[Fix] [AMD] diffusion: Prevent ComponentResidencyError after compile warmup (#43952)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
This commit is contained in:
jacky.cheng
2026-10-18 00:13:25 -07:00
committed by GitHub
co-authored by Mick Qian
parent a11fa06be7
commit f96a7326e6
3 changed files with 72 additions and 6 deletions
@@ -565,6 +565,13 @@ class DenoisingStage(PipelineStage, RolloutDenoisingMixin):
moved.append(module)
return moved
def _explicit_dit_layerwise_offload(self) -> bool:
"""True when the user asked for layerwise DiT residency, not a compile-only window."""
args = self.server_args
if args.dit_layerwise_offload:
return True
return bool(getattr(args, "is_dit_layerwise_offload_selected", False))
def _maybe_torch_compile(self, module: object) -> None:
"""
Compile a module with torch.compile, and enable inductor overlap tweak if available.
@@ -2073,9 +2080,13 @@ class DenoisingStage(PipelineStage, RolloutDenoisingMixin):
for module in moved:
module.to(device)
return
# prepare for first real request: restore the user-configured residency
for module in self._offloaded_dit_modules_for_compile:
module.disable_offload()
# Compile-only offload is undone so the real request is GPU-resident.
# An explicit layerwise request must stay enabled: disable_offload()
# clears manager.enabled, and the next use-site then raises
# ComponentResidencyError (Wan transformer_2).
if not self._explicit_dit_layerwise_offload():
for module in self._offloaded_dit_modules_for_compile:
module.disable_offload()
# clear the list for avoid overhead during real request
self._offloaded_dit_modules_for_compile.clear()
# The warmup traced the offload hooks, whose graph breaks inside the
@@ -0,0 +1,53 @@
# SPDX-License-Identifier: Apache-2.0
"""Compile warmup must not clear a layerwise DiT the user asked to keep."""
import unittest
from types import SimpleNamespace
from sglang.multimodal_gen.runtime.pipelines_core.stages.denoising import (
DenoisingStage,
)
class _OffloadModule:
def __init__(self):
self.disable_calls = 0
def disable_offload(self):
self.disable_calls += 1
class TestCompileWarmupKeepsExplicitLayerwise(unittest.TestCase):
def _stage(self, **server_args):
stage = DenoisingStage.__new__(DenoisingStage)
stage.server_args = SimpleNamespace(**server_args)
module = _OffloadModule()
stage._offloaded_dit_modules_for_compile = [module]
return stage, module
def _finish_compile_window(self, stage):
batch = SimpleNamespace(is_warmup=False)
with stage._offload_for_torch_compile_warmup(batch):
pass
def test_explicit_dit_layerwise_offload_stays_enabled(self):
stage, module = self._stage(
dit_layerwise_offload=True,
is_dit_layerwise_offload_selected=False,
)
self._finish_compile_window(stage)
self.assertEqual(module.disable_calls, 0)
self.assertEqual(stage._offloaded_dit_modules_for_compile, [])
def test_compile_only_offload_is_disabled_for_the_real_request(self):
stage, module = self._stage(
dit_layerwise_offload=False,
is_dit_layerwise_offload_selected=False,
)
self._finish_compile_window(stage)
self.assertEqual(module.disable_calls, 1)
self.assertEqual(stage._offloaded_dit_modules_for_compile, [])
if __name__ == "__main__":
unittest.main()
@@ -47,10 +47,12 @@ class TestOffloadDuringCompile(CustomTestCase):
self.dit = _DiT()
self.counter = CompileCounter()
self.dit.compile(backend=self.counter)
self.stage = SimpleNamespace(
_offloaded_dit_modules_for_compile=[self.dit],
_move_resident_components_for_warmup=lambda: [],
self.stage = DenoisingStage.__new__(DenoisingStage)
self.stage.server_args = SimpleNamespace(
dit_layerwise_offload=False, is_dit_layerwise_offload_selected=False
)
self.stage._offloaded_dit_modules_for_compile = [self.dit]
self.stage._move_resident_components_for_warmup = lambda: []
def _forward(self, *, is_warmup):
with DenoisingStage._offload_for_torch_compile_warmup(