Buckets:
| # Planned HF Jobs launch commands for the TileSparse reproduction. | |
| # Blocked until Jobs credits are added to the dmitry-rov account | |
| # (hf jobs run 402s with "Pre-paid credit balance is insufficient"). | |
| set -euo pipefail | |
| BUCKET="hf://buckets/dmitry-rov/tilesparse-repro" | |
| # --- Claim 3: empirical kernel benchmark (roofline validation) --- | |
| # Smoke test first (cheap, ~1-2 min on cpu-basic or a small GPU): | |
| hf jobs uv run scripts/attention_kernel_bench.py --quick --out /data/kernel_bench_quick.json \ | |
| --flavor a10g-large --timeout 10m \ | |
| -v "$BUCKET:/data" | |
| # Full scaled run on a real GPU (L values up to 32768, D up to 64): | |
| hf jobs uv run scripts/attention_kernel_bench.py --out /data/kernel_bench_full.json \ | |
| --flavor a10g-large --timeout 20m \ | |
| -v "$BUCKET:/data" | |
| # --- Claims 1 & 2: sparse-attention accuracy reconstruction --- | |
| # Smoke test first (2 tiny docs, short seqs, 3 configs): | |
| hf jobs uv run scripts/sparse_attention_repro.py --quick --out /data/sparse_repro_quick.json \ | |
| --flavor a10g-large --timeout 10m \ | |
| -v "$BUCKET:/data" | |
| # Full run (20 WikiText-2 test docs, up to 1024 tokens, 5 configs): | |
| hf jobs uv run scripts/sparse_attention_repro.py --out /data/sparse_repro_full.json \ | |
| --flavor a10g-large --timeout 30m \ | |
| -v "$BUCKET:/data" | |
| # --- Sync results down locally so trackio logbook auto-captures them --- | |
| hf buckets sync "$BUCKET" ./outputs --existing | |
Xet Storage Details
- Size:
- 1.41 kB
- Xet hash:
- 708015600155b9b3e38dc9abef0e3d8929684a28698ca6671325e1c3c572aeba
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.