[Perf] Share Triton JIT warmup compilation pool across registered kernels (#61422)

This commit is contained in:
Ayushman Singh
2026-10-16 08:59:42 -04:00
committed by GitHub
parent 64175d3827
commit de7eeaa87b
3 changed files with 90 additions and 39 deletions
@@ -168,23 +168,28 @@ def _warmup_store_kernel(out, VALUE: tl.constexpr):
not current_platform.is_cuda() or not hasattr(triton, "AsyncCompileMode"),
reason="Requires CUDA Triton async compilation",
)
def test_registered_warmup_parallelizes_compile_only_and_runtime_stays_serial(
def test_registered_warmup_overlaps_kernels_and_runtime_stays_serial(
monkeypatch,
) -> None:
from triton.runtime._async_compile import active_mode
monkeypatch.setattr(envs, "VLLM_TRITON_JIT_WARMUP_NUM_THREADS", 2)
def warmup_inputs():
return dict(out=TritonWarmupTensor(torch.int32), value=WarmupChoices(1, 2))
def make_store(value):
def warmup_inputs():
return dict(out=TritonWarmupTensor(torch.int32), value=value)
@triton_kernel_dispatcher_with_warmup(
kernel=_warmup_store_kernel, warmup_inputs=warmup_inputs
)
def store(out, value):
return (1,), dict(VALUE=value)
@triton_kernel_dispatcher_with_warmup(
kernel=_warmup_store_kernel, warmup_inputs=warmup_inputs
)
def store(out, value):
return (1,), dict(VALUE=value)
store.get_warmup_keys() # Initialize Triton's binder before wrapping its compiler.
# Initialize Triton's binder before wrapping its compiler.
store.get_warmup_keys()
return store
stores = [make_store(value) for value in (1, 2)]
caller = threading.get_ident()
barrier = threading.Barrier(2, timeout=10)
compiled_on = []
@@ -200,14 +205,15 @@ def test_registered_warmup_parallelizes_compile_only_and_runtime_stays_serial(
monkeypatch.setattr(_warmup_store_kernel, "compile", record_compile)
registry = JitWarmupRegistry(None)
with registry.activate():
store.register_warmup()
for store in stores:
store.register_warmup()
registry.warmup()
assert len(compiled_on) == 2
assert all(thread != caller for thread in compiled_on)
assert active_mode.get() is None
out = torch.empty(1, dtype=torch.int32, device=current_platform.device_type)
for value in (1, 2):
for store, value in zip(stores, (1, 2)):
store(out, value)
assert out.item() == value
assert len(compiled_on) == 2