User Guide

Model assets used in testing

SD/SDXL: use the checkpoint’s built-in CLIP/VAE, or point the loader to an external VAE if preferred.
Flux: external text encoders clip_l.safetensors and t5xxl_fp8_e4m3fn_scaled.safetensors, plus VAE ae.safetensors placed in models/text_encoders and models/vae, then selected in the Flux loader.

For heavy checkpoints (Flux Dev 2 FP32, etc.), run Hardware Validator (MultiGPU) first and feed its gpu_ids string into the loader plus its ok boolean into the loader’s safety_ok input.
Use the debug loaders with log_vram_snapshot when diagnosing sharding; switch to non-debug loaders for production graphs.
Pair the debug sampler with a display node to capture shard layout, per-GPU block counts, and VRAM deltas in comfyui.log.

Use the Load Checkpoint (MultiGPU) Debug or Load Flux Checkpoint (MultiGPU) Debug node to load your model; set gpu_ids (e.g., 0,1,2,3) and optionally enable log_vram_snapshot for before/after VRAM numbers.
Pair with KSampler (MultiGPU) Debug to see shard layout, per-GPU block counts, VRAM deltas, and peaks during sampling. Connect its status output to a display node to log details into comfyui.log.
For production graphs without extra outputs, use the non-debug loader/sampler counterparts.

Load Checkpoint (MultiGPU) Debug: log_vram_snapshot appends per-GPU VRAM readings (in GB) before and after sharding to the node status and comfyui.log.
KSampler (MultiGPU) Debug: reports shard counts per device, flags GPUs without assigned blocks, logs live VRAM deltas and peak usage per GPU, and lists how many blocks each GPU executed during the last sampling pass.