| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import ast |
| 4 | import inspect |
| 5 | from abc import abstractmethod |
| 6 | from collections.abc import Callable, Hashable, Iterable, Mapping |
| 7 | from concurrent.futures import ThreadPoolExecutor |
| 8 | from contextlib import nullcontext, suppress |
| 9 | from contextvars import copy_context |
| 10 | from dataclasses import dataclass, field |
| 11 | from functools import cache, cached_property, update_wrapper |
| 12 | from typing import Any, Generic, ParamSpec, Protocol, TypeVar, cast, overload |
| 13 | |
| 14 | from vllm import envs |
| 15 | from vllm.model_executor.warmup.jit_warmup import ( |
| 16 | VllmJitKernel, |
| 17 | get_ast_full_name, |
| 18 | get_function_source_node, |
| 19 | ) |
| 20 | from vllm.platforms import current_platform |
| 21 | from vllm.triton_utils import HAS_TRITON, triton |
| 22 | |
| 23 | CompileKeyT = TypeVar("CompileKeyT") |
| 24 | P = ParamSpec("P") |
| 25 | # ``kernel_launcher`` also supports a ``None`` grid when no launch is needed. |
| 26 | LaunchSpec = ( |
| 27 | tuple[tuple[int, ...] | None, dict[str, Any]] |
| 28 | | tuple[tuple[int, ...] | None, dict[str, Any], Any] |
| 29 | ) |
| 30 | # Dispatchers select a concrete launch configuration for a Triton kernel. |
| 31 | DispatchSpec = ( |
| 32 | tuple[tuple[int, ...], dict[str, Any]] | tuple[tuple[int, ...], dict[str, Any], Any] |
| 33 | ) |
| 34 | |
| 35 | |
| 36 | @dataclass(frozen=True) |
| 37 | class _LaunchBindingPlan: |
| 38 | input_targets: tuple[tuple[int, str], ...] |
| 39 | stride_targets: tuple[tuple[str, str, int], ...] |
| 40 | |
| 41 | |
| 42 | @cache |
| 43 | def _launch_binding_plan( |
| 44 | arg_names: tuple[str, ...], |
| 45 | input_names: tuple[str, ...], |
| 46 | launch_names: tuple[str, ...], |
| 47 | ) -> _LaunchBindingPlan: |
| 48 | arg_name_set = set(arg_names) |
| 49 | launch_name_set = set(launch_names) |
| 50 | |
| 51 | def kernel_arg(input_name: str) -> str | None: |
| 52 | pointer_name = f"{input_name}_ptr" |
| 53 | candidates = ( |
| 54 | input_name, |
| 55 | pointer_name, |
| 56 | input_name.upper(), |
| 57 | pointer_name.upper(), |
| 58 | ) |
| 59 | return next( |
| 60 | (candidate for candidate in candidates if candidate in arg_name_set), |
| 61 | None, |
| 62 | ) |
| 63 | |
| 64 | input_targets = tuple( |
| 65 | (index, target) |
| 66 | for index, name in enumerate(input_names) |
| 67 | if (target := kernel_arg(name)) is not None and target not in launch_name_set |
| 68 | ) |
| 69 | bound_names = launch_name_set | {target for _, target in input_targets} |
| 70 | stride_targets = [] |
| 71 | for target in arg_names: |
| 72 | if target in bound_names: |
| 73 | continue |
| 74 | name, separator, dim = target.rpartition("_stride") |
| 75 | if not separator: |
| 76 | name, separator, dim = target.rpartition("_STRIDE") |
| 77 | dim = dim.removeprefix("_") |
| 78 | if separator and (not dim or dim.isdigit()): |
| 79 | pointer = f"{name}_PTR" if separator == "_STRIDE" else f"{name}_ptr" |
| 80 | source = name if name in bound_names else pointer |
| 81 | if source in bound_names: |
| 82 | stride_targets.append((target, source, int(dim or 0))) |
| 83 | return _LaunchBindingPlan(input_targets, tuple(stride_targets)) |
| 84 | |
| 85 | |
| 86 | def triton_warmup_inputs( |
| 87 | kernel: Any, |
| 88 | *args: Any, |
| 89 | grid: tuple[int, ...], |
| 90 | pointer_dtypes: ( |
| 91 | Mapping[Any, Iterable[str]] | Iterable[tuple[Any, Iterable[str]]] | None |
| 92 | ) = None, |
| 93 | **kwargs: Any, |
| 94 | ) -> dict[str, Any]: |
| 95 | """Build launcher inputs from Triton's native positional argument order.""" |
| 96 | arg_names = tuple(kernel.arg_names) |
| 97 | if len(args) > len(arg_names): |
| 98 | raise ValueError( |
| 99 | f"Received {len(args)} positional inputs for {len(arg_names)} " |
| 100 | "Triton kernel arguments" |
| 101 | ) |
| 102 | inputs = dict(zip(arg_names, args)) |
| 103 | duplicate_names = inputs.keys() & kwargs.keys() |
| 104 | if duplicate_names: |
| 105 | raise ValueError( |
