Add TRN hybrid non-record submission (1.4942 bpb, 1x RTX 5090)#669
Open
amabito wants to merge 1 commit intoopenai:mainfrom
Open
Add TRN hybrid non-record submission (1.4942 bpb, 1x RTX 5090)#669amabito wants to merge 1 commit intoopenai:mainfrom
amabito wants to merge 1 commit intoopenai:mainfrom
Conversation
Oscillatory recurrence + attention hybrid under 16 MB constraint. 10 layers (7 TRN + 3 Attn), int5 QAT, Kogge-Stone parallel scan. Int5 collapses at 20K steps due to oscillator projection phase drift.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 submission: oscillatory recurrence + attention hybrid under the 16 MB constraint.
What this is
A 10-layer hybrid model (7 TRN layers + 3 attention layers) with int5 QAT and
zstd-22 compression. The TRN layers use a Kogge-Stone parallel prefix scan over
complex-valued oscillators -- no Triton, no custom CUDA, pure PyTorch.
Score: 1.4942 bpb (int5 roundtrip, 636 steps / 600s wallclock, 1x RTX 5090).
Artifact: 15.28 MB.
What went wrong
The model reaches 1.26 bpb in fp32 at 20K steps, but int5 quantization degrades
it to 1.93 bpb. The oscillator projection weights (d_model -> 6K, encoding
frequency and phase) accumulate O(t) phase drift from quantization errors.
At 1000 steps the error is small (+0.041); at 20K steps it collapses (+0.669).
A parameter-matched 13L Transformer shows only +0.016 int5 degradation at the
same step count. The failure is specific to oscillatory recurrence parameters.
What is included
What is not included