Non-record: Mamba-3 Hybrid + GPTQ + Late QAT + MuonEq-R — val_bpb 1.1526 (best SSM)

records/track_non_record_16mb/2026-04-04_Mamba3Hybrid_GPTQ_LateQAT_LinearWarmdown/README.md

@@ -0,0 +1,95 @@
+# Mamba-3 Hybrid + Full Hessian GPTQ + Late QAT + Linear Warmdown
+
+## Summary
+
+Best SSM submission: **1.1526 bpb** post-quantization (seed 1337, 8×H100). Closes the 174 mBPB quantization gap from our previous GPTQ run to effectively zero through late QAT with linear warmdown scheduling.
+
+- **Training bpb:** 1.1546 (BF16)
+- **Post-quant bpb:** 1.1526 (INT6 + GPTQ + Late QAT)
+- **Artifact size:** 15.78MB (within 16MB limit)
+- **Architecture:** 7× Mamba-3 SISO + 1× attention hybrid, dim=512, d_state=64, seq_len=4096
+
+---
+
+## Key Improvements Over Previous Submission (PR #1107, 1.5633 bpb)
+
+| Change | Impact |
+|--------|--------|
+| **Full Hessian GPTQ** (replaces QAT-only) | Better weight quantization — no training regression from fake quant |
+| **Late QAT** (`LATE_QAT_THRESHOLD=0.15`) | Fake quantization only in final ~500 steps when LR is low |
+| **Linear warmdown** (`WARMDOWN_SHAPE=linear, WARMDOWN_ITERS=3500`) | Fixes cosine warmdown bug that prevented late QAT from triggering |
+| **LZMA compression** | Better compression than zlib-9 |
+| Combined | **408 mBPB improvement** (1.5633 → 1.1526) |
+
+### Quantization Gap: Root Cause and Fix
+
+Our previous run used cosine warmdown with `WARMDOWN_ITERS=22000`, which meant `lr_mul ≈ 0.13` from step 1. Late QAT (which triggers when `lr_mul < threshold`) could never activate because LR was already too low to allow meaningful fine-tuning with fake quantization noise.
+
+**Fix:** `WARMDOWN_SHAPE=linear` + `WARMDOWN_ITERS=3500` keeps LR at 1.0 for the first ~33% of training, then linearly decays to 0. Late QAT triggers at `lr_mul < 0.15` (~step 4686), giving ~500 steps of QAT at low LR. Result: quantization gap goes from +174 mBPB to **-2 mBPB** (post-quant is slightly *better* than training bpb due to sliding window eval).
+
+---
+
+## Architecture
+
+8-layer U-net hybrid: 7× Mamba-3 SISO blocks + 1 attention layer at position 4.
+
+```
+[Mamba-3 SSD] + [MLP]   ×7 layers
+[Attention]   + [MLP]   ×1 layer (at layer 4)
+dim=512, d_state=64, mlp_mult=3, seq_len=4096
+```
+
+- **Mamba-3 SISO:** pure Triton kernels, chunked SSD, expand=2, headdim=64
+- **Attention:** causal GQA (8/4 heads), RoPE, q_gain, GLU values
+- **MLP:** LeakyReLU², mlp_mult=3
+- **Other:** U-net skip connections, SmearGate, BigramHash, tied embeddings, Muon optimizer
+
+---
+
+## Quantization Pipeline
+
+1. **Train** 5179 steps in BF16 with linear warmdown + late QAT (last ~500 steps)
+2. **GPTQ** with AR self-generated calibration data (32 seqs × 4096 tokens)
+3. **INT6** per-row quantization with full Hessian error minimization
+4. **LZMA** compression of quantized state dict
+5. **Selective ±1 pruning** (safety net, not activated — model already fits)
+
+---
+
+## Results
+
+| Metric | Value |
+|--------|-------|
+| Training steps | 5,179 |
+| Step time | 115.86ms |
+| Training val_bpb | 1.1546 |
+| Post-quant val_bpb | **1.1526** |
+| Quantization gap | **-2 mBPB** |
+| Artifact size | 15,781,278 bytes |
+| Hardware | 8×H100 80GB SXM |
+
+### Comparison
+
+| Submission | bpb | Type |
+|-----------|-----|------|
+| SOTA (PR #1019) | 1.1147 | Transformer |
+| Our best transformer (PR #768) | 1.1201 | Transformer |
+| **This submission** | **1.1526** | **SSM hybrid** |
+| Previous SSM (PR #1107) | 1.5633 | SSM hybrid |
+
+---
+
+## Run Command
+
+```bash
+WARMDOWN_ITERS=3500 WARMDOWN_SHAPE=linear LATE_QAT_THRESHOLD=0.15 \
+FP16_INPROJ_ROWS=0 TARGET_MB=15.9 \
+QUANT_BITS=6 QAT_START_FRAC=0.0 USE_GPTQ=1 \
+TTT_ENABLED=0 EVAL_STRIDE=32 USE_LZMA=1 EVAL_TEMP=0.9 \
+WEIGHT_DECAY=0.04 MUON_MOMENTUM=0.99 MATRIX_LR=0.025 \
+torchrun --nproc_per_node=8 train_mamba3_hybrid.py
+```
+
+## Setup
+
+See `requirements.txt`. The `mamba-ssm` wheel requires manual Mamba-3 source file installation — see [mamba repo](https://github.com/state-spaces/mamba) v2.3.1.

records/track_non_record_16mb/2026-04-04_Mamba3Hybrid_GPTQ_LateQAT_LinearWarmdown/requirements.txt

@@ -0,0 +1 @@
+mamba_ssm @ https://github.com/state-spaces/mamba/releases/download/v2.3.1/mamba_ssm-2.3.1%2Bcu12torch2.9cxx11abiTRUE-cp312-cp312-linux_x86_64.whl

records/track_non_record_16mb/2026-04-04_Mamba3Hybrid_GPTQ_LateQAT_LinearWarmdown/submission.json

@@ -0,0 +1,17 @@
+{
+  "author": "mradassaad",
+  "github_id": "mradassaad",
+  "name": "Mamba-3 Hybrid + Full Hessian GPTQ + Late QAT + Linear Warmdown",
+  "blurb": "Best SSM submission: 7x Mamba-3 SISO SSD + 1 attention hybrid with Full Hessian GPTQ, late QAT (LR-triggered fake quantization), and linear warmdown. Closes 174 mBPB quantization gap to effectively zero.",
+  "date": "2026-04-04",
+  "val_loss": 1.94610314,
+  "val_bpb": 1.15259190,
+  "bytes_total": 15781278,
+  "bytes_code": 82946,
+  "hardware": "8xH100 80GB SXM",
+  "pytorch_version": "2.9.1+cu128",
+  "cuda_version": "12.8",
+  "train_steps": 5179,
+  "step_avg_ms": 115.86,
+  "artifact_bytes": 15781278
+}