Delete app.py

app.py

@@ -1,524 +0,0 @@
-import os
-import sys
-import json
-import tempfile
-import subprocess
-import shutil
-from pathlib import Path
-
-import gradio as gr
-import spaces
-import numpy as np
-import torch
-
-# ──────────────────────────────────────────────
-# flash_attn install — must happen before any
-# lyra_2 import that pulls in flash_attn.
-# We pick the pre-built wheel that matches the
-# running torch+CUDA version automatically.
-# ──────────────────────────────────────────────
-
-def _install_flash_attn():
-    """
-    Install flash-attn from the pre-built wheels hosted on GitHub.
-    Matches the wheel to the running torch + CUDA version so the
-    .so symbols line up — which is exactly what caused the
-    'undefined symbol: _ZN3c104cuda...' error.
-    """
-    try:
-        import flash_attn          # already installed and importable → done
-        return
-    except ImportError:
-        pass
-
-    import torch, platform
-
-    torch_ver  = torch.__version__.split("+")[0].replace(".", "")   # e.g. "240"
-    cuda_ver   = torch.version.cuda.replace(".", "")                # e.g. "121"
-    py_ver     = f"cp{sys.version_info.major}{sys.version_info.minor}"  # e.g. "cp310"
-    arch       = platform.machine()                                 # "x86_64"
-
-    # Official pre-built wheel index from the flash-attn GitHub releases.
-    # Pattern: flash_attn-<fa_ver>+pt<torch>cu<cuda>-<py>-<py>-linux_<arch>.whl
-    # We try the newest FA2 release first then fall back to pip --no-build-isolation.
-    wheel_url = (
-        f"https://github.com/Dao-AILab/flash-attention/releases/download/"
-        f"v2.7.4.post1/"
-        f"flash_attn-2.7.4.post1+pt{torch_ver}cu{cuda_ver}-{py_ver}-{py_ver}"
-        f"-linux_{arch}.whl"
-    )
-
-    print(f"[Lyra] Installing flash-attn wheel: {wheel_url}")
-    result = subprocess.run(
-        [sys.executable, "-m", "pip", "install", wheel_url, "--no-deps", "-q"],
-        capture_output=True, text=True,
-    )
-    if result.returncode != 0:
-        print(f"[Lyra] Pre-built wheel not found for this env, "
-              f"falling back to pip install flash-attn --no-build-isolation ...")
-        # This compiles from source — slow (~20 min) but always works.
-        subprocess.run(
-            [sys.executable, "-m", "pip", "install", "flash-attn",
-             "--no-build-isolation", "-q"],
-            check=True,
-        )
-
-    try:
-        import flash_attn
-        print(f"[Lyra] flash_attn {flash_attn.__version__} ready.")
-    except ImportError as e:
-        raise RuntimeError(f"flash_attn install succeeded but import still fails: {e}")
-
-
-_install_flash_attn()
-
-# ──────────────────────────────────────────────
-# Paths
-# ──────────────────────────────────────────────
-REPO_ROOT = Path(__file__).parent.resolve()
-CKPT_DIR  = REPO_ROOT / "checkpoints"
-
-# Sentinel files — if any are missing we trigger a fresh download
-_REQUIRED_FILES = [
-    CKPT_DIR / "text_encoder" / "negative_prompt.pt",
-    CKPT_DIR / "model",          # directory is enough as a sentinel
-]
-
-_ckpts_ready = False
-
-
-# ──────────────────────────────────────────────
-# Checkpoint helpers
-# ──────────────────────────────────────────────
-
-def _checkpoints_present() -> bool:
-    for p in _REQUIRED_FILES:
-        if not p.exists():
-            print(f"[Lyra] Missing checkpoint path: {p}")
-            return False
-    return True
-
-
-def _download_checkpoints() -> str:
-    """
-    Download all checkpoints from nvidia/Lyra-2.0 on HuggingFace Hub.
-    Uses snapshot_download so the full checkpoints/ tree lands at
-    REPO_ROOT/checkpoints/ — exactly where the inference scripts expect them.
