[diffusion] feat: add exact component precision overrides (#36991)

docs/docs/sglang-diffusion/api/cli.mdx

@@ -131,6 +131,7 @@ pipeline's registered module name:
 | --- | --- | --- | --- |
 | Replace a component | `--component-paths.<component> {MODEL}` | `--<component>-path {MODEL}` | Load the replacement component's configuration and weights |
 | Replace only its weights | `--component-weights-paths.<component> {WEIGHTS}` | `--<component>-weights-path {WEIGHTS}` | Retain the base component configuration and replace its weights |
+| Select component precision | `--component-precisions.<component> {DTYPE}` | — | Use a loader-supported exact parameter and execution dtype |
 | Direct-load an eligible component | `--component-direct-gpu-weight-loading.<component>` | None | Use that component's audited direct-GPU loader; it must stay resident |
 | Quantize an unquantized component online | `--component-quantizations.<component> {METHOD}` | `--<component>-quantization {METHOD}` | Apply a method supported by that component's native loader |
 | Keep selected component layers unquantized | `--component-quantization-ignored-layers.<component> {PATTERN...}` | None | Pass component-local ignored-layer patterns to its online quantizer |
@@ -167,6 +168,11 @@ published, model-specific checkpoint examples; for example, all H3 sources and
 their exact overlays are kept in one
 [MiniMax-H3 compatibility table](/cookbook/diffusion/MiniMax/MiniMax-H3#checkpoint-and-adapter-formats).
 
+Exact precision overrides are capability-based. Native text and image encoders,
+standard VAE components, and native plain-state components support them;
+other component loaders reject the option instead of accepting a dtype that
+their execution stage would not honor.
+
 Direct-GPU loading is also capability-based. The existing
 `--direct-gpu-weight-loading` remains the primary DiT path. The component form
 currently supports standard native `vae` and `video_vae` state dicts on CUDA:

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

@@ -172,6 +172,18 @@ class ComponentLoader(ABC):
     ) -> dict[str, Any]:
         return {}
 
+    def component_load_precision(
+        self, server_args: ServerArgs, component_name: str
+    ) -> str | None:
+        """Return an exact precision override or reject an unsupported one."""
+        precision = server_args.component_precisions.get(component_name)
+        if precision is not None:
+            raise ComponentCheckpointUnsupportedError(
+                f"{component_name!r} does not support an exact component precision "
+                "override"
+            )
+        return None
+
     def supports_direct_gpu_weight_loading_for_component(
         self, _component_name: str
     ) -> bool:
@@ -269,6 +281,7 @@ class ComponentLoader(ABC):
 
         """
         self._native_load_manages_placement = False
+        self.component_load_precision(server_args, component_name)
         if server_args.should_direct_gpu_weight_load_component(
             component_name
         ) and not self.supports_direct_gpu_weight_loading_for_component(component_name):
@@ -582,6 +595,11 @@ class ComponentLoader(ABC):
 class PlainStateDictComponentLoader(ComponentLoader):
     """Base for native loaders whose current materializer expects plain weights."""
 
+    def component_load_precision(
+        self, server_args: ServerArgs, component_name: str
+    ) -> str | None:
+        return server_args.component_precisions.get(component_name)
+
     @staticmethod
     def ensure_plain_state_dict_checkpoint(config: object, component_name: str) -> None:
         try:

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

@@ -16,6 +16,13 @@ class ImageEncoderLoader(TextEncoderLoader):
     component_names = ["image_encoder"]
     expected_library = "transformers"
 
+    def component_load_precision(
+        self, server_args: ServerArgs, component_name: str
+    ) -> str | None:
+        return server_args.component_precisions.get(
+            component_name, server_args.pipeline_config.image_encoder_precision
+        )
+
     def load_customized(
         self,
         component_model_path: str,
@@ -56,10 +63,12 @@ class ImageEncoderLoader(TextEncoderLoader):
 
         # Always start with local device; load_model will adjust for offload if needed
         # TODO(will): add support for other dtypes
+        image_encoder_dtype = self.component_load_precision(server_args, component_name)
+        assert image_encoder_dtype is not None
         return self.load_model(
             component_weights_path,
             encoder_config,
             server_args,
-            server_args.pipeline_config.image_encoder_precision,
+            image_encoder_dtype,
             component_name=component_name,
         )

