app.py

import os
import subprocess
import sys

# Disable torch.compile / dynamo before any torch import
os.environ["TORCH_COMPILE_DISABLE"] = "1"
os.environ["TORCHDYNAMO_DISABLE"] = "1"

# Install xformers for memory-efficient attention
subprocess.run([sys.executable, "-m", "pip", "install", "xformers==0.0.32.post2", "--no-build-isolation"], check=False)

# Clone LTX-2 repo and install packages
LTX_REPO_URL = "https://github.com/Lightricks/LTX-2.git"
LTX_REPO_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "LTX-2")
LTX_COMMIT_SHA = "ae855f8538843825f9015a419cf4ba5edaf5eec2"

if not os.path.exists(LTX_REPO_DIR):
    print(f"Cloning {LTX_REPO_URL}...")
    os.makedirs(LTX_REPO_DIR)
    subprocess.run(["git", "init", LTX_REPO_DIR], check=True)
    subprocess.run(["git", "remote", "add", "origin", LTX_REPO_URL], cwd=LTX_REPO_DIR, check=True)
    subprocess.run(["git", "fetch", "--depth", "1", "origin", LTX_COMMIT_SHA], cwd=LTX_REPO_DIR, check=True)
    subprocess.run(["git", "checkout", LTX_COMMIT_SHA], cwd=LTX_REPO_DIR, check=True)


print("Installing ltx-core and ltx-pipelines from cloned repo...")
subprocess.run(
    [sys.executable, "-m", "pip", "install", "--force-reinstall", "--no-deps", "-e",
     os.path.join(LTX_REPO_DIR, "packages", "ltx-core"),
     "-e", os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines")],
    check=True,
)

sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines", "src"))
sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-core", "src"))

import logging
import random
import tempfile
from pathlib import Path

import torch
torch._dynamo.config.suppress_errors = True
torch._dynamo.config.disable = True

import spaces
import gradio as gr
import numpy as np
from huggingface_hub import hf_hub_download, snapshot_download

from ltx_core.model.video_vae import TilingConfig, get_video_chunks_number
from ltx_core.quantization import QuantizationPolicy
from ltx_pipelines.distilled import DistilledPipeline
from ltx_pipelines.utils.args import ImageConditioningInput
from ltx_pipelines.utils.media_io import encode_video

# Force-patch xformers attention into the LTX attention module.
from ltx_core.model.transformer import attention as _attn_mod
print(f"[ATTN] Before patch: memory_efficient_attention={_attn_mod.memory_efficient_attention}")
try:
    from xformers.ops import memory_efficient_attention as _mea
    _attn_mod.memory_efficient_attention = _mea
    print(f"[ATTN] After patch: memory_efficient_attention={_attn_mod.memory_efficient_attention}")
except Exception as e:
    print(f"[ATTN] xformers patch FAILED: {type(e).__name__}: {e}")

# Disable xformers FA3 dispatch: FA3 kernels are Hopper-only (sm_90a), but
# xformers' dispatcher gates them on `device_capability >= (9, 0)`, which also
# matches Blackwell (RTX PRO 6000, the ZeroGPU fleet hardware since 2026-05-12)
# and crashes at kernel launch with "invalid argument".
try:
    from xformers.ops.fmha import _set_use_fa3
    _set_use_fa3(False)
    print("[ATTN] xformers FA3 dispatch disabled (Blackwell-incompatible)")
except Exception as e:
    print(f"[ATTN] FA3 disable FAILED: {type(e).__name__}: {e}")

# FUSE/mmap workaround: SafetensorsStateDictLoader.load uses safetensors.safe_open
# under the hood, which mmap's the file. On bucket FUSE mounts that triggers a
# page-fault storm and deadlocks loading. Bypass mmap by parsing the safetensors
# header ourselves and reading each tensor's bytes directly.
import json
import struct

from ltx_core.loader.primitives import StateDict
from ltx_core.loader.sft_loader import SafetensorsStateDictLoader

_SAFETENSORS_DTYPE_MAP = {
    "F64": torch.float64,
    "F32": torch.float32,
    "F16": torch.float16,
    "BF16": torch.bfloat16,
    "F8_E5M2": torch.float8_e5m2,
    "F8_E4M3": torch.float8_e4m3fn,
    "I64": torch.int64,
    "I32": torch.int32,
    "I16": torch.int16,
    "I8": torch.int8,
    "U8": torch.uint8,
    "BOOL": torch.bool,
}


