v8: mount NVMe properly (no sudo), use train_flux_lora.py (no precompute needed)

job_ml.sh

@@ -41,7 +41,7 @@ user = os.environ.get("HF_USER", "memoryai")
 repo_id = f"{user}/4k-image-model-checkpoints"
 
 if stage == "flux":
-    ckpt_dir = Path("/data0/checkpoints/flux_lora_ds")
+    ckpt_dir = Path("/data0/checkpoints/flux_lora")
 elif stage == "sr2":
     ckpt_dir = Path("/home/adminuser/chungcat/checkpoints/sr_stage2")
 else:
@@ -77,19 +77,26 @@ BACKUP_PID=""
 trap 'echo "[TRAP] Evicted/killed"; kill $BACKUP_PID 2>/dev/null; backup_checkpoint; exit 0' SIGTERM SIGINT
 
 ###############################################################################
-# 1. STORAGE
+# 1. STORAGE - Format & mount 2x NVMe 3.5TB
 ###############################################################################
-log "=== [1/7] Storage ==="
+log "=== [1/7] Storage (NVMe 3.5TB x2) ==="
 for pair in "/dev/nvme0n1:$DATA0" "/dev/nvme1n1:$DATA1"; do
     dev="${pair%%:*}"; mnt="${pair##*:}"
-    if [ -b "$dev" ] && ! mountpoint -q "$mnt" 2>/dev/null; then
-        (command -v sudo &>/dev/null && {
-            blkid "$dev" 2>/dev/null | grep -q ext4 || sudo mkfs.ext4 -q "$dev" 2>/dev/null
-            sudo mkdir -p "$mnt" && sudo mount "$dev" "$mnt" && sudo chown "$(whoami)" "$mnt"
-        }) 2>/dev/null || true
+    if [ -b "$dev" ]; then
+        if ! mountpoint -q "$mnt" 2>/dev/null; then
+            blkid "$dev" 2>/dev/null | grep -q ext4 || mkfs.ext4 -q "$dev"
+            mkdir -p "$mnt"
+            mount "$dev" "$mnt"
+            log "  Mounted $dev -> $mnt"
+        else
+            log "  $mnt already mounted"
+        fi
+    else
+        log "  $dev not found, using local dir"
+        mkdir -p "$mnt"
     fi
-    mkdir -p "$mnt" 2>/dev/null || true
 done
+df -h "$DATA0" "$DATA1" 2>/dev/null | grep -v Filesystem || true
 mkdir -p "$DATA0"/{datasets/{raw,processed},checkpoints,models} "$DATA1"/{outputs,logs}
 mkdir -p "$PROJECT_DIR"/{scripts/{training,data_collection,serving},configs,logs,data}
 ln -sfn "$DATA0/datasets/processed" "$PROJECT_DIR/data/processed" 2>/dev/null || true
@@ -257,73 +264,15 @@ SHARD_COUNT=$(find "$SHARD_DIR" -name "*.tar" 2>/dev/null | wc -l | tr -d '[:spa
 log "  Data ready: $SHARD_COUNT shards"
 
 ###############################################################################
-# 6. PRECOMPUTE EMBEDDINGS
-###############################################################################
-log "=== [6/7] Embeddings ==="
-EMB_DIR="$DATA0/datasets/processed/flux_train/embeddings"
-EMB_COUNT=$(find "$EMB_DIR" -name "shard-*.pt" 2>/dev/null | wc -l | tr -d '[:space:]')
-EMB_COUNT=${EMB_COUNT:-0}
-
-if [ "$EMB_COUNT" -gt 0 ]; then
-    log "  Already computed ($EMB_COUNT)"
-else
-    # Download Flux model first (separate step so we see if it fails)
-    log "  Downloading Flux model (~33GB, may take 5-10 min)..."
-    python3 -u << PYEOF
-import os
-os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
-from huggingface_hub import snapshot_download
-
-snapshot_download(
-    repo_id="black-forest-labs/FLUX.1-schnell",
-    cache_dir="$DATA0/models",
-)
-print("    Flux model cached OK")
-PYEOF
-    log "  Flux model ready"
-
-    log "  Computing embeddings (this takes a while, printing progress)..."
-    python3 -u "$PROJECT_DIR/scripts/training/precompute_embeddings.py" \
-        --data-dir "$SHARD_DIR" \
-        --output-dir "$EMB_DIR" \
-        --batch-size 4 \
-        --cache-dir "$DATA0/models" \
-        --device cuda:0
-
-    EMB_COUNT=$(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 produced"
-        exit 1
-    fi
-fi
-log "  Embeddings ready: $EMB_COUNT"
-
-###############################################################################
-# 7. CONFIG + TRAIN FOREVER
+# 6. TRAIN FOREVER (skip precompute, encode on-the-fly)
 ###############################################################################
-log "=== [7/7] Train FOREVER ==="
+log "=== [6/7] Train FOREVER ==="
 
