#!/bin/bash
# Launch the ComfyUI backend (headless) for the Qwen-Image-Edit API.
# gfx906 (MI50) has no flash-attention, so use the pytorch cross-attention path.
set -e
BASE=/media/tour/APPS/comfyui
cd "$BASE/ComfyUI"
source "$BASE/venv/bin/activate"

# MI50 / Vega20 is happiest in fp16; avoid bf16 emulation.
export PYTORCH_HIP_ALLOC_CONF=expandable_segments:True
export HSA_ENABLE_SDMA=0

# Split cross-attention chunks the attention matmul -> much lower peak VRAM,
# which is what lets the 20B Q8 edit model + reference-image sequence fit in 32GB.
exec python main.py \
    --listen 127.0.0.1 \
    --port 8188 \
    --use-split-cross-attention \
    "$@"