| 106 | f"Triton inputs passed twice: {', '.join(sorted(duplicate_names))}" |
| 107 | ) |
| 108 | inputs.update(kwargs) |
| 109 | if pointer_dtypes is not None: |
| 110 | function_def = get_function_source_node(kernel) |
| 111 | if not isinstance(function_def, ast.FunctionDef): |
| 112 | raise ValueError("Expected Triton kernel to be defined as a function") |
| 113 | pointer_names = _pointer_arg_names(function_def, arg_names) |
| 114 | groups = ( |
| 115 | pointer_dtypes.items() |
| 116 | if isinstance(pointer_dtypes, Mapping) |
| 117 | else pointer_dtypes |
| 118 | ) |
| 119 | for dtype, names in groups: |
| 120 | for name in names: |
| 121 | if name not in arg_names: |
| 122 | raise ValueError(f"Unknown Triton pointer argument: {name}") |
| 123 | if name not in pointer_names: |
| 124 | raise ValueError(f"Triton argument is not a pointer: {name}") |
| 125 | if name in inputs: |
| 126 | raise ValueError(f"Triton input passed twice: {name}") |
| 127 | inputs[name] = TritonWarmupTensor(dtype) |
| 128 | missing_pointers = pointer_names - inputs.keys() |
| 129 | if missing_pointers: |
| 130 | names = ", ".join(sorted(missing_pointers)) |
| 131 | raise ValueError(f"Missing Triton pointer inputs: {names}") |
| 132 | return {"grid": grid, **inputs} |
| 133 | |
| 134 | |
| 135 | def triton_scalar_specialization_rep(value: int) -> int: |
| 136 | """Return an integer with the same default Triton JIT specialization. |
| 137 | |
| 138 | For an ordinary integer argument, Triton's cache key contains its inferred |
| 139 | type (``i32``, ``i64``, or ``u64``) and one of three value classes: |
| 140 | |
| 141 | * ``1`` is specialized as the exact constant ``1``. |
| 142 | * Multiples of 16 receive a ``tt.divisibility = 16`` attribute. |
| 143 | * All other values have no value specialization. |
| 144 | |
| 145 | Warmup only needs one concrete value for each cache-key class. This helper |
| 146 | returns ``1`` for the exact-one class and otherwise returns a divisible or |
| 147 | generic representative while preserving the inferred integer type. |
| 148 | |
| 149 | This applies only to non-``constexpr`` integer arguments using Triton's |
| 150 | default specialization. Do not use it for arguments listed in |
| 151 | ``do_not_specialize`` or ``do_not_specialize_on_alignment``. |
| 152 | """ |
| 153 | if value == 1: |
| 154 | return 1 |
| 155 | |
| 156 | if -(1 << 31) <= value < (1 << 31): |
| 157 | divisible_rep = 16 |
| 158 | generic_rep = 2 |
| 159 | elif -(1 << 63) <= value < (1 << 63): |
| 160 | divisible_rep = 1 << 31 |
| 161 | generic_rep = (1 << 31) + 1 |
| 162 | elif 0 <= value < (1 << 64): |
| 163 | divisible_rep = 1 << 63 |
| 164 | generic_rep = (1 << 63) + 1 |
| 165 | else: |
| 166 | raise OverflowError(f"Integer {value} is outside Triton's scalar range") |
| 167 | |
| 168 | return divisible_rep if value % 16 == 0 else generic_rep |
| 169 | |
| 170 | |
| 171 | @dataclass(frozen=True) |
| 172 | class TritonWarmupTensor: |
| 173 | """Compile-only tensor metadata used by Triton warmup. |
| 174 | |
| 175 | ``strides=None`` represents compact row-major storage. Pass explicit strides |
| 176 | whenever the runtime tensor can be padded, transposed, or otherwise strided. |
| 177 | """ |
| 178 | |
| 179 | dtype: Any |
| 180 | aligned: bool = True |
| 181 | shape: tuple[int, ...] = (1,) |
| 182 | strides: tuple[int, ...] | None = None |
| 183 | init: Any = 0 |
| 184 | |
| 185 | def data_ptr(self) -> int: |
| 186 | return 0 if self.aligned else 1 |
| 187 | |
| 188 | def ptr_range(self) -> int: |
| 189 | return 0 |
| 190 | |
| 191 | @property |
| 192 | def device(self) -> Any: |
| 193 | import torch |
| 194 | |
| 195 | return torch.device(current_platform.device_type) |
| 196 | |
| 197 | def stride(self, dim: int | None = None) -> int | tuple[int, ...]: |
| 198 | if self.strides is None: |
| 199 | strides: list[int] = [] |
| 200 | stride = 1 |
| 201 | for size in reversed(self.shape): |
| 202 | strides.append(stride) |
| 203 | stride *= size |
| 204 | result = tuple(reversed(strides)) |
