| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """FlashInfer autotune cache helpers.""" |
| 4 | |
| 5 | import hashlib |
| 6 | import os |
| 7 | import tempfile |
| 8 | from contextlib import suppress |
| 9 | from pathlib import Path |
| 10 | from typing import TYPE_CHECKING |
| 11 | |
| 12 | import vllm.envs as envs |
| 13 | |
| 14 | if TYPE_CHECKING: |
| 15 | from vllm.distributed.parallel_state import GroupCoordinator |
| 16 | from vllm.v1.worker.gpu_model_runner import GPUModelRunner |
| 17 | |
| 18 | |
| 19 | def flashinfer_autotune_cache_hash(runner: "GPUModelRunner") -> str: |
| 20 | config_hash = runner.vllm_config.compute_hash(include_version=False) |
| 21 | return hashlib.sha256(config_hash.encode()).hexdigest() |
| 22 | |
| 23 | |
| 24 | def resolve_flashinfer_autotune_file(runner: "GPUModelRunner") -> Path: |
| 25 | override_dir = envs.VLLM_FLASHINFER_AUTOTUNE_CACHE_DIR |
| 26 | if override_dir: |
| 27 | root = Path(override_dir).expanduser() |
| 28 | else: |
| 29 | from flashinfer.jit import env as flashinfer_jit_env |
| 30 | |
| 31 | flashinfer_workspace = flashinfer_jit_env.FLASHINFER_WORKSPACE_DIR |
| 32 | root = ( |
| 33 | Path(envs.VLLM_CACHE_ROOT) |
| 34 | / "flashinfer_autotune_cache" |
| 35 | / flashinfer_workspace.parent.name |
| 36 | / flashinfer_workspace.name |
| 37 | ) |
| 38 | |
| 39 | output_dir = root / flashinfer_autotune_cache_hash(runner) |
| 40 | output_dir.mkdir(parents=True, exist_ok=True) |
| 41 | return output_dir / "autotune_configs.json" |
| 42 | |
| 43 | |
| 44 | def sync_flashinfer_autotune_cache( |
| 45 | runner: "GPUModelRunner", |
| 46 | group: "GroupCoordinator", |
| 47 | ) -> None: |
| 48 | cache: bytes | str | None = None |
| 49 | if ( |
| 50 | group.rank_in_group == 0 |
| 51 | and runner.vllm_config.kernel_config.enable_flashinfer_autotune |
| 52 | ): |
| 53 | try: |
| 54 | from vllm.platforms import current_platform |
| 55 | from vllm.utils.flashinfer import has_flashinfer |
| 56 | |
| 57 | if has_flashinfer() and current_platform.has_device_capability(90): |
| 58 | from flashinfer.autotuner import AutoTuner |
| 59 | |
| 60 | with tempfile.TemporaryDirectory() as temp_dir: |
| 61 | path = Path(temp_dir) / "autotune_configs.json" |
| 62 | AutoTuner.get().save_configs(str(path)) |
| 63 | cache = path.read_bytes() |
| 64 | except Exception as exc: |
| 65 | cache = f"{type(exc).__name__}: {exc}" |
| 66 | |
| 67 | cache = group.broadcast_object(cache) |
| 68 | if isinstance(cache, str): |
| 69 | raise RuntimeError(f"Failed to serialize FlashInfer autotune state: {cache}") |
| 70 | if cache is None or group.rank_in_group == 0: |
| 71 | return |
| 72 | |
| 73 | from flashinfer.autotuner import AutoTuner |
| 74 | |
| 75 | with tempfile.NamedTemporaryFile() as f: |
| 76 | f.write(cache) |
| 77 | f.flush() |
| 78 | if not AutoTuner.get().load_configs(f.name): |
| 79 | raise RuntimeError("FlashInfer autotune cache is incompatible") |
| 80 | |
| 81 | |
| 82 | def write_flashinfer_autotune_cache(cache_path: Path, contents: bytes) -> None: |
| 83 | cache_path.parent.mkdir(parents=True, exist_ok=True) |
| 84 | fd, tmp_path = tempfile.mkstemp( |
| 85 | dir=cache_path.parent, suffix=".tmp", prefix=f".{cache_path.name}." |
| 86 | ) |
| 87 | try: |
| 88 | with os.fdopen(fd, "wb") as f: |
| 89 | f.write(contents) |
| 90 | os.replace(tmp_path, cache_path) |
| 91 | except BaseException: |
| 92 | with suppress(OSError): |
| 93 | os.unlink(tmp_path) |
| 94 | raise |