-# Generate config
+# Generate config - single process (script handles 2 GPU split internally)
 ACCEL_CFG="$PROJECT_DIR/configs/accelerate_job.yaml"
-DS_CFG="$PROJECT_DIR/configs/ds_zero2.json"
 mkdir -p "$PROJECT_DIR/configs"
 
-if [ "$NUM_GPUS" -gt 1 ]; then
-cat > "$ACCEL_CFG" << EOF
-compute_environment: LOCAL_MACHINE
-distributed_type: DEEPSPEED
-deepspeed_config:
-  deepspeed_config_file: $DS_CFG
-  zero3_init_flag: false
-num_machines: 1
-num_processes: $NUM_GPUS
-mixed_precision: bf16
-gpu_ids: all
-use_cpu: false
-EOF
-else
-cat > "$ACCEL_CFG" << EOF
+cat > "$ACCEL_CFG" << 'EOF'
 compute_environment: LOCAL_MACHINE
 distributed_type: "NO"
 num_machines: 1
@@ -332,24 +281,6 @@ mixed_precision: bf16
 gpu_ids: all
 use_cpu: false
 EOF
-fi
-
-cat > "$DS_CFG" << 'EOF'
-{
-  "bf16": {"enabled": true},
-  "zero_optimization": {
-    "stage": 2,
-    "overlap_comm": true,
-    "contiguous_gradients": true,
-    "reduce_bucket_size": 50000000,
-    "allgather_bucket_size": 50000000
-  },
-  "gradient_accumulation_steps": "auto",
-  "gradient_clipping": 1.0,
-  "train_micro_batch_size_per_gpu": "auto",
-  "wall_clock_breakdown": false
-}
-EOF
 
 # Auto-backup every 30 min
 (
@@ -362,54 +293,25 @@ EOF
 BACKUP_PID=$!
 log "  Auto-backup started (PID $BACKUP_PID, every 30min)"
 
-# Train with essentially infinite steps
-case "$TRAIN_STAGE" in
-    flux)
-        accelerate launch \
-            --config_file "$ACCEL_CFG" \
-            "$PROJECT_DIR/scripts/training/train_flux_deepspeed.py" \
-            --embedding-dir "$EMB_DIR" \
-            --output-dir "$DATA0/checkpoints/flux_lora_ds" \
-            --cache-dir "$DATA0/models" \
-            --batch-size 1 \
-            --gradient-accumulation 8 \
-            --learning-rate 1e-4 \
-            --lr-warmup-steps 500 \
-            --max-train-steps 999999999 \
-            --save-steps 2000 \
-            --lora-rank 128 \
-            --lora-alpha 128 \
-            --resume-from-checkpoint auto
-        ;;
-    sr2)
-        accelerate launch \
-            --config_file "$ACCEL_CFG" \
-            "$PROJECT_DIR/scripts/training/train_sr.py" \
-            --stage 2 \
-            --batch-size 4 \
-            --learning-rate 2e-4 \
-            --max-steps 999999999 \
-            --save-steps 5000 \
-            --base-channels 64 \
-            --perceptual-weight 0.1
-        ;;
-    sr3)
-        accelerate launch \
-            --config_file "$ACCEL_CFG" \
-            "$PROJECT_DIR/scripts/training/train_sr.py" \
-            --stage 3 \
-            --batch-size 2 \
-            --learning-rate 1e-4 \
-            --max-steps 999999999 \
-            --save-steps 5000 \
-            --base-channels 64 \
-            --perceptual-weight 0.1
-        ;;
-    *)
-        log "ERROR: TRAIN_STAGE must be flux|sr2|sr3"; exit 1 ;;
-esac
-
-# Should never reach here (infinite training), but just in case
+# Train using train_flux_lora.py (encodes on-the-fly, no precompute needed)
+# GPU 0: VAE + text encoders, GPU 1: transformer training
+log "  Starting Flux LoRA training (2 GPU split, encode on-the-fly)..."
+python3 -u "$PROJECT_DIR/scripts/training/train_flux_lora.py" \
+    --model-name "black-forest-labs/FLUX.1-schnell" \
+    --data-dir "$SHARD_DIR" \
+    --output-dir "$DATA0/checkpoints/flux_lora" \
+    --batch-size 1 \
+    --gradient-accumulation 8 \
+    --learning-rate 1e-4 \
+    --lr-warmup-steps 500 \
+    --max-train-steps 999999999 \
+    --save-steps 2000 \
+    --lora-rank 128 \
+    --lora-alpha 128 \
+    --encode-device cuda:0 \
+    --train-device cuda:1
+
+# Should never reach here
 kill $BACKUP_PID 2>/dev/null || true
 backup_checkpoint
 log "=== Done ==="