Meta-Llama-3.1-70B_nas_bank

Meta-Llama-3.1-70B Dense NAS Bank

This repository is the metadata entry point for a split dense-NAS weight bank for Meta-Llama-3.1-70B. The bank exposes seven selectable dense linear modules in each of 80 transformer layers (560 searchable positions). It contains 2-, 3-, and 4-bit sources from LRQ+, SliM-LLM, PAROQuant stage1, and GPTAQ. Packed GPTAQ checkpoints are published as dequantized Hugging Face linear slices. Source tensors are distributed across the repositories listed in bank_manifest.json.

Given a dense-NAS recipe, the downloader retrieves only the selected linear slices and the non-search backbone from the recipe's base source, verifies all published SHA256 hashes, and writes a Hugging Face-compatible sharded model:

python pull_recipe_model.py \
  --repo-id jingfeng0813/Meta-Llama-3.1-70B_nas_bank \
  --recipe-json ./recipes/baselines/paro_w3.json \
  --output-dir ./assembled_model

The published source manifests record tensor paths, shapes, byte sizes, and SHA256 hashes. Quantizer-only metadata and bias tensors present in the source checkpoints are retained when they are part of the Hugging Face state dict.