| 205 | else: |
| 206 | result = self.strides |
| 207 | return result if dim is None else result[dim] |
| 208 | |
| 209 | |
| 210 | class VllmTritonJitKernel(VllmJitKernel[CompileKeyT], Generic[CompileKeyT]): |
| 211 | """Triton owner whose runtime launch specification is reused for warmup.""" |
| 212 | |
| 213 | kernel: Any |
| 214 | _warming = False |
| 215 | _warming_compile_key: CompileKeyT | None = None |
| 216 | _run_autotune = False |
| 217 | |
| 218 | @abstractmethod |
| 219 | def warmup_inputs(self, compile_key: CompileKeyT) -> dict[str, Any]: |
| 220 | """Return runtime-shaped inputs that reproduce one compile key.""" |
| 221 | raise NotImplementedError |
| 222 | |
| 223 | def compile(self, compile_key: CompileKeyT) -> None: |
| 224 | inputs = self.warmup_inputs(compile_key) |
| 225 | self._warming = True |
| 226 | self._warming_compile_key = compile_key |
| 227 | try: |
| 228 | cast(Callable[..., None], self)(**inputs) |
| 229 | finally: |
| 230 | self._warming = False |
| 231 | self._warming_compile_key = None |
| 232 | |
| 233 | def compile_many(self, compile_keys: Iterable[CompileKeyT]) -> None: |
| 234 | """Compile CUDA startup warmup variants in parallel, then wait.""" |
| 235 | keys = list(compile_keys) |
| 236 | num_threads = min(envs.VLLM_TRITON_JIT_WARMUP_NUM_THREADS, len(keys)) |
| 237 | async_compile = getattr(triton, "AsyncCompileMode", None) |
| 238 | # AMD LLVM code generation can abort under concurrent compilation. |
| 239 | if ( |
| 240 | not current_platform.is_cuda() |
| 241 | or self._run_autotune |
| 242 | or async_compile is None |
| 243 | or num_threads <= 1 |
| 244 | ): |
| 245 | return super().compile_many(keys) |
| 246 | |
| 247 | def compile_all() -> None: |
| 248 | with ( |
| 249 | ThreadPoolExecutor(max_workers=num_threads) as executor, |
| 250 | async_compile(executor), |
| 251 | ): |
| 252 | # Dispatch stays on this thread: compile() mutates owner state. |
| 253 | # Only Triton's underlying compiler work runs in the pool. |
| 254 | super(VllmTritonJitKernel, self).compile_many(keys) |
| 255 | |
| 256 | # Older Triton versions leave async mode active on compilation failure. |
| 257 | # Isolate that state so it cannot affect later runtime JIT compilation. |
| 258 | copy_context().run(compile_all) |
| 259 | |
| 260 | @cached_property |
| 261 | def _kernel_arg_names(self) -> tuple[str, ...]: |
| 262 | arg_names = getattr(self.kernel, "arg_names", None) |
| 263 | if arg_names is not None: |
| 264 | return tuple(arg_names) |
| 265 | wrapped = getattr(self.kernel, "func", None) |
| 266 | if wrapped is not None: |
| 267 | return tuple(inspect.signature(wrapped).parameters) |
| 268 | raise TypeError( |
| 269 | f"Cannot inspect kernel parameters for {type(self.kernel).__name__}" |
| 270 | ) |
| 271 | |
| 272 | def _prepare_launch_kwargs( |
| 273 | self, |
| 274 | input_names: tuple[str, ...], |
| 275 | input_values: tuple[Any, ...], |
| 276 | launch_kwargs: dict[str, Any], |
| 277 | ) -> dict[str, Any]: |
| 278 | plan = _launch_binding_plan( |
| 279 | self._kernel_arg_names, input_names, tuple(launch_kwargs) |
| 280 | ) |
| 281 | for index, target in plan.input_targets: |
| 282 | launch_kwargs[target] = input_values[index] |
| 283 | for target, source, dim in plan.stride_targets: |
| 284 | launch_kwargs[target] = launch_kwargs[source].stride(dim) |
| 285 | return launch_kwargs |
| 286 | |
| 287 | def launch( |
| 288 | self, |
| 289 | launch_spec: LaunchSpec, |
| 290 | inputs: Mapping[str, Any], |
| 291 | ) -> Any: |
| 292 | if len(launch_spec) == 3: |
| 293 | grid, launch_kwargs, outputs = launch_spec |
| 294 | else: |
| 295 | grid, launch_kwargs = launch_spec |
| 296 | outputs = None |
| 297 | kwargs = self._prepare_launch_kwargs( |
| 298 | tuple(inputs), tuple(inputs.values()), dict(launch_kwargs) |
| 299 | ) |
| 300 | kernel = kwargs.pop("kernel", None) or self.kernel |
| 301 | runtime_launcher = kwargs.pop("_runtime_launcher", None) |
| 302 | runtime_launcher_arg_count = kwargs.pop("_runtime_launcher_arg_count", 0) |
