Upload job_ml.sh with huggingface_hub

job_ml.sh

@@ -1,9 +1,9 @@
 #!/bin/bash
 ###############################################################################
-# AZURE ML JOB - TRAIN FOREVER
+# AZURE ML JOB - TRAIN FOREVER (FIXED VERSION)
 #
 # Flow: setup → install → restore scripts/checkpoints → download data →
-#       precompute embeddings → train FOREVER (never exits voluntarily)
+#       train FOREVER (never exits voluntarily)
 #
 # Auto-backup every 30 min. On evict (SIGTERM) → backup → exit.
 # Next run → resume from last checkpoint.
@@ -63,11 +63,18 @@ to_upload = final if final.exists() else (checkpoints[-1] if checkpoints else No
 
 if to_upload:
     path_in_repo = f"{ckpt_dir.name}/{to_upload.name}"
-    print(f"  Uploading {to_upload.name} -> {repo_id}/{path_in_repo} (job: {os.environ.get('JOB_ID', 'default')})")
+    print(f"  Uploading {to_upload.name} -> {repo_id}/{path_in_repo}")
     upload_folder(folder_path=str(to_upload), repo_id=repo_id, path_in_repo=path_in_repo, repo_type="model")
     print("  Done!")
 else:
     print("  Nothing to upload")
+
+# Also upload samples if they exist
+samples_dir = ckpt_dir / "samples"
+if samples_dir.exists() and any(samples_dir.glob("*.png")):
+    print(f"  Uploading samples...")
+    upload_folder(folder_path=str(samples_dir), repo_id=repo_id, path_in_repo=f"{ckpt_dir.name}/samples", repo_type="model")
+    print("  Samples uploaded!")
 PYEOF
 }
 
@@ -79,7 +86,6 @@ trap 'echo "[TRAP] Evicted/killed"; kill $BACKUP_PID 2>/dev/null; backup_checkpo
 ###############################################################################
 log "=== [1/7] Storage ==="
 
-# Auto-detect all NVMe devices and mount them
 MOUNT_IDX=0
 DATA_DIRS=()
 for dev in $(lsblk -dpno NAME,TYPE 2>/dev/null | grep disk | awk '{print $1}' | grep nvme); do
@@ -96,26 +102,17 @@ for dev in $(lsblk -dpno NAME,TYPE 2>/dev/null | grep disk | awk '{print $1}' |
     MOUNT_IDX=$((MOUNT_IDX + 1))
 done
 
-# Fallback: if no NVMe found, use local dirs
 if [ ${#DATA_DIRS[@]} -eq 0 ]; then
     log "  No NVMe found, using local directories"
     mkdir -p /data0 /data1
     DATA_DIRS=("/data0" "/data1")
 fi
 
-# Use first disk for data/models, second (or same) for logs
 export DATA0="${DATA_DIRS[0]}"
 export DATA1="${DATA_DIRS[1]:-$DATA0}"
 log "  DATA0=$DATA0 DATA1=$DATA1 (${#DATA_DIRS[@]} disks)"
 df -h "${DATA_DIRS[@]}" 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" "$PROJECT_DIR/data/processed" 2>/dev/null || true
-ln -sfn "$DATA0/datasets/raw" "$PROJECT_DIR/data/raw" 2>/dev/null || true
-ln -sfn "$DATA0/checkpoints" "$PROJECT_DIR/checkpoints" 2>/dev/null || true
-ln -sfn "$DATA1/outputs" "$PROJECT_DIR/outputs" 2>/dev/null || true
-log "  Done"
 mkdir -p "$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
@@ -132,7 +129,7 @@ export HF_HOME="$DATA0/models"
 log "=== [2/7] Install ==="
 pip3 install -q --no-cache-dir \
     torch torchvision torchaudio \
-    diffusers transformers accelerate deepspeed \
+    diffusers transformers accelerate \
     peft datasets webdataset safetensors \
     opencv-python-headless tqdm huggingface_hub \
     img2dataset pandas pyarrow requests 2>&1 | tail -5
@@ -175,7 +172,7 @@ for i in range(torch.cuda.device_count()):
 "
 
 ###############################################################################
-# 5. DATA - Download COYO (100K images)
+# 5. DATA - Download shards
 ###############################################################################
 log "=== [5/7] Data ==="
 SHARD_DIR="$DATA0/datasets/processed/flux_train/shards"
@@ -184,11 +181,9 @@ mkdir -p "$SHARD_DIR"
 SHARD_COUNT=$(find "$SHARD_DIR" -name "*.tar" 2>/dev/null | wc -l | tr -d '[:space:]')
 SHARD_COUNT=${SHARD_COUNT:-0}
 
