Record: GatedAttn + Alpha-Scaled LoRA + Warm-start A + WD 1.0 — val_bpb 1.07081 (3-seed mean)#1784
Open
renqianluo wants to merge 1 commit intoopenai:mainfrom
Open
Conversation
…pb 1.07081 (3-seed mean) Stacks GatedAttn (per-head sigmoid gate on SDPA output, ported from @dexhunter PR openai#1736) on top of our LoRA-TTT stack from PR openai#1767. Two novel support changes needed for the combination: 1. Mirror the gate inside _block_with_lora and _parallel_block_with_lora (the LoRA-TTT forward path reimplements attention inline; without the mirror, TTT scoring sees a different model than training and collapses to ~1.40 BPB) 2. Per-row int8 quantization for attn_gate_w to keep the artifact under 16MB while preserving gate precision (per-tensor int8 lost 0.00112 BPB; per-row keeps it within noise) 3-seed mean 1.07081 BPB (seeds 1337, 42, 314). All seeds improve vs PR openai#1767 (1.07209).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Stacks GatedAttn (per-head sigmoid gate on SDPA output, from @dexhunter PR #1736) on top of PR #1767's LoRA-TTT stack. Two novel support changes were needed to make the combination work.
Novel changes in this PR
Mirror the gate inside
_block_with_lora/_parallel_block_with_lora— the LoRA-TTT forward path reimplements attention inline (to inject LoRA into q/k/v/out). A gate added only toCausalSelfAttention.forwardis silently dropped at TTT scoring time, so training and scoring see different models. Without this mirror, TTT collapses to 1.40 BPB. With it, TTT converges and GatedAttn adds −0.00152 BPB on the 3-seed mean.Per-row int8 quantization for
attn_gate_w— fp16 passthrough pushed the artifact from 15.93 MB to 16.01 MB, over the 16 MB cap on all 3 seeds. Per-tensor int8 saves bytes but loses 0.00112 BPB. Per-row int8 (one fp16 scale per head, 88 extra bytes per artifact) keeps precision within 0.00024 BPB of fp16 while bringing the artifact safely under 16 MB.Results
Every seed improves monotonically.
Compliance
Train ≤596s, eval 466–474s, artifact 15.98MB. Issue #1017 conditions 1–4 verified.
Attribution
@dexhunter (PR #1736), @samacqua (#1530), @bigbag (#1493, #1771), @EthanYangTW (#1523), @romeerp (#1610), @abaybektursun (#549), this author (PR #1767).