Spaces:
Paused
Paused
Commit History
-
Replace stand-in with REAL LTX-2 DiT AOT artifacts (cu128 + cu130)
Compile the actual FastVideo LTX2Transformer3DModel block (48 layers, inner_dim 4096) with the real 37.8GB LTX2-Distilled-Diffusers weights (3510 tensors, 0 missing) through aokit's AOT export path. Capture real block-0 inputs (x=(1,256,4096)) during a real forward, wrap the real block behind a flat-tensor export boundary, export -> AOTInductor -> weight-less package.pt2. Verified (cosine 0.99997, rel-err 1.9%). Ships two artifacts as requested: one built under torch 2.11+cu128 (CUDA 12.8) and one under torch 2.11+cu130 (CUDA 13.0). FA4 skipped (B200-only); TORCH_SDPA used on sm_120. Removes the prior stand-in model + its pt2. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>cbd74cc -
Add DreamVerse->aokit AOT compilation adapter + compiled pt2
Port DreamVerse's JIT torch.compile + startup-warmup strategy to aokit's ahead-of-time export flow. Regional compilation of the LTX-2-style DiT transformer block (capture -> torch.export -> AOTInductor -> weightless package.pt2), loaded onto all blocks and verified (cosine 0.999994). Compiled on RTX PRO 6000 Blackwell (sm_120); FA4 skipped (B200-only), SDPA used instead. Ships dist/dreamverse_aokit_pt2.zip (pt2 + digest). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>4bf9cd8