Upload folder using huggingface_hub

scripts/training/run_train_flux_deepspeed.sh

@@ -33,8 +33,8 @@ accelerate launch \
     "$PROJECT_DIR/scripts/training/train_flux_deepspeed.py" \
     --embedding-dir "$EMBEDDING_DIR" \
     --output-dir "/data0/checkpoints/flux_lora_ds" \
-    --batch-size 4 \
-    --gradient-accumulation 4 \
+    --batch-size 1 \
+    --gradient-accumulation 8 \
     --learning-rate 1e-4 \
     --lr-warmup-steps 500 \
     --max-train-steps 100000 \

scripts/training/train_flux_deepspeed.py

@@ -105,8 +105,8 @@ def main():
         cache_dir=args.cache_dir,
     )
 
-    # ZeRO-2 is compatible with gradient checkpointing (ZeRO-3 is not on PyTorch 2.12+)
-    transformer.enable_gradient_checkpointing()
+    # Gradient checkpointing disabled: incompatible with PEFT LoRA on PyTorch 2.12
+    # (causes illegal memory access during backward). Batch 1 fits in 96GB without it.
 
     lora_config = LoraConfig(
         r=args.lora_rank,
@@ -116,6 +116,11 @@ def main():
     )
     transformer = get_peft_model(transformer, lora_config)
 
+    # Cast LoRA params to fp32 to avoid NaN gradients in bf16 backprop
+    for name, p in transformer.named_parameters():
+        if p.requires_grad:
+            p.data = p.data.float()
+
     if accelerator.is_main_process:
         transformer.print_trainable_parameters()
 
@@ -199,7 +204,7 @@ def main():
 
                 # Target: velocity (noise - signal)
                 target = noise - packed_latents
-                loss = F.mse_loss(model_pred, target)
+                loss = F.mse_loss(model_pred.float(), target.float())
 
                 accelerator.backward(loss)