pmetal pack-experts

Pack expert weights from a local model directory so sparse MoE models can use SSD-offloaded expert storage.

Usage

pmetal pack-experts \
  --model <MODEL_DIR> \
  --output <PACKED_EXPERT_DIR> \
  [OPTIONS]

Example

pmetal pack-experts \
  --model ./Qwen3.5-122B-A10B \
  --output ./packed_experts \
  --bits 4

pmetal infer \
  --model ./Qwen3.5-122B-A10B \
  --experts-dir ./packed_experts \
  --prompt "Explain routed expert offload" \
  --chat

Parameters

Parameter	Default	Description
--model	required	Model directory containing config.json and safetensors
--output	./packed_experts	Output directory for packed expert files
--bits	—	Optional quantization bit width, usually 4 or 2

See Also

• pmetal infer — Use --experts-dir
• pmetal bench — Benchmark packed expert workloads