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

@@ -1,5 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
+import torch
+
 from sglang.multimodal_gen.runtime.loader.component_loaders.component_loader import (
     PlainStateDictComponentLoader,
 )
@@ -11,6 +13,7 @@ from sglang.multimodal_gen.runtime.loader.utils import (
 from sglang.multimodal_gen.runtime.models.registry import ModelRegistry
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
 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.utils import PRECISION_TO_TYPE
 
 logger = init_logger(__name__)
@@ -34,18 +37,19 @@ class SoundTokenizerLoader(PlainStateDictComponentLoader):
 
         server_args.model_paths[component_name] = component_model_path
 
+        dtype = resolve_component_precision(server_args, component_name)
+        if dtype is None:
             try:
-            precision = server_args.pipeline_config.vae_precision
+                dtype = PRECISION_TO_TYPE[server_args.pipeline_config.vae_precision]
             except AttributeError:
-            precision = "bf16"
+                dtype = torch.bfloat16
-        dtype = PRECISION_TO_TYPE[precision]
         target_device = self.target_device(
             server_args.should_start_component_on_cpu(component_name)
         )
 
         with set_default_torch_dtype(dtype), skip_init_modules():
             model_cls, _ = ModelRegistry.resolve_model_cls(class_name)
-            model = model_cls(config).to(target_device)
+            model = model_cls(config).to(device=target_device, dtype=dtype)
 
         loaded = load_safetensors_state_dict(component_weights_path)
         incompatible = model.load_state_dict(loaded, strict=False)

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

@@ -500,6 +500,16 @@ class TextEncoderLoader(ComponentLoader):
     expected_library = "transformers"
     supports_online_quantization_override = True
 
+    def component_load_precision(
+        self, server_args: ServerArgs, component_name: str
+    ) -> str | None:
+        override = server_args.component_precisions.get(component_name)
+        if override is not None:
+            return override
+        return server_args.pipeline_config.text_encoder_precisions[
+            self._extract_encoder_index(component_name)
+        ]
+
     def should_raise_customized_load_error(
         self, server_args: ServerArgs, component_name: str
     ) -> bool:
@@ -781,9 +791,8 @@ class TextEncoderLoader(ComponentLoader):
             server_args.encoder_parallel,
             prefer_dp=prefer_dp,
         )
-        encoder_dtype = server_args.pipeline_config.text_encoder_precisions[
+        encoder_dtype = self.component_load_precision(server_args, component_name)
-            encoder_index
+        assert encoder_dtype is not None
-        ]
         # TODO(will): add support for other dtypes
         try:
             return self.load_model(

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

@@ -16,6 +16,7 @@ from sglang.multimodal_gen.runtime.models.upsampler.latent_upsampler import (
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
 from sglang.multimodal_gen.runtime.utils.hf_diffusers_utils import maybe_download_model
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+from sglang.multimodal_gen.runtime.utils.precision import resolve_component_precision
 
 logger = init_logger(__name__)
 
@@ -226,12 +227,15 @@ class UpsamplerLoader(PlainStateDictComponentLoader):
             component_name
         )
         target_device = self.target_device(component_starts_on_cpu)
+        dtype = resolve_component_precision(server_args, component_name)
+        if dtype is None:
+            dtype = torch.bfloat16
 
         with torch.device("meta"):
             model = LatentUpsampler(**config)
 
         model.load_state_dict(state_dict, assign=True)
-        model = model.to(device=target_device, dtype=torch.bfloat16).eval()
+        model = model.to(device=target_device, dtype=dtype).eval()
 
         logger.info("Loaded LatentUpsampler to %s", target_device)
         return model

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

@@ -398,6 +398,11 @@ class VAELoader(ComponentLoader):
     ) -> bool:
         return component_name in ("vae", "video_vae")
 
+    def component_load_precision(
+        self, server_args: ServerArgs, component_name: str
+    ) -> str | None:
+        return server_args.component_precisions.get(component_name)
+
     @staticmethod
     def resolve_model_weights_path(
         component_model_path: str,

