Update config.json
Browse files- config.json +0 -35
config.json
CHANGED
|
@@ -250,41 +250,6 @@
|
|
| 250 |
"q_lora_rank": 1536,
|
| 251 |
"qk_nope_head_dim": 128,
|
| 252 |
"qk_rope_head_dim": 64,
|
| 253 |
-
"quantization_config": {
|
| 254 |
-
"config_groups": {
|
| 255 |
-
"group_0": {
|
| 256 |
-
"input_activations": null,
|
| 257 |
-
"output_activations": null,
|
| 258 |
-
"targets": [
|
| 259 |
-
"Linear"
|
| 260 |
-
],
|
| 261 |
-
"weights": {
|
| 262 |
-
"actorder": null,
|
| 263 |
-
"block_structure": null,
|
| 264 |
-
"dynamic": false,
|
| 265 |
-
"group_size": 32,
|
| 266 |
-
"num_bits": 4,
|
| 267 |
-
"observer": "minmax",
|
| 268 |
-
"observer_kwargs": {},
|
| 269 |
-
"strategy": "group",
|
| 270 |
-
"symmetric": true,
|
| 271 |
-
"type": "int"
|
| 272 |
-
}
|
| 273 |
-
}
|
| 274 |
-
},
|
| 275 |
-
"format": "pack-quantized",
|
| 276 |
-
"ignore": [
|
| 277 |
-
"re:.*self_attn.*",
|
| 278 |
-
"re:.*shared_experts.*",
|
| 279 |
-
"re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
|
| 280 |
-
"re:.*lm_head.*",
|
| 281 |
-
"re:vision_tower.*",
|
| 282 |
-
"re:mm_projector.*"
|
| 283 |
-
],
|
| 284 |
-
"kv_cache_scheme": null,
|
| 285 |
-
"quant_method": "compressed-tensors",
|
| 286 |
-
"quantization_status": "compressed"
|
| 287 |
-
},
|
| 288 |
"remove_invalid_values": false,
|
| 289 |
"repetition_penalty": 1.0,
|
| 290 |
"return_dict": true,
|
|
|
|
| 250 |
"q_lora_rank": 1536,
|
| 251 |
"qk_nope_head_dim": 128,
|
| 252 |
"qk_rope_head_dim": 64,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 253 |
"remove_invalid_values": false,
|
| 254 |
"repetition_penalty": 1.0,
|
| 255 |
"return_dict": true,
|