README.md

---

license: apache-2.0
language:
- en
datasets:
- Skylion007/openwebtext
tags:
- masked-diffusion
- diffusion-language-model
- mamba
- mamba-2
- state-space-model
- mdlm
- text-generation
- pytorch-lightning
---


# DiffMamba — Checkpoints

Training checkpoints for **DiffMamba**, a small-scale independent study of
**bidirectional Mamba-2 (state-space) denoisers for Masked Diffusion Language

Models (MDLM)**. The Transformer/DiT denoiser in MDLM is replaced with a
**bidirectional Mamba-2 backbone**, and a matched set of models is trained from
scratch on OpenWebText for a controlled quality / scaling / efficiency comparison.

> **Code, full technical report, and documentation:**
> 👉 **https://github.com/shivnarainms22/DiffMamba**
>

> This repo holds **weights only**. The GitHub repository is the source of
> truth for architecture, training recipe, evaluation, and the honest write-up
> of results and limitations.

This work builds directly on **MDLM** (Sahoo et al., NeurIPS 2024) and is a
small-scale reproduction of the research direction introduced by
**DiffuApriel / DiffuMamba** (arXiv 2511.15927). It is **not** claimed as a
novel architecture.

---

## What's in this repo

Each training run lives under `runs/<name>/checkpoints/` and holds two files:
`best.ckpt` (lowest validation loss — the weights the PPL table below reports)
and `last.ckpt` (final-step weights, for resuming). Periodic `step=<N>.ckpt`
snapshots are written during training but pruned from this repo to save space,
so only `best`/`last` are hosted. These are **PyTorch Lightning checkpoints**
from the MDLM codebase — they bundle model weights *and* EMA shadow parameters
(EMA decay 0.9999), optimizer state, and config. They are **not**
`transformers`-loadable via `from_pretrained`; load them with the training repo
(see *How to use* below).

| HF path `runs/…` | Backbone | Params | LR | Steps | Tokens | Val PPL ↓ |
|--------|----------|:------:|:--:|:-----:|:------:|:---------:|
| `runB`        | Transformer (DiT) | \~130M | 3e-4 | 76k | \~5B | **70.5** |
| `runD1`       | BiMamba-2 (SSM)   | \~130M | 3e-4 | 76k | \~5B | 85.9 |
| `runD2`       | BiMamba-2 (SSM)   | \~130M | 3e-4 | 76k | \~5B | 83.5 |
| `runD_lr1e3`  | BiMamba-2 (SSM)   | \~130M | **1e-3** | 76k | \~5B | **79.3** |
| `s100`        | BiMamba-2 (SSM)   | \~100M | 3e-4 | 60k* | \~4B | 97.5 |
| `s50`         | BiMamba-2 (SSM)   | \~50M  | 3e-4 | 30k | \~2B | 136.3 |

Val PPL = MDLM ELBO-bound validation perplexity on the OpenWebText validation
split, measured on each run's **`best.ckpt`** (EMA weights; lower is better).
`*` the 100M run's valid final checkpoint is `last.ckpt` at step 60000
(see the GitHub report for why 61k looped). Also hosted (report §6.5 LR sweep,
not shown above): `s50_lr5e4`, `s50_lr1e3`, `s50_lr2e3` — the 50M learning-rate
sweep.

### Hybrid Mamba+attention checkpoints (attention-layout ablation)

The hybrid backbone inserts full bidirectional attention among the Mamba blocks.
These runs ablate **how many** attention layers and **where**, then over-train
the winner. All share the 130M / OpenWebText / lr-3e-4 recipe and differ only in
the attention layout. Val PPL is on each run's final-step `last.ckpt`, matching
report §11.5 (grid at 76k; the winner over-trained to 150k).

