README.md

---
base_model:
- microsoft/Phi-3.5-vision-instruct
---

## Eval

```
vllm serve nm-testing/Phi-3.5-vision-instruct-W8A8-Dynamic-Per-Token --trust-remote-code --max-model-len 100000
```

```
python -m eval.run eval_vllm --model_name nm-testing/Phi-3.5-vision-instruct-W8A8-Dynamic-Per-Token --url http://0.0.0.0:8000 --output_dir output/ --eval_name "chartqa"
...
================================================================================
Metrics:
{
    "explicit_prompt_relaxed_correctness": 0.6472,
    "anywhere_in_answer_relaxed_correctness": 0.6616
}
================================================================================
```

## Creation

```python
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM

from llmcompressor.modifiers.quantization import GPTQModifier
# from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
from llmcompressor.transformers import oneshot, wrap_hf_model_class

# Select model and load it.
MODEL_ID = "microsoft/Phi-3.5-vision-instruct"
model_class = wrap_hf_model_class(AutoModelForCausalLM)
model = model_class.from_pretrained(
    MODEL_ID,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True, 
    _attn_implementation="eager",
)
processor = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)

# Select calibration dataset.
DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"

# Select number of samples. 512 samples is a good place to start.
# Increasing the number of samples can improve accuracy.
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048

# Load dataset and preprocess.
ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))


def preprocess(example):
    return {
        "text": processor.apply_chat_template(
            example["messages"],
            tokenize=False,
        )
    }


ds = ds.map(preprocess)


# Tokenize inputs.
def tokenize(sample):
    return processor(
        sample["text"],
        padding=False,
        max_length=MAX_SEQUENCE_LENGTH,
        truncation=True,
        add_special_tokens=False,
    )


ds = ds.map(tokenize, remove_columns=ds.column_names)
print(ds)

# Configure algorithms. In this case, we:
#   * apply SmoothQuant to make the activations easier to quantize
#   * quantize the weights to int8 with GPTQ (static per channel)
#   * quantize the activations to int8 (dynamic per token)
# Note: set sequential_update: true in the recipe to reduce memory
ignore=["re:.*lm_head", "re:model.vision_embed_tokens.*"]
recipe = [
    # SmoothQuantModifier(smoothing_strength=0.8, ignore=ignore),
    GPTQModifier(targets="Linear", scheme="W8A8", ignore=ignore),
]

# Apply algorithms.
oneshot(
    model=model,
    dataset=ds,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
)

# Confirm generations of the quantized model look sane.
print("\n\n")
print("========== SAMPLE GENERATION ==============")
input_ids = processor("Hello my name is", return_tensors="pt").input_ids.to("cuda")
output = model.generate(input_ids, max_new_tokens=100)
print(processor.decode(output[0]))
print("==========================================\n\n")

# Save to disk compressed.
SAVE_DIR = MODEL_ID.split("/")[1] + "-W8A8-Dynamic-Per-Token"
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
```