python/sglang/multimodal_gen/runtime/pipelines_core/stages/image_encoding.py

@@ -42,6 +42,7 @@ from sglang.multimodal_gen.runtime.utils.precision import (
     align_tensor_to_module_dtype,
     autocast_context,
     autocast_enabled,
+    resolve_component_precision_override,
     resolve_precision,
     temporary_module_dtype,
 )
@@ -156,9 +157,25 @@ class ImageEncodingStage(PipelineStage):
         stage_name = self._component_stage_name(stage_name)
         uses = []
         if self.image_encoder is not None:
-            uses.append(ComponentUse(stage_name, "image_encoder"))
+            uses.append(
+                ComponentUse(
+                    stage_name,
+                    "image_encoder",
+                    target_dtype=resolve_component_precision_override(
+                        server_args, "image_encoder"
+                    ),
+                )
+            )
         if self.text_encoder is not None:
-            uses.append(ComponentUse(stage_name, "text_encoder"))
+            uses.append(
+                ComponentUse(
+                    stage_name,
+                    "text_encoder",
+                    target_dtype=resolve_component_precision_override(
+                        server_args, "text_encoder"
+                    ),
+                )
+            )
         return uses
 
     def encoding_image_edit(self, outputs, image_inputs, pipeline_config):

python/sglang/multimodal_gen/runtime/pipelines_core/stages/text_encoding.py

@@ -41,6 +41,9 @@ from sglang.multimodal_gen.runtime.pipelines_core.stages.validators import (
 )
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+from sglang.multimodal_gen.runtime.utils.precision import (
+    resolve_component_precision_override,
+)
 
 logger = init_logger(__name__)
 
@@ -175,6 +178,10 @@ class TextEncodingStage(ConditionEncodingStage):
                 stage_name=stage_name,
                 component_name="text_encoder" if i == 0 else f"text_encoder_{i + 1}",
                 preferred_ready_after_request=i == 0,
+                target_dtype=resolve_component_precision_override(
+                    server_args,
+                    "text_encoder" if i == 0 else f"text_encoder_{i + 1}",
+                ),
             )
             for i in range(len(self.text_encoders))
         ]

python/sglang/multimodal_gen/runtime/server_args/server_args.py

@@ -81,6 +81,7 @@ from sglang.multimodal_gen.runtime.weights.source import (
     is_explicit_weight_file_reference,
 )
 from sglang.multimodal_gen.utils import (
+    PRECISION_TO_TYPE,
     FlexibleArgumentParser,
     StoreBoolean,
     expand_path_fields,
@@ -110,6 +111,23 @@ def is_ltx2_two_stage_pipeline_name(pipeline_class_name: str | None) -> bool:
     return pipeline_class_name in LTX2_TWO_STAGE_PIPELINE_NAMES
 
 
+def _normalize_component_precisions(value: object) -> dict[str, str]:
+    if not isinstance(value, dict):
+        raise ValueError("component_precisions must be a mapping")
+
+    normalized: dict[str, str] = {}
+    for component, precision in value.items():
+        component_name = str(component).strip().replace("-", "_")
+        precision_name = str(precision).strip().lower()
+        if not component_name or precision_name not in PRECISION_TO_TYPE:
+            raise ValueError(
+                "Component precision entries require a component and one of "
+                f"{sorted(PRECISION_TO_TYPE)}"
+            )
+        normalized[component_name] = precision_name
+    return normalized
+
+
 class Backend(str, Enum):
     """
     Enumeration for different model backends.
@@ -312,6 +330,9 @@ class ServerArgs(DisaggServerArgsMixin):
     # Explicit quantization override for one component. Self-describing
     # checkpoints remain auto-detected and do not need this override.
     component_quantizations: dict[str, str] = field(default_factory=dict)
+    # Exact load and execution precision overrides for components whose native
+    # loader advertises this capability.
+    component_precisions: dict[str, str] = field(default_factory=dict)
     # Component-local layer name patterns to skip during online quantization.
     component_quantization_ignored_layers: dict[str, list[str]] = field(
         default_factory=dict
@@ -1768,6 +1789,9 @@ class ServerArgs(DisaggServerArgsMixin):
             component_weights_paths[component] = path
         self.component_paths = component_paths
         self.component_weights_paths = component_weights_paths
+        self.component_precisions = _normalize_component_precisions(
+            self.component_precisions
+        )
         normalized_direct_gpu_loading: dict[str, bool] = {}
         for component, enabled in self.component_direct_gpu_weight_loading.items():
             component_name = str(component).strip().replace("-", "_")
@@ -2873,7 +2897,8 @@ class ServerArgs(DisaggServerArgsMixin):
         unknown_args: list[str],
         *,
         option_prefixes: tuple[str, ...],
-        alias_suffix: str,
+        alias_suffix: str | None,
+        expand_values: bool = True,
     ) -> tuple[dict[str, str], list[str]]:
         component_values: dict[str, str] = {}
         remaining: list[str] = []
@@ -2888,6 +2913,7 @@ class ServerArgs(DisaggServerArgsMixin):
                     break
             if (
                 component is None
+                and alias_suffix is not None
                 and key_part.startswith("--")
                 and key_part.endswith(alias_suffix)
             ):
@@ -2909,10 +2935,12 @@ class ServerArgs(DisaggServerArgsMixin):
                 remaining.append(arg)
             i += 1
 
