HeteroSD checkpoint archive

Public checkpoint archive for HeteroSD.

Built with Llama. The Llama-derived models are named Llama-3.2-1B-HeteroSD and Llama-3.2-3B-HeteroSD; directory aliases are llama1b and llama3b.

This repository contains draft models and training checkpoints. It does not contain the target models or the entire source-code workspace.

Directory layout

Directory Contents Intended use
awq/ours/<model>/ Distilled AWQ draft exports used by the dual-MI210 evaluation AWQ-capable inference runtime; includes model configuration and tokenizer
a4w4kv8/ours_best/<model>/ Selected A4W4KV8 exports used for paper evaluation HeteroSD custom FPGA-aligned parameter loader; not a standard Hugging Face model
a4w4kv8/osd_step24000/<model>/ A4W4KV8 exports distilled with OSD, step 24000 HeteroSD custom parameter loader
training_checkpoints/a4w4kv8_ours_epoch3_final/<model>/ Saved final epoch-3 training state and metrics Training recovery; these are not the intermediate best checkpoints used in evaluation

Model directory names: llama1b, llama3b, qwen05b, qwen15b.

The selected A4W4KV8 evaluation exports are epoch 3, step 6000 for Llama 1B/3B and epoch 3, step 4000 for Qwen 0.5B/1.5B. Step numbers are within the epoch. The corresponding intermediate optimizer-state files were pruned during training; the exported inference parameters are preserved separately.

source_manifest.json records the source files for the home-resident artifacts. best_checkpoint_manifest.json, when present, records the source node, original paths, and SHA-256 checksums of the selected A4W4KV8 exports after verified backup.

During this archive's initial upload, a directory may not yet be present. An UPLOAD_VERIFIED.json record is added only after all intended checkpoint files have been verified against the Hub.

Upstream models

The draft families are Llama 3.2 1B/3B and Qwen 2.5 0.5B/1.5B. Applicable upstream model licenses and usage terms remain in force. Llama derivatives use the Llama 3.2 Community License; Qwen 0.5B/1.5B derivatives retain their upstream Apache 2.0 licenses. Copies are included in licenses/ and in each model directory. No new blanket license is assigned to these derived weights by this archive.

HeteroSD modifications include quantization, target-guided distillation, and export to the formats documented above.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support