| 303 | if self._warming: |
| 304 | if self._run_autotune: |
| 305 | assert grid is not None |
| 306 | kernel[grid](**kwargs) |
| 307 | return outputs |
| 308 | kwargs = { |
| 309 | name: _triton_metadata_arg(value) for name, value in kwargs.items() |
| 310 | } |
| 311 | if ( |
| 312 | self._warming_compile_key is not None |
| 313 | and "launch_pdl" in kwargs |
| 314 | and hasattr(self._warming_compile_key, "launch_pdl") |
| 315 | ): |
| 316 | kwargs["launch_pdl"] = self._warming_compile_key.launch_pdl |
| 317 | warmup = getattr(kernel, "warmup", None) |
| 318 | assert warmup is not None |
| 319 | warmup(grid=(1,), **kwargs) |
| 320 | return outputs |
| 321 | if grid is None: |
| 322 | return outputs |
| 323 | if runtime_launcher is not None: |
| 324 | regular_args = [ |
| 325 | kwargs.pop(name) |
| 326 | for name in self._kernel_arg_names[:runtime_launcher_arg_count] |
| 327 | ] |
| 328 | runtime_launcher(kernel, grid, *regular_args, **kwargs) |
| 329 | else: |
| 330 | kernel[grid](**kwargs) |
| 331 | return outputs |
| 332 | |
| 333 | |
| 334 | @dataclass(frozen=True) |
| 335 | class TritonJitKey: |
| 336 | """Process-local identity of one Triton JIT specialization.""" |
| 337 | |
| 338 | jit_function_id: int |
| 339 | jit_function_key: Hashable |
| 340 | device: Hashable |
| 341 | cache_key: Hashable |
| 342 | |
| 343 | |
| 344 | @dataclass(frozen=True) |
| 345 | class TritonCompileKey: |
| 346 | """Triton-derived compile key with one non-comparing replay input.""" |
| 347 | |
| 348 | jit_keys: frozenset[TritonJitKey] |
| 349 | inputs: tuple[tuple[str, Any], ...] = field(compare=False, hash=False, repr=False) |
| 350 | |
| 351 | |
| 352 | def _triton_key_deriver( |
| 353 | kernel: Any, |
| 354 | ) -> Callable[[Mapping[str, Any]], set[TritonJitKey]]: |
| 355 | """Prepare Triton's key derivation once for one kernel and device.""" |
| 356 | knobs = triton.knobs |
| 357 | Autotuner = triton.runtime.autotuner.Autotuner |
| 358 | Heuristics = triton.runtime.autotuner.Heuristics |
| 359 | driver = triton.runtime.driver |
| 360 | JITFunction = triton.runtime.jit.JITFunction |
| 361 | compute_cache_key = triton.runtime.jit.compute_cache_key |
| 362 | |
| 363 | if isinstance(kernel, Heuristics): |
| 364 | derive_inner = _triton_key_deriver(kernel.fn) |
| 365 | |
| 366 | def derive_heuristic(kwargs: Mapping[str, Any]) -> set[TritonJitKey]: |
| 367 | heuristic_kwargs = dict(kwargs) |
| 368 | for name, heuristic in kernel.values.items(): |
| 369 | heuristic_kwargs[name] = heuristic(heuristic_kwargs) |
| 370 | return derive_inner(heuristic_kwargs) |
| 371 | |
| 372 | return derive_heuristic |
| 373 | |
| 374 | if isinstance(kernel, Autotuner): |
| 375 | derive_inner = _triton_key_deriver(kernel.fn) |
| 376 | |
| 377 | def derive_autotuned(kwargs: Mapping[str, Any]) -> set[TritonJitKey]: |
| 378 | previous_nargs = getattr(kernel, "nargs", None) |
| 379 | kernel.nargs = {} |
| 380 | try: |
| 381 | configs = kernel.prune_configs(dict(kwargs)) |
| 382 | finally: |
| 383 | kernel.nargs = previous_nargs |
| 384 | keys: set[TritonJitKey] = set() |
| 385 | for config in configs: |
| 386 | conflicts = kwargs.keys() & config.kwargs.keys() |
| 387 | if conflicts: |
| 388 | names = ", ".join(sorted(conflicts)) |
| 389 | raise ValueError(f"Conflicting autotune parameters: {names}") |
| 390 | keys.update(derive_inner(dict(kwargs) | config.all_kwargs())) |
| 391 | return keys |
| 392 | |
| 393 | return derive_autotuned |
| 394 | |
| 395 | if not isinstance(kernel, JITFunction): |
| 396 | raise TypeError(f"Unsupported Triton kernel wrapper: {type(kernel).__name__}") |
| 397 | |
| 398 | device = cast(Hashable, driver.active.get_current_device()) |
| 399 | _, kernel_key_cache, _, _, binder = kernel.device_caches[device] |
| 400 | jit_function_id = id(kernel) |
| 401 | jit_function_key = cast(Hashable, kernel.cache_key) |
| 402 | debug = kernel.debug or knobs.runtime.debug |
| 403 | instrumentation_mode = knobs.compilation.instrumentation_mode |
