Non-record: GDN Hybrid (E2E TTT / State-Space Model) — val_bpb 1.14502#1479
Open
andrewbaggio1 wants to merge 1 commit intoopenai:mainfrom
Open
Non-record: GDN Hybrid (E2E TTT / State-Space Model) — val_bpb 1.14502#1479andrewbaggio1 wants to merge 1 commit intoopenai:mainfrom
andrewbaggio1 wants to merge 1 commit intoopenai:mainfrom
Conversation
8 Gated DeltaNet layers + 2 softmax attention layers. GDN is mathematically equivalent to E2E TTT-Linear with MSE loss. First competitive GDN hybrid in the 10-min budget. Targets bounty items: E2E TTT + State-space models. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Non-record: GDN Hybrid — Gated DeltaNet as E2E TTT / State-Space Model
val_bpb: 1.14502 (seed 1234, 8xH100, 600s)
Summary
Replaces 8 of 10 attention layers with Gated DeltaNet (Yang et al., ICLR 2025). GDN is mathematically equivalent to E2E TTT-Linear with MSE loss — the delta rule update
S_t = α·S·(I - β·k·kᵀ) + β·v·kᵀis exactly one step of SGD onL = 0.5·‖S·k - v‖², trained end-to-end.Targets bounty items: E2E TTT + State-space models.
Architecture
GatedDeltaNetwith chunk-parallel Triton kernelsResults
Not competitive with softmax attention at 10-min budget: 4.91M tok/s (GDN) vs 6.93M tok/s (attention), yielding 3673 vs 4624 steps. The 20% training deficit is not compensated by GDN's per-step learning advantage at this scale. However, training is stable, GPTQ works cleanly, and PR #1370 showed 1.003 BPB is achievable with unlimited compute.
Credits
Builds on @clarkkev's #1394, FLA library by @sustcsonglin, and PureGDN work by @Christopher-Lee-McClendon (#1370).