File size: 5,667 Bytes
bd12cae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
# Gemma-4-31B × Augmental SFT — Run Summary (2026-07-17/18)

Full-parameter SFT of `google/gemma-4-31B` (base) on `Heralax/Augmental-Dataset` (7,831 rows of
visual-novel multi-character roleplay). 8× H200, DeepSpeed ZeRO-3, bf16, axolotl.

## Results

| Checkpoint | eval_loss | eval_ppl | Where |
|---|---|---|---|
| base (step 0) | 1.753 | 5.77 | — |
| epoch 1 (step 240) | 1.572 | 4.81 | `out/epoch1-checkpoint-240` + [HF -ep1](https://huggingface.co/nshuster/gemma4-31b-augmental-sft-ep1) |
| **epoch 2 (step 480)** | **1.521 (best)** | **4.58** | `out/checkpoint-480` + [HF -ep2](https://huggingface.co/nshuster/gemma4-31b-augmental-sft-ep2) |
| epoch 3 (step 720, final) | 1.707 | 5.51 | `out/checkpoint-720` + [HF -ep3](https://huggingface.co/nshuster/gemma4-31b-augmental-sft-ep3) |

- 720 steps in **1h08m49s** (~5.2–5.8 s/step, ~29.3k real tokens/step, ~17–22% MFU)
- Epoch 3 overfits (train loss 0.42 vs eval 1.71) — expected/accepted for style-soak; **ep2 is the
  best held-out model**, ep3 the strongest style imitator.
- Peak VRAM ~74/141 GiB per GPU; grad norms spiked during warmup (up to 1927 pre-clip) then settled.

## Prompt format (no chat template — screenplay style)

```
Scenario: {setup}

{Speaker A}: *action* "dialogue"
{Speaker B}: *action* "dialogue"
{target speaker}:
```
Model writes one turn for the trailing speaker tag and stops at `<eos>`. Prompt is masked in
training (`train_on_inputs: false`); only completions + `<eos>` are trained. Verified against the
real tokenizer: single `<bos>` (id 2), trained `<eos>` (id 1), no double-BOS.

## Problems hit & fixes (in order)

1. **Checkpoint size vs 500 GB volume quota.** Full ZeRO-3 checkpoints are ~434 GB (fp32 Adam m+v
   248 + fp32 master weights 124 + bf16 model 62); quota had ~426 GB free. → `save_only_model: true`
   (weights-only, ~62 GB/ckpt). Note: `df` on the volume shows the whole 378 TB MooseFS pool, not
   your quota — the 500 GB limit is enforced externally by RunPod.
2. **Crash: `No module named 'torchvision'` → `Gemma4Processor` import failure.** Gemma-4 is a
   VLM; its processor needs torchvision, which `axolotl[deepspeed,flash-attn]` doesn't pull in.
   → `uv pip install torchvision` (added to setup_pod.sh + Dockerfile).
3. **Crash: `FlashAttention forward only supports head dimension at most 256`.** Structural:
   Gemma-4 is hybrid-attention — 50 sliding layers (head_dim 256) + 10 global layers
   (**global_head_dim 512**, `attention_k_eq_v: true`). FA2 hard-caps at 256; flash-attn-4 is
   beta-only. → `attn_implementation: sdpa`.
4. **save_total_limit: 2 would have pruned the epoch-1 checkpoint** at the final save. Avoided a
   restart by renaming `checkpoint-240 → epoch1-checkpoint-240` after its save completed (the
   Trainer's pruner only globs `checkpoint-*`). yaml now says 3 for future runs.
5. **HF cache confusion:** pod presets `HF_HUB_CACHE=/workspace/data/huggingface-cache/hub`, which
   overrides the script's `HF_HOME=/workspace/hf`. Model was already fully cached there (62.6 GB) —
   no re-download ever needed. Launch commands pin `HF_HUB_CACHE` explicitly.

## What I'd change next run (ranked by impact)

1. **`sample_packing: true`** — avg sample is ~917 tokens in a 4096 window → ~75% of compute was
   padding. Packing ≈ 2× compute win. Must smoke-test with SDPA (axolotl packing has historically
   preferred flash-attention var-len kernels).
2. **Tune ZeRO-3 comms** — fetched `zero3_bf16.json` uses `stage3_max_live_parameters: 0` /
   `max_reuse_distance: 0` (maximally re-gathers). ~70 GiB/GPU sat idle; raising these cuts
   all-gather traffic. Est. 10–25%.
3. **`micro_batch_size: 4`, accum 1** (same global 32) — fewer accumulation passes. Est. 10–20%.
4. **Bake the env into a Docker image** (see Dockerfile/start.sh) — venv on the network fuse mount
   made every 8-rank launch pay ~9 min of imports; local-disk venv cuts startup to ~2 min, and
   skips the 30-min flash-attn compile forever. Drop `flash-attn` from extras (SDPA is in use).
5. **Checkpoint to local NVMe then async-copy to the volume** — each 62 GB save blocked training
   ~3 min writing to MooseFS.
6. **1-GPU smoke test before the 8-rank launch** (`load model + one forward`) — would have caught
   crashes #2 and #3 in ~3 min instead of ~25 min of failed distributed launches.
7. Consider lr ~2–3e-6 if targeting best held-out (warmup grad spikes + fast ep2→ep3 overfit
   suggest 5e-6 is hot for full-FT), or stop at 2 epochs.
8. Optional/speculative: torch.compile, flash-attn-4 beta (Hopper supports it; 5–15% maybe).

## Files in this bundle

| File | What |
|---|---|
| `RUN_SUMMARY.md` | this document |
| `setup_pod.sh` | pod bootstrap (idempotent; now installs torchvision) |
| `augmental-fullsft.yaml` | axolotl config (final: sdpa, save_only_model, limit 3) |
| `prep_augmental.py` | dataset → axolotl input_output JSONL converter |
| `Dockerfile` + `start.sh` | bake-the-env image for instant pod startup (RunPod notes inside) |
| `zero3_bf16.json` | DeepSpeed config used |
| `README-ep{1,2,3}.md` | HF model cards as published |
| `train.log` | full training log (final successful run) |
| `setup_pod.run.log` | environment build log |

## Artifacts on the pod volume (/workspace, ~500 GB quota)

- `out/epoch1-checkpoint-240`, `out/checkpoint-480`, `out/checkpoint-720` — ~62 GB each
- `out/` root also holds a duplicate final save (= checkpoint-720); delete one pair member to
  reclaim 62 GB
- `data/huggingface-cache` — base model cache (62.6 GB, re-downloadable)
- All three checkpoints are also on HF (public): nshuster/gemma4-31b-augmental-sft-ep{1,2,3}