-        return {
+        if expand_values:
+            component_values = {
                 component: os.path.expanduser(value)
                 for component, value in component_values.items()
-        }, remaining
+            }
+        return component_values, remaining
 
     @classmethod
     def _extract_component_paths(
@@ -3002,6 +3030,19 @@ class ServerArgs(DisaggServerArgsMixin):
             alias_suffix="-quantization",
         )
 
+    @classmethod
+    def _extract_component_precisions(
+        cls,
+        unknown_args: list[str],
+    ) -> tuple[dict[str, str], list[str]]:
+        """Extract exact component precision overrides."""
+        return cls._extract_dynamic_component_map(
+            unknown_args,
+            option_prefixes=("--component-precisions.", "--component_precisions."),
+            alias_suffix=None,
+            expand_values=False,
+        )
+
     @staticmethod
     def _extract_component_quantization_ignored_layers(
         unknown_args: list[str],
@@ -3092,6 +3133,7 @@ class ServerArgs(DisaggServerArgsMixin):
         dynamic_quantizations, remaining = cls._extract_component_quantizations(
             unknown_args
         )
+        dynamic_precisions, remaining = cls._extract_component_precisions(remaining)
         dynamic_direct_gpu_loading, remaining = (
             cls._extract_component_direct_gpu_weight_loading(remaining)
         )
@@ -3138,6 +3180,11 @@ class ServerArgs(DisaggServerArgsMixin):
             existing.update(dynamic_quantizations)
             provided_args["component_quantizations"] = existing
             explicit_arg_names.add("component_quantizations")
+        if dynamic_precisions:
+            existing = dict(provided_args.get("component_precisions") or {})
+            existing.update(dynamic_precisions)
+            provided_args["component_precisions"] = existing
+            explicit_arg_names.add("component_precisions")
         if dynamic_direct_gpu_loading:
             existing = dict(
                 provided_args.get("component_direct_gpu_weight_loading") or {}

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

@@ -24,6 +24,14 @@ def resolve_precision(
     precision_attr: Optional[str] = None,
     field_name: Optional[str] = None,
 ) -> torch.dtype:
+    component_precision = server_args.component_precisions.get(
+        component_or_precision_attr
+    )
+    if component_precision is not None:
+        return precision_to_dtype(
+            component_precision,
+            f"component_precisions.{component_or_precision_attr}",
+        )
     precision_attr = precision_attr or component_or_precision_attr
     precision = getattr(server_args.pipeline_config, precision_attr)
     return precision_to_dtype(precision, field_name or precision_attr)
@@ -35,6 +43,12 @@ def resolve_decode_precision(
     *,
     quality: str | None = None,
 ) -> torch.dtype:
+    component_precision = server_args.component_precisions.get(component_name)
+    if component_precision is not None:
+        return precision_to_dtype(
+            component_precision, f"component_precisions.{component_name}"
+        )
+
     pipeline_config = server_args.pipeline_config
     if component_name in ("audio_vae", "vocoder"):
         return resolve_precision(
@@ -58,10 +72,21 @@ def resolve_decode_precision(
     )
 
 
-def resolve_component_precision(server_args, module_name: str) -> Optional[torch.dtype]:
+def resolve_component_precision_override(
-    pipeline_config = getattr(server_args, "pipeline_config", None)
+    server_args, module_name: str
-    if pipeline_config is None:
+) -> Optional[torch.dtype]:
+    exact_precision = server_args.component_precisions.get(module_name)
+    if exact_precision is None:
         return None
+    return precision_to_dtype(exact_precision, f"component_precisions.{module_name}")
+
+
+def resolve_component_precision(server_args, module_name: str) -> Optional[torch.dtype]:
+    exact_precision = resolve_component_precision_override(server_args, module_name)
+    if exact_precision is not None:
+        return exact_precision
+
+    pipeline_config = server_args.pipeline_config
 
     if module_name in ("audio_vae", "vocoder"):
         precision_attr = "audio_vae_precision"

