Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
@@ -0,0 +1,67 @@
# Submission: 8xH100 QK5.25 TTT-disabled PR2069 rerun

Candidate: `best4x_ttt_disabled_qk525`
Mean val_bpb: `1.23485583`
Artifact bytes: `15843310`
Hardware: `8xNVIDIA H100 80GB HBM3`

> Note: this generated package contains fewer than 3 seeds. For a new SOTA record PR, rerun the selected candidate with additional seeds or submit as a non-record result.

## Results

| Seed | val_bpb | val_loss | artifact bytes | train seconds |
|---|---:|---:|---:|---:|
| 1337 | 1.23485583 | 2.08500235 | 15843310 | 702.656 |

## Technique

4xH100 promotion of the best 1xH100 TTT-disabled QK5.25 control

Candidate environment is captured in `candidate_env.json`; source batch rows are copied into `run_result*.json`.

## Reproduction

From this record folder:

```bash
export ARTIFACT_BUDGET_STRICT="0"
export CANDIDATE_IMPL="autoregressive_gpt"
export MAX_WALLCLOCK_SECONDS="600"
export QK_GAIN_INIT="5.25"
export SKIP_PRE_QUANT_FINAL_EVAL="1"
export TRAIN_BATCH_TOKENS="2097152"
export TRAIN_LOG_EVERY="100"
export TTT_ENABLED="0"
export VAL_LOSS_EVERY="0"
export VOCAB_SIZE="1024"
export NPROC_PER_NODE="${NPROC_PER_NODE:-8}"
export DATA_PATH="${DATA_PATH:-../../data/datasets/fineweb10B_sp1024}"
export TOKENIZER_PATH="${TOKENIZER_PATH:-../../data/tokenizers/fineweb_1024_bpe.model}"
export RUN_ID="${RUN_ID:-submission_rerun}"
torchrun --standalone --nproc_per_node="$NPROC_PER_NODE" train_gpt.py
```

For a 3-seed verification run:

```bash
for SEED in 42 314 1234; do
SEED=$SEED RUN_ID=submission_seed${SEED} \
torchrun --standalone --nproc_per_node=8 train_gpt.py > train_seed${SEED}_rerun.log 2>&1
done
```

## Compliance Notes

- Training under 600s in packaged run(s): `False`
- Artifact under 16,000,000 bytes in packaged run(s): `True`
- Three-seed evidence included: `False`

## Included Files

- `README.md`
- `submission.json`
- `train_gpt.py`
- `train.log`
- `candidate_env.json`
- `train*.log`
- `run_result*.json`
Original file line number Diff line number Diff line change
@@ -0,0 +1,16 @@
{
"ARTIFACT_BUDGET_STRICT": "0",
"CANDIDATE_IMPL": "autoregressive_gpt",
"DATA_PATH": "/workspace/parameter-golf/data/datasets/fineweb10B_sp1024",
"MAX_WALLCLOCK_SECONDS": "600",
"QK_GAIN_INIT": "5.25",
"RUN_ID": "h100_batch_20260501T153233Z_01_best4x_ttt_disabled_qk525",
"SKIP_PRE_QUANT_FINAL_EVAL": "1",
"SUBMISSION_CODE_PATHS": "/workspace/parameter-golf/train_gpt.py,/workspace/parameter-golf/candidates,/workspace/parameter-golf/fastest/scripts/artifact_budget_analyzer.py",
"TOKENIZER_PATH": "/workspace/parameter-golf/data/tokenizers/fineweb_1024_bpe.model",
"TRAIN_BATCH_TOKENS": "2097152",
"TRAIN_LOG_EVERY": "100",
"TTT_ENABLED": "0",
"VAL_LOSS_EVERY": "0",
"VOCAB_SIZE": "1024"
}
Loading