FlashSVD / README.md
Zishan-Shao's picture
Document complete exact Table 1 Source migration and pinned downloader
a2d28cb verified
|
Raw History Blame Contribute Delete
3.57 kB
# FlashSVD checkpoints and BASIS experiment inputs
## BASIS: download experiment inputs here
For the [Additional Experiments](https://github.com/Zishan-Shao/Basis/tree/handoff/paper-20260915/experiments/intern), start with the [BASIS data guide](https://huggingface.co/Duke-CEI-SVD/FlashSVD/blob/main/basis/README.md).
| You need… | Go to… |
|---|---|
| **C4 training data** | [Six frozen C4 streams: two backbones × three seeds](https://huggingface.co/Duke-CEI-SVD/FlashSVD/tree/main/basis/data/c4) |
| Q64/Q128 probe and held-out tensors | [Probe inputs](https://huggingface.co/Duke-CEI-SVD/FlashSVD/tree/main/basis/data/probes) |
| Cached per-window gradients | [Q128](https://huggingface.co/Duke-CEI-SVD/FlashSVD/tree/main/basis/gradients/q128) · [Original Q64](https://huggingface.co/Duke-CEI-SVD/FlashSVD/tree/main/basis/gradients/q64) |
| Exact frozen compressed Sources | [33 hosted native checkpoints, including all 24 exact Table 1 Sources](https://huggingface.co/Duke-CEI-SVD/FlashSVD/tree/main/basis/sources) |
| Tokenizers, original paths, and checksums | [Tokenizers](https://huggingface.co/Duke-CEI-SVD/FlashSVD/tree/main/basis/tokenizers) · [Artifact manifest](https://huggingface.co/Duke-CEI-SVD/FlashSVD/blob/main/basis/manifest.json) |
The shared data and caches total approximately **4.4 GiB**; the 33 hosted native checkpoints total **353.5 GiB**. Download only the Source files needed for your task. The guide includes selective download commands, loader instructions, checkpoint coverage, and the seven newly supplied Task 05 GFWSVD/Swift-SVD ratio Sources. Tasks 01–02 already include their numerical inputs in GitHub.
**Table 1, verified 2026-09-23:** All 24 exact paper Source checkpoints and their loader records are hosted in `basis/sources/`. The 120 checkpoint files passed remote size/hash audit. Keep the original `.pt` or sharded format and load with the [BASIS handoff code](https://github.com/Zishan-Shao/Basis/tree/handoff/paper-20260915). [Download and verification guide](basis/releases/table1_20260922/README.md).
**Added 2026-09-17:** [GFWSVD r0.5/r0.7 (Llama/Qwen) and Swift-SVD r0.4/r0.5/r0.7 (Qwen), with download instructions and checksums](basis/releases/source_ratios_20260917/README.md).
## Original FlashSVD example checkpoints
This repository contains a small set of curated example low-rank checkpoints for FlashSVD-v1.5 benchmarking and integration tests.
The current examples cover SVD-LLM v1, Basis Sharing, and DobiSVD variants.
Properties of these original example checkpoints:
- all uploaded checkpoints remain factorized / low-rank; they are not fused into dense tensors
- removable RoPE runtime buffers are omitted from the SVD-LLM examples
- each example is a self-contained HF export subfolder
Example loader:
```python
from utils.model_utils import get_model_from_source
import torch
source = "Duke-CEI-SVD/FlashSVD/llama_7b/svdllm_v1/whitening_only_0.5_hf"
model, tok = get_model_from_source(source)
model = model.eval().to("cuda", dtype=torch.bfloat16)
```
## BASIS intern code
Use the [intern GitHub branch](https://github.com/Zishan-Shao/Basis/tree/handoff/paper-20260915) for installation, Source loading/checking, compression, and the five assignments. [Tested code revision](https://github.com/Zishan-Shao/Basis/tree/57218ca8c09e12464adaddddc4924394eac668f5): `57218ca`. The [current handoff downloader](https://github.com/Zishan-Shao/Basis/tree/2801812) pins the verified complete checkpoint revision `4ada2a6`; download weights/data here and run the commands from that branch.