-    Returns a status string for display in the UI.
-    """
-    if _checkpoints_present():
-        return "✅ Checkpoints already present — skipping download."
-
-    print("[Lyra] Checkpoints not found — downloading from nvidia/Lyra-2.0 …")
-
-    try:
-        from huggingface_hub import snapshot_download
-    except ImportError:
-        subprocess.run(
-            [sys.executable, "-m", "pip", "install", "huggingface_hub", "-q"],
-            check=True,
-        )
-        from huggingface_hub import snapshot_download
-
-    snapshot_download(
-        repo_id="nvidia/Lyra-2.0",
-        allow_patterns=["checkpoints/**"],   # only grab the checkpoints subtree
-        local_dir=str(REPO_ROOT),            # → REPO_ROOT/checkpoints/...
-        local_dir_use_symlinks=False,
-    )
-
-    if _checkpoints_present():
-        msg = "✅ Checkpoints downloaded successfully."
-        print(f"[Lyra] {msg}")
-        return msg
-
-    msg = ("❌ Download finished but sentinel files are still missing. "
-           "Check Space logs and available disk space (~50 GB required).")
-    print(f"[Lyra] {msg}")
-    return msg
-
-
-def _ensure_models():
-    """Called before every inference run — downloads checkpoints if needed."""
-    global _ckpts_ready
-    if _ckpts_ready:
-        return
-    _download_checkpoints()
-    _ckpts_ready = _checkpoints_present()
-    if not _ckpts_ready:
-        raise RuntimeError(
-            "Checkpoints unavailable. Please click 'Download Checkpoints', "
-            "check the Space logs, and ensure ~50 GB of storage is available."
-        )
-
-
-# ── Kick off download at module load so it's done before the first request ──
-print("[Lyra] Startup checkpoint check …")
-_download_checkpoints()
-_ckpts_ready = _checkpoints_present()
-
-
-# ──────────────────────────────────────────────
-# Subprocess helper
-# ──────────────────────────────────────────────
-
-def _run(cmd: list, desc: str = "") -> tuple:
-    """
-    Run a `python -m …` command with:
-      • sys.executable  → correct venv interpreter
-      • PYTHONPATH      → REPO_ROOT so lyra_2._src.* is importable
-      • cwd             → REPO_ROOT so 'checkpoints/model' resolves correctly
-    """
-    if cmd[0] in ("python", "python3"):
-        cmd = [sys.executable] + cmd[1:]
-
-    run_env = {
-        **os.environ,
-        "PYTHONPATH": str(REPO_ROOT),
-        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
-    }
-    print(f"[Lyra] {desc}: {' '.join(str(c) for c in cmd)}")
-    result = subprocess.run(
-        cmd,
-        capture_output=True,
-        text=True,
-        env=run_env,
-        cwd=str(REPO_ROOT),
-    )
-    log = result.stdout + "\n" + result.stderr
-    return result.returncode == 0, log
-
-
-# ──────────────────────────────────────────────
-# Zoom-in / Zoom-out trajectory  (Option 1)
-# ──────────────────────────────────────────────
-
-@spaces.GPU(duration=900)
-def run_zoomgs(
-    image,
-    caption,
-    zoom_in_strength,
-    zoom_out_strength,
-    num_frames_in,
-    num_frames_out,
-    use_dmd,
-    run_reconstruction,
-):
-    _ensure_models()
-
-    with tempfile.TemporaryDirectory() as tmp:
-        from PIL import Image as PILImage
-
-        img_path     = Path(tmp) / "input.png"
-        caption_path = Path(tmp) / "input.txt"
-        PILImage.fromarray(image).save(img_path)
-        caption_path.write_text(caption.strip() or "A scenic outdoor environment.")
