54 lines
1.9 KiB
Markdown
54 lines
1.9 KiB
Markdown
---
|
|
base_model: meta-llama/Meta-Llama-3-8B
|
|
language:
|
|
- en
|
|
license: llama3
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
tags:
|
|
- kronq
|
|
- quantization
|
|
- fake-quant
|
|
- fp16
|
|
---
|
|
|
|
# Meta-Llama-3-8B — KronQ W3A16 (fake-quant fp16)
|
|
|
|
**Paper:** [arXiv:2607.07964](https://arxiv.org/abs/2607.07964) · **Code:** [GitHub](https://github.com/Intelligent-Computing-Lab-Panda/KronQ)
|
|
|
|
> ⚠️ **Fake-quant fp16 checkpoint.** The 3-bit weights are stored in **fp16**
|
|
> (KronQ does not pack int3), so this repo is the **same size as bf16** — no
|
|
> compression or speedup, for **PPL / accuracy reproduction only**. For
|
|
> deployable low-bit, see the W4A16 (packed int4) repo.
|
|
|
|
[Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B) quantized to 3-bit weights with **KronQ** (Kronecker-factored Hessian quantization), exported as a standard fp16 model.
|
|
|
|
## Results (WikiText-2, seqlen 2048)
|
|
|
|
**Perplexity:** **7.09**
|
|
|
|
**Zero-shot accuracy:**
|
|
|
|
| PIQA | ARC-E | ARC-C | HellaSwag | WinoGrande | BoolQ | OBQA | Average |
|
|
|---|---|---|---|---|---|---|---|
|
|
| 77.53 | 74.54 | 50.17 | 74.92 | 71.74 | 81.13 | 41.20 | **67.32** |
|
|
|
|
(lm-evaluation-harness, 0-shot. `acc_norm` for PIQA/HellaSwag/ARC/OBQA, `acc` for WinoGrande/BoolQ.)
|
|
|
|
## Usage
|
|
|
|
Loads as a **standard fp16 model** — no KronQ code needed:
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
m = AutoModelForCausalLM.from_pretrained("donghyunli/Meta-Llama-3-8B-KronQ-W3A16-fake", torch_dtype="float16").cuda()
|
|
tok = AutoTokenizer.from_pretrained("donghyunli/Meta-Llama-3-8B-KronQ-W3A16-fake")
|
|
```
|
|
|
|
## Recipe
|
|
|
|
Per-channel asymmetric W3, weight-only (a_bits=16), `--alpha 0.25`, bidirectional incoherence processing (BiIP), `act_order`, raw H_G. Calibrated on 128 WikiText-2 sequences.
|
|
|
|
## License
|
|
|
|
Derivative of Meta-Llama-3-8B — subject to the [Llama 3 Community License](https://llama.meta.com/llama3/license/). |