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

@@ -39,6 +39,13 @@ class _TestLoader(PlainStateDictComponentLoader):
 
 
 class TestComponentQuantizationAdmission(unittest.TestCase):
+    def test_plain_loader_admits_its_exact_precision(self):
+        server_args = SimpleNamespace(component_precisions={"vocoder": "fp16"})
+
+        self.assertEqual(
+            _TestLoader().component_load_precision(server_args, "vocoder"), "fp16"
+        )
+
     def test_direct_gpu_selection_requires_a_declared_component(self):
         server_args = SimpleNamespace(
             component_direct_gpu_weight_loading={"missing_vae": True}
@@ -51,6 +58,7 @@ class TestComponentQuantizationAdmission(unittest.TestCase):
 
     def test_direct_gpu_selector_is_rejected_by_unqualified_loader(self):
         server_args = SimpleNamespace(
+            component_precisions={},
             component_quantizations={},
             should_direct_gpu_weight_load_component=lambda component: component
             == "vocoder",
@@ -65,6 +73,7 @@ class TestComponentQuantizationAdmission(unittest.TestCase):
 
     def test_direct_gpu_selector_is_rejected_by_unqualified_component(self):
         server_args = SimpleNamespace(
+            component_precisions={},
             component_quantizations={},
             should_direct_gpu_weight_load_component=lambda component: component
             == "audio_vae",

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

@@ -16,6 +16,9 @@ from sglang.multimodal_gen.runtime.managers.memory_managers.component_residency_
     ComponentOffloadStrategy,
     ResidentStrategy,
 )
+from sglang.multimodal_gen.runtime.pipelines_core.stages.image_encoding import (
+    ImageEncodingStage,
+)
 from sglang.multimodal_gen.runtime.pipelines_core.stages.realtime.text_encoding import (
     RealtimeTextEncodingStage,
 )
@@ -253,13 +256,48 @@ def test_realtime_text_encoder_use_starts_at_call_site():
     stage.text_encoders = [None]
     stage._registered_stage_name = None
 
-    uses = stage.component_uses(SimpleNamespace(), "RealtimeTextEncodingStage")
+    uses = stage.component_uses(
+        SimpleNamespace(component_precisions={}, pipeline_config=None),
+        "RealtimeTextEncodingStage",
+    )
 
     assert len(uses) == 1
     assert uses[0].component_name == "text_encoder"
     assert uses[0].start_at_stage_entry is False
 
 
+def test_image_encoder_use_has_exact_precision():
+    stage = ImageEncodingStage.__new__(ImageEncodingStage)
+    stage.image_encoder = object()
+    stage.text_encoder = None
+    stage._registered_stage_name = None
+
+    uses = stage.component_uses(
+        SimpleNamespace(component_precisions={"image_encoder": "fp16"}),
+        "ImageEncodingStage",
+    )
+
+    assert [(use.component_name, use.target_dtype) for use in uses] == [
+        ("image_encoder", torch.float16)
+    ]
+
+
+def test_image_encoder_use_preserves_loaded_dtype_without_override():
+    stage = ImageEncodingStage.__new__(ImageEncodingStage)
+    stage.image_encoder = object()
+    stage.text_encoder = None
+    stage._registered_stage_name = None
+
+    uses = stage.component_uses(
+        SimpleNamespace(component_precisions={}),
+        "ImageEncodingStage",
+    )
+
+    assert [(use.component_name, use.target_dtype) for use in uses] == [
+        ("image_encoder", None)
+    ]
+
+
 def test_qwen_layered_uses_loaded_text_encoder(monkeypatch):
     from sglang.multimodal_gen.runtime.pipelines import qwen_image
 

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

@@ -23,16 +23,31 @@ class TestDecodingStageParallelism(unittest.TestCase):
     def test_component_use_honors_decode_precision_override(self):
         stage = DecodingStage(FakeVAE())
         server_args = SimpleNamespace(
+            component_precisions={},
             pipeline_config=SimpleNamespace(
                 vae_precision="fp32",
                 vae_decode_precision="bf16",
-            )
+            ),
         )
 
         [component_use] = stage.component_uses(server_args)
 
         self.assertEqual(component_use.target_dtype, torch.bfloat16)
 
