Training Status: No Blocking Issues | Generated by AI

Home 2026.07

Obstacles Encountered & Status

1. Missing Python Dev Headers ✅ Fixed

Symptom: Triton crashed at startup — gcc couldn’t compile cuda_utils.c because Python.h was missing for Python 3.10. Fix: apt-get install python3.10-dev Root cause: uv sync --extra gpu created the venv with Python 3.10, but the system only had Python 3.11 dev headers installed.

2. Disk Space ✅ Fixed

Symptom: Only 16G free on /workspace — not enough for training + checkpoints. Fix: Deleted stale tokenized .npy files (67G) + duplicate dataset dir (8.9G). Now 176G free.

3. uv Sync Timeout ✅ Resolved (user ran it)

Symptom: uv sync --extra gpu timed out at 600s — torch 2.9.1+cu128 is a large download. Workaround: User ran it themselves.

4. FA3 flash_attn Not Callable in Tests ⚠️ Non-blocking

Symptom: Calling flash_attn(q, k, v) directly fails with 'SimpleNamespace' object is not callable. Why it’s OK: The training uses it successfully (see 128K tok/s, 62% MFU). The direct API call is a wrapper issue — the actual attention kernel works via the GPT model’s code path. FA3 prints ✓ Using Flash Attention 3 (Hopper GPU) at startup.

5. Wandb Offline Mode ⚠️ Note

Symptom: Metrics logged to wandb/offline-run-* — not synced to cloud. Workaround: Run wandb sync /workspace/nanochat/wandb/offline-run-<id> later, or re-run with WANDB_MODE=online and a logged-in wandb account.

6. FineWeb-Edu ≠ ClimbMix-400B ⚠️ Data Quality Gap

Issue: nanochat’s speedrun records (1.65h to GPT-2) use ClimbMix-400B, a better-curated dataset. FineWeb-Edu is older, lower quality — expect slightly worse final loss/CORE scores. Mitigation: We have 35.7B tokens — enough for d24 (15.2B) and d30 (29B). Results will be decent, just not SOTA.

7. FP8 API Check Fails But Training Works ⚠️ Non-blocking

Symptom: torch.cuda.is_fp8_supported() raises error (API not exposed in this torch build). Reality: Training shows ✓ FP8 training enabled (tensorwise scaling) - converted 145/158 linear layers — FP8 is working on H200 hardware via the custom fp8.py module, not the standard torch API.


Current Training Health

Check Status
Triton compilation ✅ Working
FA3 on H200 ✅ Active
FP8 matmuls ✅ Enabled (145 layers)
Data loading ✅ 44 parquet files, streaming
Loss curve ✅ Clean descent: 10.4 → 4.9 at step 108
Throughput ✅ ~128K tok/s
MFU ✅ ~62% (excellent for single GPU)
ETA ~33h total, ~32h remaining

Bottom line: No blocking issues. The training will complete with good results on FineWeb-Edu. For the 1.5B (d30) run after this, consider downloading ClimbMix-400B for better quality, or just scale up the same pipeline with --depth=30.


Back Donate