def _patched_load(self, path, sd_ops, device=None):
    sd = {}
    size = 0
    dtype = set()
    device = device or torch.device("cpu")
    model_paths = path if isinstance(path, list) else [path]
    for shard_path in model_paths:
        with open(shard_path, "rb") as f:
            header_len = struct.unpack("<Q", f.read(8))[0]
            header = json.loads(f.read(header_len).decode("utf-8"))
            data_base = 8 + header_len
            for name, meta in header.items():
                if name == "__metadata__":
                    continue
                expected_name = name if sd_ops is None else sd_ops.apply_to_key(name)
                if expected_name is None:
                    continue
                start, end = meta["data_offsets"]
                f.seek(data_base + start)
                buf = f.read(end - start)
                t = torch.frombuffer(
                    bytearray(buf), dtype=_SAFETENSORS_DTYPE_MAP[meta["dtype"]]
                ).reshape(meta["shape"])
                t = t.to(device=device, non_blocking=True, copy=False)
                kvs = (
                    ((expected_name, t),)
                    if sd_ops is None
                    else sd_ops.apply_to_key_value(expected_name, t)
                )
                for key, v in kvs:
                    size += v.nbytes
                    dtype.add(v.dtype)
                    sd[key] = v
    return StateDict(sd=sd, device=device, size=size, dtype=dtype)


SafetensorsStateDictLoader.load = _patched_load
print("[FUSE-PATCH] SafetensorsStateDictLoader.load replaced (chunked-read)")

logging.getLogger().setLevel(logging.INFO)

MAX_SEED = np.iinfo(np.int32).max
DEFAULT_PROMPT = (
    "An astronaut hatches from a fragile egg on the surface of the Moon, "
    "the shell cracking and peeling apart in gentle low-gravity motion. "
    "Fine lunar dust lifts and drifts outward with each movement, floating "
    "in slow arcs before settling back onto the ground."
)
DEFAULT_FRAME_RATE = 24.0

# Resolution presets: (width, height)
RESOLUTIONS = {
    "high": {"16:9": (1536, 1024), "9:16": (1024, 1536), "1:1": (1024, 1024)},
    "low": {"16:9": (768, 512), "9:16": (512, 768), "1:1": (768, 768)},
}

LTX_REPO_ID = "Lightricks/LTX-2.3"
# Lightricks' own ungated mirror of the Gemma text encoder — no license gate,
# no HF_TOKEN needed.
GEMMA_REPO_ID = "Lightricks/gemma-3-12b-it-qat-q4_0-unquantized"

DISTILLED_FILENAME = "ltx-2.3-22b-distilled-1.1.safetensors"
UPSCALER_FILENAME = "ltx-2.3-spatial-upscaler-x2-1.1.safetensors"

# Checkpoints come from the Hub instead of the upstream space's bucket mounts.
distilled_checkpoint_path = hf_hub_download(repo_id=LTX_REPO_ID, filename=DISTILLED_FILENAME)
spatial_upsampler_path = hf_hub_download(repo_id=LTX_REPO_ID, filename=UPSCALER_FILENAME)
gemma_root = snapshot_download(repo_id=GEMMA_REPO_ID)

# Initialize pipeline WITH text encoder
pipeline = DistilledPipeline(
    distilled_checkpoint_path=distilled_checkpoint_path,
    spatial_upsampler_path=spatial_upsampler_path,
    gemma_root=gemma_root,
    loras=[],
    quantization=QuantizationPolicy.fp8_cast(),
)

# Preload all models for ZeroGPU tensor packing. Gemma is loaded and
# quantized FIRST: in bf16 it is ~24G, and with the fp8 transformer (~22G)
# and the rest already resident, the transient peak before quantization is
# the full ~50G packed ledger — right at the default ZeroGPU slice's limit,
# so startup survived or died depending on host packing (OOM-killed silently
# at this very print). Quantizing before anything else is resident caps the
# peak at the bf16 Gemma alone.
print("Preloading all models (including Gemma)...")
ledger = pipeline.model_ledger
_text_encoder = ledger.text_encoder()
# Quantize Gemma to int8 weight-only. int8 (~12G) fits the default slice, and
# the conditioning path consumes raw hidden states from every layer through a
# projection trained on bf16 activations, so int8 is the least-lossy option
# that fits (int4 would degrade those activations further for VRAM we don't
# need; the boot peak is the bf16 Gemma before quantization either way).
from torchao.quantization import quantize_ as _quantize_, Int8WeightOnlyConfig as _Int8
_quantize_(_text_encoder, _Int8())
print("[QUANT] Gemma text encoder quantized to int8")
_embeddings_processor = ledger.gemma_embeddings_processor()
_transformer = ledger.transformer()
_video_encoder = ledger.video_encoder()
_video_decoder = ledger.video_decoder()
_audio_decoder = ledger.audio_decoder()
_vocoder = ledger.vocoder()
_spatial_upsampler = ledger.spatial_upsampler()

ledger.transformer = lambda: _transformer
ledger.video_encoder = lambda: _video_encoder
ledger.video_decoder = lambda: _video_decoder
ledger.audio_decoder = lambda: _audio_decoder
ledger.vocoder = lambda: _vocoder
ledger.spatial_upsampler = lambda: _spatial_upsampler
ledger.text_encoder = lambda: _text_encoder
ledger.gemma_embeddings_processor = lambda: _embeddings_processor
print("All models preloaded (including Gemma text encoder)!")

