=== BUILD LOG TAIL ===
rbi_serve/weight_quant/exl3/csrc. Rebuild the image, or run against the sources the image was built from.
BUILD OK <module 'arbi_serve_exl3_i8_gemm_v1' from '/ext/torch_extensions/arbi_serve_exl3_i8_gemm_v1/arbi_serve_exl3_i8_gemm_v1.so'> in 1228.6s

=== ORACLE GATE: repack -> kernel -> oracle decomposition ===
prebaked extension 'arbi_serve_exl3_i8_gemm_v1' is stale against the installed sources and is set aside; building it from those sources instead (a first-boot cost, cached under TORCH_EXTENSIONS_DIR). prebaked extension 'arbi_serve_exl3_i8_gemm_v1' was built from different sources than are installed: baked 9ef796f20e4fe5df, installed f9b956d2fa650678. The binary at /opt/cache-baked/torch_extensions/arbi_serve_exl3_i8_gemm_v1/arbi_serve_exl3_i8_gemm_v1.so would not match the sources under /w/arbi_serve/weight_quant/exl3/csrc. Rebuild the image, or run against the sources the image was built from.
  PASS IQ1_M   cb=28 bits=4 nibble   : 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ2_S   cb=26 bits=4 nibble   : 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ2_XS  cb=26 bits=4 nibble   : 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ2_XXS cb=29 bits=2 grid:iq2xxs: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ3_S   cb=21 bits=4 nibble   : 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ3_XXS cb=27 bits=4 nibble   : 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ4_NL  cb=20 bits=4 nibble   : 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ4_XS  cb=20 bits=4 nibble   : 512x256 through repack -> kernel -> oracle fields, exactly
  PASS Q2_K    cb=22 bits=2 bitstream: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS Q3_K    cb=23 bits=3 bitstream: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS Q4_0    cb=23 bits=4 bitstream: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS Q4_K    cb=22 bits=4 bitstream: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS Q5_K    cb=22 bits=5 bitstream: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS Q6_K    cb=23 bits=6 bitstream: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS Q8_0    cb=23 bits=8 bitstream: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ2_S   cb=35 bits=3 (twin spelling) grid:iq2s: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ2_XS  cb=34 bits=3 (twin spelling) grid:iq2xs: 512x256 through repack -> kernel -> oracle fields, exactly
  PASS IQ2_XXS cb=26 bits=4 (twin spelling) nibble   : 512x256 through repack -> kernel -> oracle fields, exactly

INTEGRATION GATE PASSED
  18 type(s) compared against a decomposition that is itself checked
  against llama.cpp's C and gguf-py, so this is NOT two of ours agreeing.

[exited with code 0]
