[BugFix] Allow model_loader_extra_config with remote_instance + modelexpress backend (#34639)
Signed-off-by: joeltg <joel@reflection.ai> Co-authored-by: mmangkad <176301910+mmangkad@users.noreply.github.com>
This commit is contained in:
co-authored by
mmangkad
parent
78eef34356
commit
0a585d5bb1
@@ -333,6 +333,28 @@ class BaseModelLoader(ABC):
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
def _validate_default_loader_extra_config(
|
||||
*, extra_config: dict, load_format: LoadFormat
|
||||
) -> None:
|
||||
allowed_keys = {"enable_multithread_load", "num_threads"}
|
||||
if load_format == LoadFormat.FASTSAFETENSORS:
|
||||
allowed_keys.add("enable_gds")
|
||||
if "enable_gds" in extra_config and not isinstance(
|
||||
extra_config["enable_gds"], bool
|
||||
):
|
||||
raise ValueError(
|
||||
"enable_gds in --model-loader-extra-config must be a boolean"
|
||||
)
|
||||
|
||||
unexpected_keys = set(extra_config.keys()) - allowed_keys
|
||||
if unexpected_keys:
|
||||
raise ValueError(
|
||||
f"Unexpected extra config keys for load format "
|
||||
f"{load_format}: "
|
||||
f"{unexpected_keys}"
|
||||
)
|
||||
|
||||
|
||||
class DefaultModelLoader(BaseModelLoader):
|
||||
"""Model loader that can load different file types from disk."""
|
||||
|
||||
@@ -384,24 +406,10 @@ class DefaultModelLoader(BaseModelLoader):
|
||||
|
||||
def __init__(self, load_config: LoadConfig):
|
||||
super().__init__(load_config)
|
||||
extra_config = load_config.model_loader_extra_config
|
||||
allowed_keys = {"enable_multithread_load", "num_threads"}
|
||||
if load_config.load_format == LoadFormat.FASTSAFETENSORS:
|
||||
allowed_keys.add("enable_gds")
|
||||
if "enable_gds" in extra_config and not isinstance(
|
||||
extra_config["enable_gds"], bool
|
||||
):
|
||||
raise ValueError(
|
||||
"enable_gds in --model-loader-extra-config must be a boolean"
|
||||
)
|
||||
unexpected_keys = set(extra_config.keys()) - allowed_keys
|
||||
|
||||
if unexpected_keys:
|
||||
raise ValueError(
|
||||
f"Unexpected extra config keys for load format "
|
||||
f"{load_config.load_format}: "
|
||||
f"{unexpected_keys}"
|
||||
)
|
||||
_validate_default_loader_extra_config(
|
||||
extra_config=load_config.model_loader_extra_config,
|
||||
load_format=load_config.load_format,
|
||||
)
|
||||
|
||||
def _maybe_download_from_modelscope(
|
||||
self, model: str, revision: Optional[str]
|
||||
@@ -3230,10 +3238,27 @@ class RemoteInstanceModelLoader(BaseModelLoader):
|
||||
def __init__(self, load_config: LoadConfig):
|
||||
super().__init__(load_config)
|
||||
if load_config.model_loader_extra_config:
|
||||
raise ValueError(
|
||||
f"Model loader extra config is not supported for "
|
||||
f"load format {load_config.load_format}"
|
||||
)
|
||||
if (
|
||||
load_config.remote_instance_weight_loader_backend
|
||||
== RemoteInstanceWeightLoaderBackend.MODELEXPRESS
|
||||
):
|
||||
# ModelExpress falls back to a DefaultModelLoader whenever no
|
||||
# peer holds the weights, so it consumes this config; validate
|
||||
# the keys here so a bad one fails on every rank instead of only
|
||||
# on the ranks that end up taking the fallback.
|
||||
_validate_default_loader_extra_config(
|
||||
extra_config=load_config.model_loader_extra_config,
|
||||
load_format=load_config.load_format,
|
||||
)
|
||||
else:
|
||||
# nccl and transfer_engine replace the native loader outright,
|
||||
# so nothing would ever read the config.
|
||||
raise ValueError(
|
||||
f"Model loader extra config is not supported for "
|
||||
f"load format {load_config.load_format} with "
|
||||
f"remote instance weight loader backend "
|
||||
f"{load_config.remote_instance_weight_loader_backend}"
|
||||
)
|
||||
self.remote_instance_transfer_engine_weight_info = None
|
||||
|
||||
def download_model(self, model_config: ModelConfig) -> None:
|
||||
|
||||
@@ -0,0 +1,77 @@
|
||||
"""Unit tests for RemoteInstanceModelLoader construction - no server, no weights."""