| 404 | fpsan_homomorphic_casts = getattr( |
| 405 | knobs.compilation, "fpsan_homomorphic_casts", None |
| 406 | ) |
| 407 | inspection_hook = knobs.runtime.add_stages_inspection_hook |
| 408 | |
| 409 | def derive_jit(kwargs: Mapping[str, Any]) -> set[TritonJitKey]: |
| 410 | binder_kwargs = dict(kwargs) |
| 411 | binder_kwargs["debug"] = binder_kwargs.get("debug", debug) or debug |
| 412 | binder_kwargs["instrumentation_mode"] = instrumentation_mode |
| 413 | if fpsan_homomorphic_casts is not None: |
| 414 | binder_kwargs["fpsan_homomorphic_casts"] = fpsan_homomorphic_casts |
| 415 | _, specialization, options = binder(**binder_kwargs) |
| 416 | if inspection_hook is not None: |
| 417 | _, inspection_hash = inspection_hook() |
| 418 | specialization.append(f'("custom_pipeline", {inspection_hash})') |
| 419 | cache_key = cast( |
| 420 | Hashable, |
| 421 | compute_cache_key(kernel_key_cache, specialization, options), |
| 422 | ) |
| 423 | return { |
| 424 | TritonJitKey( |
| 425 | jit_function_id, |
| 426 | jit_function_key, |
| 427 | device, |
| 428 | cache_key, |
| 429 | ) |
| 430 | } |
| 431 | |
| 432 | return derive_jit |
| 433 | |
| 434 | |
| 435 | WarmupCases = Mapping[str, Any] | Iterable[Mapping[str, Any]] |
| 436 | |
| 437 | |
| 438 | class _AutomaticTritonJitKernel(VllmTritonJitKernel[TritonCompileKey]): |
| 439 | def warmup_inputs(self, compile_key: TritonCompileKey) -> dict[str, Any]: |
| 440 | return dict(compile_key.inputs) |
| 441 | |
| 442 | def _provider_cases( |
| 443 | self, |
| 444 | provider: Callable[..., WarmupCases], |
| 445 | *args: Any, |
| 446 | **kwargs: Any, |
| 447 | ) -> Iterable[Mapping[str, Any]]: |
| 448 | cases_node = get_function_source_node(provider) |
| 449 | if any( |
| 450 | isinstance(node, ast.Call) |
| 451 | and (get_ast_full_name(node.func) or "").split(".")[-1] |
| 452 | in {"WarmupIntRange", "WarmupChoices", "_when"} |
| 453 | for node in ast.walk(cases_node) |
| 454 | ): |
| 455 | return self._expand_warmup_cases(provider, *args, **kwargs) |
| 456 | cases = provider(*args, **kwargs) |
| 457 | return (cases,) if isinstance(cases, Mapping) else cases |
| 458 | |
| 459 | |
| 460 | def _is_autotuned(kernel: Any) -> bool: |
| 461 | if not HAS_TRITON: |
| 462 | return False |
| 463 | Autotuner = triton.runtime.autotuner.Autotuner |
| 464 | |
| 465 | while kernel is not None: |
| 466 | if isinstance(kernel, Autotuner): |
| 467 | return True |
| 468 | kernel = getattr(kernel, "fn", None) |
| 469 | return False |
| 470 | |
| 471 | |
| 472 | def _triton_metadata_arg(value: Any) -> Any: |
| 473 | from torch._subclasses.fake_tensor import FakeTensor |
| 474 | |
| 475 | if not isinstance(value, FakeTensor): |
| 476 | return value |
| 477 | return TritonWarmupTensor( |
| 478 | value.dtype, |
| 479 | aligned=getattr(value, "_vllm_warmup_aligned", True), |
| 480 | shape=tuple(value.shape), |
| 481 | strides=tuple(value.stride()), |
| 482 | ) |
| 483 | |
| 484 | |
| 485 | def _materialize_warmup_case( |
| 486 | case: Mapping[str, Any], |
| 487 | *, |
| 488 | real: bool, |
| 489 | mode: Any, |
| 490 | cache: dict[TritonWarmupTensor, Any], |
| 491 | ) -> dict[str, Any]: |
| 492 | import torch |
| 493 | |
| 494 | def materialize(value: Any) -> Any: |
| 495 | if not isinstance(value, TritonWarmupTensor): |
| 496 | return value |
| 497 | if not real: |
| 498 | try: |
| 499 | return cache[value] |
| 500 | except KeyError: |
| 501 | pass |
| 502 | except TypeError: |
| 503 | # Unhashable initializers are uncommon and safe to materialize directly. |
| 504 | pass |
| 505 | |
| 506 | strides = cast(tuple[int, ...], value.stride()) |
| 507 | with mode: |
| 508 | tensor = torch.empty_strided( |
| 509 | value.shape, |
| 510 | strides, |
| 511 | dtype=value.dtype, |
| 512 | device=value.device, |
| 513 | ) |
| 514 | if real: |
| 515 | if callable(value.init): |
| 516 | value.init(tensor) |
| 517 | else: |
| 518 | tensor.fill_(value.init) |
| 519 | else: |
| 520 | tensor._vllm_warmup_aligned = value.aligned |
| 521 | with suppress(TypeError): |
