SuperGemma4-26B DFlash Draft (pilot / PoC)

This is a proof-of-concept DFlash block-diffusion drafter trained against

AEON-7/supergemma4-26b-abliterated-multimodal-nvfp4

(the NVFP4-quantized SuperGemma4 26B Abliterated Multimodal model, bf16 twin).

What is this?

A small draft model used for speculative decoding: instead of the big 26B

target generating one token per step, the drafter proposes multiple tokens in

parallel using block diffusion, and the target verifies them in a single pass.

This can 2-3× generation throughput once the drafter is accurate enough.

This release is a pilot (5K samples, 1 epoch, ~28 min on 1× RTX PRO 6000

Blackwell). Top-1 match rate with the target is only ~5.8%, too low to actually

speed up generation. The artifact exists to validate the full training +

export + deployment pipeline end-to-end. See Roadmap below.

Architecture

| Field | Value |

|---|---|

| Type | DFlash block-diffusion drafter (Qwen3-style) |

| Layers | 5 |

| Hidden | 2816 (matches target) |

| Heads | 22 attention / 22 KV |

| Head dim | 128 (vLLM-compatible) |

| Intermediate | 9728 |

| Vocab | 262144 |

| Max pos | 262144 |

| Block size | 8 |

| Target layer anchors | [1, 8, 14, 20, 27] |

| Parameters | ~570M |

| Dtype | bfloat16 |

Training

| Field | Value |

|---|---|

| Target | SuperGemma4 26B (bf16, transformers 5.5.4 layout) |

| Data | HuggingFaceH4/ultrachat_200k train_sft, first 5000 conversations |

| Seq length | 2048 |

| Optimizer | AdamW (fused) |

| LR | 6e-4 (linear decay, 50 warmup steps) |

| Epochs | 1 |

| Steps | 5000 (bs=1, grad_accum=1) |

| Precision | bf16 mixed |

| Hardware | 1× NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB) |

| Wall time | 27:48 |

| Loss (final step) | ~5.9 (noisy; distillation + self-logit) |

| Train acc @ step 5000 | 5.79% (parallel_0_step_0 top-1) |

Training infra notes

Built with NVIDIA/TensorRT-Model-Optimizer#1211

(DFlash training mode, merged in modelopt 0.43.0rc2.dev). Required two

local patches:

1. EagleTrainerWithAccLog._save override to save drafter-only (1.2 GB)

instead of the full 50 GB frozen target + drafter tree — stock HF Trainer

would OOM the filesystem.

2. modeling_gemma4.py line 2027 — bypass the

mm_token_type_ids required when training check for text-only training.

Deployment with vLLM

vllm serve AEON-7/supergemma4-26b-abliterated-multimodal-nvfp4 \
    --speculative-config '{
        "method":"draft_model",
        "model":"AEON-7/supergemma4-26b-dflash-pilot",
        "num_speculative_tokens":5
    }' \
    --quantization modelopt \
    --max-model-len 65536

⚠️ With only 5.8% top-1, expect negative speedup (verification overhead

swamps any gain). This pilot is for plumbing validation, not perf.

Roadmap

This pilot proves the training stack. To make it actually fast:

| Stage | Data | Epochs | Expected top-1 | ETA |

|---|---|---|---|---|

| Pilot (this) | 5K | 1 | 5.8% | 28 min × 1 GPU |

| Small | 50K | 3 | ~25% | ~15 hr × 1 GPU |

| Medium | 500K | 3 | ~55% | ~6 days × 1 GPU (or 18 hr × 8 GPUs) |

| Production | 2M | 5-10 | 70-80% | ~1 week × 8 GPUs |

Production-quality domain-general data (mix of ShareGPT, UltraChat,

Magpie, LMSYS Chat, code) rather than UltraChat alone is also key.

Files

• model.safetensors — 1.2 GB, 58 tensors (5-layer drafter)
• config.json — Qwen3-style DFlashDraftModel config
• tokenizer.json, tokenizer_config.json, chat_template.jinja — from target
• trainer_state.json — loss/acc history per step
• training_args.bin — exact training config for reproducibility

License

Apache 2.0. This is derived work from the SuperGemma4 26B target.

---

☕ Support the work

If this release has been useful, tips are deeply appreciated — they go directly toward more compute, more models, and more open releases.

₿ Bitcoin (BTC)

bc1q09xmzn00q4z3c5raene0f3pzn9d9pvawfm0py4

Ξ Ethereum (ETH)

0x1512667F6D61454ad531d2E45C0a5d1fd82D0500

◎ Solana (SOL)

DgQsjHdAnT5PNLQTNpJdpLS3tYGpVcsHQCkpoiAKsw8t

ⓜ Monero (XMR)

836XrSKw4R76vNi3QPJ5Fa9ugcyvE2cWmKSPv3AhpTNNKvqP8v5ba9JRL4Vh7UnFNjDz3E2GXZDVVenu3rkZaNdUFhjAvgd

Ethereum L2s (Base, Arbitrum, Optimism, Polygon, etc.) and EVM-compatible tokens can be sent to the same Ethereum address.