| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Warm b12x JIT kernels used by a loaded model.""" |
| 4 | |
| 5 | from collections import Counter |
| 6 | from collections.abc import Iterable |
| 7 | from typing import TYPE_CHECKING |
| 8 | |
| 9 | import torch |
| 10 | |
| 11 | from vllm.logger import init_logger |
| 12 | from vllm.platforms import current_platform |
| 13 | from vllm.utils.b12x import B12xWarmupUnit, b12x_warmup_token_counts |
| 14 | |
| 15 | if TYPE_CHECKING: |
| 16 | from vllm.v1.worker.gpu_worker import Worker |
| 17 | |
| 18 | logger = init_logger(__name__) |
| 19 | |
| 20 | |
| 21 | def _collect_warmup_units( |
| 22 | model: torch.nn.Module, |
| 23 | token_counts: tuple[int, ...], |
| 24 | output_dtype: torch.dtype, |
| 25 | ) -> Iterable[B12xWarmupUnit]: |
| 26 | units: dict[object, B12xWarmupUnit] = {} |
| 27 | for layer in model.modules(): |
| 28 | provider = getattr(layer, "b12x_warmup_provider", None) |
| 29 | get_unit = getattr(provider, "get_b12x_warmup_unit", None) |
| 30 | if not callable(get_unit): |
| 31 | continue |
| 32 | unit = get_unit(layer, token_counts, output_dtype) |
| 33 | assert isinstance(unit, B12xWarmupUnit) |
| 34 | units.setdefault(unit.key, unit) |
| 35 | return units.values() |
| 36 | |
| 37 | |
| 38 | def _compile_warmup_units( |
| 39 | units: Iterable[B12xWarmupUnit], |
| 40 | ) -> Counter[str]: |
| 41 | warmed: Counter[str] = Counter() |
| 42 | with torch.inference_mode(): |
| 43 | for unit in units: |
| 44 | unit.compile() |
| 45 | warmed[unit.name] += 1 |
| 46 | if warmed: |
| 47 | torch.accelerator.synchronize() |
| 48 | return warmed |
| 49 | |
| 50 | |
| 51 | def b12x_warmup(worker: "Worker", cudagraph_capture_sizes: list[int]) -> None: |
| 52 | if not current_platform.is_cuda(): |
| 53 | return |
| 54 | if not current_platform.is_device_capability_family(120): |
| 55 | return |
| 56 | |
| 57 | output_dtype = getattr( |
| 58 | getattr(worker, "model_config", None), |
| 59 | "dtype", |
| 60 | torch.bfloat16, |
| 61 | ) |
| 62 | if output_dtype not in (torch.bfloat16, torch.float16): |
| 63 | output_dtype = torch.bfloat16 |
| 64 | compile_sizes = worker.vllm_config.compilation_config.compile_sizes or [] |
| 65 | max_num_batched_tokens = worker.scheduler_config.max_num_batched_tokens |
| 66 | serving_sizes = [ |
| 67 | max_num_batched_tokens, |
| 68 | *cudagraph_capture_sizes, |
| 69 | *(size for size in compile_sizes if isinstance(size, int)), |
| 70 | ] |
| 71 | max_tokens = max_num_batched_tokens |
| 72 | max_num_scheduled_tokens = worker.scheduler_config.max_num_scheduled_tokens |
| 73 | if max_num_scheduled_tokens is not None: |
| 74 | max_tokens = max(max_tokens, max_num_scheduled_tokens) |
| 75 | token_counts = b12x_warmup_token_counts( |
| 76 | max_tokens=max_tokens, |
| 77 | cudagraph_capture_sizes=serving_sizes, |
| 78 | ) |
| 79 | units = _collect_warmup_units( |
| 80 | worker.get_model(), |
| 81 | token_counts, |
| 82 | output_dtype, |
| 83 | ) |
| 84 | for name, count in _compile_warmup_units(units).items(): |
| 85 | logger.info_once( |
| 86 | "Warmed up %d b12x %s kernel signature(s).", |
| 87 | count, |
| 88 | name, |
| 89 | ) |