Download config.json from intelava/QuadNet-2B: direct link, hf CLI and curl.
- Browser
- Download file 828 Bytes
-
https://huggingface.co/intelava/QuadNet-2B/resolve/main/config.json
- Command line
-
hf download hf://intelava/QuadNet-2B/config.json
-
curl -L -o config.json https://huggingface.co/intelava/QuadNet-2B/resolve/main/config.json
828 Bytes
| { | |
| "architectures": [ | |
| "QuadNetForCausalLM" | |
| ], | |
| "auto_map": { | |
| "AutoConfig": "configuration_bitnet.BitNetConfig", | |
| "AutoModelForCausalLM": "modeling_bitnet.BitNetForCausalLM" | |
| }, | |
| "bos_token_id": 128000, | |
| "eos_token_id": 128001, | |
| "hidden_act": "relu2", | |
| "hidden_size": 2560, | |
| "initializer_range": 0.02, | |
| "intermediate_size": 6912, | |
| "max_position_embeddings": 4096, | |
| "model_type": "quadnet", | |
| "rms_norm_eps": 1e-05, | |
| "num_attention_heads": 20, | |
| "num_hidden_layers": 30, | |
| "num_key_value_heads": 5, | |
| "rope_theta": 500000.0, | |
| "tie_word_embeddings": true, | |
| "torch_dtype": "bfloat16", | |
| "use_cache": true, | |
| "vocab_size": 128256, | |
| "quantization_config": { | |
| "quant_method": "bitnet", | |
| "linear_class": "autobitlinear", | |
| "quantization_mode": "offline" | |
| }, | |
| "quantization": "2-bit-packed-uint16" | |
| } |