---
license: apache-2.0
base_model: microsoft/phi-4
library_name: transformers
tags:
- dashq
- quantized
- post-training-quantization
---
# phi-4-DASHQ-INT3-g32

This repository contains a DASH-Q packed quantized checkpoint for `microsoft/phi-4`.

DASH-Q checkpoints require the lightweight DASH-Q runtime package for loading. They are not plain Transformers checkpoints because linear layers are stored as `PackedQuantizedLinear` modules.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "jkim96/phi-4-DASHQ-INT3-g32", trust_remote_code=True, device_map="cuda", dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("jkim96/phi-4-DASHQ-INT3-g32")

messages = [{"role": "user", "content": "Explain 2-bit quantization in one sentence."}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=256)[0]))
```

`trust_remote_code=True` is required: the checkpoint ships its quantized-layer
implementation (`modeling_dashq.py`) and Triton kernels (`dashq_kernel.py`).
Without Triton, or on CPU, it falls back to dequantize-and-matmul in PyTorch.

### Requirements

| Package | Minimum | Verified with |
| --- | --- | --- |
| `torch` | 2.4 | 2.12.1+cu130 |
| `transformers` | 5.8 | 5.9.0 |
| `triton` | 3.0 (Linux; bundled with CUDA builds of PyTorch) | 3.7.1 |
| `huggingface_hub` | 1.5 (pulled in by transformers) | 1.15.0 |

## Quantization

| Field | Value |
| --- | --- |
| Base model | `microsoft/phi-4` |
| Bits | `3` |
| Group size | `32` |
| Scale/zero dtype | `float16` |
| Calibration dataset | `wikitext2` |
| Calibration samples | `128` |
| Sequence length | `2048` |
| Original size | `29.3190 GB` |
| Quantized size | `9.2126 GB` |

## Evaluation

| Metric | Value |
| --- | ---: |
| `wikitext2_ppl` | 6.6497 |
| `zero-shot accuracy avg` | 69.2339 |
| `arc_challenge` | 57.2526 |
| `arc_easy` | 74.9158 |
| `commonsense_qa` | 75.5119 |
| `hellaswag` | 80.5218 |
| `lambada_openai` | 73.9375 |
| `openbookqa` | 45.8000 |
| `piqa` | 81.1208 |
| `truthfulqa_mc2` | 59.5376 |
| `winogrande` | 74.5067 |