+    def test_component_use_honors_exact_component_precision(self):
+        stage = DecodingStage(FakeVAE())
+        server_args = SimpleNamespace(
+            component_precisions={"vae": "fp16"},
+            pipeline_config=SimpleNamespace(
+                vae_precision="fp32",
+                vae_decode_precision="bf16",
+            ),
+        )
+
+        [component_use] = stage.component_uses(server_args)
+
+        self.assertEqual(component_use.target_dtype, torch.float16)
+
     def test_cfg_parallel_uses_replicated_decode_when_decode_group_has_multiple_ranks(
         self,
     ):

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

@@ -520,7 +520,8 @@ class TestImageVAEEncodingStageComponentName(_GlobalStageArgsMixin, unittest.Tes
     def test_component_name_can_follow_non_default_vae_key(self):
         stage = ImageVAEEncodingStage(vae=object(), component_name="video_vae")
         server_args = SimpleNamespace(
-            pipeline_config=SimpleNamespace(vae_precision="bf16")
+            component_precisions={},
+            pipeline_config=SimpleNamespace(vae_precision="bf16"),
         )
 
         uses = stage.component_uses(server_args, "image_vae_encoding")

python/sglang/multimodal_gen/test/unit/test_ideogram4.py

@@ -186,6 +186,7 @@ def _fake_server_args(cfg=None):
         disable_autocast=False,
         enable_cfg_parallel=False,
         attention_backend_config=None,
+        component_precisions={},
         kv_gather_degree=1,
         sp_split_auto=False,
     )

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

@@ -8,6 +8,7 @@ from torch import nn
 from sglang.multimodal_gen.configs.models.encoders.clip import CLIPVisionConfig
 from sglang.multimodal_gen.runtime.loader.component_loaders.component_loader import (
     ComponentCheckpointUnsupportedError,
+    ComponentLoader,
     NativeComponentLoaderRequired,
 )
 from sglang.multimodal_gen.runtime.loader.component_loaders.image_encoder_loader import (
@@ -39,6 +40,7 @@ class TestImageEncoderQuantizationAdmission(unittest.TestCase):
             ),
             component_weights_paths={},
             component_quantizations={},
+            component_precisions={},
             encoder_parallel="replicate",
             resolve_component_attention_backend=lambda _name: (None, None),
             should_direct_gpu_weight_load_component=lambda _name: False,
@@ -83,6 +85,20 @@ class TestImageEncoderQuantizationAdmission(unittest.TestCase):
             {"qkv_proj": ["q_proj", "k_proj", "v_proj"]},
         )
 
+    def test_exact_image_encoder_precision_override(self):
+        self.server_args.component_precisions["image_encoder"] = "fp16"
+
+        self.assertEqual(
+            self.loader.component_load_precision(self.server_args, "image_encoder"),
+            "fp16",
+        )
+
+    def test_unadmitted_component_precision_fails_closed(self):
+        with self.assertRaises(ComponentCheckpointUnsupportedError):
+            ComponentLoader().component_load_precision(
+                SimpleNamespace(component_precisions={"vae": "fp16"}), "vae"
+            )
+
     def test_unknown_transformers_quantized_architecture_falls_back(self):
         config = self._component_config("UnknownVisionModel", quantized=True)
         with self._config_patch(config):
@@ -131,6 +147,7 @@ class TestImageEncoderNativeLoading(unittest.TestCase):
             from_pretrained=mock.Mock(return_value=loaded_encoder)
         )
         server_args = SimpleNamespace(
+            component_precisions={},
             pipeline_config=SimpleNamespace(image_encoder_precision="bf16"),
             explicit_residency_mode=mock.Mock(return_value=None),
             require_component_resident=mock.Mock(),
@@ -182,6 +199,7 @@ class TestImageEncoderNativeLoading(unittest.TestCase):
         )
         model_class = SimpleNamespace(from_pretrained=mock.Mock())
         server_args = SimpleNamespace(
+            component_precisions={},
             pipeline_config=SimpleNamespace(image_encoder_precision="bf16"),
             explicit_residency_mode=mock.Mock(return_value=COMPONENT_OFFLOAD),
             require_component_resident=mock.Mock(),
@@ -228,6 +246,7 @@ class TestImageEncoderNativeLoading(unittest.TestCase):
         )
         server_args = SimpleNamespace(
             component_quantizations={},
+            component_precisions={},
             pipeline_config=SimpleNamespace(
                 image_encoder_precision="bf16",
                 native_only_components=(),

