[diffusion] chore: fix ZMQ binding and model loading for FastWan compatibility (#13978)

Co-authored-by: Han Yu <hyu5@dt-login01.delta.ncsa.illinois.edu>
Co-authored-by: Mick <mickjagger19@icloud.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

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

@@ -728,6 +728,7 @@ class TransformerLoader(ComponentLoader):
             param_dtype=torch.bfloat16,
             reduce_dtype=torch.float32,
             output_dtype=None,
+            strict=False,
         )
 
         total_params = sum(p.numel() for p in model.parameters())

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

@@ -79,6 +79,7 @@ def maybe_load_fsdp_model(
     fsdp_inference: bool = False,
     output_dtype: torch.dtype | None = None,
     pin_cpu_memory: bool = True,
+    strict: bool = True,
 ) -> torch.nn.Module:
     """
     Load the model with FSDP if is training, else load the model without FSDP.
@@ -138,7 +139,7 @@ def maybe_load_fsdp_model(
         weight_iterator,
         device,
         default_dtype,
-        strict=True,
+        strict=strict,
         cpu_offload=cpu_offload,
         param_names_mapping=param_names_mapping_fn,
     )
@@ -255,9 +256,15 @@ def load_model_from_full_model_state_dict(
     for target_param_name, full_tensor in custom_param_sd.items():
         meta_sharded_param = meta_sd.get(target_param_name)
         if meta_sharded_param is None:
-            raise ValueError(
-                f"Parameter {target_param_name} not found in custom model state dict. The hf to custom mapping may be incorrect."
-            )
+            if strict:
+                raise ValueError(
+                    f"Parameter {target_param_name} not found in custom model state dict. The hf to custom mapping may be incorrect."
+                )
+            else:
+                logger.warning(
+                    f"Parameter '{target_param_name}' from checkpoint not found in model; skipping. This is expected for optional parameters."
+                )
+                continue
         if not hasattr(meta_sharded_param, "device_mesh"):
             full_tensor = full_tensor.to(device=device, dtype=param_dtype)
             actual_param = param_dict.get(target_param_name)