Update README.md
Browse files
README.md
CHANGED
|
@@ -2,4 +2,31 @@
|
|
| 2 |
license: apache-2.0
|
| 3 |
base_model:
|
| 4 |
- Qwen/Qwen3-32B
|
| 5 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
license: apache-2.0
|
| 3 |
base_model:
|
| 4 |
- Qwen/Qwen3-32B
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
This model is created with the following code:
|
| 8 |
+
|
| 9 |
+
```Python
|
| 10 |
+
from datasets import load_dataset
|
| 11 |
+
from gptqmodel import GPTQModel, QuantizeConfig
|
| 12 |
+
from huggingface_hub import constants
|
| 13 |
+
|
| 14 |
+
model_id = "Qwen/Qwen3-32B"
|
| 15 |
+
# Save the quantized model in the HF cache directory
|
| 16 |
+
cache_dir = constants.HF_HUB_CACHE
|
| 17 |
+
quant_path = os.path.join(cache_dir, "models--quantized--" + model_id.replace("/", "--"))
|
| 18 |
+
os.makedirs(quant_path, exist_ok=True)
|
| 19 |
+
|
| 20 |
+
# Load calibration data (1024 samples from C4)
|
| 21 |
+
calibration_dataset = load_dataset(
|
| 22 |
+
"allenai/c4",
|
| 23 |
+
data_files="en/c4-train.00001-of-01024.json.gz",
|
| 24 |
+
split="train"
|
| 25 |
+
).select(range(1024))["text"]
|
| 26 |
+
|
| 27 |
+
# Configure and run quantization
|
| 28 |
+
quant_config = QuantizeConfig(bits=4, group_size=128)
|
| 29 |
+
model = GPTQModel.load(model_id, quant_config)
|
| 30 |
+
model.quantize(calibration_dataset, batch_size=2)
|
| 31 |
+
model.save(quant_path)
|
| 32 |
+
```
|