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

@@ -113,7 +113,9 @@ class TestDiffusionPrecisionConsistency(unittest.TestCase):
             "text_encoder_precisions": ["fp16", "bf16"],
         }
         config.update(overrides)
-        return SimpleNamespace(pipeline_config=SimpleNamespace(**config))
+        return SimpleNamespace(
+            component_precisions={}, pipeline_config=SimpleNamespace(**config)
+        )
 
     def test_precision_lookup(self):
         server_args = self._server_args()
@@ -163,8 +165,27 @@ class TestDiffusionPrecisionConsistency(unittest.TestCase):
                 self._server_args(vae_decode_precision_high="fp8"), quality="high"
             )
 
+    def test_exact_vae_precision_overrides_load_and_decode_defaults(self):
+        server_args = self._server_args(vae_decode_precision="bf16")
+        server_args.component_precisions["vae"] = "fp16"
+        server_args.component_precisions["video_vae"] = "bf16"
+
+        self.assertEqual(
+            resolve_precision(server_args, "vae", precision_attr="vae_precision"),
+            torch.float16,
+        )
+        self.assertEqual(resolve_decode_precision(server_args, "vae"), torch.float16)
+        self.assertEqual(
+            resolve_precision(server_args, "video_vae", precision_attr="vae_precision"),
+            torch.bfloat16,
+        )
+        self.assertEqual(
+            resolve_decode_precision(server_args, "video_vae"), torch.bfloat16
+        )
+
     def test_component_precision_mapping(self):
         server_args = self._server_args()
+        server_args.component_precisions["text_encoder_2"] = "fp32"
         expected = {
             "vae": torch.float16,
             "video_vae": torch.float16,
@@ -178,7 +199,7 @@ class TestDiffusionPrecisionConsistency(unittest.TestCase):
             "dual_tower_bridge": torch.float32,
             "image_encoder": torch.float16,
             "text_encoder": torch.float16,
-            "text_encoder_2": torch.bfloat16,
+            "text_encoder_2": torch.float32,
         }
 
         for module_name, expected_dtype in expected.items():
@@ -188,7 +209,11 @@ class TestDiffusionPrecisionConsistency(unittest.TestCase):
                 module_name,
             )
 
-        self.assertIsNone(resolve_component_precision(SimpleNamespace(), "vae"))
+        self.assertIsNone(
+            resolve_component_precision(
+                SimpleNamespace(component_precisions={}, pipeline_config=None), "vae"
+            )
+        )
         self.assertIsNone(
             resolve_component_precision(server_args, "unregistered_component")
         )

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

@@ -313,6 +313,28 @@ class TestServerArgsPathExpansion(unittest.TestCase):
             server_args.component_attention_backends, {"text_encoder": "torch_sdpa"}
         )
 
+    def test_dynamic_component_precision_cli_args(self):
+        parser = FlexibleArgumentParser()
+        ServerArgs.add_cli_args(parser)
+        argv = [
+            "--model-path",
+            "/fake",
+            "--component-precisions.text-encoder-2",
+            "fp32",
+        ]
+
+        with (
+            patch.object(sys, "argv", ["sglang"] + argv),
+            patch.object(
+                PipelineConfig, "from_kwargs", return_value=QwenImagePipelineConfig()
+            ),
+            _mock_cuda_platform(),
+        ):
+            args, unknown_args = parser.parse_known_args(argv)
+            server_args = ServerArgs.from_cli_args(args, unknown_args)
+
+        self.assertEqual(server_args.component_precisions, {"text_encoder_2": "fp32"})
+
     def test_layerwise_offload_components_imply_layerwise(self):
         args = self._from_dict_without_model_resolution(
             {

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

@@ -150,6 +150,7 @@ class TestTextEncoderClassResolution(unittest.TestCase):
             from_pretrained=mock.Mock(return_value=loaded_encoder)
         )
         server_args = SimpleNamespace(
+            component_precisions={},
             pipeline_config=SimpleNamespace(text_encoder_precisions=["bf16"]),
             explicit_residency_mode=mock.Mock(return_value=None),
             require_component_resident=mock.Mock(),

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

@@ -78,6 +78,25 @@ def test_negative_text_cache_key_tracks_encode_options():
     assert stage.calls == 3
 
