File size: 1,397 Bytes
7845694 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 | {
"version": "mlx-validation-v1",
"reference": "Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
"runtime": {
"mlx": "0.32.2",
"mlx-lm": "0.31.3"
},
"exact_contracts": [
"prompt text",
"input token IDs",
"answer token IDs",
"decision IDs",
"option IDs",
"finite normalized probabilities",
"complete fixture coverage"
],
"tiny_fp32_cache_probability_atol": 0.0001,
"bf16_probability_difference_review_threshold": 0.05,
"bf16_policy": "Report every changed choice. Investigate differences above the review threshold; it is a diagnostic trigger, not an accuracy or calibration guarantee. Do not require bit-identical CUDA and Metal outputs.",
"pilot": "results/mlx/2026-09-17-bf16-pilot/diagnostic.json",
"pilot_observation": "Seven decisions: no changed choices, at most 0.030491 probability movement after prefix splitting, and exactly unchanged shared scores after reversing row order.",
"quantization_policy": "Evaluate 8-bit and 4-bit separately against the unquantized MLX baseline. Never assume calibration or preserve CUDA claims by changing tolerances.",
"production_runtime": {
"mlx": "0.32.2",
"mlx-lm": "0.32.0",
"mlx_lm_commit": "a63e24c389382619eb6d9af656e3b46024be217a"
},
"runtime_note": "Original pilot used 0.31.3. Production pins the upstream Qwen L2 epsilon fix; tolerances are unchanged."
}
|