VAE 也做 BF16

这个权重不做修改更好精度

scripts/evaluation/world_model_interaction.py

@@ -664,6 +664,10 @@ def get_latent_z(model, videos: Tensor) -> Tensor:
     with profiler.profile_section("get_latent_z/encode"):
         b, c, t, h, w = videos.shape
         x = rearrange(videos, 'b c t h w -> (b t) c h w')
+        vae_ctx = nullcontext()
+        if getattr(model, "vae_bf16", False) and model.device.type == "cuda":
+            vae_ctx = torch.autocast("cuda", dtype=torch.bfloat16)
+        with vae_ctx:
             z = model.encode_first_stage(x)
         z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t)
     return z
@@ -879,6 +883,15 @@ def image_guided_synthesis_sim_mode(
 
         # Reconstruct from latent to pixel space
         with profiler.profile_section("synthesis/decode_first_stage"):
+            if getattr(model, "vae_bf16", False):
+                if samples.dtype != torch.bfloat16:
+                    samples = samples.to(dtype=torch.bfloat16)
+                vae_ctx = nullcontext()
+                if model.device.type == "cuda":
+                    vae_ctx = torch.autocast("cuda", dtype=torch.bfloat16)
+                with vae_ctx:
+                    batch_images = model.decode_first_stage(samples)
+            else:
                 if samples.dtype != torch.float32:
                     samples = samples.float()
                 batch_images = model.decode_first_stage(samples)
@@ -944,6 +957,14 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
         diffusion_autocast_dtype = torch.bfloat16
         print(">>> diffusion backbone set to bfloat16")
 
+    if hasattr(model, "first_stage_model") and model.first_stage_model is not None:
+        if args.vae_dtype == "bf16":
+            model.first_stage_model.to(dtype=torch.bfloat16)
+        else:
+            model.first_stage_model.to(dtype=torch.float32)
+    model.vae_bf16 = args.vae_dtype == "bf16"
+    print(f">>> VAE dtype set to {args.vae_dtype}")
+
     encoder_mode = args.encoder_mode
     encoder_bf16 = encoder_mode in ("autocast", "bf16_full")
     encoder_weight_dtype = torch.bfloat16 if encoder_mode == "bf16_full" else torch.float32
@@ -957,9 +978,21 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
         f">>> encoder mode set to {encoder_mode} (weights={encoder_weight_dtype})"
     )
 
+    projector_mode = args.projector_mode
+    projector_bf16 = projector_mode in ("autocast", "bf16_full")
+    projector_weight_dtype = torch.bfloat16 if projector_mode == "bf16_full" else torch.float32
+    if hasattr(model, "image_proj_model") and model.image_proj_model is not None:
+        model.image_proj_model.to(dtype=projector_weight_dtype)
+    if hasattr(model, "state_projector") and model.state_projector is not None:
+        model.state_projector.to(dtype=projector_weight_dtype)
+    if hasattr(model, "action_projector") and model.action_projector is not None:
+        model.action_projector.to(dtype=projector_weight_dtype)
     if hasattr(model, "projector_bf16"):
-        model.projector_bf16 = args.projector_dtype == "bf16"
+        model.projector_bf16 = projector_bf16
-        print(f">>> projector dtype set to {args.projector_dtype}")
+    model.projector_mode = projector_mode
+    print(
+        f">>> projector mode set to {projector_mode} (weights={projector_weight_dtype})"
+    )
 
     log_inference_precision(model)
 
@@ -1305,11 +1338,14 @@ def get_parser():
         help="Dtype for diffusion backbone weights and sampling autocast."
     )
     parser.add_argument(
-        "--projector_dtype",
+        "--projector_mode",
         type=str,
-        choices=["fp32", "bf16"],
+        choices=["fp32", "autocast", "bf16_full"],
         default="fp32",
-        help="Dtype for image/state/action projectors (autocast in forward)."
+        help=
+        "Projector precision mode for image/state/action projectors: "
+        "fp32=full fp32, autocast=fp32 weights + bf16 autocast in forward, "
+        "bf16_full=bf16 weights + bf16 forward."
     )
     parser.add_argument(
         "--encoder_mode",
@@ -1321,6 +1357,13 @@ def get_parser():
         "fp32=full fp32, autocast=fp32 weights + bf16 autocast in forward, "
         "bf16_full=bf16 weights + bf16 forward."
     )
+    parser.add_argument(
+        "--vae_dtype",
+        type=str,
+        choices=["fp32", "bf16"],
+        default="fp32",
+        help="Dtype for VAE/first_stage_model weights and forward autocast."
+    )
     parser.add_argument(
         "--n_action_steps",
         type=int,

src/unifolm_wma/models/ddpms.py

@@ -2032,6 +2032,13 @@ class LatentVisualDiffusion(LatentDiffusion):
                            target_dtype: torch.dtype | None) -> Tensor:
         use_bf16 = (self.projector_bf16 and x.device.type == "cuda"
                     and torch.cuda.is_bf16_supported())
+        if not use_bf16:
+            weight_dtype = None
+            for param in projector.parameters():
+                weight_dtype = param.dtype
+                break
+            if weight_dtype is not None and x.dtype != weight_dtype:
+                x = x.to(dtype=weight_dtype)
         if use_bf16:
             with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
                 out = projector(x)

unitree_g1_pack_camera/case1/run_world_model_interaction.sh

@@ -22,6 +22,7 @@ dataset="unitree_g1_pack_camera"
         --guidance_rescale 0.7 \
         --perframe_ae \
         --diffusion_dtype bf16 \
-        --projector_dtype bf16 \
+        --projector_mode autocast \
-        --encoder_mode autocast #fp32/autocast/bf16_full
+        --encoder_mode bf16_full \
+        --vae_dtype bf16
 } 2>&1 | tee "${res_dir}/output.log"