inferno-glm-5.2-q2-expertpack

Inferno ExpertPack v1 for GLM-5.2 Q2

This repository contains the Inferno ExpertPack for GLM-5.2-UD-Q2_K_RoutedQ2K.gguf.

The ExpertPack is a lossless physical reordering of the routed Q2 expert payloads published by Antirez. It keeps the quantized expert bytes unchanged and places each expert's gate, up, and down matrices next to each other. This layout is designed for demand-driven SSD streaming by Inferno on Apple Silicon.

Important

This artifact is not a standalone model and does not replace the source GGUF. Inferno also requires:

• GLM-5.2-UD-Q2_K_RoutedQ2K.gguf
• config.json
• tokenizer.json

These source files are available from antirez/glm-5.2-gguf.

Artifact

The file starts with a 4 KiB header followed by fixed-size expert records in layer-major, expert-major order:

header -> [layer][expert][gate][up][down]

This lets Inferno issue large contiguous SSD reads for selected experts instead of relying on scattered page faults into the component-major GGUF layout.

Intended Use

The ExpertPack is intended only for running the target GLM-5.2 Q2 artifact with Inferno on Apple Silicon. It contains routed-expert weights only. Embeddings, attention weights, dense layers, normalization weights, the output head, configuration, and tokenizer remain in their original files.

Setup

Place the source files and ExpertPack in one directory:

models/glm-5.2/
  config.json
  tokenizer.json
  GLM-5.2-UD-Q2_K_RoutedQ2K.gguf
  GLM-5.2-UD-Q2_K_RoutedQ2K-Inferno-ExpertPack-v1.q2pack

Build and start Inferno:

git clone https://github.com/nektarlabs/inferno.git
cd inferno
cargo build --release
target/release/inferno

Inferno discovers the ExpertPack in the model directory and validates its version, source layout fingerprint, dimensions, component strides, and exact file size before use. A pack generated from a different GGUF is rejected.

Download

After this repository is published, the artifact can be downloaded with:

hf download allemanfredi/inferno-glm-5.2-q2-expertpack \
  GLM-5.2-UD-Q2_K_RoutedQ2K-Inferno-ExpertPack-v1.q2pack \
  --local-dir models/glm-5.2

Reproducibility

The ExpertPack can be recreated losslessly from the source GGUF:

cargo run --release -p inferno --example pack_experts -- \
  --model models/glm-5.2

Creation is resumable at complete expert-record boundaries. The final file is published atomically after its structure and size have been validated.

Limitations

• This format is specific to Inferno and is not a GGUF replacement.
• Compatibility is limited to the exact source layout recorded in the header.
• Runtime behavior and output quality remain those of the source quantized model because expert values are not modified.
• This work is under active development and has not been validated on every Apple Silicon configuration.

Credits

GLM-5.2 was created and released by Z.ai.

The Q2 GGUF used as the direct source for this artifact was produced and published by Antirez in antirez/glm-5.2-gguf.

Inferno changes only the physical ordering of routed-expert payloads required for its SSD streaming path. It does not modify their quantized values.

License

This derived artifact is distributed under the MIT License, following the license of the source GLM-5.2 model. See the source model repositories and the Inferno license for details.