[AMD] [Issue] FLUX Dev initialisis

[AI-JUNAITa]

Why aren’t FLUX Dev U-net ckpt being initialized? But are being frozen to initialize?

Workflow: https://pastebin.com/qfKUrJc9

Comfy command,

HSA_OVERRIDE_GFX_VERSION=9.0.0 \
        CL_DEVICE_GLOBAL_FREE_MEMORY_AMD=24396768 \
        CL_DEVICE_GLOBAL_MEM_SIZE=25189056512 \
        CL_DEVICE_MAX_MEM_ALLOC_SIZE=21410698032 \
PYTORCH_HIP_ALLOC_CONF=garbage_collection_threshold:0.45,max_split_size_mb:2560 \
        python3.11 main.py --gpu-only --disable-smart-memory --disable-cuda-malloc \
        --listen 127.0.0.1 --auto-launch --disable-xformers \
        --force-upcast-attention --use-split-cross-attention \
        --verbose DEBUG \
        --port $((RANDOM % (65535 - 4097 + 1) + 4097)) --preview-method latent2rgb \
        --output-directory /opt/local/.8A/terminal/ComfyAI/Downloads/04112025

HW,

ASRock AMD DeskMini X300 × 14.9GiB UniVRAM’s × AMD R4750G APU
OS: ROCm 6.3.x-OpenCL-AUR on Artix Linux

Freeze logs at 7% after CLIP unloads,

Requested to load FluxClipModel_
lowvram: loaded module regularly
t5xxl.transformer.shared Embedding(32128, 4096)
lowvram: loaded module regularly 
clip_l.transformer.text_model.embeddings.token_embedding Embedding(49408, 768) lowvram: loaded module regularly
t5xxl.transformer.encoder.block.9.layer.1.DenseReluDense.wo
Linear(in_features=10240, out_features=4096, bias=False)
lowvram: loaded module regularly t5xxl.transformer.encoder.block.9.layer.0.SelfAttention.v 
Linear(in_features=4096, out_features=4096, bias=False)
lowvram: loaded module regularly clip_l.transformer.text_model.encoder.layers.9.mlp.fc1
Linear(in_features=768, out_features=3072, bias=True)
lowvram: loaded module regularly clip_l.transformer.text_model.encoder.layers.9.self_attn.v_proj
Linear(in_features=768, out_features=768, bias=True) lowvram: loaded module regularly clip_l.transformer.text_projection Linear(in_features=768, out_features=768, bias=False)
lowvram: loaded module regularly
clip_l.transformer.text_model.embeddings.position_embedding Embedding(77, 768) lowvram: loaded module regularly t5xxl.transformer.encoder.final_layer_norm 
T5LayerNorm()
lowvram: loaded module regularly
t5xxl.transformer.encoder.block.9.layer.1.layer_norm T5LayerNorm()
lowvram: loaded module regularly
t5xxl.transformer.encoder.block.0.layer.0.SelfAttention.relative_attention_bias Embedding(32, 64)
lowvram: loaded module regularly
clip_l.transformer.text_model.final_layer_norm LayerNorm((768,), eps=1e-05, elementwise_affine=True)
lowvram: loaded module regularly
clip_l.transformer.text_model.encoder.layers.9.layer_norm2 LayerNorm((768,), eps=1e-05, elementwise_affine=True) loaded completely 9.5367431640625e+25 4903.231597900391
True CLIP model load device: cuda:0, offload device: cuda:0, current: cuda:0