| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from dataclasses import dataclass |
| 5 | from typing import Any |
| 6 | |
| 7 | from vllm.model_executor.warmup import jit_warmup_cutedsl_helper |
| 8 | from vllm.model_executor.warmup.jit_warmup import kernel_launcher |
| 9 | from vllm.model_executor.warmup.jit_warmup_cutedsl_helper import ( |
| 10 | CuTeDSLLaunchSpec, |
| 11 | VllmCuTeDSLJitKernel, |
| 12 | ) |
| 13 | |
| 14 | |
| 15 | class _TestCuTeDSLKernel(VllmCuTeDSLJitKernel["_TestCuTeDSLKernel.CompileKey"]): |
| 16 | @dataclass(frozen=True) |
| 17 | class CompileKey: |
| 18 | variant: int |
| 19 | |
| 20 | @staticmethod |
| 21 | def kernel(compile_key: CompileKey) -> tuple[str, int]: |
| 22 | return "entry", compile_key.variant |
| 23 | |
| 24 | def dispatch(self, *, variant: int) -> CompileKey: |
| 25 | return self.CompileKey(variant=variant) |
| 26 | |
| 27 | def get_warmup_keys(self) -> list[CompileKey]: |
| 28 | return self._trace_dispatch(self.dispatch)(variant=(1, 2)) |
| 29 | |
| 30 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 31 | return (f"fake-{compile_key.variant}",) |
| 32 | |
| 33 | @kernel_launcher |
| 34 | def __call__( |
| 35 | self, |
| 36 | payload: str, |
| 37 | *, |
| 38 | variant: int, |
| 39 | ) -> CuTeDSLLaunchSpec[CompileKey]: |
| 40 | compile_key = self.dispatch(variant=variant) |
| 41 | return compile_key, (payload,) |
| 42 | |
| 43 | |
| 44 | def test_cutedsl_launcher_reuses_compiled_executor(monkeypatch) -> None: |
| 45 | compile_calls: list[tuple[Any, tuple[Any, ...]]] = [] |
| 46 | launch_calls: list[tuple[int, tuple[Any, ...]]] = [] |
| 47 | |
| 48 | def compile_cutedsl(entry: tuple[str, int], *args: Any) -> Any: |
| 49 | compile_calls.append((entry, args)) |
| 50 | variant = entry[1] |
| 51 | |
| 52 | def executor(*runtime_args: Any) -> tuple[int, tuple[Any, ...]]: |
| 53 | launch_calls.append((variant, runtime_args)) |
| 54 | return variant, runtime_args |
| 55 | |
| 56 | return executor |
| 57 | |
| 58 | monkeypatch.setattr( |
| 59 | jit_warmup_cutedsl_helper, |
| 60 | "compile_cutedsl", |
| 61 | compile_cutedsl, |
| 62 | ) |
| 63 | kernel = _TestCuTeDSLKernel() |
| 64 | compile_key = kernel.CompileKey(variant=1) |
| 65 | |
| 66 | kernel.compile(compile_key) |
| 67 | kernel.compile(compile_key) |
| 68 | assert compile_calls == [(("entry", 1), ("fake-1",))] |
| 69 | |
| 70 | assert kernel("runtime", variant=1) == (1, ("runtime",)) |
| 71 | assert kernel("other", variant=2) == (2, ("other",)) |
| 72 | assert compile_calls == [ |
| 73 | (("entry", 1), ("fake-1",)), |
| 74 | (("entry", 2), ("fake-2",)), |
| 75 | ] |
| 76 | assert launch_calls == [(1, ("runtime",)), (2, ("other",))] |
| 77 | |
| 78 | |
| 79 | def test_cutedsl_launcher_runs_post_launch_finalizer(monkeypatch) -> None: |
| 80 | events: list[str] = [] |
| 81 | |
| 82 | def compile_cutedsl(entry: tuple[str, int], *args: Any) -> Any: |
| 83 | def executor(*runtime_args: Any) -> None: |
| 84 | events.append("launch") |
| 85 | |
| 86 | return executor |
| 87 | |
| 88 | class FinalizedKernel(_TestCuTeDSLKernel): |
| 89 | @kernel_launcher |
| 90 | def __call__( |
| 91 | self, |
| 92 | payload: str, |
| 93 | *, |
| 94 | variant: int, |
| 95 | ) -> CuTeDSLLaunchSpec[_TestCuTeDSLKernel.CompileKey]: |
| 96 | compile_key = self.dispatch(variant=variant) |
| 97 | |
| 98 | def finalize() -> str: |
| 99 | events.append("finalize") |
| 100 | return "output" |
| 101 | |
| 102 | return compile_key, (payload,), "unused", finalize |
| 103 | |
| 104 | monkeypatch.setattr( |
| 105 | jit_warmup_cutedsl_helper, |
| 106 | "compile_cutedsl", |
| 107 | compile_cutedsl, |
| 108 | ) |
| 109 | |
| 110 | assert FinalizedKernel()("runtime", variant=1) == "output" |
| 111 | assert events == ["launch", "finalize"] |