| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Compile-only helpers for TileLang JIT warmup.""" |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | import logging |
| 8 | from abc import abstractmethod |
| 9 | from collections.abc import Callable, Iterator, Mapping, MutableMapping |
| 10 | from contextlib import contextmanager |
| 11 | from dataclasses import dataclass, field |
| 12 | from functools import wraps |
| 13 | from typing import Any, ClassVar, Generic, TypeAlias, TypeVar, cast |
| 14 | |
| 15 | import torch |
| 16 | |
| 17 | from vllm.model_executor.warmup.jit_warmup import VllmJitKernel |
| 18 | |
| 19 | CompileKeyT = TypeVar("CompileKeyT") |
| 20 | # (kernel_args, launch_args), optionally followed by keywords and a result. |
| 21 | TileLangLaunchSpec: TypeAlias = ( |
| 22 | tuple[tuple[Any, ...], tuple[Any, ...]] |
| 23 | | tuple[tuple[Any, ...], tuple[Any, ...], Mapping[str, Any] | None] |
| 24 | | tuple[tuple[Any, ...], tuple[Any, ...], Mapping[str, Any] | None, Any] |
| 25 | ) |
| 26 | |
| 27 | |
| 28 | @contextmanager |
| 29 | def _quiet_tilelang_warmup_logs() -> Iterator[None]: |
| 30 | loggers = [ |
| 31 | logging.getLogger("tilelang.cache.kernel_cache"), |
| 32 | logging.getLogger("tilelang.jit.kernel"), |
| 33 | ] |
| 34 | previous_levels = [logger.level for logger in loggers] |
| 35 | try: |
| 36 | for logger in loggers: |
| 37 | logger.setLevel(logging.ERROR) |
| 38 | yield |
| 39 | finally: |
| 40 | for logger, level in zip(loggers, previous_levels): |
| 41 | logger.setLevel(level) |
| 42 | |
| 43 | |
| 44 | @dataclass(frozen=True) |
| 45 | class TileLangWarmupTensor: |
| 46 | """Minimal tensor-like object accepted by TileLang compile(). |
| 47 | |
| 48 | TileLang builds its cache key and TIR from tensor dtype, shape and stride. |
| 49 | This object deliberately has no storage, so compile-only warmup does not |
| 50 | allocate GPU memory and does not launch the kernel. |
| 51 | """ |
| 52 | |
| 53 | dtype: torch.dtype |
| 54 | shape: tuple[int, ...] = (1,) |
| 55 | strides: tuple[int, ...] | None = field(default=None) |
| 56 | |
| 57 | def stride(self) -> tuple[int, ...]: |
| 58 | if self.strides is not None: |
| 59 | return self.strides |
| 60 | |
| 61 | strides: list[int] = [] |
| 62 | stride = 1 |
| 63 | for size in reversed(self.shape): |
| 64 | strides.append(stride) |
| 65 | stride *= size |
| 66 | return tuple(reversed(strides)) |
| 67 | |
| 68 | def new_empty( |
| 69 | self, |
| 70 | shape: tuple[int, ...], |
| 71 | *, |
| 72 | dtype: torch.dtype | None = None, |
| 73 | device: Any = None, |
| 74 | ) -> TileLangWarmupTensor: |
| 75 | return TileLangWarmupTensor(dtype or self.dtype, shape) |
| 76 | |
| 77 | |
| 78 | def make_tilelang_warmup_tensor( |
| 79 | dtype: torch.dtype, |
| 80 | *shape: int, |
| 81 | strides: tuple[int, ...] | None = None, |
| 82 | ) -> TileLangWarmupTensor: |
| 83 | return TileLangWarmupTensor(dtype=dtype, shape=tuple(shape), strides=strides) |
| 84 | |
| 85 | |
| 86 | def _tilelang_call_kwargs(kwargs: dict[str, Any]) -> dict[str, Any]: |
| 87 | call_kwargs = dict(kwargs) |
| 88 | tune_params = call_kwargs.pop("__tune_params", {}) |
| 89 | if isinstance(tune_params, dict): |
| 90 | call_kwargs.update(tune_params) |
| 91 | return call_kwargs |
| 92 | |
