Upload app.py with huggingface_hub

app.py

@@ -45,10 +45,43 @@ _tokenizer_path = os.path.join(_base_dir, "google", "umt5-xxl")
 from env.loop_utils import load_pipeline_and_ckpt
 from env.run_replay_loop_two_chunk import run_one_chunk, encode_context_frames_per_frame
 from env.memory_baseline_runtime import MemoryProfile, infer_memory_profile_spec
-from inference.unified_inference import resolve_memory_profile, apply_profile_to_pipe
 from diffsynth import save_video
 from src.model_training.fov_retrieval import compute_rotation_list
 
+
+# ── Inline helpers from inference/unified_inference.py ──────────────────────
+def resolve_memory_profile(memory_type: str, ckpt_path: str) -> MemoryProfile:
+    """Resolve memory_type to a MemoryProfile. context_k* use default pipe flags."""
+    _CONTEXT_K_PROFILES = {
+        "context_k1": MemoryProfile(context_override=1),
+        "context_k5": MemoryProfile(context_override=5),
+        "context_k20": MemoryProfile(context_override=20),
+    }
+    if memory_type in _CONTEXT_K_PROFILES:
+        print(f"[app] Using context learning profile: {memory_type}")
+        return _CONTEXT_K_PROFILES[memory_type]
+    spec = infer_memory_profile_spec(ckpt_path)
+    if spec is not None:
+        print(f"[app] Auto-detected memory profile: {spec.profile_id}")
+        return spec.profile
+    return MemoryProfile()
+
+
+def apply_profile_to_pipe(pipe, profile: MemoryProfile) -> None:
+    """Apply a MemoryProfile directly to the pipeline object."""
+    pipe.use_framepack_memory = bool(profile.use_framepack_memory)
+    pipe.context_temporal_decay = float(profile.context_temporal_decay or 1.0)
+    pipe.context_attention_weight = float(profile.context_attention_weight or 1.0)
+    pipe.use_framepack_length_compress = bool(profile.use_framepack_length_compress)
+    pipe.framepack_ratio = int(profile.framepack_ratio or 2)
+    pipe.use_spatial_memory = bool(profile.use_spatial_memory)
+    pipe.spatial_memory_tokens = int(profile.spatial_memory_tokens or 64)
+    if profile.spatial_memory_inject_mode:
+        pipe.spatial_memory_inject_mode = str(profile.spatial_memory_inject_mode)
+    pipe.use_spatial_memory_legacy = bool(profile.use_spatial_memory_legacy)
+    pipe.use_block_wise_ssm = bool(getattr(profile, "use_block_wise_ssm", False))
+    pipe.use_videossm_hybrid = bool(getattr(profile, "use_videossm_hybrid", False))
+
 print("[app] Loading pipeline (DiT -> cuda)...")
 pipe = load_pipeline_and_ckpt(
     ckpt_path=_ckpt_path,