[diffusion] fix: honor the lazy transformer loader contract (#15215)
Signed-off-by: Zerohertz <ohg3417@gmail.com> Co-authored-by: mickqian <mickqian@users.noreply.github.com>
This commit is contained in:
co-authored by
mickqian
parent
b0dff4009f
commit
1d019ce1c8
@@ -15,8 +15,12 @@ def load_transformer_if_needed(stage: Any, server_args: ServerArgs) -> bool:
|
||||
"""Load transformer onto stage if not marked loaded. Returns True if freshly loaded."""
|
||||
if server_args.model_loaded["transformer"]:
|
||||
return False
|
||||
stage.transformer = TransformerLoader().load(
|
||||
server_args.model_paths["transformer"], server_args, "transformer"
|
||||
loader = TransformerLoader()
|
||||
stage.transformer, _ = loader.load(
|
||||
server_args.model_paths["transformer"],
|
||||
server_args,
|
||||
component_name="transformer",
|
||||
transformers_or_diffusers=loader.expected_library,
|
||||
)
|
||||
return True
|
||||
|
||||
|
||||
@@ -19,15 +19,25 @@ from sglang.multimodal_gen.runtime.loader.component_loaders.component_loader imp
|
||||
PipelineComponentLoader,
|
||||
)
|
||||
from sglang.multimodal_gen.runtime.loader.component_loaders.transformer_loader import (
|
||||
TransformerLoader,
|
||||
_server_args_for_transformer_component,
|
||||
)
|
||||
from sglang.multimodal_gen.runtime.loader.component_loaders.vae_loader import VAELoader
|
||||
from sglang.multimodal_gen.runtime.loader.component_loaders.vl_encoder_loader import (
|
||||
VisionLanguageEncoderLoader,
|
||||
)
|
||||
from sglang.multimodal_gen.runtime.managers.memory_managers.component_residency import (
|
||||
COMPONENT_OFFLOAD,
|
||||
)
|
||||
from sglang.multimodal_gen.runtime.pipelines_core.component_loading import (
|
||||
load_transformer_if_needed,
|
||||
register_loaded_transformer,
|
||||
)
|
||||
from sglang.multimodal_gen.runtime.pipelines_core.composed_pipeline_base import (
|
||||
ComposedPipelineBase,
|
||||
)
|
||||
from sglang.multimodal_gen.runtime.platforms import current_platform
|
||||
from sglang.multimodal_gen.runtime.server_args import ServerArgs
|
||||
|
||||
|
||||
class _AliasPipeline(ComposedPipelineBase):
|
||||
@@ -39,6 +49,85 @@ class _AliasPipeline(ComposedPipelineBase):
|
||||
|
||||
|
||||
class TestComponentLoaderIdentity(unittest.TestCase):
|
||||
def test_lazy_transformer_load_register_and_reuse(self):
|
||||
for native_fallback in (False, True):
|
||||
with self.subTest(native_fallback=native_fallback):
|
||||
server_args = ServerArgs(
|
||||
model_path="x",
|
||||
component_residency={"transformer": COMPONENT_OFFLOAD},
|
||||
)
|
||||
server_args.model_loaded = {"transformer": False}
|
||||
server_args.model_paths = {"transformer": "/model/transformer"}
|
||||
stage = SimpleNamespace(transformer=None)
|
||||
pipeline = object.__new__(_AliasPipeline)
|
||||
pipeline.modules = {}
|
||||
transformer = nn.Linear(2, 2)
|
||||
|
||||
with (
|
||||
patch.object(
|
||||
current_platform, "get_available_gpu_memory", return_value=16.0
|
||||
),
|
||||
patch.object(
|
||||
TransformerLoader,
|
||||
"load_customized",
|
||||
autospec=True,
|
||||
return_value=transformer,
|
||||
side_effect=(
|
||||
ValueError("Unsupported model architecture")
|
||||
if native_fallback
|
||||
else None
|
||||
),
|
||||
) as customized,
|
||||
patch.object(
|
||||
TransformerLoader,
|
||||
"load_native",
|
||||
autospec=True,
|
||||
return_value=transformer,
|
||||
) as native,
|
||||
):
|
||||
self.assertTrue(load_transformer_if_needed(stage, server_args))
|
||||
self.assertIs(stage.transformer, transformer)
|
||||
self.assertFalse(server_args.model_loaded["transformer"])
|
||||
register_loaded_transformer(stage, server_args, pipeline)
|
||||
self.assertTrue(server_args.model_loaded["transformer"])
|
||||
self.assertIs(pipeline.get_module("transformer"), transformer)
|
||||
self.assertFalse(load_transformer_if_needed(stage, server_args))
|
||||
customized.assert_called_once()
|
||||
if native_fallback:
|
||||
self.assertEqual(
|
||||
native.call_args.args[1:],
|
||||
(
|
||||
"/model/transformer",
|
||||
server_args,
|
||||
"diffusers",
|
||||
"transformer",
|
||||
),
|
||||
)
|
||||
else:
|
||||
native.assert_not_called()
|
||||
self.assertFalse(transformer.training)
|
||||
self.assertEqual(stage.transformer(torch.ones(1, 2)).shape, (1, 2))
|
||||
|
||||
def test_lazy_transformer_load_failure_preserves_state(self):
|
||||
server_args = ServerArgs(model_path="x")
|
||||
server_args.model_loaded = {"transformer": False}
|
||||
server_args.model_paths = {"transformer": "/missing/transformer"}
|
||||
stage = SimpleNamespace(transformer=None)
|
||||
with (
|
||||
patch.object(
|
||||
current_platform, "get_available_gpu_memory", return_value=16.0
|
||||
),
|
||||
patch.object(
|
||||
TransformerLoader, "load_customized", side_effect=FileNotFoundError
|
||||
),
|
||||
patch.object(TransformerLoader, "load_native") as native,
|
||||
):
|
||||
with self.assertRaises(FileNotFoundError):
|
||||
load_transformer_if_needed(stage, server_args)
|
||||
self.assertIsNone(stage.transformer)
|
||||
self.assertFalse(server_args.model_loaded["transformer"])
|
||||
native.assert_not_called()
|
||||
|
||||
def test_structural_identity_selects_roles_and_preserves_exact_policy_keys(self):
|
||||
aliases = {
|
||||
"conditioning": "text_encoder_2",
|
||||
|
||||
@@ -301,7 +301,8 @@ def test_bcg_joint_cfg_output_lifetime_and_text_shape_fallback(monkeypatch, use_
|
||||
batch.is_warmup = True
|
||||
stage._predict_joint_velocity(cfg_policy=policies[0], **inputs)
|
||||
batch.is_warmup = False
|
||||
runner = stage._bcg_runners[id(dit)]
|
||||
runner = stage._maybe_get_bcg_runner(dit)
|
||||
assert runner is not None
|
||||
replay = runner.replay
|
||||
replay_calls = 0
|
||||
|
||||
|
||||
Reference in New Issue
Block a user