Skip to content

Split Q/K/V inputs with per-head attention results #1257

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

Merged
jlarson4 merged 7 commits into from
Apr 16, 2026
Merged

Split Q/K/V inputs with per-head attention results #1257

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
02bae3d
Fixing issue with exposed hooks
jlarson4 Apr 15, 2026
ca70d7b
verification
jlarson4 Apr 15, 2026
00f6592
Fix typing and formating
jlarson4 Apr 15, 2026
ce00661
Improved QKV split attention/block system
jlarson4 Apr 15, 2026
05360f6
Merge latest dev-3.x-canary
jlarson4 Apr 16, 2026
88f4de8
Resolving failures
jlarson4 Apr 16, 2026
a90fd2e
fixed hook error
jlarson4 Apr 16, 2026
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
27 changes: 14 additions & 13 deletions demos/ARENA_Content.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,7 +76,10 @@
" \"gpt2\",\n",
" device=device,\n",
")\n",
"reference_gpt2.enable_compatibility_mode(disable_warnings=True)"
"reference_gpt2.enable_compatibility_mode(disable_warnings=True)",
"\n",
"reference_gpt2.set_use_split_qkv_input(True)\n",
"reference_gpt2.set_use_attn_result(True)"
]
},
{
Expand All @@ -98,7 +101,7 @@
"source": [
"\n",
"# [1.1] Transformer From Scratch\n",
"# 1️⃣ UNDERSTANDING INPUTS & OUTPUTS OF A TRANSFORMER\n",
"# 1\ufe0f\u20e3 UNDERSTANDING INPUTS & OUTPUTS OF A TRANSFORMER\n",
"\n",
"sorted_vocab = sorted(list(reference_gpt2.tokenizer.vocab.items()), key=lambda n: n[1])\n",
"first_vocab = sorted_vocab[0]\n",
Expand Down Expand Up @@ -287,7 +290,7 @@
}
],
"source": [
"# 2️⃣ CLEAN TRANSFORMER IMPLEMENTATION\n",
"# 2\ufe0f\u20e3 CLEAN TRANSFORMER IMPLEMENTATION\n",
"\n",
"layer_0_hooks = [\n",
" (name, tuple(tensor.shape)) for name, tensor in cache.items() if \".0.\" in name\n",
Expand All @@ -311,32 +314,30 @@
" ('blocks.0.attn.hook_hidden_states', (1, 35, 768)),\n",
" ('blocks.0.attn.hook_in', (1, 35, 768)),\n",
" ('blocks.0.attn.hook_k', (1, 35, 12, 64)),\n",
" ('blocks.0.attn.hook_k_input', (1, 35, 12, 768)),\n",
" ('blocks.0.attn.hook_out', (1, 35, 768)),\n",
" ('blocks.0.attn.hook_pattern', (1, 12, 35, 35)),\n",
" ('blocks.0.attn.hook_q', (1, 35, 12, 64)),\n",
" ('blocks.0.attn.hook_result', (1, 35, 768)),\n",
" ('blocks.0.attn.hook_q_input', (1, 35, 12, 768)),\n",
" ('blocks.0.attn.hook_result', (1, 35, 12, 768)),\n",
" ('blocks.0.attn.hook_v', (1, 35, 12, 64)),\n",
" ('blocks.0.attn.hook_v_input', (1, 35, 12, 768)),\n",
" ('blocks.0.attn.hook_z', (1, 35, 12, 64)),\n",
" ('blocks.0.attn.k.hook_in', (1, 35, 768)),\n",
" ('blocks.0.attn.k.hook_out', (1, 35, 12, 64)),\n",
" ('blocks.0.attn.o.hook_in', (1, 35, 12, 64)),\n",
" ('blocks.0.attn.o.hook_out', (1, 35, 768)),\n",
" ('blocks.0.attn.q.hook_in', (1, 35, 768)),\n",
" ('blocks.0.attn.q.hook_out', (1, 35, 12, 64)),\n",
" ('blocks.0.attn.v.hook_in', (1, 35, 768)),\n",
" ('blocks.0.attn.v.hook_out', (1, 35, 12, 64)),\n",
" ('blocks.0.hook_attn_in', (1, 35, 768)),\n",
" ('blocks.0.hook_attn_out', (1, 35, 768)),\n",
" ('blocks.0.hook_in', (1, 35, 768)),\n",
" ('blocks.0.hook_k_input', (1, 35, 768)),\n",
" ('blocks.0.hook_k_input', (1, 35, 12, 768)),\n",
" ('blocks.0.hook_mlp_in', (1, 35, 768)),\n",
" ('blocks.0.hook_mlp_out', (1, 35, 768)),\n",
" ('blocks.0.hook_out', (1, 35, 768)),\n",
" ('blocks.0.hook_q_input', (1, 35, 768)),\n",
" ('blocks.0.hook_q_input', (1, 35, 12, 768)),\n",
" ('blocks.0.hook_resid_mid', (1, 35, 768)),\n",
" ('blocks.0.hook_resid_post', (1, 35, 768)),\n",
" ('blocks.0.hook_resid_pre', (1, 35, 768)),\n",
" ('blocks.0.hook_v_input', (1, 35, 768)),\n",
" ('blocks.0.hook_v_input', (1, 35, 12, 768)),\n",
" ('blocks.0.ln1.hook_in', (1, 35, 768)),\n",
" ('blocks.0.ln1.hook_normalized', (1, 35, 768)),\n",
" ('blocks.0.ln1.hook_out', (1, 35, 768)),\n",
Expand Down Expand Up @@ -373,7 +374,7 @@
"source": [
"# NBVAL_IGNORE_OUTPUT\n",
"# [1.2] Intro to mech interp\n",
"# 2️⃣ FINDING INDUCTION HEADS\n",
"# 2\ufe0f\u20e3 FINDING INDUCTION HEADS\n",
"\n",
"if not IN_GITHUB:\n",
" # Cannot run in CI\n",
Expand Down
Loading
Loading