embedder权重改成bf16

似乎因为权重的处理更慢了，整体速度反而变慢了一点点

scripts/evaluation/world_model_interaction.py

@@ -772,7 +772,11 @@ def image_guided_synthesis_sim_mode(
     with profiler.profile_section("synthesis/conditioning_prep"):
         img = observation['observation.images.top'].permute(0, 2, 1, 3, 4)
         cond_img = rearrange(img, 'b o c h w -> (b o) c h w')[-1:]
-        cond_img_emb = model.embedder(cond_img)
+        embedder_ctx = nullcontext()
+        if getattr(model, "encoder_bf16", False) and model.device.type == "cuda":
+            embedder_ctx = torch.autocast("cuda", dtype=torch.bfloat16)
+        with embedder_ctx:
+            cond_img_emb = model.embedder(cond_img)
 
         if model.model.conditioning_key == 'hybrid':
             z = get_latent_z(model, img.permute(0, 2, 1, 3, 4))
@@ -912,6 +916,16 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
         diffusion_autocast_dtype = torch.bfloat16
         print(">>> diffusion backbone set to bfloat16")
 
+    encoder_dtype = torch.float32
+    if args.encoder_dtype == "bf16":
+        encoder_dtype = torch.bfloat16
+    if hasattr(model, "cond_stage_model") and model.cond_stage_model is not None:
+        model.cond_stage_model.to(dtype=encoder_dtype)
+    if hasattr(model, "embedder") and model.embedder is not None:
+        model.embedder.to(dtype=encoder_dtype)
+    model.encoder_bf16 = args.encoder_dtype == "bf16"
+    print(f">>> encoder dtype set to {args.encoder_dtype}")
+
     if hasattr(model, "projector_bf16"):
         model.projector_bf16 = args.projector_dtype == "bf16"
         print(f">>> projector dtype set to {args.projector_dtype}")
@@ -1266,6 +1280,13 @@ def get_parser():
         default="fp32",
         help="Dtype for image/state/action projectors (autocast in forward)."
     )
+    parser.add_argument(
+        "--encoder_dtype",
+        type=str,
+        choices=["fp32", "bf16"],
+        default="fp32",
+        help="Dtype for text/image encoders (cond_stage_model/embedder)."
+    )
     parser.add_argument(
         "--n_action_steps",
         type=int,

unitree_g1_pack_camera/case1/run_world_model_interaction.sh

@@ -22,5 +22,6 @@ dataset="unitree_g1_pack_camera"
         --guidance_rescale 0.7 \
         --perframe_ae \
         --diffusion_dtype bf16 \
-        --projector_dtype bf16
+        --projector_dtype bf16 \
+        --encoder_dtype bf16
 } 2>&1 | tee "${res_dir}/output.log"

useful.sh

@@ -78,4 +78,7 @@ BF16 projector比FP32 projector更准的可能原因：
   - 训练分布匹配：训练时你是 precision:16，projector 长期在低精度环境下被优化；推理用 FP32 反而可能偏离训练时的统计特性。
   - LayerNorm/Softmax 敏感：Resampler/MLP 里 LN/Softmax 对精度很敏感，FP32 计算后再降精度，数值边界更容易“硬截断”；BF16 全程计算可能更平滑。
 
-  这也解释了为什么你看到 BF16 projector 反而更准。
+  这也解释了为什么你看到 BF16 projector 反而更准。
+
+embedder：
+  改成 autocast only（权重 FP32，预处理 FP32，仅主干 BF16）