v6: fix shard threshold, separate Flux model download, better error logs

job_ml.sh

@@ -156,7 +156,7 @@ mkdir -p "$SHARD_DIR"
 SHARD_COUNT=$(find "$SHARD_DIR" -name "*.tar" 2>/dev/null | wc -l | tr -d '[:space:]')
 SHARD_COUNT=${SHARD_COUNT:-0}
 
-if [ "$SHARD_COUNT" -ge 5 ]; then
+if [ "$SHARD_COUNT" -ge 2 ]; then
     log "  Shards ready ($SHARD_COUNT)"
 else
     # Try pull from HF first
@@ -178,7 +178,7 @@ PYEOF
     SHARD_COUNT=$(find "$SHARD_DIR" -name "*.tar" 2>/dev/null | wc -l | tr -d '[:space:]')
     SHARD_COUNT=${SHARD_COUNT:-0}
 
-    if [ "$SHARD_COUNT" -lt 5 ]; then
+    if [ "$SHARD_COUNT" -lt 2 ]; then
         # Download from COYO
         log "  Downloading from COYO (1 partition, top 20K images)..."
         COYO_RAW="$DATA0/datasets/raw/coyo"
@@ -267,17 +267,30 @@ EMB_COUNT=${EMB_COUNT:-0}
 if [ "$EMB_COUNT" -gt 0 ]; then
     log "  Already computed ($EMB_COUNT)"
 else
-    log "  Computing..."
+    # Download Flux model first (separate step so we see if it fails)
+    log "  Downloading Flux model..."
+    python3 -u << PYEOF
+from huggingface_hub import snapshot_download
+snapshot_download(
+    repo_id="black-forest-labs/FLUX.1-schnell",
+    cache_dir="$DATA0/models",
+)
+print("    Flux model cached")
+PYEOF
+
+    log "  Computing embeddings..."
     python3 -u "$PROJECT_DIR/scripts/training/precompute_embeddings.py" \
         --data-dir "$SHARD_DIR" \
         --output-dir "$EMB_DIR" \
         --batch-size 4 \
         --cache-dir "$DATA0/models" \
-        --device cuda:0
+        --device cuda:0 2>&1 | tee "$DATA1/logs/precompute.log" | tail -30
+
     EMB_COUNT=$(find "$EMB_DIR" -name "shard-*.pt" 2>/dev/null | wc -l | tr -d '[:space:]')
     EMB_COUNT=${EMB_COUNT:-0}
     if [ "$EMB_COUNT" -eq 0 ]; then
-        log "  FATAL: No embeddings"
+        log "  FATAL: No embeddings. Last 20 lines of log:"
+        tail -20 "$DATA1/logs/precompute.log" 2>/dev/null || true
         exit 1
     fi
 fi