc_all_reduce, |
| 246 | ) |
| 247 | |
| 248 | flashinfer_pcie_ipc_all_reduce.warmup_flashinfer_pcie_ipc_allreduce(worker) |
| 249 | |
| 250 | enable_flashinfer_autotune = ( |
| 251 | worker.vllm_config.kernel_config.enable_flashinfer_autotune |
| 252 | ) |
| 253 | # FlashInfer autotune for Hopper (SM 9.0) and Blackwell (SM 10.0) GPUs |
| 254 | if enable_flashinfer_autotune is False: |
| 255 | logger.info_once("Skipping FlashInfer autotune because it is disabled.") |
| 256 | elif has_flashinfer() and current_platform.has_device_capability(90): |
| 257 | flashinfer_autotune(worker.model_runner) |
| 258 | |
| 259 | # FlashInfer attention warmup |
| 260 | # Only warmup if the model has FlashInfer attention groups |
| 261 | # and is not a pooling model |
| 262 | def _is_flashinfer_backend(backend): |
| 263 | try: |
| 264 | return backend.get_name() == "FLASHINFER" |
| 265 | except NotImplementedError: |
| 266 | return False |
| 267 | |
| 268 | if ( |
| 269 | not worker.model_runner.is_pooling_model |
| 270 | and worker.model_runner.attn_groups |
| 271 | # NOTE: This should be `any` instead of `all` but other hybrid attention |
| 272 | # backends don't support this dummy run. Once we remove |
| 273 | # `build_for_cudagraph_capture`, we can change it to `any`. |
| 274 | and all( |
| 275 | _is_flashinfer_backend(group.backend) |
| 276 | for groups in worker.model_runner.attn_groups |
| 277 | for group in groups |
| 278 | ) |
| 279 | ): |
| 280 | logger.info_once("Warming up FlashInfer attention.") |
| 281 | # Warmup with mixed batch containing both prefill and decode tokens |
| 282 | # This is to warm up both prefill and decode attention kernels |
| 283 | worker.model_runner._dummy_run( |
| 284 | num_tokens=16, |
| 285 | skip_eplb=True, |
| 286 | is_profile=True, |
| 287 | force_attention=True, |
| 288 | create_mixed_batch=True, |
| 289 | ) |
| 290 | |
| 291 | |
| 292 | def _flashinfer_autotune_skip_ops(runner: "GPUModelRunner") -> set[str] | None: |
| 293 | if envs.VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS is not None: |
| 294 | return set(envs.VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS) or None |
| 295 | |
| 296 | from vllm.model_executor.kernels.linear import ( |
| 297 | FlashInferCuteDslNvFp4LinearKernel, |
| 298 | ) |
| 299 | |
| 300 | for module in runner.get_model().modules(): |
| 301 | for holder_name in ("quant_method", "scheme"): |
| 302 | kernel = getattr(getattr(module, holder_name, None), "kernel", None) |
| 303 | # CuTe-DSL mm_fp4 tuning JIT-compiles every tactic and its |
| 304 | # fallback is already the heuristic; all mm_fp4 backends share |
| 305 | # the "fp4_gemm" op name, so skip only when cute-dsl is selected. |
| 306 | if isinstance(kernel, FlashInferCuteDslNvFp4LinearKernel): |
| 307 | return {"fp4_gemm"} |
| 308 | return None |
| 309 | |
| 310 | |
| 311 | _FLASHINFER_BF16_AUTOTUNE_MAX_TOKENS = 32 |
| 312 | |
| 313 | |
| 314 | def _flashinfer_deferred_moe_token_counts( |
| 315 | runner: "GPUModelRunner", |
| 316 | ) -> tuple[int, ...]: |
| 317 | """Return bounded token counts that exercise deferred MoE finalization.""" |
| 318 | from vllm.model_executor.layers.fused_moe import MoERunner |
| 319 | |
| 320 | max_tokens = runner.scheduler_config.max_num_batched_tokens |
| 321 | token_counts: list[int] = [] |
| 322 | for module in runner.get_model().modules(): |
| 323 | if not isinstance(module, MoERunner): |
| 324 | continue |
| 325 | |
| 326 | moe_config = module.moe_config |
| 327 | max_deferred_tokens = moe_config.defer_moe_finalize_max_num_tokens |
| 328 | if moe_config.use_deferred_moe_finalize and max_deferred_tokens > 0: |
