赞
踩
大模型太卷了,总感觉天天出新的模型。今天看公众的号,看到阿里Qwen发布一款总模型7B,推理时,激活参数2.7B的大模型。仔细看了一下他的技术报告,记录一下。
技术博客:Qwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters
Qwen1.5-MoE-A2.7B。它仅拥有27亿个激活参数,但其性能却能与当前最先进的70亿参数模型,如Mistral 7B和Qwen1.5-7B相媲美。相较于包含65亿个Non-Embedding参数的Qwen1.5-7B,Qwen1.5-MoE-A2.7B只有20亿个Non-Embedding参数,约为原模型大小的三分之一。此外,相比Qwen1.5-7B,Qwen1.5-MoE-A2.7B的训练成本降低了75%,推理速度则提升至1.74倍。
参考资料:huggingface
{ "architectures": [ "Qwen2MoeForCausalLM" ], "attention_dropout": 0.0, "bos_token_id": 151643, "eos_token_id": 151643, "hidden_act": "silu", "hidden_size": 2048, "initializer_range": 0.02, "intermediate_size": 5632, "max_position_embeddings": 8192, "max_window_layers": 21, "model_type": "qwen2_moe", "num_attention_heads": 16, "num_hidden_layers": 24, "num_key_value_heads": 16, "rms_norm_eps": 1e-06, "rope_theta": 1000000.0, "sliding_window": 32768, "tie_word_embeddings": false, "torch_dtype": "bfloat16", "transformers_version": "4.39.0.dev0", "use_cache": true, "use_sliding_window": false, "vocab_size": 151936, "decoder_sparse_step": 1, "moe_intermediate_size": 1408, "shared_expert_intermediate_size": 5632, "num_experts_per_tok": 4, "num_experts": 60, "norm_topk_prob": false, "output_router_logits": false, "router_aux_loss_coef": 0.001 }
词汇表,多头注意力机制, FFN
主要方面创新: Finegrained experts,初始化,新的routing机制
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。