[Perf] diffusion: Replay Triton launches without the runner's per-launch hook metadata (#44471)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Mick
2026-10-18 09:58:54 +08:00
committed by GitHub
co-authored by Mick Qian Claude Opus 5.5
parent 949e1a2080
commit 8d02b69c11
2 changed files with 43 additions and 5 deletions
@@ -33,6 +33,9 @@ from typing import Any, Callable
import torch
from torch.utils._python_dispatch import TorchDispatchMode
from torch.utils._pytree import tree_leaves
from triton import knobs
from triton.knobs import HookChain
from triton.runtime import driver
from triton.runtime.jit import JITFunction
logger = logging.getLogger(__name__)
@@ -125,8 +128,10 @@ class _Plan:
def __init__(self, outputs: list, result: str, launches: list, device: int):
self.outputs = outputs # (size, stride, dtype, device), or an input index
self.result = result # "none", "tensor" or "tuple"
self.launches = launches # (launcher, constant args, tensor slots)
# (kernel, grid, run, function, metadata, constant args, tensor slots)
self.launches = launches
self.device = device
self.current_stream = driver.active.get_current_stream
def replay(self, inputs: list[torch.Tensor]):
outs = [
@@ -140,16 +145,29 @@ class _Plan:
for spec in self.outputs
]
sources = (inputs, outs)
for launcher, constants, slots in self.launches:
stream = None if _launch_hooked() else self.current_stream(self.device)
for compiled, grid, run, function, metadata, constants, slots in self.launches:
values = list(constants)
for position, source, index in slots:
values[position] = sources[source][index]
launcher(*values)
if stream is None:
compiled[grid](*values)
else:
# Triton's runner, minus the hook metadata it builds per launch.
run(*grid, stream, function, metadata, None, None, None, *values)
if self.result == "tensor":
return outs[0]
return tuple(outs) if self.result == "tuple" else None
def _launch_hooked() -> bool:
"""A profiler such as proton hooks Triton launches; those go through the runner."""
return any(
hook is not None and not (isinstance(hook, HookChain) and not hook.calls)
for hook in (knobs.runtime.launch_enter_hook, knobs.runtime.launch_exit_hook)
)
def _covers_storage(t: torch.Tensor) -> bool:
extent = 1 + sum((n - 1) * s for n, s in zip(t.shape, t.stride()))
return (
@@ -209,8 +227,9 @@ def _plan(out: Any, inputs: list[torch.Tensor], launches: list) -> Any:
slots.append((position, 0, input_at[pointer][0]))
else: # a temporary, or an ambiguous input
return _NORMAL
launcher = compiled[tuple(grid) + (1,) * (3 - len(grid))]
replays.append((launcher, tuple(constants), tuple(slots)))
grid = tuple(grid) + (1,) * (3 - len(grid))
launcher = (compiled.run, compiled.function, compiled.packed_metadata)
replays.append((compiled, grid, *launcher, tuple(constants), tuple(slots)))
devices.add(device)
if len(devices) != 1:
return _NORMAL