-
-        output_dir = Path(tmp) / "outputs" / "zoomgs"
-        output_dir.mkdir(parents=True, exist_ok=True)
-
-        cmd = [
-            "python", "-m", "lyra_2._src.inference.lyra2_zoomgs_inference",
-            "--input_image_path", str(tmp),
-            "--sample_id", "0",
-            "--experiment", "lyra2",
-            "--checkpoint_dir", str(CKPT_DIR / "model"),
-            "--prompt_dir",     str(tmp),
-            "--output_path",    str(output_dir),
-            "--num_frames_zoom_in",  str(int(num_frames_in)),
-            "--num_frames_zoom_out", str(int(num_frames_out)),
-            "--zoom_in_strength",    str(zoom_in_strength),
-            "--zoom_out_strength",   str(zoom_out_strength),
-        ]
-        if use_dmd:
-            cmd.append("--use_dmd")
-
-        ok, log = _run(cmd, "ZoomGS video generation")
-
-        # The script writes to <output_dir>/<sample_id>/videos/<sample_id>.mp4
-        video_path = output_dir / "0" / "videos" / "0.mp4"
-        if not video_path.exists():
-            candidates = list(output_dir.rglob("*.mp4"))
-            video_path = candidates[0] if candidates else None
-
-        gs_video = None
-        if run_reconstruction and video_path and Path(video_path).exists():
-            ok2, log2 = _run(
-                ["python", "-m", "lyra_2._src.inference.vipe_da3_gs_recon",
-                 "--input_video_path", str(video_path)],
-                "GS reconstruction",
-            )
-            log += "\n" + log2
-            gs_candidates = list(output_dir.rglob("gs_trajectory.mp4"))
-            if gs_candidates:
-                gs_video = str(gs_candidates[0])
-
-        return (
-            str(video_path) if video_path and Path(video_path).exists() else None,
-            gs_video,
-            log[-4000:],
-        )
-
-
-# ──────────────────────────────────────────────
-# Custom trajectory  (Option 2)
-# ────────────────────────────────��─────────────
-
-@spaces.GPU(duration=900)
-def run_custom_traj(
-    image,
-    trajectory_file,
-    captions_json,
-    num_frames,
-    pose_scale,
-    use_dmd,
-    run_reconstruction,
-):
-    _ensure_models()
-
-    with tempfile.TemporaryDirectory() as tmp:
-        from PIL import Image as PILImage
-
-        img_path = Path(tmp) / "first_frame.png"
-        PILImage.fromarray(image).save(img_path)
-
-        traj_path = Path(tmp) / "trajectory.npz"
-        shutil.copy(trajectory_file.name, traj_path)
-
-        captions_path = Path(tmp) / "captions.json"
-        try:
-            json.loads(captions_json)
-            captions_path.write_text(captions_json)
-        except (json.JSONDecodeError, TypeError):
-            captions_path.write_text(json.dumps({"0": captions_json or ""}))
-
-        output_dir = Path(tmp) / "outputs" / "custom"
-        output_dir.mkdir(parents=True, exist_ok=True)
-
-        cmd = [
-            "python", "-m", "lyra_2._src.inference.lyra2_custom_traj_inference",
-            "--input_image_path", str(img_path),
-            "--trajectory_path",  str(traj_path),
-            "--experiment", "lyra2",
-            "--checkpoint_dir", str(CKPT_DIR / "model"),
-            "--captions_path",  str(captions_path),
-            "--num_frames",     str(int(num_frames)),
-            "--output_path",    str(output_dir),
-            "--pose_scale",     str(pose_scale),
-        ]
-        if use_dmd:
-            cmd.append("--use_dmd")
-
-        ok, log = _run(cmd, "Custom trajectory video generation")
-
-        video_candidates = list(output_dir.rglob("*.mp4"))
-        video_path = video_candidates[0] if video_candidates else None
-
-        gs_video = None
-        if run_reconstruction and video_path:
-            ok2, log2 = _run(
-                ["python", "-m", "lyra_2._src.inference.vipe_da3_gs_recon",
-                 "--input_video_path", str(video_path)],
-                "GS reconstruction",
-            )
-            log += "\n" + log2
-            gs_candidates = list(output_dir.rglob("gs_trajectory.mp4"))
-            if gs_candidates:
-                gs_video = str(gs_candidates[0])
-
-        return (