print("=" * 80)
print("Pipeline ready!")
print("=" * 80)


def log_memory(tag: str):
    if torch.cuda.is_available():
        allocated = torch.cuda.memory_allocated() / 1024**3
        peak = torch.cuda.max_memory_allocated() / 1024**3
        free, total = torch.cuda.mem_get_info()
        print(f"[VRAM {tag}] allocated={allocated:.2f}GB peak={peak:.2f}GB free={free / 1024**3:.2f}GB total={total / 1024**3:.2f}GB")


def detect_aspect_ratio(image) -> str:
    """Detect the closest aspect ratio (16:9, 9:16, or 1:1) from an image."""
    if image is None:
        return "16:9"
    if hasattr(image, "size"):
        w, h = image.size
    elif hasattr(image, "shape"):
        h, w = image.shape[:2]
    else:
        return "16:9"
    ratio = w / h
    candidates = {"16:9": 16 / 9, "9:16": 9 / 16, "1:1": 1.0}
    return min(candidates, key=lambda k: abs(ratio - candidates[k]))


def on_image_upload(image, high_res):
    """Auto-set resolution when image is uploaded."""
    aspect = detect_aspect_ratio(image)
    tier = "high" if high_res else "low"
    w, h = RESOLUTIONS[tier][aspect]
    return gr.update(value=w), gr.update(value=h)


def on_highres_toggle(image, high_res):
    """Update resolution when high-res toggle changes."""
    aspect = detect_aspect_ratio(image)
    tier = "high" if high_res else "low"
    w, h = RESOLUTIONS[tier][aspect]
    return gr.update(value=w), gr.update(value=h)


def _gpu_duration(input_image, prompt, duration, enhance_prompt=True, seed=10,
                  randomize_seed=True, height=1024, width=1536, *args, **kwargs):
    # ZeroGPU pre-checks the REQUESTED duration against the visitor's remaining
    # quota (and shorter requests rank higher in the node queue), so this must
    # sit close to reality. Measured over ~100 production runs: wall time is
    # ~0.33s per million (pixels x frames), near-perfectly linear across low
    # and high res. 0.55 gives >50% headroom for outliers plus tensor attach.
    frames = int(float(duration) * DEFAULT_FRAME_RATE) + 1
    frames = ((frames - 1 + 7) // 8) * 8 + 1
    mpx_frames = int(height) * int(width) * frames / 1e6
    return int(min(110, max(20, 15 + 0.55 * mpx_frames)))

@spaces.GPU(duration=_gpu_duration)
@torch.inference_mode()
def generate_video(
    input_image,
    prompt: str,
    duration: float,
    enhance_prompt: bool = True,
    seed: int = 42,
    randomize_seed: bool = True,
    height: int = 1024,
    width: int = 1536,
    progress=gr.Progress(track_tqdm=True),
):
    try:
        torch.cuda.reset_peak_memory_stats()
        log_memory("start")

        current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)

        frame_rate = DEFAULT_FRAME_RATE
        num_frames = int(duration * frame_rate) + 1
        num_frames = ((num_frames - 1 + 7) // 8) * 8 + 1

        print(f"Generating: {height}x{width}, {num_frames} frames ({duration}s), seed={current_seed}")

        images = []
        if input_image is not None:
            output_dir = Path("outputs")
            output_dir.mkdir(exist_ok=True)
            temp_image_path = output_dir / f"temp_input_{current_seed}.jpg"
            if hasattr(input_image, "save"):
                input_image.save(temp_image_path)
            else:
                temp_image_path = Path(input_image)
            images = [ImageConditioningInput(path=str(temp_image_path), frame_idx=0, strength=1.0)]

        tiling_config = TilingConfig.default()
        video_chunks_number = get_video_chunks_number(num_frames, tiling_config)

        log_memory("before pipeline call")

        video, audio = pipeline(
            prompt=prompt,
            seed=current_seed,
            height=int(height),
            width=int(width),
            num_frames=num_frames,
            frame_rate=frame_rate,
            images=images,
            tiling_config=tiling_config,
            enhance_prompt=enhance_prompt,
        )

        log_memory("after pipeline call")

        output_path = tempfile.mktemp(suffix=".mp4")
        encode_video(
            video=video,
            fps=frame_rate,
            audio=audio,
            output_path=output_path,
            video_chunks_number=video_chunks_number,
        )

        log_memory("after encode_video")
        return str(output_path), current_seed

    except Exception as e:
        import traceback
        log_memory("on error")
        print(f"Error: {str(e)}\n{traceback.format_exc()}")
        return None, current_seed


