[diffusion] feat: add plain component weight overrides (#36086)

python/sglang/multimodal_gen/runtime/loader/component_loaders/adapter_loader.py

@@ -44,6 +44,9 @@ class AdapterLoader(PlainStateDictComponentLoader):
         *args,
     ):
         config = self.load_component_config(component_model_path, component_name)
+        component_weights_path = self.resolve_component_weights_path(
+            component_model_path, server_args, component_name
+        )
 
         cls_name = config.pop("_class_name", None)
         if cls_name is None:
@@ -74,7 +77,7 @@ class AdapterLoader(PlainStateDictComponentLoader):
             adapter_cfg.update_model_arch(config)
             model = model_cls(adapter_cfg).to(device=target_device, dtype=default_dtype)
 
-        loaded = load_safetensors_state_dict(component_model_path)
+        loaded = load_safetensors_state_dict(component_weights_path)
         mapping = adapter_cfg.arch_config.param_names_mapping
         loaded = {_remap_connector_key(k, mapping): v for k, v in loaded.items()}
 
@@ -84,7 +87,7 @@ class AdapterLoader(PlainStateDictComponentLoader):
         # else uninitialized would surface later as garbage embeddings.
         if missing or unexpected:
             raise ValueError(
-                f"Adapter weights at '{component_model_path}' do not match the "
+                f"Adapter weights at '{component_weights_path}' do not match the "
                 f"instantiated {cls_name}. Missing: {sorted(missing)}. "
                 f"Unexpected: {sorted(unexpected)}. This usually means the "
                 "adapter config or its weight-name mapping is wrong."

python/sglang/multimodal_gen/runtime/loader/component_loaders/component_loader.py

@@ -46,6 +46,10 @@ from sglang.multimodal_gen.runtime.utils.hf_diffusers_utils import (
 )
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
 from sglang.multimodal_gen.runtime.utils.precision import resolve_component_precision
+from sglang.multimodal_gen.runtime.weights.source import (
+    materialize_weight,
+    resolve_weight,
+)
 from sglang.srt.model_loader.checkpoint_quantization import (
     resolve_checkpoint_quant_spec,
 )
@@ -518,6 +522,19 @@ class PlainStateDictComponentLoader(ComponentLoader):
         self.ensure_plain_state_dict_checkpoint(config, component_name)
         return config
 
+    def resolve_component_weights_path(
+        self,
+        component_model_path: str,
+        server_args: ServerArgs,
+        component_name: str,
+    ) -> str:
+        override = server_args.component_weights_paths.get(component_name)
+        if override is None:
+            return component_model_path
+        weights_path = materialize_weight(resolve_weight(override))
+        logger.info("Using weight override for %s: %s", component_name, weights_path)
+        return weights_path
+
 
 class ImageProcessorLoader(ComponentLoader):
     """Loader for image processor."""

python/sglang/multimodal_gen/runtime/loader/component_loaders/diffusion_decoder_loader.py

@@ -30,6 +30,9 @@ class DiffusionDecoderLoader(PlainStateDictComponentLoader):
         *args,
     ):
         config = self.load_component_config(component_model_path, component_name)
+        component_weights_path = self.resolve_component_weights_path(
+            component_model_path, server_args, component_name
+        )
         class_name = config.pop("_class_name", None)
         if class_name is None:
             raise ValueError(
@@ -54,6 +57,6 @@ class DiffusionDecoderLoader(PlainStateDictComponentLoader):
             model = model_cls(decoder_config).to(device=target_device, dtype=dtype)
 
         model.load_state_dict(
-            load_safetensors_state_dict(component_model_path), strict=True
+            load_safetensors_state_dict(component_weights_path), strict=True
         )
         return model

python/sglang/multimodal_gen/runtime/loader/component_loaders/sound_tokenizer_loader.py

@@ -1,11 +1,10 @@
 # SPDX-License-Identifier: Apache-2.0
-from safetensors.torch import load_file as safetensors_load_file
 
 from sglang.multimodal_gen.runtime.loader.component_loaders.component_loader import (
     PlainStateDictComponentLoader,
 )
 from sglang.multimodal_gen.runtime.loader.utils import (
-    _list_safetensors_files,
+    load_safetensors_state_dict,
     set_default_torch_dtype,
     skip_init_modules,
 )
@@ -25,6 +24,9 @@ class SoundTokenizerLoader(PlainStateDictComponentLoader):
         self, component_model_path: str, server_args: ServerArgs, component_name: str
     ):
         config = self.load_component_config(component_model_path, component_name)
+        component_weights_path = self.resolve_component_weights_path(
+            component_model_path, server_args, component_name
+        )
         class_name = config.pop("_class_name", None) or self.component_architecture
         assert (
             class_name is not None
@@ -45,11 +47,7 @@ class SoundTokenizerLoader(PlainStateDictComponentLoader):
             model_cls, _ = ModelRegistry.resolve_model_cls(class_name)
             model = model_cls(config).to(target_device)
 