 
+def test_component_uses_exact_encoder_precision():
+    with patch(_GLOBAL_ARGS_PATCH) as mock_global_args:
+        mock_global_args.return_value = MagicMock()
+        stage = TextEncodingStage(text_encoders=[object(), object()], tokenizers=[])
+    server_args = make_server_args(
+        component_precisions={"text_encoder_2": "fp32"},
+        pipeline_config=SimpleNamespace(
+            text_encoder_configs=[], text_encoder_precisions=["bf16", "bf16"]
+        ),
+    )
+
+    uses = stage.component_uses(server_args)
+
+    assert [(use.component_name, use.target_dtype) for use in uses] == [
+        ("text_encoder", None),
+        ("text_encoder_2", torch.float32),
+    ]
+
+
 def test_negative_text_cache_skips_warmup():
     stage = DummyTextEncodingStage()
     server_args = make_server_args()

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

@@ -36,6 +36,7 @@ class _TinyVAE(nn.Module):
 
 def _server_args():
     return SimpleNamespace(
+        component_precisions={},
         pipeline_config=SimpleNamespace(
             vae_decode_precision="fp16",
             vae_precision="fp32",

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

@@ -36,6 +36,9 @@ from sglang.multimodal_gen.runtime.managers.memory_managers import (
     host_memory_budget,
 )
 from sglang.multimodal_gen.runtime.models.vaes import wanvae
+from sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.ltx_2.decoding_av import (
+    LTX2AVDecodingStage,
+)
 
 
 class _FakeServerArgs:
@@ -47,6 +50,7 @@ class _FakeServerArgs:
         self.trust_remote_code = True
         self.layerwise_components = set()
         self.component_quantizations = {}
+        self.component_precisions = {}
         self.component_direct_gpu_weight_loading = {}
 
     def resolve_component_attention_backend(self, _component_name):
@@ -262,6 +266,56 @@ class TestDirectGPUVAEState(unittest.TestCase):
 
 
 class TestVAELoader(unittest.TestCase):
+    def test_exact_precision_is_admitted_for_every_vae_component(self):
+        loader = vae_loader.VAELoader()
+        server_args = _FakeServerArgs(QwenImagePipelineConfig())
+        server_args.component_precisions = {"vae": "bf16"}
+
+        self.assertEqual(loader.component_load_precision(server_args, "vae"), "bf16")
+
+        server_args.component_precisions = {"audio_vae": "bf16"}
+        self.assertEqual(
+            loader.component_load_precision(server_args, "audio_vae"), "bf16"
+        )
+
+    def test_exact_audio_vae_precision_reaches_customized_loader(self):
+        loader = vae_loader.VAELoader()
+        server_args = _FakeServerArgs(LTX2PipelineConfig())
+        server_args.component_precisions = {"audio_vae": "bf16"}
+        audio_vae = nn.Identity()
+
+        with (
+            patch.object(loader, "load_customized", return_value=audio_vae) as load,
+            patch.object(
+                vae_loader.current_platform,
+                "get_available_gpu_memory",
+                side_effect=[10.0, 9.0],
+            ),
+            patch(
+                "sglang.multimodal_gen.runtime.loader.component_loaders."
+                "component_loader.get_memory_usage_of_component",
+                return_value=1.0,
+            ),
+        ):
+            loaded, _ = loader.load(
+                "/component/audio_vae", server_args, "audio_vae", "diffusers"
+            )
+
+        self.assertIs(loaded, audio_vae)
+        load.assert_called_once_with("/component/audio_vae", server_args, "audio_vae")
+
+    def test_ltx_audio_vae_use_honors_exact_component_precision(self):
+        stage = LTX2AVDecodingStage(
+            vae=torch.nn.Identity(),
+            audio_vae=torch.nn.Identity(),
+            vocoder=torch.nn.Identity(),
+        )
+        server_args = _FakeServerArgs(LTX2PipelineConfig())
+        server_args.component_precisions = {"audio_vae": "fp32"}
+
+        uses = {use.component_name: use for use in stage.component_uses(server_args)}
+        self.assertEqual(uses["audio_vae"].target_dtype, torch.float32)
+
     def test_weights_override_keeps_base_component_config(self):
         loader = vae_loader.VAELoader()
         server_args = _FakeServerArgs(QwenImagePipelineConfig())

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

@@ -27,6 +27,7 @@ class _RecordingVAE:
 
 def _server_args(decode_precision="fp16", disable_autocast=False):
     return SimpleNamespace(
+        component_precisions={},
         pipeline_config=SimpleNamespace(
             vae_decode_precision=decode_precision,
             vae_precision="fp32",