-            str(video_path) if video_path else None,
-            gs_video,
-            log[-4000:],
-        )
-
-
-# ──────────────────────────────────────────────
-# Manual download button handler
-# ──────────────────────────────────────────────
-
-def manual_download():
-    global _ckpts_ready
-    msg = _download_checkpoints()
-    _ckpts_ready = _checkpoints_present()
-    return msg
-
-
-# ──────────────────────────────────────────────
-# UI
-# ──────────────────────────────────────────────
-
-CSS = """
-@import url('https://fonts.googleapis.com/css2?family=Syne:wght@400;600;700;800&family=DM+Mono:wght@300;400;500&display=swap');
-:root {
-    --bg:#0a0c10; --surface:#111318; --border:#1e2230;
-    --accent:#5affb0; --accent2:#a78bfa; --text:#e8eaf0; --muted:#5a5f72;
-    --radius:12px; --font-head:'Syne',sans-serif; --font-mono:'DM Mono',monospace;
-}
-body,.gradio-container{background:var(--bg)!important;color:var(--text)!important;font-family:var(--font-head)!important;}
-#header{background:linear-gradient(135deg,#0d1117 0%,#161b27 60%,#0f1520 100%);border:1px solid var(--border);border-radius:var(--radius);padding:32px 40px 28px;margin-bottom:24px;position:relative;overflow:hidden;}
-#header::before{content:'';position:absolute;inset:0;background:radial-gradient(ellipse 70% 60% at 80% 50%,rgba(94,255,176,.06) 0%,transparent 70%),radial-gradient(ellipse 50% 80% at 20% 80%,rgba(167,139,250,.06) 0%,transparent 70%);pointer-events:none;}
-#header h1{font-size:2.4rem;font-weight:800;letter-spacing:-.02em;margin:0 0 8px;background:linear-gradient(90deg,var(--accent) 0%,var(--accent2) 100%);-webkit-background-clip:text;-webkit-text-fill-color:transparent;}
-#header p{color:var(--muted);font-family:var(--font-mono);font-size:.85rem;margin:0;}
-#header .badge{display:inline-block;margin-right:8px;padding:3px 10px;background:rgba(94,255,176,.1);border:1px solid rgba(94,255,176,.25);border-radius:20px;color:var(--accent);font-size:.75rem;font-family:var(--font-mono);}
-.tab-nav button{background:transparent!important;border:none!important;border-bottom:2px solid transparent!important;color:var(--muted)!important;font-family:var(--font-head)!important;font-weight:600!important;font-size:.95rem!important;padding:10px 20px!important;transition:all .2s!important;}
-.tab-nav button.selected,.tab-nav button:hover{color:var(--accent)!important;border-bottom-color:var(--accent)!important;background:transparent!important;}
-.gr-panel,.gr-box,.gradio-group{background:var(--surface)!important;border:1px solid var(--border)!important;border-radius:var(--radius)!important;}
-input,textarea,.gr-input,.gr-textbox textarea{background:#0d0f14!important;border:1px solid var(--border)!important;color:var(--text)!important;font-family:var(--font-mono)!important;border-radius:8px!important;}
-input:focus,textarea:focus{border-color:var(--accent)!important;box-shadow:0 0 0 2px rgba(94,255,176,.12)!important;}
-input[type=range]{accent-color:var(--accent)!important;}
-button.primary,.gr-button-primary{background:linear-gradient(135deg,var(--accent) 0%,#38d9a9 100%)!important;color:#0a0c10!important;font-family:var(--font-head)!important;font-weight:700!important;border:none!important;border-radius:8px!important;padding:12px 28px!important;font-size:.95rem!important;transition:opacity .2s!important;}
-button.primary:hover{opacity:.85!important;}
-button.secondary,.gr-button-secondary{background:transparent!important;border:1px solid var(--border)!important;color:var(--muted)!important;font-family:var(--font-head)!important;border-radius:8px!important;}
-label,.gr-form>label,.block>label span{color:var(--muted)!important;font-family:var(--font-mono)!important;font-size:.8rem!important;letter-spacing:.04em!important;text-transform:uppercase!important;}
-#log-box textarea{font-size:.78rem!important;color:#7af0b0!important;background:#060709!important;}
-.info-note{background:rgba(167,139,250,.07);border:1px solid rgba(167,139,250,.2);border-radius:8px;padding:12px 16px;font-family:var(--font-mono);font-size:.8rem;color:#c4b5fd;line-height:1.6;}
-"""
-
-
-def build_app():
-    initial_status = (
-        "✅ Checkpoints ready."
