Saga-AI/bdh-k20-gx10-probe

K20 GB10 compile and fit probe

This dataset preserves the completed GB10 compile/fit preflight for the first K20 Estonian growth step. It is hardware and provenance evidence for the later K20 ladder, not an additional K20 scientific result.

What was measured

• Host: gx10-50ef (NVIDIA GB10, 124,544 MiB unified memory)
• Start checkpoint: the pre-K20 et base, MD5 8d9af6d6f2a37b1e7e6a0f19e0ef4840
• Growth: multiplier 480 to 512; width 30,720 to 32,768
• Trainable growth: +2,048 neurons per head; old neurons, embedding, and LM head frozen
• Route-aware loss: 0.9 prefix + 0.1 full-width
• 2,000 steps, approximately 1,049 ms/step
• Parameters: 402,915,328; dtype: BF16
• Best validation PPL: 4.49; final validation PPL: 4.55; test PPL: 4.37
• Minimum available unified memory: 85,768 MiB
• Minimum free unified memory: 45,107 MiB
• Final verdict: FIT_MEASURED

The successful run ended at 2026-09-24 04:04 CEST. The full K20 Arm A et phase began at 04:42 CEST, so this dataset is the immediate compile/fit preflight rather than part of the final ladder.

Earlier failed attempts

probe_summary.txt records two incomplete attempts: one exited after 9 seconds and one exited after 938 seconds without a final checkpoint. The preserved log is the successful compile run. Their exact failure causes were not fully captured and are therefore not inferred here.

Contents

• Two final BF16 PyTorch checkpoints (best and last)
• Successful compile-run training log
• Unified-memory telemetry (meminfo_samples.csv)
• Machine-readable run summary (probe_summary.txt)
• MANIFEST.sha256 for all five source payloads

Source payload bytes: 9666113353 Source payload count: 5 Manifest SHA-256: efc42e806215a23995ca6b0e019ed063f897db35980cf167fedda5d7918e8fbf Uploaded repository: Saga-AI/bdh-k20-gx10-probe

The source files were hardlink-staged and were not modified during upload.