bert4torch_config / Qwen2-1.5B /bert4torch_config.json
Tongjilibo's picture
修改flash_attention为_attn_implementation,增加deepseek
35d10b1
raw
history blame contribute delete
No virus
878 Bytes
{
"model": "qwen2",
"hidden_act": "silu",
"attention_probs_dropout_prob": 0.0,
"bos_token_id": 151643,
"eos_token_id": 151645,
"intermediate_size": 8960,
"initializer_range": 0.02,
"layer_norm_eps": 1e-06,
"hidden_size": 1536,
"num_attention_heads": 12,
"num_hidden_layers": 28,
"num_key_value_heads": 2,
"rope_theta": 1000000.0,
"tie_word_embeddings": true,
"torch_dtype": "bfloat16",
"_attn_implementation": "sdpa",
"is_causal": true,
"vocab_size": 151936,
"segment_vocab_size": 0,
"skip_init": true,
"rope_rank": "updown",
"max_position_embeddings": 32768,
"sliding_window": 32768,
"max_window_layers": 28,
"convert_lm_logits_dtype": "float32",
"generation_config": {"tokenizer_config": {"skip_special_tokens": true}, "eos_token_id": [151643],
"max_length": 32768}
}