add motus to loongforge vla

Change-Id: I8b7d9541e7b41f5d73b3a0fc6b719918b08cd03b

Author: fearblackcat

configs/models/embodied/motus.yaml

@@ -0,0 +1,71 @@
+# Motus VLA (WAN video + Qwen3-VL + Action Expert three-modal MoT) embodied config.
+# Values mirror the source repo configs/lerobot.yaml (model + dataset sections).
+model:
+  model_type: motus
+
+  # WAN video backbone
+  wan_config_path: "/workspace/motus/models/hf/Wan2.2-TI2V-5B"
+  wan_checkpoint_path: "/workspace/motus/models/hf/Wan2.2-TI2V-5B"
+  vae_path: "/workspace/motus/models/hf/Wan2.2-TI2V-5B/Wan2.2_VAE.pth"
+  video_precision: bfloat16
+
+  # VLM (frozen)
+  vlm_checkpoint_path: "/workspace/motus/models/hf/Qwen3-VL-2B-Instruct"
+  vlm_precision: bfloat16
+
+  # Task dimensions (shared with data side)
+  action_dim: 14
+  state_dim: 14
+
+  # Action expert
+  action_expert_hidden_size: 1024
+  action_expert_ffn_dim_multiplier: 4
+  action_expert_norm_eps: 1.0e-5
+
+  # Understanding expert
+  und_expert_hidden_size: 512
+  und_expert_ffn_dim_multiplier: 4
+  und_expert_norm_eps: 1.0e-5
+  vlm_adapter_input_dim: 2048
+  vlm_adapter_projector_type: mlp3x_silu
+
+  # MoT backbone depth
+  num_layers: 30
+
+  # Video / sampling geometry (shared with data side)
+  num_video_frames: 8
+  video_action_freq_ratio: 6
+  global_downsample_rate: 1
+  video_height: 384
+  video_width: 320
+
+  # Loss weights
+  video_loss_weight: 1.0
+  action_loss_weight: 1.0
+
+  # Training mode / switches
+  training_mode: finetune
+  load_pretrained_backbones: null
+
+data:
+  # LeRobot dataset location
+  repo_id: "/workspace/motus/data/aloha_mobile_cabinet"
+  root: "/workspace/motus/data/aloha_mobile_cabinet"
+
+  # Task selection
+  task_mode: single
+  task_name: null
+  max_episodes: null
+  image_aug: false
+
+  # Normalization / embodiment
+  embodiment_type: aloha_agilex_2
+
+  # Video decoding
+  video_backend: torchcodec
+
+  # T5 language embedding (on-the-fly encode + cache fallback)
+  enable_t5_fallback: true
+  t5_wan_path: "/workspace/motus/models/hf/Wan2.2-TI2V-5B/"
+  t5_folder_name: t5_embedding
+  t5_text_len: 512

examples/embodied/motus/diff_online_dumps.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+# Copyright 2026 The LoongForge Authors.
+# SPDX-License-Identifier: Apache-2.0
+"""Online-vs-online VAE-input parity check (DATA link).
+
+Compares the LATENT_DUMP dumps of TWO separate online training runs (both
+launched with the SAME PARITY_DATA_SEED so the sampler visits identical anchors
+in identical order). For each matching rankR_stepN.pt it bit-compares:
+
+  - first_frame        (decode + normalization + H2D)
+  - video_frames       (decode + normalization + H2D)
+  - clean_full_latent  (end-to-end: data + VAE compute)
+
+A `torch.equal` pass on first_frame / video_frames proves the data pipeline
+(torchcodec decode + transform + collate) is bit-deterministic ACROSS PROCESSES
+-- i.e. re-fetching the same (episode, condition_frame) anchor in the offline
+precompute will get the exact same input pixels. Combined with
+verify_offline_latents.py (compute link), a pass on both scripts means the
+offline latent cache is end-to-end bit-identical to the online encode.
+
+Usage
+-----
+  python examples/embodied/motus/diff_online_dumps.py \
+      --dir-a /tmp/latent_dump_run1 --dir-b /tmp/latent_dump_run2
+"""
+from __future__ import annotations
+
+import argparse
+import glob
+import os
+
+import torch
+
+
+def _cmp(a: torch.Tensor, b: torch.Tensor) -> str:
+    if a.shape != b.shape:
+        return f"SHAPE-MISMATCH {tuple(a.shape)} vs {tuple(b.shape)}"
+    if torch.equal(a, b):
+        return "equal"
+    diff = (a.float() - b.float()).abs().max().item()
+    return f"DIFFERS max|abs|={diff:.3e}"
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--dir-a", required=True, help="dump dir of online run #1")
+    ap.add_argument("--dir-b", required=True, help="dump dir of online run #2")
+    args = ap.parse_args()
+
+    files_a = sorted(glob.glob(os.path.join(args.dir_a, "rank*_step*.pt")))
+    if not files_a:
+        print(f"[FAIL] no dumps in {args.dir_a}")
+        return 2
+
+    all_ok = True
+    for fa in files_a:
+        name = os.path.basename(fa)
+        fb = os.path.join(args.dir_b, name)
+        if not os.path.exists(fb):
+            print(f"{name}: MISSING in dir-b")
+            all_ok = False
+            continue
+        da = torch.load(fa, map_location="cpu")
+        db = torch.load(fb, map_location="cpu")
+        ff = _cmp(da["first_frame"], db["first_frame"])
+        vf = _cmp(da["video_frames"], db["video_frames"])
+        lat = _cmp(da["clean_full_latent"], db["clean_full_latent"])
+        ok = ff == "equal" and vf == "equal"
+        all_ok = all_ok and ok
+        print(f"{name}: first_frame={ff} | video_frames={vf} | latent={lat}")
+
+    print()
+    if all_ok:
+        print("[PASS] data pipeline is bit-deterministic across processes.")
+        print("       -> offline precompute will get identical VAE inputs per anchor.")
+        return 0
+    print("[WARN] VAE inputs differ across runs -> decode/transform is not")
+    print("       cross-process deterministic; offline cache keys must pin the")
+    print("       exact decoded frames (store inputs, not just anchor ids).")
+    return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())