| 522 | cache[value] = tensor |
| 523 | return tensor |
| 524 | |
| 525 | return {name: materialize(value) for name, value in case.items()} |
| 526 | |
| 527 | |
| 528 | class _DecoratedTritonJitKernel(_AutomaticTritonJitKernel): |
| 529 | def __init__( |
| 530 | self, |
| 531 | kernel: Any, |
| 532 | warmup_inputs: Callable[..., WarmupCases], |
| 533 | dispatch: Callable[..., DispatchSpec] | None, |
| 534 | ) -> None: |
| 535 | self.kernel = kernel |
| 536 | self._run_autotune = _is_autotuned(kernel) |
| 537 | self._warmup_inputs_fn = warmup_inputs |
| 538 | self._dispatch_fn = dispatch |
| 539 | self._dispatch_arg_names: tuple[str, ...] = () |
| 540 | self._dispatch_defaults: dict[str, Any] = {} |
| 541 | if dispatch is not None: |
| 542 | parameters = tuple(inspect.signature(dispatch).parameters.values()) |
| 543 | if any( |
| 544 | parameter.kind |
| 545 | in (inspect.Parameter.VAR_POSITIONAL, inspect.Parameter.VAR_KEYWORD) |
| 546 | for parameter in parameters |
| 547 | ): |
| 548 | raise TypeError("Triton dispatchers require explicit parameters") |
| 549 | self._dispatch_arg_names = tuple(parameter.name for parameter in parameters) |
| 550 | self._dispatch_defaults = { |
| 551 | parameter.name: parameter.default |
| 552 | for parameter in parameters |
| 553 | if parameter.default is not inspect.Parameter.empty |
| 554 | } |
| 555 | super().__init__() |
| 556 | |
| 557 | def __call__(self, *args: Any, **kwargs: Any) -> Any: |
| 558 | if self._dispatch_fn is None: |
| 559 | grid = kwargs.pop("grid") if not args else args[0] |
| 560 | kernel_args = () if not args else args[1:] |
| 561 | inputs = triton_warmup_inputs( |
| 562 | self.kernel, *kernel_args, grid=grid, **kwargs |
| 563 | ) |
| 564 | inputs.pop("grid") |
| 565 | return self.launch((grid, inputs), {}) |
| 566 | spec = self._dispatch_fn(*args, **kwargs) |
| 567 | input_values = args |
| 568 | if len(args) != len(self._dispatch_arg_names): |
| 569 | input_values += tuple( |
| 570 | kwargs[name] if name in kwargs else self._dispatch_defaults[name] |
| 571 | for name in self._dispatch_arg_names[len(args) :] |
| 572 | ) |
| 573 | return self.launch( |
| 574 | spec, |
| 575 | dict(zip(self._dispatch_arg_names, input_values)), |
| 576 | ) |
| 577 | |
| 578 | def __getitem__(self, grid: tuple[int, ...]) -> Callable[..., Any]: |
| 579 | return cast(Callable[..., Any], self.kernel[grid]) |
| 580 | |
| 581 | def __getattr__(self, name: str) -> Any: |
| 582 | return getattr(self.kernel, name) |
| 583 | |
| 584 | def get_warmup_keys( |
| 585 | self, *, vllm_config: Any | None = None, **kwargs: Any |
| 586 | ) -> list[TritonCompileKey]: |
| 587 | keys: list[TritonCompileKey] = [] |
| 588 | seen_jit_keys: set[frozenset[TritonJitKey]] = set() |
| 589 | if "vllm_config" in inspect.signature(self._warmup_inputs_fn).parameters: |
| 590 | kwargs.setdefault("vllm_config", vllm_config) |
| 591 | cases = self._provider_cases(self._warmup_inputs_fn, **kwargs) |
| 592 | derive_keys = _triton_key_deriver(self.kernel) |
| 593 | materialize = self._run_autotune |
| 594 | mode: Any = nullcontext() if self._run_autotune else None |
| 595 | materialized: dict[TritonWarmupTensor, Any] = {} |
| 596 | for case in cases: |
| 597 | inputs = ( |
| 598 | _materialize_warmup_case( |
| 599 | case, |
| 600 | real=self._run_autotune, |
| 601 | mode=mode, |
| 602 | cache=materialized, |
| 603 | ) |
| 604 | if materialize |
| 605 | else case |
| 606 | ) |
| 607 | if self._dispatch_fn is None: |
| 608 | prepared = dict(inputs) |
| 609 | prepared.pop("grid") |
| 610 | else: |
| 611 | dispatch_inputs = { |
| 612 | name: inputs[name] |
| 613 | for name in self._dispatch_arg_names |
| 614 | if name in inputs |
| 615 | } |
| 616 | try: |
| 617 | _, launch_kwargs = self._dispatch_fn(**dispatch_inputs)[:2] |
| 618 | except (AssertionError, AttributeError, TypeError): |
| 619 | if materialize: |
| 620 | raise |
| 621 | from torch._subclasses.fake_tensor import FakeTensorMode |
| 622 | |
| 623 | materialize = True |
| 624 | mode = FakeTensorMode() |
