DrCubix/aevox-diffusion-drafter-nemotron3-super

AeVox Diffusion Drafter — LoRA for Nemotron-Labs-Diffusion-3B

Created by Daniel Rodd / AeVox.Ai.

A LoRA adapter that aligns NVIDIA's `Nemotron-Labs-Diffusion-3B` to the token distribution of `nvidia/NVIDIA-Nemotron-3-Super-120B-A12B`, so the 3B can serve as a cross-model speculative-decoding drafter for the 120B.

What it does

Speculative decoding accepts more tokens per expensive 120B forward when the drafter predicts the 120B's tokens well. This adapter raises the 3B drafter's accepted tokens per target forward:

With an NVFP4-quantized drafter this projects to ~2.5× decode speedup over autoregressive (acceptance measured offline; tok/s projected — see the repo's docs/04-serving.md).

Usage

import torch
from transformers import AutoModel
from peft import PeftModel

m = AutoModel.from_pretrained("nvidia/Nemotron-Labs-Diffusion-3B",
                              trust_remote_code=True, dtype=torch.bfloat16).to("cuda").eval()
m = PeftModel.from_pretrained(m, "DrCubix/aevox-diffusion-drafter-nemotron3-super").merge_and_unload()
# m.encoder / m.diffusion_head = the aligned drafter.
# See github.com/DRCubix/aevox-diffusion-drafter (src/eval_lora_acceptance.py) for the spec-decode loop.

Requires transformers>=5.0 and trust_remote_code=True. The drafter and 120B share a byte-identical tokenizer (vocab 131,072), so draft token IDs are directly verifiable by the 120B.

Training

• —Data: 10K (prompt, 120B-completion) pairs (DrCubix/nemotron3-super-120b-distill), code-heavy, full-reasoning, temp=1.0.
• —Objective: the 3B's native bidirectional diffusion CE on masked completion tokens.
• —LoRA: rank 16 on q/k/v/o + gate/up/down (24.7M params, 0.64%), 3 epochs, lr 5e-5, eps-floored loss for stability.

Limitations & honest notes

• —Acceptance is measured; the ~2.5× speedup is projected (needs a custom vLLM drafter integration + NVFP4 quantization).
• —The acceptance ceiling for this token-only, single-step, standalone drafter is ~2.8. Pushing to 3–4× needs hidden-state conditioning (DFlash) or tree drafting — see the repo's findings doc.

Citation

See CITATION.cff in the repo. Created by Daniel Rodd / AeVox.Ai. Built on NVIDIA Nemotron models under the NVIDIA Open Model License.

Repo: https://github.com/DRCubix/aevox-diffusion-drafter