File size: 3,599 Bytes
d6ba0d1 4ce6750 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 |
---
base_model:
- mistralai/Mistral-Small-Instruct-2409
---
# Mistral-Small-Instruct CTranslate2 Model
This repository contains a CTranslate2 version of the [Mistral-Small-Instruct model](https://huggingface.co/mistralai/Mistral-Small-Instruct-2409). The conversion process involved AWQ quantization followed by CTranslate2 format conversion.
## Quantization Parameters
The following AWQ parameters were used:
```zero_point=true```
```q_group_size=128```
```w_bit=4```
```version=gemv```
## Quantization Process
The quantization was performed using the [AutoAWQ library](https://casper-hansen.github.io/AutoAWQ/examples/). AutoAWQ supports two quantization approaches:
1. **Without calibration data**:
- Quick process (~few minutes)
- Uses standard quantization schema
- Suitable for general use cases
2. **With calibration data**:
- Longer process (3-4 hours on RTX 4090)
- Preserves full precision for task-specific weights
- Slightly better performance for targeted tasks
## Calibration Details
This model was quantized with calibration data. Specifically, the [cosmopedia-100k](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia-100k) dataset was used, which is good for overall QA and instruction-following.
Key parameters:
- `max_calib_seq_len`: 8192 (enables long-form responses)
- `text_token_length`: 2048 (minimum input token length during quantization)
While these parameters don't fundamentally alter the model's architecture, they fine-tune its behavior for specific input-output length patterns and topic domains.
## Requirements
```torch 2.2.2```
```ctranslate2 4.4.0```
- NOTE: The soon-to-be-released ```ctranslate2 4.5.0``` will support ```torch``` greater than version 2.2.2. These instructions will be updated when that occurs.
## Sample Script
```
import os
import sys
import ctranslate2
import gc
import torch
from transformers import AutoTokenizer
system_message = "You are a helpful person who answers questions."
user_message = "Hello, how are you today? I'd like you to write me a funny poem that is a parody of Milton's Paradise Lost if you are familiar with that famous epic poem?"
model_dir = r"D:\Scripts\bench_chat\models\mistralai--Mistral-Small-Instruct-2409-AWQ-ct2-awq" # uses ~13.8 GB
def build_prompt_mistral_small():
prompt = f"""<s>
[INST] {system_message}
{user_message}[/INST]"""
return prompt
def main():
model_name = os.path.basename(model_dir)
print(f"\033[32mLoading the model: {model_name}...\033[0m")
intra_threads = max(os.cpu_count() - 4, 4)
generator = ctranslate2.Generator(
model_dir,
device="cuda",
# compute_type="int8_bfloat16", # NOTE...YOU DO NOT USE THIS AT ALL WHEN USING AWQ/CTRANSLATE2 MODELS
intra_threads=intra_threads
)
tokenizer = AutoTokenizer.from_pretrained(model_dir, add_prefix_space=None)
prompt = build_prompt_mistral_small()
tokens = tokenizer.convert_ids_to_tokens(tokenizer.encode(prompt))
print(f"\nRun 1 (Beam Size: {beam_size}):")
results_batch = generator.generate_batch(
[tokens],
include_prompt_in_result=False,
max_batch_size=4096,
batch_type="tokens",
beam_size=1,
num_hypotheses=1,
max_length=512,
sampling_temperature=0.0,
)
output = tokenizer.decode(results_batch[0].sequences_ids[0])
print("\nGenerated response:")
print(output)
del generator
del tokenizer
torch.cuda.empty_cache()
gc.collect()
if __name__ == "__main__":
main()
``` |