| HF path `runs/…` | Attention layers | # | Steps | Val PPL ↓ | `+experiment=` |
|--------|--------|:--:|:--:|:--:|--------|
| `hybrid_130m` | `[3,7,11]`   | 3 | 76k  | 69.5 | `hybrid_130m` |
| **`hyb_e3`**  | `[2,5,8,11]` | 4 | **150k** | **60.9** ← best | `hyb_e3` |

| `hyb_e3_s2`   | `[2,5,8,11]` | 4 | 150k | 61.5 (seed 2) | `hyb_e3` |

| `hyb_e6`      | `[5,11]`     | 2 | 76k  | 71.4 | `hyb_e6` |

| `hyb_e12`     | `[11]`       | 1 | 76k  | 75.7 | `hyb_e12` |

| `hyb_early`   | `[0,1,2]`    | 3 | 76k  | 80.8 | `hyb_early` |

| `hyb_mid`     | `[4,5,6]`    | 3 | 76k  | 73.9 | `hyb_mid` |

| `hyb_late`    | `[9,10,11]`  | 3 | 76k  | 73.0 | `hyb_late` |



At matched 76k, more attention lowers PPL (4 → 68.1, 3 → 69.5, 2 → 71.4, 1 → 75.7)

and **placement matters more than count** — distributed `[3,7,11]` (69.5) beats

every clustered layout, and clustering early is catastrophic (80.8). The winner

`hyb_e3` (`[2,5,8,11]`) over-trained to 150k reaches **61.2 mean ±0.3** (seeds
60.9 / 61.5) — the best quality in this study, at ~2× the 76k compute.

### Results at a glance

- **Quality.** With the MDLM (Transformer-tuned) recipe at matched 130M / \~5B
  tokens, the Transformer denoiser (70.5) is modestly but consistently stronger
  than pure BiMamba-2. BiMamba prefers a **\~3.3× higher learning rate**; a
  50M LR sweep found `1e-3` best, and retraining 130M at `1e-3` (the
  `runD_lr1e3` checkpoints) closes **\~43%** of the gap (85.9 → 79.3)
  but does not close it.
- **Scaling** (BiMamba, lr 3e-4): 50M → 136.3, 100M → 97.5, 130M → 84.7 —
  clean, monotonic, seed-stable (Δ≈2.4 between seeds).
- **Efficiency.** Forward-pass latency is **textbook-linear** in sequence
  length for BiMamba vs. empirically O(L^1.55) for DiT (with FlashAttention);
  crossover at \~3K tokens, **3.12× faster at 32K**.
- **Honest finding:** *pure* BiMamba-2 trades quality for long-context
  throughput — consistent with DiffuApriel, where a *hybrid* Mamba+attention
  model is what recovers quality.
- **Hybrid backbone + attention ablation (best quality here).** Inserting sparse
  bidirectional attention recovers DiT-class quality: at matched 130M / 76k /
  lr 3e-4, **3 of 12 layers as attention** (`hybrid_130m`, `[3,7,11]`) reaches
  **69.5** — matching the DiT (70.5). An attention-layout ablation (grid runs
  `hyb_e3`/`hyb_e6`/`hyb_e12`/`hyb_early`, all hosted here at 76k) shows
  **placement matters more than count**: distribute attention through depth
  (clustering it early is catastrophic, 80.8), and **4 evenly-spread layers**
  (`hyb_e3`, `[2,5,8,11]`) is best (68.1 at 76k). Over-training that winner to
  150k steps reaches **61.2** val PPL (2-seed mean ±0.3) — the strongest result
  here, though at ~2× the compute of the matched table above (not a
  matched-compute claim vs the 70.5 DiT). All hybrid checkpoints are hosted (see
  the table above); full grid + over-train detail:
  [report §11.5](https://github.com/shivnarainms22/DiffMamba/blob/master/DiffMamba_VLM_Report.md).

Full numbers, caveats, and the LR-fairness analysis are in the
[technical report on GitHub](https://github.com/shivnarainms22/DiffMamba/blob/master/DiffMamba_Report.md).

