Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Extending Support for Additional Bloom Models (up to 7b) #447

Merged
merged 3 commits into from
Jan 17, 2024
Merged

Extending Support for Additional Bloom Models (up to 7b) #447

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
eb146b5
add todos
SeuperHakkerJa Nov 14, 2023
cf2a0ef
add alias, pass notebook test, unit test tba
SeuperHakkerJa Nov 18, 2023
203ff1d
Merge branch 'neelnanda-io:main' into main
SeuperHakkerJa Dec 13, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 10 additions & 4 deletions transformer_lens/loading_from_pretrained.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -137,6 +137,10 @@
"stabilityai/stablelm-tuned-alpha-3b",
"stabilityai/stablelm-tuned-alpha-7b",
"bigscience/bloom-560m",
"bigscience/bloom-1b1",
"bigscience/bloom-1b7",
"bigscience/bloom-3b",
"bigscience/bloom-7b1",
"bigcode/santacoder",
]
"""Official model names for models on HuggingFace."""
Expand Down Expand Up @@ -497,6 +501,10 @@
"stablelm-tuned-7b",
],
"bigscience/bloom-560m": ["bloom-560m"],
"bigscience/bloom-1b1": ["bloom-1b1"],
"bigscience/bloom-1b7": ["bloom-1b7"],
"bigscience/bloom-3b": ["bloom-3b"],
"bigscience/bloom-7b1": ["bloom-7b1"],
"bigcode/santacoder": ["santacoder"],
}
"""Model aliases for models on HuggingFace."""
Expand Down Expand Up @@ -1684,10 +1692,8 @@ def convert_bloom_weights(bloom, cfg: HookedTransformerConfig):
state_dict[f"blocks.{l}.ln1.w"] = bloom.transformer.h[l].input_layernorm.weight
state_dict[f"blocks.{l}.ln1.b"] = bloom.transformer.h[l].input_layernorm.bias

# Bloom attn weight is stored as a fused matrx. BloomAttn: Linear(in=1024, out=3072)
# The .weight returned matrix will be in shape (3072, 1024)
W = bloom.transformer.h[l].self_attention.query_key_value.weight
# First transpose -> (1024, 3072), then split into (d_model, n_heads, 3, d_head)

W_split = W.T.reshape(cfg.d_model, cfg.n_heads, 3, cfg.d_head)

W_Q, W_K, W_V = W_split[..., 0, :], W_split[..., 1, :], W_split[..., 2, :]
Expand Down Expand Up @@ -1732,7 +1738,7 @@ def convert_bloom_weights(bloom, cfg: HookedTransformerConfig):
state_dict[f"blocks.{l}.mlp.b_out"] = bloom.transformer.h[
l
].mlp.dense_4h_to_h.bias
state_dict["unembed.W_U"] = bloom.lm_head.weight.T # transpose to match shape
state_dict["unembed.W_U"] = bloom.lm_head.weight.T

state_dict["ln_final.w"] = bloom.transformer.ln_f.weight
state_dict["ln_final.b"] = bloom.transformer.ln_f.bias
Expand Down