fix(quant): load offline-quantized compressed-tensors checkpoints directly on CPU under offload

The diffusion loader treated any quant config without the MXFP8/
is_checkpoint_quantized markers as online quantization: it loaded every
component on the accelerator before offloading, so serving the int8
weight-only compressed-tensors artifact OOM'd at model construction
(encoder 17 GiB + VAE + packed DiT skeleton > 24 GiB). compressed-tensors
is always a serialized format with no on-load quantization pass — detect
it (through component wrappers too) and load on the host, round-tripping
each quantized module through the accelerator only for the post-load
Marlin repack, which has no CPU path.

vllm_omni/quantization/component_config.py

@@ -33,6 +33,26 @@ PRE_QUANTIZED_METHODS: frozenset[str] = frozenset(
 )
 
 
+def is_prequantized_config(quant_config: QuantizationConfig | None) -> bool:
+    """True when the config can only describe pre-quantized (serialized) weights.
+
+    Such checkpoints need no accelerator-side quantization pass at load, so a
+    CPU-offload load can read weights directly on the host. compressed-tensors
+    is the canonical case: its linear methods consume packed checkpoint tensors
+    (weight_packed/weight_scale) directly and vLLM has no on-load quantization
+    path for the format. A component wrapper is pre-quantized when every routed
+    component (and the default) is itself pre-quantized or unquantized.
+    """
+    if quant_config is None:
+        return False
+    if isinstance(quant_config, ComponentQuantizationConfig):
+        routed = [*quant_config.component_configs.values(), quant_config.default_config]
+        return any(c is not None for c in routed) and all(
+            c is None or is_prequantized_config(c) for c in routed
+        )
+    return quant_config.get_name() == "compressed-tensors"
+
+
 def resolve_component_quant_config(
     quant_config: QuantizationConfig | None,
     component: str,