| { |
| "calibration_sequences": 128, |
| "covariance_passes": 3, |
| "hardware": { |
| "cuda_available": true, |
| "cuda_runtime": "12.8", |
| "gpu_capability": [ |
| 8, |
| 6 |
| ], |
| "gpu_name": "NVIDIA A40", |
| "hostname": "gpu-node", |
| "nvidia_smi": [ |
| "0, NVIDIA A40, GPU-REDACTED, 46068 MiB, 43932 MiB, 580.173.02" |
| ], |
| "python": "3.11.16 | packaged by conda-forge | (main, Aug 21 2026, 22:44:51) [GCC 14.4.0]", |
| "slurm_job_gpus": "6", |
| "slurm_job_id": "143327", |
| "torch": "2.11.0+cu128" |
| }, |
| "model": "qwen3_0.6b_base", |
| "model_id": "Qwen/Qwen3-0.6B-Base", |
| "r1": "single global pooled post-RMSNorm NAR, group-128, separately frozen k=8 and k=n/128", |
| "r1_rows": [ |
| { |
| "active_reflectors": 8, |
| "anchor_error": 1.1920928955078125e-07, |
| "k": 8 |
| } |
| ], |
| "r2": "per-layer, per-head V NAR, head_dim=128, group-128, k=1", |
| "r4": "reuse frozen E5 per-layer down-input NAR factors", |
| "seed": 0, |
| "sequence_length": 2048 |
| } |
|
|