---

## Model details

- **Framework:** MDLM — absorbing-state discrete diffusion, SUBS
  parameterization, loglinear noise schedule, continuous time (T=0).
- **Tokenizer:** GPT-2 BPE (vocab 50257 + 1 mask token).
- **Sequence length:** 1024.
- **BiMamba-2 backbone** (`models/dimamba.py`): forward + flipped-reverse
  Mamba-2 with weight-tied projections and **AdaLN** noise-level conditioning,
  Mamba-2 defaults `d_state=64`, `headdim=64`, `cond_dim=128`, dropout 0.1.
  - 130M = hidden 768 / 12 blocks · 100M = hidden 640 / 10 blocks ·
    50M = hidden 512 / 8 blocks.

- **Transformer baseline** (`models/dit.py`): DiT, hidden 768 / 12 blocks /

  12 heads.

- **Training:** AdamW (wd 0.01, β=(0.9, 0.999), eps 1e-8), constant LR with

  warmup, gradient clip 1.0, `bf16-mixed`, global batch 64 (micro-batch 16 ×

  grad-accum 4), EMA 0.9999, single A100 per run on an academic SLURM cluster

  with 8-hour-wall checkpoint/resume job-chaining.

- **Data:** OpenWebText (`Skylion007/openwebtext`), GPT-2-tokenized,

  \~40:1 tokens-per-parameter recipe.


## How to use

These are Lightning checkpoints for the [DiffMamba / MDLM codebase](https://github.com/shivnarainms22/DiffMamba),
not `from_pretrained`-loadable. To evaluate or resume:

```bash

git clone https://github.com/shivnarainms22/DiffMamba

cd DiffMamba

# set up the environment (see requirements.yaml / scripts/)



# download a checkpoint, e.g. the LR-tuned BiMamba-130M (best.ckpt reproduces

# the table PPL; --local-dir preserves the runs/…/checkpoints/ path)

huggingface-cli download Shiv-22/diffmamba-checkpoints \

  runs/runD_lr1e3/checkpoints/best.ckpt --local-dir ./ckpts



# validation perplexity (EMA), matching the table above

python main.py mode=ppl_eval +experiment=runD_130m \

  eval.checkpoint_path=./ckpts/runs/runD_lr1e3/checkpoints/best.ckpt \

  data.cache_dir=<path>/data loader.eval_batch_size=32



# generate samples

python main.py mode=sample_eval +experiment=runD_130m \

  eval.checkpoint_path=./ckpts/runs/runD_lr1e3/checkpoints/best.ckpt \

  loader.eval_batch_size=4

```

Use the matching `+experiment=` for each run: `runD_130m` (BiMamba-130M `runD1`/
`runD2` and the LR-tuned `runD_lr1e3`), `runB_transformer_130m` (DiT `runB`),
`scaling_100m` (`s100`), `scaling_50m` (`s50`). For the hybrid runs use
`hybrid_130m` / `hyb_e3` / `hyb_e6` / `hyb_e12` (see below).

## Limitations

Small scale (50–130M, ≤5B tokens), single-GPU academic compute, forward-pass-only
efficiency benchmark, and a Transformer-tuned training recipe that BiMamba is
shown to be undertuned for. Pure BiMamba-2 does **not** match the Transformer on
quality at this scale. Treat these as a reproduction/portfolio artifact, not a
production model. See the GitHub report for the full limitations section.

## Citation & attribution

Built on **MDLM** (Sahoo et al., *Simple and Effective Masked Diffusion Language

Models*, NeurIPS 2024; [code](https://github.com/kuleshov-group/mdlm)) and
reproduces the direction of **DiffuApriel / DiffuMamba**
(*High-Throughput Diffusion LMs with Mamba Backbone*, arXiv 2511.15927, 2025).

```bibtex

@inproceedings{sahoo2024simple,

  title={Simple and Effective Masked Diffusion Language Models},

  author={Subham Sekhar Sahoo and Marianne Arriola and Aaron Gokaslan and Edgar Mariano Marroquin and Alexander M Rush and Yair Schiff and Justin T Chiu and Volodymyr Kuleshov},

  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},

  year={2024},

  url={https://openreview.net/forum?id=L4uaAR4ArM}

}

```

License: Apache-2.0 (inherited from MDLM).