Converter_Protocol

Weight Packing Utility

Convert Hugging Face checkpoints into the Disk-LLM packed format: layer-oriented memmap shards designed for predictable, inspectable disk access during inference experiments.

The converter and manifest flow are already solid for text-path packing. The next Qwen 3.5 milestone after conversion is the native linear_attention runtime adapter, not the packing stage itself.

01. Source Input

STATUS: READY

Local Checkpoint Path

02. Sharding Config

Shard Strategy Layer-Oriented

Layer-Oriented Packing

Subgraph Filtering (Text-only)

inventory_2

03. Output Manifest

Shard Layout

embeddings/ layers/ final/

Manifest

JSON

CONVERSION_CONSOLE

Ready to convert weights...