-# Need at least 20 shards (~20K images) to train properly
 if [ "$SHARD_COUNT" -ge 20 ]; then
     log "  Shards ready ($SHARD_COUNT)"
 else
-    # Pull existing shards from HF first
     log "  Pulling existing shards from HF..."
     python3 -u << PYEOF || true
 from huggingface_hub import snapshot_download
@@ -208,7 +203,6 @@ PYEOF
     SHARD_COUNT=${SHARD_COUNT:-0}
 
     if [ "$SHARD_COUNT" -lt 20 ]; then
-        # Download from COYO - 5 partitions, filter top 100K images
         log "  Downloading COYO metadata (5 partitions)..."
         COYO_RAW="$DATA0/datasets/raw/coyo"
         COYO_FILTERED="$DATA0/datasets/raw/coyo_filtered"
@@ -231,7 +225,6 @@ snapshot_download(
 print('    Metadata OK')
 PYEOF
 
-        # Filter - top 100K aesthetic images
         mkdir -p "$COYO_FILTERED"
         python3 -u "$PROJECT_DIR/scripts/data_collection/filter_coyo.py" \
             --input-dir "$COYO_RAW/data" \
@@ -241,14 +234,13 @@ PYEOF
             --max-watermark 0.7 \
             --max-records 100000
 
-        # img2dataset - download images
         python3 -u -c "
 import pandas as pd
 df = pd.read_parquet('$COYO_FILTERED/coyo_aesthetic.parquet')
 df[['url','text']].rename(columns={'url':'URL','text':'TEXT'}).to_parquet('$SHARD_DIR/_urls.parquet', index=False)
 print(f'    {len(df)} URLs ready for download')
 "
-        log "  Running img2dataset (100K images, ~30-60 min)..."
+        log "  Running img2dataset (100K images)..."
         img2dataset \
             --url_list "$SHARD_DIR/_urls.parquet" \
             --input_format parquet \
@@ -273,8 +265,7 @@ print(f'    {len(df)} URLs ready for download')
             exit 1
         fi
 
-        # Upload shards to HF for next run
-        log "  Uploading shards to HF (for faster resume)..."
+        log "  Uploading shards to HF..."
         python3 -u << PYEOF || true
 from huggingface_hub import HfApi, upload_folder
 api = HfApi()
@@ -292,24 +283,10 @@ SHARD_COUNT=$(find "$SHARD_DIR" -name "*.tar" 2>/dev/null | wc -l | tr -d '[:spa
 log "  Data ready: $SHARD_COUNT shards"
 
 ###############################################################################
-# 6. TRAIN FOREVER (skip precompute, encode on-the-fly)
+# 6. TRAIN FOREVER
 ###############################################################################
 log "=== [6/7] Train FOREVER ==="
 
-# Generate config - single process (script handles 2 GPU split internally)
-ACCEL_CFG="$PROJECT_DIR/configs/accelerate_job.yaml"
-mkdir -p "$PROJECT_DIR/configs"
-
-cat > "$ACCEL_CFG" << 'EOF'
-compute_environment: LOCAL_MACHINE
-distributed_type: "NO"
-num_machines: 1
-num_processes: 1
-mixed_precision: bf16
-gpu_ids: all
-use_cpu: false
-EOF
-
 # Auto-backup every 30 min
 (
     while true; do
@@ -321,8 +298,7 @@ EOF
 BACKUP_PID=$!
 log "  Auto-backup started (PID $BACKUP_PID, every 30min)"
 
-# Train using train_flux_lora.py (encodes on-the-fly, no precompute needed)
-# GPU 0: VAE + text encoders, GPU 1: transformer training
+# Train using FIXED train_flux_lora.py
 log "  Starting Flux LoRA training (job: $JOB_ID)..."
 python3 -u "$PROJECT_DIR/scripts/training/train_flux_lora.py" \
     --model-name "black-forest-labs/FLUX.1-schnell" \
@@ -331,11 +307,15 @@ python3 -u "$PROJECT_DIR/scripts/training/train_flux_lora.py" \
     --batch-size 1 \
     --gradient-accumulation 8 \
     --learning-rate 1e-4 \
-    --lr-warmup-steps 500 \
+    --lr-scheduler constant \
+    --lr-warmup-steps 100 \
     --max-train-steps 999999999 \
     --save-steps 2000 \
+    --sample-steps 2000 \
     --lora-rank 128 \
-    --lora-alpha 128 \
+    --lora-alpha 64 \
+    --weighting-scheme none \
+    --max-grad-norm 1.0 \
     --encode-device cuda:0 \
     --train-device cuda:1