Coverage for transformer_lens/pretrained/weight_conversions/gptj.py: 13%

1import einops 

2import torch 

3 

4from transformer_lens.HookedTransformerConfig import HookedTransformerConfig 

5 

6 

7def convert_gptj_weights(gptj, cfg: HookedTransformerConfig): 

8 state_dict = {} 

9 

10 state_dict["embed.W_E"] = gptj.transformer.wte.weight 

11 

12 for l in range(cfg.n_layers): 

13 state_dict[f"blocks.{l}.ln1.w"] = gptj.transformer.h[l].ln_1.weight 

14 state_dict[f"blocks.{l}.ln1.b"] = gptj.transformer.h[l].ln_1.bias 

15 

16 W_Q = gptj.transformer.h[l].attn.q_proj.weight 

17 W_K = gptj.transformer.h[l].attn.k_proj.weight 

18 W_V = gptj.transformer.h[l].attn.v_proj.weight 

19 W_Q = einops.rearrange(W_Q, "(i h) m->i m h", i=cfg.n_heads) 

20 W_K = einops.rearrange(W_K, "(i h) m->i m h", i=cfg.n_heads) 

21 W_V = einops.rearrange(W_V, "(i h) m->i m h", i=cfg.n_heads) 

22 state_dict[f"blocks.{l}.attn.W_Q"] = W_Q 

23 state_dict[f"blocks.{l}.attn.W_K"] = W_K 

24 state_dict[f"blocks.{l}.attn.W_V"] = W_V 

25 

26 state_dict[f"blocks.{l}.attn.b_Q"] = torch.zeros(cfg.n_heads, cfg.d_head, dtype=cfg.dtype) 

27 state_dict[f"blocks.{l}.attn.b_K"] = torch.zeros(cfg.n_heads, cfg.d_head, dtype=cfg.dtype) 

28 state_dict[f"blocks.{l}.attn.b_V"] = torch.zeros(cfg.n_heads, cfg.d_head, dtype=cfg.dtype) 

29 

30 W_O = gptj.transformer.h[l].attn.out_proj.weight 

31 W_O = einops.rearrange(W_O, "m (i h)->i h m", i=cfg.n_heads) 

32 state_dict[f"blocks.{l}.attn.W_O"] = W_O 

33 state_dict[f"blocks.{l}.attn.b_O"] = torch.zeros(cfg.d_model, dtype=cfg.dtype) 

34 

35 # Layer Norm 1 and 2 are tied. 

36 state_dict[f"blocks.{l}.ln2.w"] = state_dict[f"blocks.{l}.ln1.w"] 

37 state_dict[f"blocks.{l}.ln2.b"] = state_dict[f"blocks.{l}.ln1.b"] 

38 

39 state_dict[f"blocks.{l}.mlp.W_in"] = gptj.transformer.h[l].mlp.fc_in.weight.T 

40 state_dict[f"blocks.{l}.mlp.b_in"] = gptj.transformer.h[l].mlp.fc_in.bias 

41 

42 state_dict[f"blocks.{l}.mlp.W_out"] = gptj.transformer.h[l].mlp.fc_out.weight.T 

43 state_dict[f"blocks.{l}.mlp.b_out"] = gptj.transformer.h[l].mlp.fc_out.bias 

44 state_dict["ln_final.w"] = gptj.transformer.ln_f.weight 

45 state_dict["ln_final.b"] = gptj.transformer.ln_f.bias 

46 

47 state_dict["unembed.W_U"] = gptj.lm_head.weight.T 

48 # Contains a bias, for some reason? 

49 state_dict["unembed.b_U"] = gptj.lm_head.bias 

50 return state_dict