feat(rollout): forward --sglang-text-encoder-precisions to the engine

miles/backends/fsdp_utils/configs/sd3.py

@@ -26,6 +26,7 @@ class SD3TrainPipelineConfig(TrainPipelineConfig):
         "attn.to_add_out",
     ]
     needs_timestep_scaling = False
+    input_dtype_policy = {"latents": "default", "cond": None, "timestep": "fp32"}
 
     def prepare_cond_kwargs(self, cond: CondKwargs | None, device: torch.device) -> dict:
         if cond is None:

miles/backends/sglang_diffusion_utils/sglang_diffusion_engine.py

@@ -340,7 +340,7 @@ def _compute_server_args(args, host, port, nccl_port):
     # dit_precision / vae_precision are PipelineConfig fields, not ServerArgs, so forward them explicitly (only when changed from the class default, to avoid clobbering a subclass override).
     from sglang.multimodal_gen.configs.pipeline_configs.base import PipelineConfig
 
-    for field_name in ("dit_precision", "vae_precision"):
+    for field_name in ("dit_precision", "vae_precision", "text_encoder_precisions"):
         val = getattr(args, f"sglang_{field_name}", None)
         if val is not None and val != getattr(PipelineConfig, field_name, None):
             kwargs[field_name] = val

miles/utils/arguments.py

@@ -1420,6 +1420,7 @@ def add_sglang_tp_size():
         )
 
         parser.set_defaults(sglang_tensor_parallel_size=add_sglang_tp_size())
+        parser.set_defaults(sglang_text_encoder_precisions=None)
         return parser
 
     return add_miles_arguments