97 lines
2.6 KiB
JSON
97 lines
2.6 KiB
JSON
{
|
|
"architectures": [
|
|
"LlamaForCausalLM"
|
|
],
|
|
"attention_bias": false,
|
|
"attention_dropout": 0.0,
|
|
"bos_token_id": 0,
|
|
"dtype": "float16",
|
|
"eos_token_id": [
|
|
1,
|
|
130073
|
|
],
|
|
"head_dim": 128,
|
|
"hidden_act": "silu",
|
|
"hidden_size": 2048,
|
|
"initializer_range": 0.02,
|
|
"intermediate_size": 6144,
|
|
"max_position_embeddings": 131072,
|
|
"mlp_bias": false,
|
|
"model_type": "llama",
|
|
"num_attention_heads": 16,
|
|
"num_hidden_layers": 42,
|
|
"num_key_value_heads": 2,
|
|
"pad_token_id": 1,
|
|
"pretraining_tp": 1,
|
|
"quantization_config": {
|
|
"bits": 4,
|
|
"checkpoint_format": "gemm",
|
|
"desc_act": false,
|
|
"format": "gemm",
|
|
"group_size": 128,
|
|
"lm_head": false,
|
|
"method": "awq",
|
|
"pack_dtype": "int32",
|
|
"quant_method": "awq",
|
|
"version": "gemm",
|
|
"zero_point": true,
|
|
"meta": {
|
|
"act_group_aware": true,
|
|
"auto_forward_data_parallel": true,
|
|
"damp_auto_increment": 0.01,
|
|
"damp_percent": 0.01,
|
|
"dense_vram_strategy": "exclusive",
|
|
"dense_vram_strategy_devices": null,
|
|
"fallback": {
|
|
"smooth": null,
|
|
"strategy": "rtn",
|
|
"threshold": "0.5%"
|
|
},
|
|
"foem": null,
|
|
"gc_mode": "interval",
|
|
"gptaq": {
|
|
"alpha": 0.25,
|
|
"device": "auto"
|
|
},
|
|
"hessian": {
|
|
"chunk_bytes": null,
|
|
"chunk_size": null,
|
|
"staging_dtype": "float32"
|
|
},
|
|
"mock_quantization": false,
|
|
"moe_vram_strategy": "exclusive",
|
|
"moe_vram_strategy_devices": null,
|
|
"mse": 0.0,
|
|
"offload_to_disk": true,
|
|
"offload_to_disk_path": "/tmp/gptqmodel_vq422ji5",
|
|
"pack_impl": "cpu",
|
|
"quantizer": [
|
|
"gptqmodel:7.3.5"
|
|
],
|
|
"static_groups": false,
|
|
"true_sequential": true,
|
|
"uri": "https://github.com/modelcloud/gptqmodel",
|
|
"wait_for_submodule_finalizers": false,
|
|
"repacked_from": {
|
|
"source_dir": "target_model/rotquant-gptaq-w4-g128-asym-fullmix_scaled_t183",
|
|
"source_quant_method": "gptq",
|
|
"source_format": "gptq",
|
|
"source_sym": false,
|
|
"converter": "scripts/repack_gptq_asym_to_awq.py",
|
|
"note": "Layout-only re-encode of an asymmetric GPTQ checkpoint into AWQ GEMM packing. Weights are bit-identical after dequantization; no AWQ scale search was run. Done because vLLM's AutoGPTQConfig.TYPE_MAP has no (4, False) entry and raises on load.",
|
|
"dequant_verified": true
|
|
}
|
|
}
|
|
},
|
|
"rms_norm_eps": 1e-06,
|
|
"rope_parameters": {
|
|
"rope_theta": 5000000,
|
|
"rope_type": "default"
|
|
},
|
|
"tie_word_embeddings": false,
|
|
"transformers_version": "5.16.1",
|
|
"use_cache": true,
|
|
"vocab_size": 130560,
|
|
"torch_dtype": "float16"
|
|
}
|