internlm

/

internlm2-chat-7b

like 82

Comparing this v2 model with the older one, we can find lots of meaningless renaming of variables, such as:

• lm_head -> output
• embed_tokens -> tok_embeddings
• post_attention_layernorm -> ffn_norm
• mlp -> feed_forward
• self_attn -> attention
• o_proj -> wo
• gate_proj -> w1
• down_proj -> w2
• up_proj -> w3

I fully respect the hardwork and kindness of sharing the model. But, I still want to say, these modifications are truly meaningless, and may hurt the community.