Multilingual-Multimodal-NLP/IndustrialCoder-32B-FP8

1{2    "architectures": [3        "IQuestCoderForCausalLM"4    ],5    "attention_bias": false,6    "attention_dropout": 0.0,7    "auto_map": {8        "AutoConfig": "configuration_iquestcoder.IQuestCoderConfig",9        "AutoModel": "modeling_iquestcoder.IQuestCoderModel",10        "AutoModelForCausalLM": "modeling_iquestcoder.IQuestCoderForCausalLM",11        "AutoModelForQuestionAnswering": "modeling_iquestcoder.IQuestCoderForQuestionAnswering",12        "AutoModelForSequenceClassification": "modeling_iquestcoder.IQuestCoderForSequenceClassification",13        "AutoModelForTokenClassification": "modeling_iquestcoder.IQuestCoderForTokenClassification"14    },15    "bos_token_id": 1,16    "clip_qkv": null,17    "eos_token_id": [18        2,19        75864,20        7586921    ],22    "head_dim": 128,23    "hidden_act": "silu",24    "hidden_size": 5120,25    "initializer_range": 0.02,26    "intermediate_size": 27648,27    "max_position_embeddings": 131072,28    "max_window_layers": 0,29    "mlp_bias": false,30    "model_type": "iquestcoder",31    "num_attention_heads": 40,32    "num_hidden_layers": 64,33    "num_key_value_heads": 8,34    "pretraining_tp": 1,35    "rms_norm_eps": 1e-05,36    "rope_scaling": null,37    "rope_theta": 500000.0,38    "sliding_window": null,39    "tie_word_embeddings": false,40    "torch_dtype": "bfloat16",41    "transformers_version": "4.52.0",42    "use_cache": false,43    "use_sliding_window": false,44    "vocab_size": 76800,45    "compression_config": {46        "config_groups": {47            "group_0": {48                "targets": [49                    "Linear"50                ],51                "input_activations": {52                    "dynamic": true,53                    "group_size": null,54                    "num_bits": 8,55                    "observer": "minmax",56                    "observer_kwargs": {},57                    "strategy": "token",58                    "symmetric": true,59                    "type": "float"60                },61                "weights": {62                    "dynamic": false,63                    "group_size": null,64                    "num_bits": 8,65                    "observer": "minmax",66                    "observer_kwargs": {},67                    "strategy": "channel",68                    "symmetric": true,69                    "type": "float"70                }71            }72        },73        "format": "float-quantized",74        "ignore": [75            "lm_head"76        ],77        "quant_method": "compressed-tensors"78    }79}