hf2jns_mtp

hf2jns_mtp main.jns shard.safetensors... out.jnshf2jns_mtp main.jns hf:owner/repo[@revision] out.jnshf2jns_mtp main.jns head.gguf out.jns

hf2jns_mtp writes a one-layer JNS file that holds a model's multi-token prediction (MTP) block. GGUF conversions of Qwen3-Next and similar models drop that block. The original checkpoint keeps it under `mtp.*` (bf16). This tool quantizes those matrices to Q4-K with Janas's quantizer, writes routed experts as expert slots (or a dense feed-forward as the one slot, as gguf2jns does), and everything else as resident tensors named like the main model's layer tensors (`blk.0.*`) plus `mtp.*` for the block's own parts.The metadata region is copied from the main model's JNS: hyperparameters match. Whether the model is dense, and how many experts it has, comes from the main file; the shapes come from the checkpoint.Three input forms:1. One or more safetensors shards that contain the block (or a file holding only those tensors). Give every shard that has a part of it. At most 64 shard arguments.2. hf:owner/repo[@revision]: the block's tensors alone are read from Hugging Face with range requests into a temporary file named the output path plus `.mtp.safetensors`, then converted and deleted. One `hf:` repository per run.3. A GGUF draft head whose last block is the prediction block (block count minus the nextn predict layers). Tensors are copied in their types (`blk.N.*` as `blk.0.*`, `blk.N.nextn.*` as `mtp.*`). Token embeddings and the output head are left out; the engine uses the main model's. The head's architecture, expert count, and width must match the main file.janas-get runs this command when a catalog model has an MTP extra. Wrong usage prints a short help (the safetensors and GGUF forms) and exits 2. Failures print `hf2jns_mtp:` and exit 1. Success prints each tensor mapping, expert progress every 64 experts, and `written` with sizes.Checkpoint RMS norms are treated as zero-centred (the effective scale is 1 + w): they are stored plus one, matching GGUF. Only BF16 safetensors tensors are accepted.

main.jns

Already converted trunk model. Its architecture and expert layout are required.

Hugging Face shard (or a file of only the block's tensors). Repeat for every shard that holds `mtp.*`.

Fetch `mtp.*` tensors from that repository. Optional @revision selects a git revision.

Draft-head GGUF whose last block is the prediction block. Detected when there are exactly three arguments and this path ends in `.gguf`.

Output one-layer JNS path.

The stderr usage line omits the hf: form; it still works. Qwen3.5 and Qwen3.6 keep their prediction block inside the main GGUF, so gguf2jns converts it with no extra step. Qwen3-Next still needs this tool (or a catalog file from janas-get). The GGUF-head path copies types as they are; the safetensors path requantizes to Q4-K. Built into `bin/` by Janas's `./build.sh` from `tools/hf2jns_mtp.c`. It is not a separate package.

hf2jns_mtp ships with Janas, written by Maurizio "camauri" Cammalleri (first public release September 2026, GNU GPL 3 or later).

gguf2jns(1), janas-get(1), jns_planes(1), janas-chat(1), janas-bench(1), huggingface-cli(1)

Source code · Documentation