scripts/run_jobs.sh

#!/usr/bin/env bash
# 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