-        if _ckpts_ready else
-        "⚠️  Checkpoints not found locally. Click 'Download Checkpoints' or they will be fetched automatically on first inference."
-    )
-
-    with gr.Blocks() as demo:
-
-        gr.HTML("""
-        <div id="header">
-          <h1>✦ Lyra 2.0</h1>
-          <p>
-            <span class="badge">NVIDIA Research</span>
-            <span class="badge">3D Gaussian Splatting</span>
-            <span class="badge">arXiv 2604.13036</span>
-          </p>
-          <p style="margin-top:14px;color:#8892a4;font-size:.9rem;font-family:'Syne',sans-serif;">
-            Generate persistent, explorable 3D worlds from a single image —
-            no spatial forgetting, no temporal drift.
-          </p>
-        </div>
-        """)
-
-        # ── Checkpoint status bar ────────────────────
-        with gr.Row():
-            ckpt_status = gr.Textbox(
-                value=initial_status, label="Checkpoint Status",
-                interactive=False, scale=4,
-            )
-            ckpt_btn = gr.Button("⬇️ Download Checkpoints", variant="secondary", scale=1)
-        ckpt_btn.click(fn=manual_download, outputs=ckpt_status)
-
-        with gr.Tabs():
-
-            # ── Tab 1: Zoom Trajectory ───────────────
-            with gr.Tab("🔭 Zoom Trajectory"):
-                gr.HTML('<div class="info-note">Generate a zoom-in → zoom-out exploration video from a single image, then optionally lift it to a 3D Gaussian Splatting scene.</div>')
-                with gr.Row():
-                    with gr.Column(scale=1):
-                        z_image   = gr.Image(label="Input Image", type="numpy", height=280)
-                        z_caption = gr.Textbox(
-                            label="Scene Caption",
-                            placeholder="A sunlit forest clearing with tall pine trees…",
-                            lines=2,
-                        )
-                        with gr.Accordion("Advanced Options", open=False):
-                            with gr.Row():
-                                z_in_str  = gr.Slider(0.1, 3.0, value=0.5,  step=0.1, label="Zoom-in Strength")
-                                z_out_str = gr.Slider(0.1, 3.0, value=1.5,  step=0.1, label="Zoom-out Strength")
-                            with gr.Row():
-                                z_frames_in  = gr.Slider(81, 401, value=81,  step=80, label="Frames Zoom-in (1+80k)")
-                                z_frames_out = gr.Slider(81, 401, value=241, step=80, label="Frames Zoom-out (1+80k)")
-                            with gr.Row():
-                                z_dmd   = gr.Checkbox(label="⚡ Fast Mode (DMD ×15 speedup, lower quality)", value=False)
-                                z_recon = gr.Checkbox(label="🧊 Run 3DGS Reconstruction after video", value=True)
-                        z_btn = gr.Button("Generate World", variant="primary")
-                    with gr.Column(scale=1):
-                        z_video  = gr.Video(label="Generated Exploration Video", height=280)
-                        z_gs_vid = gr.Video(label="3DGS Flythrough", height=280)
-                        z_log    = gr.Textbox(label="Log", lines=8, interactive=False, elem_id="log-box")
-
-                z_btn.click(
-                    fn=run_zoomgs,
-                    inputs=[z_image, z_caption,
-                            z_in_str, z_out_str, z_frames_in, z_frames_out,
-                            z_dmd, z_recon],
-                    outputs=[z_video, z_gs_vid, z_log],
-                )
-
-            # ── Tab 2: Custom Trajectory ─────────────
-            with gr.Tab("🎮 Custom Trajectory"):