-        safetensors_list = _list_safetensors_files(component_model_path)
+        loaded = load_safetensors_state_dict(component_weights_path)
-        assert (
-            len(safetensors_list) == 1
-        ), f"Found {len(safetensors_list)} safetensors files in {component_model_path}"
-        loaded = safetensors_load_file(safetensors_list[0])
         incompatible = model.load_state_dict(loaded, strict=False)
         missing = getattr(incompatible, "missing_keys", [])
         # The tokenizer is decoder-only; the checkpoint's encoder weights are

python/sglang/multimodal_gen/runtime/loader/component_loaders/vocoder_loader.py

@@ -1,12 +1,10 @@
 import re
 
-from safetensors.torch import load_file as safetensors_load_file
-
 from sglang.multimodal_gen.runtime.loader.component_loaders.component_loader import (
     PlainStateDictComponentLoader,
 )
 from sglang.multimodal_gen.runtime.loader.utils import (
-    _list_safetensors_files,
+    load_safetensors_state_dict,
     set_default_torch_dtype,
     skip_init_modules,
 )
@@ -27,6 +25,9 @@ class VocoderLoader(PlainStateDictComponentLoader):
         self, component_model_path: str, server_args: ServerArgs, component_name: str
     ):
         config = self.load_component_config(component_model_path, component_name)
+        component_weights_path = self.resolve_component_weights_path(
+            component_model_path, server_args, component_name
+        )
         class_name = config.pop("_class_name", None) or self.component_architecture
         assert (
             class_name is not None
@@ -57,11 +58,7 @@ class VocoderLoader(PlainStateDictComponentLoader):
             vocoder_cls, _ = ModelRegistry.resolve_model_cls(class_name)
             vocoder = vocoder_cls(vocoder_config).to(target_device)
 
-        safetensors_list = _list_safetensors_files(component_model_path)
+        loaded = load_safetensors_state_dict(component_weights_path)
-        assert (
-            len(safetensors_list) == 1
-        ), f"Found {len(safetensors_list)} safetensors files in {component_model_path}"
-        loaded = safetensors_load_file(safetensors_list[0])
         mapping = vocoder_config.arch_config.param_names_mapping
         loaded = {_remap_vocoder_key(k, mapping): v for k, v in loaded.items()}
 
@@ -69,7 +66,7 @@ class VocoderLoader(PlainStateDictComponentLoader):
         # A half-loaded vocoder produces plausible but wrong audio.
         if missing_keys or unexpected_keys:
             raise ValueError(
-                f"Vocoder weights at '{component_model_path}' do not match the "
+                f"Vocoder weights at '{component_weights_path}' do not match the "
                 f"instantiated {class_name}. Missing: {sorted(missing_keys)}. "
                 f"Unexpected: {sorted(unexpected_keys)}."
             )

python/sglang/multimodal_gen/runtime/loader/utils.py

@@ -297,6 +297,9 @@ def _list_safetensors_files(model_path: str) -> list[str]:
     automatically via HuggingFace Hub (if the path is an HF cache entry);
     if repair fails a clear RuntimeError is raised.
     """
+    if os.path.isfile(model_path):
+        return [str(model_path)] if str(model_path).endswith(".safetensors") else []
+
     found = sorted(glob.glob(os.path.join(str(model_path), "*.safetensors")))
 
     index_path = os.path.join(

python/sglang/multimodal_gen/test/unit/test_component_quantization_admission.py

@@ -2,6 +2,7 @@
 
 import re
 import unittest
+from types import SimpleNamespace
 from unittest.mock import patch
 
 from sglang.multimodal_gen.runtime.loader.component_loaders.adapter_loader import (
@@ -33,6 +34,29 @@ class _TestLoader(PlainStateDictComponentLoader):
 
 
 class TestComponentQuantizationAdmission(unittest.TestCase):
+    def test_plain_loader_resolves_weights_separately_from_config(self):
+        server_args = SimpleNamespace(
+            component_weights_paths={"vocoder": "owner/repo/vocoder.safetensors"}
+        )
+        with (
+            patch(
+                "sglang.multimodal_gen.runtime.loader.component_loaders."
+                "component_loader.resolve_weight",
+                return_value="resolved",
+            ),
+            patch(
+                "sglang.multimodal_gen.runtime.loader.component_loaders."
+                "component_loader.materialize_weight",
+                return_value="/cache/vocoder.safetensors",
+            ),
+        ):
+            self.assertEqual(
+                _TestLoader().resolve_component_weights_path(
+                    "/base/vocoder", server_args, "vocoder"
+                ),
+                "/cache/vocoder.safetensors",
+            )
+
     def test_plain_checkpoint_config_is_accepted(self):
         config = {"_class_name": "TestModel"}