jingfengya/Qwen3-32B_nas_bank

Qwen3-32B Dense NAS Bank

This repository is the metadata entry point for a split dense-NAS weight bank. The bank contains 64 transformer layers and seven selectable dense linear modules per layer. The source checkpoints are sharded across the linked ModelScope repositories listed in bank_manifest.json.

Use a recipe to download only the selected backbone and linear slices:

python pull_recipe_model.py \
  --repo-id jingfeng0813/Qwen3-32B_nas_bank \
  --recipe-json ./recipes/baselines/paro_w3.json \
  --output-dir ./assembled_model

The downloader verifies the published SHA256 values before writing a HF-style model.safetensors.index.json.

Canonical Legacy Quantization Sources

This bank also publishes 2-, 3-, and 4-bit linear sources from awq, gptq, omni, quip. Each source is converted to ordinary Hugging Face float16 linear tensors. Source-specific fused RMSNorm scales are folded into the affected q/k/v and gate/up projections against the original pretrained backbone, so recipes can mix these published sources without requiring a runtime quantization package or an additional LET transformation.

The extension weights are stored in jingfeng0813/Qwen3-32B_nas_bank_part5, jingfeng0813/Qwen3-32B_nas_bank_part6, jingfeng0813/Qwen3-32B_nas_bank_part7, jingfeng0813/Qwen3-32B_nas_bank_part8; source manifests and recipes are indexed from this main repository.

Source Repair Records

The records/repairs/ directory documents narrowly scoped source repairs and their finite-tensor and remote-hash audits. These records are part of the published provenance for the affected source; the original quantized checkpoint is not modified.