Update app.py

app.py

@@ -427,6 +427,17 @@ def load_models() -> str | None:
         print(f"[gen] loading {[c.name for c in blocks.expected_components]} from {MODEL_REPO} ...", flush=True)
         pipe = blocks.init_pipeline(MODEL_REPO, components_manager=manager, collection="h3")
         pipe.load_components(dtype=torch.bfloat16)
+
+        import h3_eros
+
+        eros_status = h3_eros.apply_eros(pipe.transformer)
+        if eros_status:
+            print(f"[gen] {eros_status}", flush=True)
+
+        import h3_proj_cache
+
+        h3_proj_cache.patch(pipe.transformer)
+
         pipe.transformer.set_attention_backend(ATTENTION)
 
         # --- Diagnostic: dump the LoRA files' key names/shapes and the transformer's own shapes to the Space
@@ -696,6 +707,7 @@ def _generate(
     maximize_gpu,
     video_shift,
     audio_shift,
+    proj_cache,
     sampler,
     total_steps,
     stage_from,
@@ -734,6 +746,15 @@ def _generate(
         }
         PIPE.transformer.set_adapters(list(active), weights=list(active.values()))
 
+    # Toggled per request, not patched per-call: `h3_proj_cache.patch` wrapped `forward` once in `load_models()`,
+    # and this just flips the attribute that wrapped method checks. `MOMENTUM_PIPE.transformer` is the same
+    # object as `PIPE.transformer` (see `load_models()`'s `update_components` call), so this covers the
+    # momentum path too, regardless of which `active_pipe` ends up running below. Called unconditionally, same
+    # reasoning as `set_adapters` above.
+    import h3_proj_cache
+
+    h3_proj_cache.set_enabled(PIPE.transformer, proj_cache)
+
     if PLACEMENT == "lazy":
         PIPE.to("cuda")
     elif PLACEMENT == "pack":
@@ -888,6 +909,7 @@ def generate(
     semantic_bridge_alpha=0.10,
     semantic_bridge_magnitude_match="per_token",
     semantic_bridge_adapter="semantic_bridge_v1",
+    proj_cache=True,
     progress=gr.Progress(track_tqdm=True),
     *,
     advance: bool = False,
@@ -1037,6 +1059,7 @@ def generate(
         bool(maximize_gpu),
         float(video_shift),
         float(audio_shift),
+        bool(proj_cache),
         SAMPLERS.get(sampler, "euler"),
         int(target_steps) if stage_enabled else None,
         steps_done if advance else 0,
@@ -1394,42 +1417,22 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
                     value=DEFAULT_INTERPOLATION,
                     info="MiniMax-H3 generates 24 fps; FILM synthesizes the frames in between.",
                 )
-                seed = gr.Number(label="Seed", value=42, precision=0)
                 upsample = gr.Checkbox(
                     label="Upsample prompt",
                     value=False,
                     info="Rewrite the prompt on the conditioner Space first, MiniMax's Context-IR style.",
                 )
-                with gr.Accordion("Semantic Bridge", open=False):
-                    gr.Markdown(
-                        "A small, independent adapter that nudges text conditioning toward better constraint "
-                        "adherence — left/right ordering, which hand holds what, material and reflection "
-                        "correctness — without touching the video transformer. Runs on the conditioner Space. "
-                        "Experimental; may help some prompts, do little on others."
-                    )
-                    semantic_bridge = gr.Checkbox(label="Enable Semantic Bridge", value=False)
-                    semantic_bridge_alpha = gr.Slider(
-                        label="alpha", minimum=0.0, maximum=1.0, step=0.01, value=0.10, visible=False
-                    )
-                    semantic_bridge_magnitude_match = gr.Dropdown(
-                        label="magnitude_match",
-                        choices=["per_token", "global", "none"],
-                        value="per_token",
-                        visible=False,
-                    )
-                    semantic_bridge_adapter = gr.Dropdown(
-                        label="adapter",
-                        choices=["semantic_bridge_v1", "bunny_action_logic_v1"],
-                        value="semantic_bridge_v1",
-                        visible=False,
-                        info="semantic_bridge_v1: broad (composition, spatial, materials). "
-                        "bunny_action_logic_v1: action/combat, character-action ownership in multi-character scenes.",
-                    )
                 maximize_gpu = gr.Checkbox(
                     label="Maximize Free Tier ZeroGPU (150 seconds)",
                     value=False,
                     info="Forces this request to book exactly 140s (plus 8s on the conditioner) for debugging purposes; does not prevent timeouts.",
                 )
+                proj_cache = gr.Checkbox(
+                    label="Cache keyframe projection",
+                    value=True,
+                    info="Skips recomputing the unchanging keyframe rows' proj_in output on every denoising step. "
+                    "Lossless — same output either way. Turn off only if you suspect it of causing an issue.",
+                )
 
         with gr.Column():
             video = gr.Video(label="Video + soundtrack", elem_id="h3-generated-video")
@@ -1572,7 +1575,33 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
                 )
                 advance_btn = gr.Button("Advance", variant="secondary", visible=False)
 
-            with gr.Accordion("Chunked Generation", open=False):
+            with gr.Accordion("Semantic Bridge", open=False):
+                gr.Markdown(
+                    "A small, independent adapter that nudges text conditioning toward better constraint "
+                    "adherence — left/right ordering, which hand holds what, material and reflection "
+                    "correctness — without touching the video transformer. Runs on the conditioner Space. "
+                    "Experimental; may help some prompts, do little on others."
+                )
+                semantic_bridge = gr.Checkbox(label="Enable Semantic Bridge", value=False)
+                semantic_bridge_alpha = gr.Slider(
+                    label="alpha", minimum=0.0, maximum=1.0, step=0.01, value=0.10, visible=False
+                )
+                semantic_bridge_magnitude_match = gr.Dropdown(
+                    label="magnitude_match",
+                    choices=["per_token", "global", "none"],
+                    value="per_token",
+                    visible=False,
+                )
+                semantic_bridge_adapter = gr.Dropdown(
+                    label="adapter",
+                    choices=["semantic_bridge_v1", "bunny_action_logic_v1"],
+                    value="semantic_bridge_v1",
+                    visible=False,
+                    info="semantic_bridge_v1: broad (composition, spatial, materials). "
+                    "bunny_action_logic_v1: action/combat, character-action ownership in multi-character scenes.",
+                )
+
+            with gr.Accordion("Chunked Generation", open=False, visible=False):
                 gr.Markdown(
                     "Splits a longer product into independent, full-quality chunks joined afterward — each "
                     "press is a complete generation, not a partial one. The previous chunk's last frame carries "
@@ -1676,6 +1705,7 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
         semantic_bridge_alpha,
         semantic_bridge_magnitude_match,
         semantic_bridge_adapter,
+        proj_cache,
     ]
 
     # `functools.partial` binds `advance`/`chunk_advance` by keyword regardless of their position in `generate`'s