Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: llama3.2
|
| 3 |
+
tags:
|
| 4 |
+
- quantization
|
| 5 |
+
- 2-bit
|
| 6 |
+
- llama
|
| 7 |
+
- qat
|
| 8 |
+
pretty_name: Trellis vs Scalar 2-Bit QAT — Single-Layer Snapshots
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Single-Layer 2-Bit QAT Weight Snapshots
|
| 12 |
+
|
| 13 |
+
Per-layer 2-bit quantized weight snapshots for the controlled trellis-vs-scalar
|
| 14 |
+
comparison in **"Small GPU, Small Edge: A Controlled Study of Trellis vs. Scalar
|
| 15 |
+
2-Bit QAT."** These reproduce **Tables 1, 3, and 4** of the paper.
|
| 16 |
+
|
| 17 |
+
Built on **meta-llama/Llama-3.2-1B** and subject to the **Llama 3.2 Community
|
| 18 |
+
License**. Each snapshot quantizes a *single* decoder layer's 7 projections to
|
| 19 |
+
2 bits; all other layers stay FP16. **These are experiment artifacts, not a
|
| 20 |
+
deployable model.**
|
| 21 |
+
|
| 22 |
+
## Contents
|
| 23 |
+
| file | layer | arm | init → train |
|
| 24 |
+
|---|---|---|---|
|
| 25 |
+
| `snapshots/layer04_bcjr_seed{0,1,2}.pt` | 4 (mild) | BCJR-trellis (soft) | PTQ → 10-step BCJR |
|
| 26 |
+
| `snapshots/layer04_ste_trellis_seed0.pt` | 4 | STE-trellis | FP → 10-step STE |
|
| 27 |
+
| `snapshots/layer04_scalar_ctrl_seed0.pt` | 4 | scalar-ctrl (uniform) | FP → 10-step STE |
|
| 28 |
+
| `snapshots/layer01_bcjr_seed0.pt` | 1 (sensitive) | BCJR-trellis | PTQ → 10-step BCJR |
|
| 29 |
+
| `snapshots/layer01_ste_trellis_seed0.pt` | 1 | STE-trellis | FP → 10-step STE |
|
| 30 |
+
| `snapshots/layer01_scalar_ctrl_seed0.pt` | 1 | scalar-ctrl | FP → 10-step STE |
|
| 31 |
+
|
| 32 |
+
The trellis-PTQ and scalar-PTQ arms are **not** included — they carry no training
|
| 33 |
+
and are recomputed on the fly from the FP weights.
|
| 34 |
+
|
| 35 |
+
## Format
|
| 36 |
+
Each `.pt` is a `torch.save` dict `{proj_name: W_q}` (fp32/bf16, original
|
| 37 |
+
un-rotated basis), with keys `attn_{q,k,v,o}_proj` and `mlp_{gate,up,down}_proj`.
|
| 38 |
+
|
| 39 |
+
## Usage (reproduce the tables)
|
| 40 |
+
```python
|
| 41 |
+
import torch
|
| 42 |
+
from transformers import AutoModelForCausalLM
|
| 43 |
+
from src.qat.eval_llama_layer import install_layer_weights # from the code repo
|
| 44 |
+
|
| 45 |
+
m = AutoModelForCausalLM.from_pretrained(
|
| 46 |
+
"meta-llama/Llama-3.2-1B", torch_dtype=torch.bfloat16, device_map="cuda")
|
| 47 |
+
snap = torch.load("snapshots/layer04_bcjr_seed0.pt", weights_only=True)
|
| 48 |
+
install_layer_weights(m, 4, snap, dtype=torch.bfloat16)
|
| 49 |
+
# then eval WikiText-2 PPL / lm-eval 0-shot -> Tables 1, 3, 4
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
- **Code:** https://github.com/Venugopalan2610/quant-2bit
|
| 53 |
+
- **Paper:** *Small GPU, Small Edge* (arXiv link TBD)
|