ApacheOne's picture
Upload Wan Animate-2 OrbitQuant packed W4A4 model
f2c0505 verified
|
Raw
History Blame Contribute Delete
822 Bytes

Validation

Package-build-host validation:

  • Python compileall: PASS
  • pytest: PASS (17 tests)
  • direct-source target rule: 40 x 12 = 480
  • nibble pack/unpack: PASS
  • direct-source W4 packing unit test: PASS, BF16 row norms
  • CPU packed GEMM reference/layout tests: PASS
  • runtime static invariants: PASS
  • no dense target weight Parameter in OrbitQuantPackedLinear: PASS
  • CPU K/V cache streaming static checks: PASS
  • Sol/flex separation static checks: PASS

Not claimable on the build host:

  • Triton JIT compilation on the user's Colab GPU
  • numerical CUDA parity for d=5120 / 13824
  • full 32.8-GB transformer quantization
  • full Wan-Animate-2 generation

These are mandatory gates in COLAB_ONE_CELL.py; generation is not allowed to proceed until the actual packed artifact passes the target-GPU kernel gate.