Halo Strix 1500 step LoRA for Comfyui Flux 1 dev GMTEK EVO2 128gb total 64gb for VRAM

[bkpaine1]

[RANK 0] 2025-12-27 05:35:43,274 [INFO] Moving AutoencoderKL to accelerator, converting from torch.float32 to torch.bfloat16
[RANK 0] 2025-12-27 05:35:43,547 [INFO] [Rank 0] Processing 1 local validation work items (distributed_mode=False, total=1, work_item_prompts=[''])
Epoch 1/1, Steps: 60%|███████▏ | 900/1500 [11:06:47<7:17:18, 43.73s/it, grad_absmax=0.000372, lr=0.0001, step_loss=0.292][RANK 0] 2025-12-27 06:49:08,997 [INFO] Loading AutoencoderKL from black-forest-labs/FLUX.1-dev
[RANK 0] 2025-12-27 06:49:09,333 [INFO] Moving AutoencoderKL to accelerator, converting from torch.float32 to torch.bfloat16
[RANK 0] 2025-12-27 06:49:09,617 [INFO] [Rank 0] Processing 1 local validation work items (distributed_mode=False, total=1, work_item_prompts=[''])
Epoch 1/1, Steps: 66%|████████▌ | 993/1500 [12:15:07<6:09:20, 43.71s/it, grad_absmax=0.00028, lr=0.0001, step_loss=0.248]

Works just great! startup script if you don't want to fight the gui.

cat > ~/venvbc/SimpleTuner/config/config.json << 'EOF'
{
"--data_backend_config": "config/multidatabackend.json",
"--output_dir": "/home/test/venvbc/ST-OUTPUT/test_v1",
"--model_type": "lora",
"--model_family": "flux",
"--pretrained_model_name_or_path": "black-forest-labs/FLUX.1-dev",
"--train_batch_size": 1,
"--max_train_steps": 1500,
"--num_train_epochs": 0,
"--learning_rate": "1e-4",
"--lora_rank": 16,
"--mixed_precision": "bf16",
"--gradient_checkpointing": true,
"--resolution": 1024,
"--seed": 42,
"--optimizer": "adamw_bf16"
}
EOF

cd ~/venvbc/SimpleTuner
source ~/venvbc/bin/activate
unset PYTORCH_HIP_ALLOC_CONF
unset PYTORCH_ALLOC_CONF
export HSA_ENABLE_SDMA=0
export HSA_USE_SVM=0

python << 'PYTHONCODE'
import torch
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.set_float32_matmul_precision('medium')

import runpy
runpy.run_module('simpletuner.train', run_name='main', alter_sys=True)
PYTHONCODE

[bghira]

have you tried enabling torch inductor?

also, set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1

[bkpaine1]

No but I will. TheROCK7.10+ has changed the game this is a CUDA killer at this price point. will report back after this lora run love this tool though! KILLER! GUI a bit overload and some caching sticks but otherwise a HUGE win for TheROCK thank you for your efforts!