| 329 | token_counts.append(min(max_tokens, max_deferred_tokens)) |
| 330 | |
| 331 | return tuple(dict.fromkeys(token_counts)) |
| 332 | |
| 333 | |
| 334 | def _flashinfer_autotune_token_counts( |
| 335 | runner: "GPUModelRunner", *, include_bf16: bool = True |
| 336 | ) -> tuple[int, ...]: |
| 337 | max_tokens = runner.scheduler_config.max_num_batched_tokens |
| 338 | # Tune the widest bucket set first so bounded passes reuse its configs. |
| 339 | token_counts = [max_tokens] |
| 340 | linear_backend = runner.vllm_config.kernel_config.linear_backend |
| 341 | if ( |
| 342 | include_bf16 |
| 343 | and linear_backend == "flashinfer_cutedsl" |
| 344 | and max_tokens > _FLASHINFER_BF16_AUTOTUNE_MAX_TOKENS |
| 345 | ): |
| 346 | token_counts.append(_FLASHINFER_BF16_AUTOTUNE_MAX_TOKENS) |
| 347 | token_counts.extend(_flashinfer_deferred_moe_token_counts(runner)) |
| 348 | return tuple(dict.fromkeys(token_counts)) |
| 349 | |
| 350 | |
| 351 | def _run_flashinfer_autotune_dummy_runs( |
| 352 | runner: "GPUModelRunner", *, skip_attn: bool = False |
| 353 | ) -> None: |
| 354 | import vllm.utils.flashinfer as fi_utils |
| 355 | |
| 356 | dummy_run_kwargs = {"skip_attn": True} if skip_attn else {} |
| 357 | for num_tokens in _flashinfer_autotune_token_counts(runner, include_bf16=False): |
| 358 | tuning_buckets = fi_utils.flashinfer_get_hybrid_num_tokens_buckets(num_tokens) |
| 359 | logger.info( |
| 360 | "Running FlashInfer autotune with %d tokens and token buckets %s.", |
| 361 | num_tokens, |
| 362 | tuning_buckets, |
| 363 | ) |
| 364 | with fi_utils.autotune(tuning_buckets=tuning_buckets): |
| 365 | runner._dummy_run( |
| 366 | num_tokens=num_tokens, |
| 367 | skip_eplb=True, |
| 368 | is_profile=True, |
| 369 | randomize_inputs=True, |
| 370 | **dummy_run_kwargs, |
| 371 | ) |
| 372 | |
| 373 | |
| 374 | def _run_flashinfer_bf16_autotune_dummy_run( |
| 375 | runner: "GPUModelRunner", |
| 376 | *, |
| 377 | skip_ops: set[str] | None = None, |
| 378 | skip_attn: bool = False, |
| 379 | ) -> None: |
| 380 | import vllm.utils.flashinfer as fi_utils |
| 381 | |
| 382 | if ( |
| 383 | runner.vllm_config.kernel_config.linear_backend != "flashinfer_cutedsl" |
| 384 | or runner.scheduler_config.max_num_batched_tokens |
| 385 | <= _FLASHINFER_BF16_AUTOTUNE_MAX_TOKENS |
| 386 | ): |
| 387 | return |
| 388 | |
| 389 | num_tokens = _FLASHINFER_BF16_AUTOTUNE_MAX_TOKENS |
| 390 | tuning_buckets = fi_utils.flashinfer_get_hybrid_num_tokens_buckets(num_tokens) |
| 391 | logger.info("Running FlashInfer BF16-only autotune with %d tokens.", num_tokens) |
| 392 | # Other ops execute from their existing cache or fallback, without the |
| 393 | # BF16 bucket cap. No full-model autotune context may enclose this pass. |
| 394 | with fi_utils.autotune_bf16_only(tuning_buckets, skip_ops=skip_ops): |
| 395 | runner._dummy_run( |
| 396 | num_tokens=num_tokens, |
| 397 | skip_eplb=True, |
| 398 | is_profile=True, |
| 399 | randomize_inputs=True, |
| 400 | **({"skip_attn": True} if skip_attn else {}), |
| 401 | ) |
| 402 | |
| 403 | |
| 404 | def flashinfer_autotune(runner: "GPUModelRunner") -> None: |
| 405 | """Autotune FlashInfer operations. |
| 406 | FlashInfer have many implementations for the same operation, |
| 407 | autotuning runs benchmarks for each implementation and stores |
| 408 | the results. The results are cached transparently and |
| 409 | future calls to FlashInfer will use the best implementation. |
| 410 | Without autotuning, FlashInfer will rely on heuristics, which may |
| 411 | be significantly slower. |
| 412 | |