| 93 | |
| 94 | def compile_tilelang(jit_impl: Any, *args: Any, **kwargs: Any) -> None: |
| 95 | """Compile one TileLang specialization and populate its call cache. |
| 96 | |
| 97 | TileLang's ``compile()`` materializes the kernel without launching it. |
| 98 | We also store the compiled kernel in ``_kernel_cache`` using the same |
| 99 | parsed key as ``__call__`` so runtime does not report a cache miss for an |
| 100 | already materialized specialization. |
| 101 | """ |
| 102 | with _quiet_tilelang_warmup_logs(): |
| 103 | compiled = jit_impl.compile(*args, **kwargs) |
| 104 | func = getattr(jit_impl, "func", None) |
| 105 | parse_args = getattr(func, "parse_args", None) |
| 106 | cache = getattr(jit_impl, "_kernel_cache", None) |
| 107 | if not callable(parse_args) or not isinstance(cache, MutableMapping): |
| 108 | return |
| 109 | |
| 110 | key, _ = parse_args(*args, **_tilelang_call_kwargs(kwargs)) |
| 111 | cache[key] = compiled |
| 112 | |
| 113 | |
| 114 | class VllmTileLangJitKernel(VllmJitKernel[CompileKeyT], Generic[CompileKeyT]): |
| 115 | """TileLang owner whose runtime launch specification is reused for warmup.""" |
| 116 | |
| 117 | kernel: ClassVar[Any] |
| 118 | _warming_key: CompileKeyT | None = None |
| 119 | |
| 120 | @abstractmethod |
| 121 | def warmup_inputs(self, compile_key: CompileKeyT) -> dict[str, Any]: |
| 122 | """Return runtime-shaped inputs that reproduce one compile key.""" |
| 123 | raise NotImplementedError |
| 124 | |
| 125 | def compile(self, compile_key: CompileKeyT) -> None: |
| 126 | self._warming_key = compile_key |
| 127 | try: |
| 128 | cast(Callable[..., Any], self)(**self.warmup_inputs(compile_key)) |
| 129 | finally: |
| 130 | self._warming_key = None |
| 131 | |
| 132 | def launch( |
| 133 | self, |
| 134 | jit_impl: Any, |
| 135 | args: tuple[Any, ...], |
| 136 | kwargs: Mapping[str, Any] | None = None, |
| 137 | ) -> Any: |
| 138 | call_kwargs = dict(kwargs or {}) |
| 139 | if self._warming_key is not None: |
| 140 | return compile_tilelang(jit_impl, *args, **call_kwargs) |
| 141 | return jit_impl(*args, **call_kwargs) |
| 142 | |
| 143 | |
| 144 | def kernel_launcher( |
| 145 | call_fn: Callable[..., TileLangLaunchSpec], |
| 146 | ) -> Callable[..., Any]: |
| 147 | """Launch TileLang from declarative kernel and runtime argument tuples.""" |
| 148 | |
| 149 | @wraps(call_fn) |
| 150 | def wrapper( |
| 151 | self: VllmTileLangJitKernel[Any], |
| 152 | *args: Any, |
| 153 | **kwargs: Any, |
| 154 | ) -> Any: |
| 155 | launch_spec = call_fn(self, *args, **kwargs) |
| 156 | kernel_args, launch_args = launch_spec[:2] |
| 157 | if self._warming_key is not None and kernel_args not in ( |
| 158 | (), |
| 159 | (self._warming_key,), |
| 160 | ): |
| 161 | raise RuntimeError( |
| 162 | "TileLang warmup inputs produced a different compile key: " |
| 163 | f"expected {self._warming_key!r}, got kernel arguments " |
| 164 | f"{kernel_args!r}" |
| 165 | ) |
| 166 | launch_kwargs = launch_spec[2] if len(launch_spec) > 2 else None |
| 167 | output = launch_spec[3] if len(launch_spec) > 3 else None |
| 168 | result = self.launch(self.kernel(*kernel_args), launch_args, launch_kwargs) |
| 169 | return output if len(launch_spec) > 3 else result |
| 170 | |
| 171 | return wrapper |