Spaces:
Runtime error
Runtime error
File size: 3,848 Bytes
de071e9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 |
"""
Module for permuting weights in a Llama model architecture.
This enables exploring model connectivity and representation properties
by applying consistent neuron permutations throughout the network.
"""
def permute_model(model, mlp_permutation, emb_permutation, n_blocks=32):
"""
Apply permutations to a Llama model's weights to maintain functional equivalence.
Args:
model: The Llama model to permute
mlp_permutation: Permutation indices for MLP hidden dimensions
emb_permutation: Permutation indices for embedding dimensions
n_blocks: Number of transformer blocks in the model (default: 32)
Returns:
None: Modifies the model in-place
"""
permute_embedding_layer(model, emb_permutation)
permute_transformer_blocks(model, mlp_permutation, emb_permutation)
permute_output_layer(model, emb_permutation)
def permute_transformer_blocks(model, mlp_permutation, emb_permutation):
"""
Apply permutations to transformer block weights in a Llama model.
Permutes attention layers, MLP layers, and normalization layers according to
the provided permutation indices to maintain functional equivalence.
Args:
model: The Llama model to permute
mlp_permutation: Permutation indices for MLP hidden dimensions
emb_permutation: Permutation indices for embedding dimensions
Returns:
None: Modifies the model in-place
"""
weights = model.state_dict()
# Permuting the Self attention layers
for key in weights:
if "self_attn" not in key:
continue
if "o_proj" in key:
weights[key] = weights[key][emb_permutation]
else:
weights[key] = weights[key][:, emb_permutation]
# Permuting the mlp projection layers
for key in weights:
if "mlp" not in key:
continue
if len(weights[key].shape) != 2:
continue
dim_0 = weights[key].size(0)
dim_1 = weights[key].size(1)
if dim_0 == len(mlp_permutation):
weights[key] = weights[key][mlp_permutation]
elif dim_1 == len(mlp_permutation):
weights[key] = weights[key][:, mlp_permutation]
if dim_0 == len(emb_permutation):
weights[key] = weights[key][emb_permutation]
elif dim_1 == len(emb_permutation):
weights[key] = weights[key][:, emb_permutation]
# input_layernorm, post_attention_layernorm
for key in weights:
if "model.layers" not in key:
continue
if len(weights[key].shape) != 1 or len(weights[key]) != len(emb_permutation):
continue
weights[key] = weights[key][emb_permutation]
model.load_state_dict(weights)
def permute_embedding_layer(model, emb_permutation):
"""
Apply permutation to embedding layer weights in a Llama model.
Args:
model: The Llama model to permute
emb_permutation: Permutation indices for embedding dimensions
Returns:
None: Modifies the model in-place
"""
weights = model.state_dict()
weights["model.embed_tokens.weight"] = weights["model.embed_tokens.weight"][:, emb_permutation]
model.load_state_dict(weights)
def permute_output_layer(model, emb_permutation):
"""
Apply permutation to output layer weights in a Llama model.
Permutes the language model head and final normalization layer.
Args:
model: The Llama model to permute
emb_permutation: Permutation indices for embedding dimensions
Returns:
None: Modifies the model in-place
"""
weights = model.state_dict()
weights["lm_head.weight"] = weights["lm_head.weight"][:, emb_permutation]
weights["model.norm.weight"] = weights["model.norm.weight"][emb_permutation]
model.load_state_dict(weights)
|