| 413 | With PP > 1, stages run different layers and may profile different ops, |
| 414 | so each stage's TP group tunes separately with its own cache file; |
| 415 | otherwise the world group tunes together. Per-tactic timings are |
| 416 | averaged over the tuning group so all its ranks select the same tactic. |
| 417 | """ |
| 418 | from flashinfer.autotuner import AutoTuner, set_autotune_process_group |
| 419 | |
| 420 | import vllm.utils.flashinfer as fi_utils |
| 421 | from vllm.distributed.parallel_state import ( |
| 422 | get_pp_group, |
| 423 | get_tp_group, |
| 424 | get_world_group, |
| 425 | ) |
| 426 | |
| 427 | world = get_world_group() |
| 428 | pp_size = get_pp_group().world_size |
| 429 | tune_group = get_tp_group() if pp_size > 1 else world |
| 430 | is_leader = tune_group.rank_in_group == 0 |
| 431 | tuner = AutoTuner.get() |
| 432 | |
| 433 | autotune_kwargs: dict = {} |
| 434 | skip_ops = _flashinfer_autotune_skip_ops(runner) |
| 435 | if skip_ops: |
| 436 | logger.info_once( |
| 437 | "Skipping FlashInfer autotuning for ops %s", |
| 438 | tuple(sorted(skip_ops)), |
| 439 | ) |
| 440 | autotune_kwargs["skip_ops"] = skip_ops |
| 441 | |
| 442 | cache_path = resolve_flashinfer_autotune_file(runner) |
| 443 | if pp_size > 1: |
| 444 | ranks = "-".join(str(rank) for rank in tune_group.ranks) |
| 445 | cache_path = cache_path.with_name( |
| 446 | f"{cache_path.stem}_tp_{ranks}{cache_path.suffix}" |
| 447 | ) |
| 448 | if is_leader: |
| 449 | logger.info_once("Using FlashInfer autotune cache file: %s", cache_path) |
| 450 | |
| 451 | # We skip EPLB here since we don't want to record dummy metrics. |
| 452 | # Randomize inputs to avoid every token pick the same experts, |
| 453 | # which lead to some EP ranks receiving no tokens and skipping their |
| 454 | # MoE kernel entirely, and cause hang due to all-reduce collective |
| 455 | # during synchronized autotuning. |
| 456 | # Read cached autotune results and broadcast within the tuning group. |
| 457 | cached_results: bytes | None = None |
| 458 | if is_leader and cache_path.exists(): |
| 459 | with open(cache_path, "rb") as f: |
| 460 | cached_results = f.read() |
| 461 | cached_results = tune_group.broadcast_object(cached_results, src=0) |
| 462 | if cached_results is not None: |
| 463 | write_flashinfer_autotune_cache(cache_path, cached_results) |
| 464 | tune_group.barrier() |
| 465 | tuner.load_configs(str(cache_path)) |
| 466 | |
| 467 | group = tune_group.cpu_group if tune_group.world_size > 1 else None |
| 468 | set_autotune_process_group(group) |
| 469 | try: |
| 470 | with ( |
| 471 | torch.inference_mode(), |
| 472 | fi_utils.autotune(tune_mode=True, **autotune_kwargs), |
| 473 | ): |
| 474 | hisparse_enabled = ( |
| 475 | runner.vllm_config.attention_config.hisparse_config is not None |
| 476 | ) |
| 477 | if hisparse_enabled: |
| 478 | # HiSparse hot-buffer attention is bounded by decode batch |
| 479 | # size, not the prefill-sized batch used for the full model. |
| 480 | autotune_hisparse_flashinfer_attention(runner) |
| 481 | _run_flashinfer_autotune_dummy_runs(runner, skip_attn=hisparse_enabled) |
| 482 | replayssm_autotune_warmup(runner) |
| 483 | _autotune_kimi_k3_kda_qkvg(runner.get_model()) |
| 484 | with torch.inference_mode(): |
| 485 | _run_flashinfer_bf16_autotune_dummy_run( |
| 486 | runner, skip_ops=skip_ops, skip_attn=hisparse_enabled |
| 487 | ) |
| 488 | finally: |
| 489 | set_autotune_process_group(None) |
| 490 | |
| 491 | if world.world_size > 1: |
| 492 | world.barrier() |
| 493 | if is_leader: |
| 494 | tuner.save_configs(str(cache_path)) |