| { |
| "schema_version": 1, |
| "id": "minimax-h3-r3-packed-exact-tuning", |
| "title": "Exact QK/FC2 tuning survives lossless BF16 storage on M5 Pro", |
| "tags": [ |
| "packed-weights", |
| "bf16", |
| "memory", |
| "m5", |
| "dynamic-inputs", |
| "gemm" |
| ], |
| "decision": "accepted", |
| "evidence_level": "observed", |
| "mechanism": "Compressed arrays remain explicit dynamic inputs to each compiled block, then reconstruct the original BF16 bits before native projections or the selected FC2 TensorOps kernel. Exact QK fusion operates on the same reconstructed projection outputs.", |
| "recommendation": "Use existing packed12 storage when lower physical footprint is needed, together with the qualified QK/FC2 mode. Compare against its own packed control and retain native dense storage for the lowest observed latency. Reuse the source-bound cache rather than making another model copy.", |
| "counterexample": "Faster vector reconstruction alone did not establish a useful decode-plus-GEMM gain, and is not shipped. Dense and packed timings from different campaigns are not a paired storage-speed comparison.", |
| "retest_when": "Retest after codec, MLX compiler constant-folding, Metal matrix dispatch, hardware or token-range changes. Inspect full physical footprint as well as output equality; a retained dense constant can erase the storage benefit.", |
| "scope": { |
| "projects": [ |
| "minimax-h3-image-mlx" |
| ], |
| "hardware": [ |
| "Apple M5 Pro, 16 GPU cores, 64GiB unified memory" |
| ], |
| "runtime": [ |
| "MLX/MLX Metal 0.32.4; macOS27.0.1" |
| ], |
| "workloads": [ |
| "Four reused full-image/trajectory quality cases; product and portrait three paired warm repeats each. 3.5667% latency reduction vs packed control; 31.83GiB maximum timed physical peak, 32.56GiB across all candidate cases." |
| ] |
| }, |
| "evidence": [ |
| { |
| "project": "minimax-h3-image-mlx", |
| "path": "reports/optimization-r3/compact-deployment/comparison.json", |
| "sha256": "f04c312015382fdfd7cd021df6168d7729344594ee7c4ac2143b18f4b503c66c" |
| }, |
| { |
| "project": "minimax-h3-image-mlx", |
| "path": "reports/optimization-r3/compact-deployment/quality-audit.json", |
| "sha256": "57ae4aef9579058528381de211636fe9ff02ad9b2f793f65b7b217e91c6e7f16" |
| }, |
| { |
| "project": "minimax-h3-image-mlx", |
| "path": "reports/optimization-r3/wheel-cli.json", |
| "sha256": "813384da8643af0ea86432df03aaaab840fbac9f1133e84e9d9497116ef95fd5" |
| }, |
| { |
| "project": "minimax-h3-image-mlx", |
| "path": "reports/optimization-r3/screen.json", |
| "sha256": "58675fb0df998d08b52c9fe4258cedf708ce638c914f628eb0b417ee27993c8f" |
| } |
| ], |
| "supersedes": [] |
| } |
|
|
