Skip to content

Non-record: XSA-all + mHC + Full QAT (val_bpb=1.1211) #928

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

Open
autocode-rayes wants to merge 1 commit into
base: main
Choose a base branch
from
+2,238 −0
Open

Non-record: XSA-all + mHC + Full QAT (val_bpb=1.1211) #928

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
40 changes: 40 additions & 0 deletions ...ck_non_record_16mb/2026-03-27_XSA-all_mHC_FullQAT_ParallelMuon_1.1211/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,40 @@
# Non-record: XSA-all + mHC + Full QAT + Parallel Muon

**val_bpb: 1.1211** (seed 1337) | 15.95 MB | 8xH100 SXM, 600s train + 482s eval

## Method

Three changes on top of PR #549's LeakyReLU_LegalTTT_ParallelMuon stack:

| Change | Impact | Extra params |
|--------|--------|-------------|
| **XSA on all 11 layers** | Cross-sequence attention from layer 0 instead of last 4. Forces cross-position mixing earlier. | 0 |
| **Manifold-constrained Hyper-Connections (mHC)** | Learnable alpha/beta residual mixing per block with norm constraint (alpha^2 + beta^2 = 2). Each layer learns how much of its own output vs input to keep. | 22 (2 per layer) |
| **Full-training QAT** | Int6 fake quantization from step 1 (LATE_QAT_THRESHOLD=1.0) instead of late QAT. Model learns quantization robustness throughout training. | 0 |

## Results (seed 1337)

| Metric | Value |
|--------|-------|
| Training steps | 6558 |
| Step avg | 91.5 ms |
| Post-EMA val_bpb | 1.1389 |
| Int6 roundtrip | 1.1463 |
| Sliding window (stride=64) | 1.1229 |
| **Legal TTT** | **1.1211** |
| Artifact size | 15.95 MB |

## Reproduction

```bash
RUN_ID=mhc_xsa_s1337 SEED=1337 \
TTT_ENABLED=1 QAT_ENABLED=1 LATE_QAT_THRESHOLD=1.0 \
torchrun --standalone --nproc_per_node=8 train_gpt.py
```

XSA_LAST_N defaults to 11 in this script. No other env vars needed.

## Environment

- torch 2.11+cu126, flash-attn-3, cuda-toolkit 13.0.2
- RunPod 8xH100 SXM 80GB, driver 580+
6 changes: 6 additions & 0 deletions ...track_non_record_16mb/2026-03-27_XSA-all_mHC_FullQAT_ParallelMuon_1.1211/requirements.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,6 @@
torch>=2.11
flash-attn-3>=3.0.0
cuda-toolkit[cudart]==13.0.2
sentencepiece
numpy
zstandard
12 changes: 12 additions & 0 deletions .../track_non_record_16mb/2026-03-27_XSA-all_mHC_FullQAT_ParallelMuon_1.1211/submission.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,12 @@
{
"track": "10min_16mb",
"val_bpb": 1.1211,
"val_bpb_sliding_window": 1.1229,
"val_bpb_ttt": 1.1211,
"artifact_bytes": 15953842,
"training_seconds": 600,
"eval_seconds": 482,
"hardware": "8xH100 SXM 80GB",
"seeds": [1337],
"method": "XSA-all + mHC + Full-training QAT + LeakyReLU_LegalTTT_ParallelMuon"
}
Loading