| 625 | inputs = _materialize_warmup_case( |
| 626 | case, |
| 627 | real=False, |
| 628 | mode=mode, |
| 629 | cache=materialized, |
| 630 | ) |
| 631 | dispatch_inputs = { |
| 632 | name: inputs[name] |
| 633 | for name in self._dispatch_arg_names |
| 634 | if name in inputs |
| 635 | } |
| 636 | _, launch_kwargs = self._dispatch_fn(**dispatch_inputs)[:2] |
| 637 | prepared = self._prepare_launch_kwargs( |
| 638 | tuple(inputs), |
| 639 | tuple(inputs.values()), |
| 640 | dict(launch_kwargs), |
| 641 | ) |
| 642 | if materialize: |
| 643 | prepared = { |
| 644 | name: _triton_metadata_arg(value) |
| 645 | for name, value in prepared.items() |
| 646 | } |
| 647 | jit_keys = frozenset(derive_keys(prepared)) |
| 648 | if jit_keys and jit_keys not in seen_jit_keys: |
| 649 | seen_jit_keys.add(jit_keys) |
| 650 | keys.append(TritonCompileKey(jit_keys, tuple(inputs.items()))) |
| 651 | return keys |
| 652 | |
| 653 | |
| 654 | class TritonKernelDispatcher(Protocol[P]): |
| 655 | """Callable Triton launcher with compile-time warmup support.""" |
| 656 | |
| 657 | def __call__(self, *args: P.args, **kwargs: P.kwargs) -> Any: ... |
| 658 | |
| 659 | def register_warmup(self, *args: Any, **kwargs: Any) -> None: ... |
| 660 | |
| 661 | |
| 662 | @overload |
| 663 | def triton_kernel_dispatcher_with_warmup( |
| 664 | *, |
| 665 | warmup_inputs: Callable[..., WarmupCases], |
| 666 | kernel: None = None, |
| 667 | ) -> Callable[[Any], Any]: ... |
| 668 | |
| 669 | |
| 670 | @overload |
| 671 | def triton_kernel_dispatcher_with_warmup( |
| 672 | *, |
| 673 | warmup_inputs: Callable[..., WarmupCases], |
| 674 | kernel: Any, |
| 675 | ) -> Callable[[Callable[P, DispatchSpec]], TritonKernelDispatcher[P]]: ... |
| 676 | |
| 677 | |
| 678 | def triton_kernel_dispatcher_with_warmup( |
| 679 | *, |
| 680 | warmup_inputs: Callable[..., WarmupCases], |
| 681 | kernel: Any | None = None, |
| 682 | ) -> Callable[[Any], Any]: |
| 683 | """Decorate a dispatch function or a native Triton kernel for warmup.""" |
| 684 | |
| 685 | def decorate(dispatch: Any) -> Any: |
| 686 | owner = _DecoratedTritonJitKernel( |
| 687 | dispatch if kernel is None else kernel, |
| 688 | warmup_inputs, |
| 689 | None if kernel is None else dispatch, |
| 690 | ) |
| 691 | update_wrapper(owner, dispatch, updated=()) |
| 692 | return owner |
| 693 | |
| 694 | return decorate |
| 695 | |
| 696 | |
| 697 | @dataclass(frozen=True) |
| 698 | class TritonPointerInputVariant: |
| 699 | # Named pointer-alignment variant for compile-only Triton warmup. |
| 700 | alignments: tuple[tuple[str, bool], ...] |
| 701 | |
| 702 | @classmethod |
| 703 | def from_alignment(cls, **aligned: bool) -> "TritonPointerInputVariant": |
| 704 | return cls(tuple(aligned.items())) |
| 705 | |
| 706 | def is_aligned(self, name: str) -> bool: |
| 707 | for alignment_name, aligned in self.alignments: |
| 708 | if alignment_name == name: |
| 709 | return aligned |
| 710 | raise KeyError(f"Unknown Triton pointer input variant: {name}") |
| 711 | |
| 712 | def pointer( |
| 713 | self, |
| 714 | name: str, |
| 715 | dtype: Any, |
| 716 | shape: tuple[int, ...] = (1,), |
| 717 | ) -> TritonWarmupTensor: |
| 718 | return TritonWarmupTensor(dtype, aligned=self.is_aligned(name), shape=shape) |
| 719 | |
| 720 | |
| 721 | def _literal_str_refs(node: ast.AST) -> tuple[str | int, ...]: |
| 722 | if isinstance(node, ast.Constant) and isinstance(node.value, str | int): |
| 723 | return (node.value,) |
| 724 | if isinstance(node, ast.List | ast.Tuple): |
| 725 | refs: list[str | int] = [] |
| 726 | for elt in node.elts: |
| 727 | if isinstance(elt, ast.Constant) and isinstance(elt.value, str | int): |
| 728 | refs.append(elt.value) |
| 729 | else: |
| 730 | raise ValueError( |
| 731 | f"Unsupported Triton specialization ref: {ast.dump(elt)}" |
| 732 | ) |
| 733 | return tuple(refs) |
| 734 | raise ValueError(f"Unsupported Triton specialization refs: {ast.dump(node)}") |
| 735 | |
| 736 | |
| 737 | def _normalize_arg_refs( |
| 738 | refs: tuple[str | int, ...], |