-                gr.HTML('<div class="info-note">Provide a custom camera trajectory (.npz with <code>w2c</code>, <code>intrinsics</code>, <code>image_height</code>, <code>image_width</code>) and per-chunk captions as JSON keyed by frame index.</div>')
-                with gr.Row():
-                    with gr.Column(scale=1):
-                        c_image    = gr.Image(label="First Frame", type="numpy", height=240)
-                        c_traj     = gr.File(label="Trajectory (.npz)", file_types=[".npz"])
-                        c_captions = gr.Textbox(
-                            label='Per-chunk Captions (JSON or single string)',
-                            placeholder='{"0": "A grand hall interior", "81": "Corridor leading outside"}',
-                            lines=3,
-                        )
-                        with gr.Accordion("Advanced Options", open=False):
-                            with gr.Row():
-                                c_frames     = gr.Slider(81, 961, value=481, step=80, label="Num Frames (1+80k)")
-                                c_pose_scale = gr.Slider(0.1, 5.0, value=1.0, step=0.1, label="Pose Scale")
-                            with gr.Row():
-                                c_dmd   = gr.Checkbox(label="⚡ Fast Mode (DMD)", value=False)
-                                c_recon = gr.Checkbox(label="🧊 Run 3DGS Reconstruction", value=True)
-                        c_btn = gr.Button("Generate World", variant="primary")
-                    with gr.Column(scale=1):
-                        c_video  = gr.Video(label="Generated Video", height=260)
-                        c_gs_vid = gr.Video(label="3DGS Flythrough", height=260)
-                        c_log    = gr.Textbox(label="Log", lines=8, interactive=False, elem_id="log-box")
-
-                c_btn.click(
-                    fn=run_custom_traj,
-                    inputs=[c_image, c_traj, c_captions,
-                            c_frames, c_pose_scale, c_dmd, c_recon],
-                    outputs=[c_video, c_gs_vid, c_log],
-                )
-
-            # ── Tab 3: About ─────────────────────────
-            with gr.Tab("ℹ️ About"):
-                gr.Markdown("""
-## Lyra 2.0 — Explorable Generative 3D Worlds
-
-**NVIDIA Research** · [Paper](https://arxiv.org/abs/2604.13036) · [Project Page](https://research.nvidia.com/labs/sil/projects/lyra2/) · [HuggingFace](https://huggingface.co/nvidia/Lyra-2.0)
-
-### How it works
-| Problem | Solution |
-|---|---|
-| **Spatial Forgetting** — revisited regions hallucinated when outside temporal context | Per-frame 3D geometry used for routing: retrieve past frames + dense correspondences |
-| **Temporal Drifting** — autoregressive errors accumulate over long trajectories | Self-augmented histories teach the model to correct drift rather than propagate it |
-
-### Checkpoint layout
-```
-checkpoints/
-├── model/                        ← main generation model weights
-└── text_encoder/
-    └── negative_prompt.pt        ← required for CFG negative guidance
-```
-Checkpoints are auto-downloaded (~50 GB) from `nvidia/Lyra-2.0` on HuggingFace at startup.
-
-### GPU requirements
-- Recommended: **H100 80 GB** or A100 80 GB
-- ~9 min / 80 frames full quality · ~35 s with DMD fast mode
-- GS reconstruction adds ~1 min
-
-### Citation
-```bibtex
-@article{shen2026lyra2,
-  title   = {Lyra 2.0: Explorable Generative 3D Worlds},
-  author  = {Shen, Tianchang and Bahmani, Sherwin and He, Kai and others},
-  journal = {arXiv preprint arXiv:2604.13036},
-  year    = {2026}
-}
-```
-                """)
-
-    return demo
-
-
-if __name__ == "__main__":
-    demo = build_app()
-    demo.launch(css=CSS)