|
||||
|
||||
import unittest
|
||||
|
||||
import sglang.srt.model_loader.loader as loader_mod
|
||||
from sglang.srt.configs.load_config import LoadConfig, LoadFormat
|
||||
from sglang.srt.model_loader.remote_instance_weight_loader_utils import (
|
||||
RemoteInstanceWeightLoaderBackend,
|
||||
)
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
from sglang.test.test_utils import CustomTestCase
|
||||
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
EXTRA_CONFIG = {"enable_multithread_load": True, "num_threads": 64}
|
||||
|
||||
# ServerArgs forwards the backend as a plain str, so every case covers both that
|
||||
# and the enum member; the loader's comparison holds only via the str mixin.
|
||||
MODELEXPRESS_BACKENDS = (
|
||||
RemoteInstanceWeightLoaderBackend.MODELEXPRESS,
|
||||
"modelexpress",
|
||||
)
|
||||
OTHER_BACKENDS = (
|
||||
RemoteInstanceWeightLoaderBackend.NCCL,
|
||||
"nccl",
|
||||
RemoteInstanceWeightLoaderBackend.TRANSFER_ENGINE,
|
||||
"transfer_engine",
|
||||
)
|
||||
ALL_BACKENDS = MODELEXPRESS_BACKENDS + OTHER_BACKENDS
|
||||
|
||||
|
||||
def _load_config(backend, extra_config=None):
|
||||
return LoadConfig(
|
||||
load_format=LoadFormat.REMOTE_INSTANCE,
|
||||
model_loader_extra_config=extra_config or {},
|
||||
remote_instance_weight_loader_backend=backend,
|
||||
)
|
||||
|
||||
|
||||
class TestRemoteInstanceModelLoaderExtraConfig(CustomTestCase):
|
||||
def test_modelexpress_backend_accepts_extra_config(self):
|
||||
for backend in MODELEXPRESS_BACKENDS:
|
||||
with self.subTest(backend=backend):
|
||||
loader = loader_mod.RemoteInstanceModelLoader(
|
||||
_load_config(backend, EXTRA_CONFIG)
|
||||
)
|
||||
# ModelExpress hands the extra config to the DefaultModelLoader
|
||||
# it falls back to, so it must survive construction unmodified.
|
||||
self.assertEqual(
|
||||
loader.load_config.model_loader_extra_config, EXTRA_CONFIG
|
||||
)
|
||||
|
||||
def test_modelexpress_backend_rejects_unknown_extra_config_key(self):
|
||||
for backend in MODELEXPRESS_BACKENDS:
|
||||
with self.subTest(backend=backend):
|
||||
with self.assertRaises(ValueError):
|
||||
loader_mod.RemoteInstanceModelLoader(
|
||||
_load_config(backend, {"num_thread": 64})
|
||||
)
|
||||
|
||||
def test_other_backends_reject_extra_config(self):
|
||||
for backend in OTHER_BACKENDS:
|
||||
with self.subTest(backend=backend):
|
||||
with self.assertRaises(ValueError):
|
||||
loader_mod.RemoteInstanceModelLoader(
|
||||
_load_config(backend, EXTRA_CONFIG)
|
||||
)
|
||||
|
||||
def test_all_backends_construct_without_extra_config(self):
|
||||
for backend in ALL_BACKENDS:
|
||||
with self.subTest(backend=backend):
|
||||
loader = loader_mod.RemoteInstanceModelLoader(_load_config(backend))
|
||||
self.assertFalse(loader.load_config.model_loader_extra_config)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user