SDNQ uint4 JoyCaption — single safetensors file (layer-by-layer CPU quant)
Browse files- config.json +82 -0
config.json
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"LlavaForConditionalGeneration"
|
| 4 |
+
],
|
| 5 |
+
"image_seq_length": 729,
|
| 6 |
+
"image_token_index": 128077,
|
| 7 |
+
"model_type": "llava",
|
| 8 |
+
"multimodal_projector_bias": true,
|
| 9 |
+
"projector_hidden_act": "gelu",
|
| 10 |
+
"text_config": {
|
| 11 |
+
"_name_or_path": "meta-llama/Llama-3.1-8B-Instruct",
|
| 12 |
+
"architectures": [
|
| 13 |
+
"LlamaForCausalLM"
|
| 14 |
+
],
|
| 15 |
+
"attention_bias": false,
|
| 16 |
+
"attention_dropout": 0.0,
|
| 17 |
+
"bos_token_id": 128000,
|
| 18 |
+
"eos_token_id": [
|
| 19 |
+
128001,
|
| 20 |
+
128008,
|
| 21 |
+
128009
|
| 22 |
+
],
|
| 23 |
+
"head_dim": 128,
|
| 24 |
+
"hidden_act": "silu",
|
| 25 |
+
"hidden_size": 4096,
|
| 26 |
+
"initializer_range": 0.02,
|
| 27 |
+
"intermediate_size": 14336,
|
| 28 |
+
"max_position_embeddings": 131072,
|
| 29 |
+
"mlp_bias": false,
|
| 30 |
+
"model_type": "llama",
|
| 31 |
+
"num_attention_heads": 32,
|
| 32 |
+
"num_hidden_layers": 32,
|
| 33 |
+
"num_key_value_heads": 8,
|
| 34 |
+
"pretraining_tp": 1,
|
| 35 |
+
"rms_norm_eps": 1e-05,
|
| 36 |
+
"rope_scaling": {
|
| 37 |
+
"factor": 8.0,
|
| 38 |
+
"high_freq_factor": 4.0,
|
| 39 |
+
"low_freq_factor": 1.0,
|
| 40 |
+
"original_max_position_embeddings": 8192,
|
| 41 |
+
"rope_type": "llama3"
|
| 42 |
+
},
|
| 43 |
+
"rope_theta": 500000.0,
|
| 44 |
+
"torch_dtype": "bfloat16",
|
| 45 |
+
"use_cache": true,
|
| 46 |
+
"vocab_size": 128256
|
| 47 |
+
},
|
| 48 |
+
"torch_dtype": "bfloat16",
|
| 49 |
+
"transformers_version": "4.51.0.dev0",
|
| 50 |
+
"vision_config": {
|
| 51 |
+
"_name_or_path": "google/siglip2-so400m-patch14-384",
|
| 52 |
+
"architectures": [
|
| 53 |
+
"SiglipVisionModel"
|
| 54 |
+
],
|
| 55 |
+
"attention_dropout": 0.0,
|
| 56 |
+
"hidden_act": "gelu_pytorch_tanh",
|
| 57 |
+
"hidden_size": 1152,
|
| 58 |
+
"image_size": 384,
|
| 59 |
+
"intermediate_size": 4304,
|
| 60 |
+
"layer_norm_eps": 1e-06,
|
| 61 |
+
"model_type": "siglip_vision_model",
|
| 62 |
+
"num_attention_heads": 16,
|
| 63 |
+
"num_channels": 3,
|
| 64 |
+
"num_hidden_layers": 27,
|
| 65 |
+
"patch_size": 14,
|
| 66 |
+
"torch_dtype": "bfloat16"
|
| 67 |
+
},
|
| 68 |
+
"vision_feature_layer": -2,
|
| 69 |
+
"vision_feature_select_strategy": "full",
|
| 70 |
+
"quantization_config": {
|
| 71 |
+
"quant_method": "sdnq",
|
| 72 |
+
"weights_dtype": "uint4",
|
| 73 |
+
"group_size": 128,
|
| 74 |
+
"quantization_device": "cpu",
|
| 75 |
+
"return_device": "cpu",
|
| 76 |
+
"use_svd": false,
|
| 77 |
+
"quant_conv": false,
|
| 78 |
+
"quant_embedding": false,
|
| 79 |
+
"use_quantized_matmul": false,
|
| 80 |
+
"dequantize_fp32": true
|
| 81 |
+
}
|
| 82 |
+
}
|