Drop module-load pre-warming (Triton/fp8+LoRA needs real CUDA)

The pre-warm block hit torch.AcceleratorError: no CUDA-capable device on ZeroGPU because fp8_cast's LoRA fusion launches a Triton kernel (_fused_add_round_launch) at build time, and ZeroGPU only provides a real CUDA device inside @spaces.GPU-decorated functions.

The Lightricks/LTX-2-3 Space gets away with module-load pre-warm only because it uses loras=[] — no LoRA fusion path, no Triton. We always have the HDR LoRA fused in, so we can't pre-warm at module load.

Keeping only the scene-emb .pt revert from this PR. Each generate call will re-build transformer + VAE components inside the GPU slot; slower than cached, but correct.

app.py

@@ -67,8 +67,6 @@ import logging
 import random
 import tempfile
 import zipfile
-from collections.abc import Iterator
-from contextlib import contextmanager
 
 import torch
 torch._dynamo.config.suppress_errors = True
@@ -79,7 +77,6 @@ import gradio as gr
 import numpy as np
 from huggingface_hub import hf_hub_download
 
-from ltx_core.model.upsampler import upsample_video
 from ltx_core.model.video_vae import TilingConfig
 from ltx_core.quantization import QuantizationPolicy
 from ltx_pipelines.hdr_ic_lora import HDRICLoraPipeline, _make_tiling_config
@@ -156,83 +153,6 @@ pipeline = HDRICLoraPipeline(
 )
 print(f"HDRICLoraPipeline ready. HDR transform: {pipeline.hdr_transform}, "
       f"ref_downscale={pipeline.reference_downscale_factor}")
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# Pre-warm for ZeroGPU: build each component once at module load (ZeroGPU
-# tensor-packing captures the weights), then replace the pipeline's blocks
-# with tiny wrappers that reuse the cached models and skip gpu_model's
-# meta-device freeing. Avoids re-reading the 22B checkpoint + re-fusing
-# the LoRA + re-fp8-casting on every @spaces.GPU invocation.
-# ─────────────────────────────────────────────────────────────────────────────
-print("Pre-warming models (one-shot build)...")
-
-_cached_image_encoder = pipeline.image_conditioner._build_encoder()
-_cached_transformer = pipeline.stage_1._build_transformer()
-_cached_upsampler_encoder = (
-    pipeline.upsampler._encoder_builder
-    .build(device=_DEVICE, dtype=_DTYPE).to(_DEVICE).eval()
-)
-_cached_upsampler = (
-    pipeline.upsampler._upsampler_builder
-    .build(device=_DEVICE, dtype=_DTYPE).to(_DEVICE).eval()
-)
-_cached_video_decoder = (
-    pipeline.video_decoder._decoder_builder
-    .build(device=_DEVICE, dtype=_DTYPE).to(_DEVICE).eval()
-)
-
-
-@contextmanager
-def _yield_cached(model):
-    """Drop-in for gpu_model that does NOT move params to meta on exit."""
-    yield model
-
-
-# Patch the transformer context manager on both stages to yield the cached
-# transformer without freeing. stage_1 uses _transformer_ctx inside __call__;
-# stage_2 uses model_context() -> _transformer_ctx.
-def _cached_stage_ctx(**_kwargs):
-    return _yield_cached(_cached_transformer)
-
-
-pipeline.stage_1._transformer_ctx = _cached_stage_ctx
-pipeline.stage_2._transformer_ctx = _cached_stage_ctx
-
-
-class _CachedImageConditioner:
-    def __call__(self, fn):
-        return fn(_cached_image_encoder)
-
-
-class _CachedVideoUpsampler:
-    def __call__(self, latent):
-        return upsample_video(
-            latent=latent,
-            video_encoder=_cached_upsampler_encoder,
-            upsampler=_cached_upsampler,
-        )
-
-
-class _CachedVideoDecoder:
-    def __call__(
-        self,
-        latent: torch.Tensor,
-        tiling_config=None,
-        generator=None,
-        *,
-        output_dtype: torch.dtype = torch.uint8,
-    ) -> Iterator[torch.Tensor]:
-        return _cached_video_decoder.decode_video(
-            latent, tiling_config, generator, output_dtype=output_dtype,
-        )
-
-
-pipeline.image_conditioner = _CachedImageConditioner()
-pipeline.upsampler = _CachedVideoUpsampler()
-pipeline.video_decoder = _CachedVideoDecoder()
-
-print("Pre-warm complete.")
 print("=" * 80)