UPSAMPLER_THEME = gr.themes.Soft(
    primary_hue=gr.themes.colors.indigo,
    secondary_hue=gr.themes.colors.purple,
    neutral_hue=gr.themes.colors.slate,
).set(
    button_primary_background_fill="linear-gradient(135deg, #6366f1, #a855f7)",
    button_primary_background_fill_hover="linear-gradient(135deg, #5457e5, #9333ea)",
    button_primary_text_color="#ffffff",
    button_primary_border_color="*primary_500",
)
UPSAMPLER_CSS = """
footer{display:none !important}
.gradio-container{max-width:1000px !important; margin:0 auto !important}
h1,h2,h3{font-family:system-ui,-apple-system,'Segoe UI',sans-serif}
"""
_header = """<div style="max-width:760px;margin:0 auto;text-align:center;padding:20px 16px 2px;font-family:system-ui,-apple-system,'Segoe UI',sans-serif">
  <h1 style="font-size:1.7rem;font-weight:700;margin:0 0 6px;letter-spacing:-.02em">LTX 2.3 22B Distilled</h1>
  <p style="font-size:1rem;line-height:1.5;opacity:.6;margin:0">Fast video with synchronized audio from a text prompt or an image.</p>
</div>"""
_footer = """<div style="max-width:640px;margin:2rem auto .4rem;text-align:center;font-family:system-ui,-apple-system,'Segoe UI',sans-serif">
  <p style="font-size:.85rem;line-height:1.6;opacity:.5;margin:0 0 10px">LTX 2.3 by Lightricks is one of the fastest open text-to-video models, generating HD clips with synchronized audio from a written prompt or a single image. Dialogue, sound effects, and camera motion come out of a single pass, so a finished video with sound takes well under a minute on this space.</p>
  <p style="font-size:.85rem;line-height:1.6;opacity:.65;margin:0">Maintained by <a href="https://upsampler.com" target="_blank" rel="noopener" style="color:#8b7cf6;font-weight:600;text-decoration:none">Upsampler</a>. Check out the <a href="https://upsampler.com/free-video-generator-no-signup" target="_blank" rel="noopener" style="color:#8b7cf6;font-weight:600;text-decoration:none">free AI video generator</a>, no sign-up required.</p>
</div>"""

with gr.Blocks(title="LTX 2.3 - Text or Image to Video") as demo:
    gr.HTML(_header)

    with gr.Row():
        with gr.Column():
            input_image = gr.Image(label="Input Image (Optional)", type="pil")
            prompt = gr.Textbox(
                label="Prompt",
                info="for best results - make it as elaborate as possible",
                value="Make this image come alive with cinematic motion, smooth animation",
                lines=3,
                placeholder="Describe the motion and animation you want...",
            )
            
            with gr.Row():
                duration = gr.Slider(label="Duration (seconds)", minimum=1.0, maximum=10.0, value=3.0, step=0.1)
                with gr.Column():
                    enhance_prompt = gr.Checkbox(label="Enhance Prompt", value=False)
                    high_res = gr.Checkbox(label="High Resolution", value=True)

            generate_btn = gr.Button("Generate Video", variant="primary", size="lg")

            with gr.Accordion("Advanced Settings", open=False):
                seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, value=10, step=1)
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
                with gr.Row():
                    width = gr.Number(label="Width", value=1536, precision=0)
                    height = gr.Number(label="Height", value=1024, precision=0)

        with gr.Column():
            output_video = gr.Video(label="Generated Video", autoplay=True)

    # Auto-detect aspect ratio from uploaded image and set resolution
    input_image.change(
        fn=on_image_upload,
        inputs=[input_image, high_res],
        outputs=[width, height],
    )

    # Update resolution when high-res toggle changes
    high_res.change(
        fn=on_highres_toggle,
        inputs=[input_image, high_res],
        outputs=[width, height],
    )

    generate_btn.click(
        fn=generate_video,
        inputs=[
            input_image, prompt, duration, enhance_prompt,
            seed, randomize_seed, height, width,
        ],
        outputs=[output_video, seed],
    )

    gr.HTML(_footer)


css = """
.fillable{max-width: 1200px !important}
.progress-text {color: white}
"""

if __name__ == "__main__":
    demo.launch(theme=UPSAMPLER_THEME, css=UPSAMPLER_CSS + css, ssr_mode=False, show_error=True)