auto_round format
Browse files- quantization_config.json +24 -0
quantization_config.json
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bits": 2,
|
| 3 |
+
"group_size": 64,
|
| 4 |
+
"sym": true,
|
| 5 |
+
"data_type": "int",
|
| 6 |
+
"enable_quanted_input": true,
|
| 7 |
+
"enable_minmax_tuning": true,
|
| 8 |
+
"seqlen": 2048,
|
| 9 |
+
"batch_size": 8,
|
| 10 |
+
"scale_dtype": "torch.float16",
|
| 11 |
+
"lr": 0.0005,
|
| 12 |
+
"minmax_lr": 0.0005,
|
| 13 |
+
"gradient_accumulate_steps": 1,
|
| 14 |
+
"iters": 2000,
|
| 15 |
+
"amp": true,
|
| 16 |
+
"nsamples": 1024,
|
| 17 |
+
"low_gpu_mem_usage": true,
|
| 18 |
+
"to_quant_block_names": "model.layers",
|
| 19 |
+
"enable_norm_bias_tuning": false,
|
| 20 |
+
"dataset": "NeelNanda/pile-10k",
|
| 21 |
+
"autoround_version": "0.4.2",
|
| 22 |
+
"quant_method": "intel/auto-round",
|
| 23 |
+
"backend": "auto_round:gptq:exllamav2"
|
| 24 |
+
}
|