Coverage for transformer_lens/model_bridge/supported_architectures/gemma1.py: 100%

1"""Gemma1 architecture adapter.""" 

2 

3from typing import Any 

4 

5from transformer_lens.conversion_utils.conversion_steps import TransposeTensorConversion 

6from transformer_lens.conversion_utils.param_processing_conversion import ( 

7 ParamProcessingConversion, 

8) 

9from transformer_lens.model_bridge.architecture_adapter import ArchitectureAdapter 

10from transformer_lens.model_bridge.generalized_components import ( 

11 BlockBridge, 

12 EmbeddingBridge, 

13 LinearBridge, 

14 PositionEmbeddingsAttentionBridge, 

15 RMSNormalizationBridge, 

16 RotaryEmbeddingBridge, 

17 UnembeddingBridge, 

18) 

19 

20 

21class Gemma1ArchitectureAdapter(ArchitectureAdapter): 

22 """Architecture adapter for Gemma1 models.""" 

23 

24 def __init__(self, cfg: Any) -> None: 

25 """Initialize the Gemma1 architecture adapter.""" 

26 super().__init__(cfg) 

27 

28 self._set_rms_rotary_defaults() 

29 

30 # Gemma models use BOS tokens (tokenizer prepends BOS by default) 

31 # Matches HookedTransformer behavior (default_prepend_bos = True) 

32 self.cfg.default_prepend_bos = True 

33 # Gemma models use (1.0 + weight) in RMSNorm instead of just weight 

34 # See: https://github.com/huggingface/transformers/pull/29402 

35 self.cfg.rmsnorm_uses_offset = True 

36 

37 self.weight_processing_conversions = { 

38 # NOTE: Gemma1 scales embeddings by sqrt(d_model) at RUNTIME inside 

39 # GemmaTextScaledWordEmbedding.forward() (HF transformers >= 5.0). 

40 # That layer is what bridge.embed wraps, so embed.hook_out already 

41 # captures the scaled value — matching HookedTransformer's hook_embed 

42 # (which uses pre-scaled W_E). We must NOT pre-scale weights here and 

43 # we must NOT install a runtime hook_conversion that re-scales. 

44 # 

45 # Attention weight conversions 

46 **self._qkvo_weight_conversions(), 

47 # Norm weights carry no conversion entry: cfg.rmsnorm_uses_offset is the 

48 # only offset signal — weight processing and the runtime norm read it 

49 # directly and apply (1 + w) themselves. 

50 # MLP weight conversions - transpose from [out, in] to [in, out] 

51 "blocks.{i}.mlp.gate.weight": ParamProcessingConversion( 

52 tensor_conversion=TransposeTensorConversion(), 

53 ), 

54 "blocks.{i}.mlp.in.weight": ParamProcessingConversion( 

55 tensor_conversion=TransposeTensorConversion(), 

56 ), 

57 "blocks.{i}.mlp.out.weight": ParamProcessingConversion( 

58 tensor_conversion=TransposeTensorConversion(), 

59 ), 

60 # Unembed weight conversion - transpose from [vocab, d_model] to [d_model, vocab] 

61 "unembed.weight": ParamProcessingConversion( 

62 tensor_conversion=TransposeTensorConversion(), 

63 ), 

64 } 

65 

66 self.component_mapping = { 

67 "embed": EmbeddingBridge(name="model.embed_tokens"), 

68 "rotary_emb": RotaryEmbeddingBridge(name="model.rotary_emb", config=self.cfg), 

69 "blocks": BlockBridge( 

70 name="model.layers", 

71 submodules={ 

72 "ln1": RMSNormalizationBridge(name="input_layernorm", config=self.cfg), 

73 "ln2": RMSNormalizationBridge(name="post_attention_layernorm", config=self.cfg), 

74 "attn": PositionEmbeddingsAttentionBridge( 

75 name="self_attn", 

76 config=self.cfg, 

77 submodules={ 

78 "q": LinearBridge(name="q_proj"), 

79 "k": LinearBridge(name="k_proj"), 

80 "v": LinearBridge(name="v_proj"), 

81 "o": LinearBridge(name="o_proj"), 

82 }, 

83 requires_attention_mask=True, 

84 requires_position_embeddings=True, 

85 ), 

86 "mlp": self._gated_mlp(), 

87 }, 

88 ), 

89 "ln_final": RMSNormalizationBridge(name="model.norm", config=self.cfg), 

90 "unembed": UnembeddingBridge(name="lm_head"), 

91 }