| 739 | arg_names: tuple[str, ...], |
| 740 | ) -> frozenset[str]: |
| 741 | names: set[str] = set() |
| 742 | for ref in refs: |
| 743 | if isinstance(ref, int): |
| 744 | names.add(arg_names[ref]) |
| 745 | else: |
| 746 | names.add(ref) |
| 747 | return frozenset(names) |
| 748 | |
| 749 | |
| 750 | def _decorator_keyword_refs( |
| 751 | function_def: ast.FunctionDef, |
| 752 | keyword_name: str, |
| 753 | ) -> tuple[str | int, ...]: |
| 754 | for decorator in function_def.decorator_list: |
| 755 | if not isinstance(decorator, ast.Call): |
| 756 | continue |
| 757 | decorator_name = get_ast_full_name(decorator.func) |
| 758 | if decorator_name not in ("triton.jit", "jit"): |
| 759 | continue |
| 760 | for keyword in decorator.keywords: |
| 761 | if keyword.arg == keyword_name: |
| 762 | return _literal_str_refs(keyword.value) |
| 763 | return () |
| 764 | |
| 765 | |
| 766 | def _triton_do_not_specialize_args( |
| 767 | kernel: Callable[..., Any], |
| 768 | function_def: ast.FunctionDef, |
| 769 | arg_names: tuple[str, ...], |
| 770 | ) -> frozenset[str]: |
| 771 | refs = getattr(kernel, "do_not_specialize", None) |
| 772 | if refs is not None: |
| 773 | return _normalize_arg_refs(tuple(refs), arg_names) |
| 774 | return _normalize_arg_refs( |
| 775 | _decorator_keyword_refs(function_def, "do_not_specialize"), |
| 776 | arg_names, |
| 777 | ) |
| 778 | |
| 779 | |
| 780 | def _triton_constexpr_arg_names( |
| 781 | kernel: Callable[..., Any], |
| 782 | function_def: ast.FunctionDef, |
| 783 | arg_names: tuple[str, ...], |
| 784 | ) -> frozenset[str]: |
| 785 | constexprs = getattr(kernel, "constexprs", None) |
| 786 | if constexprs is not None: |
| 787 | return frozenset(arg_names[index] for index in constexprs) |
| 788 | |
| 789 | names: set[str] = set() |
| 790 | for arg in function_def.args.args + function_def.args.kwonlyargs: |
| 791 | if arg.annotation is None: |
| 792 | continue |
| 793 | annotation = get_ast_full_name(arg.annotation) |
| 794 | if annotation in ("tl.constexpr", "triton.language.constexpr", "constexpr"): |
| 795 | names.add(arg.arg) |
| 796 | return frozenset(names) |
| 797 | |
| 798 | |
| 799 | def _leftmost_name(node: ast.AST) -> str | None: |
| 800 | if isinstance(node, ast.Name): |
| 801 | return node.id |
| 802 | if isinstance(node, ast.BinOp): |
| 803 | return _leftmost_name(node.left) |
| 804 | return None |
| 805 | |
| 806 | |
| 807 | def _pointer_arg_names( |
| 808 | function_def: ast.FunctionDef, |
| 809 | arg_names: tuple[str, ...], |
| 810 | ) -> frozenset[str]: |
| 811 | candidate_names = set(arg_names) |
| 812 | pointer_names = {name for name in arg_names if name.endswith("_ptr")} |
| 813 | for node in ast.walk(function_def): |
| 814 | if not isinstance(node, ast.Call): |
| 815 | continue |
| 816 | if get_ast_full_name(node.func) not in ("tl.load", "tl.store"): |
| 817 | continue |
| 818 | if not node.args: |
| 819 | continue |
| 820 | name = _leftmost_name(node.args[0]) |
| 821 | if name in candidate_names: |
| 822 | pointer_names.add(name) |
| 823 | return frozenset(pointer_names) |
| 824 | |
| 825 | |
| 826 | def trace_triton_kernel_specialization_args( |
| 827 | kernel: Callable[..., Any], |
| 828 | ) -> tuple[str, ...]: |
| 829 | function_def = get_function_source_node(kernel) |
| 830 | if not isinstance(function_def, ast.FunctionDef): |
| 831 | raise ValueError("Expected Triton kernel to be defined as a function") |
| 832 | source_fn = getattr(kernel, "fn", kernel) |
| 833 | arg_names = tuple(inspect.signature(source_fn).parameters) |
| 834 | constexpr_args = _triton_constexpr_arg_names(kernel, function_def, arg_names) |
| 835 | do_not_specialize_args = _triton_do_not_specialize_args( |
| 836 | kernel, function_def, arg_names |
| 837 | ) |
| 838 | pointer_args = _pointer_arg_names(function_def, arg_names) |
| 839 | |
| 840 | return tuple( |
| 841 | name |
| 842 | for name in arg_names |
| 843 | if name in constexpr_args |
| 844 | or (name not in pointer_args and name not in do_not_specialize_args) |
| 845 | ) |