Credits & Attribution

May 25, 2026 · View on GitHub

dlmserve builds on the ideas and research from the following works. Every adapted technique is listed here per the licensing protocol documented in CREDITS.md.

LLaDA — Large Language Diffusion with mAsking

  • Paper: "LLaDA: Large Language Diffusion with mAsking", Nie et al., arXiv:2502.09992 (2025)
  • Source repo: https://github.com/ML-GSAI/LLaDA
  • License: MIT (Copyright 2025 NieShenRuc) — full text in reference/LICENSE-LLaDA
  • Adapted in:
    • reference/llada_reference.py — near-verbatim copy of upstream generate.py, used only as the reference-match ground truth. main() removed; otherwise unchanged. Not imported by the runtime engine.
    • dlmserve/denoise_loop.py — independent reimplementation of the absorb-and-resample inference procedure (paper §3) for batch serving. Matches the reference behavior token-exactly under the deterministic path (see tests/test_reference_match.py).
    • dlmserve/sampler.py — confidence top-k commit, derived from the same procedure.
  • What we used: Absorb-and-resample inference procedure (paper §3); per-step token-count schedule (get_num_transfer_tokens integer formula); low-confidence remasking strategy; mask token id 126336.
  • Modifications in dlmserve's reimplementation: structured for batched multi-request serving; separate KV-cache layer; scheduler outside the denoise loop. No CFG, no semi-AR variants beyond the upstream block_length knob in v0.1.

DiffuLLaMA — Scaling Masked Diffusion Language Models from LLaMA

  • Paper: Gong et al., arXiv:2410.12327 (2024) — "Scaling Diffusion Language Models via Adaptation from Autoregressive Models"
  • Source repo: https://huggingface.co/diffusionfamily/diffullama
  • License: Apache 2.0 (confirmed on HF model card, 2026-05-23)
  • Adapted in:
    • dlmserve/models/diffullama.py — model loader for DiffuLLaMA-7B (~13.5 GB bf16). Supported as an alternative model; tests default to LLaDA-8B-Instruct.
  • What we used: Model architecture and HF checkpoint for integration test coverage. No inference-math adapted — loader pattern mirrors llada.py.
  • Modifications: None — loading checkpoint as-is from HF hub.

Dream-7B — Diffusion Large Language Model

  • Paper: Lin et al., "Dream 7B: Diffusion Large Language Models", arXiv:2508.15487 (Aug 2025)
  • Source repo: https://github.com/HKUNLP/Dream
  • Model: https://huggingface.co/Dream-org/Dream-v0-Instruct-7B
  • License: Apache 2.0 (confirmed on HF model card and GitHub repo, 2026-05-25)
  • Adapted in (v0.2.0, planned):
    • dlmserve/models/dream.py — model loader for Dream-7B (~5.6 GB INT4).
    • dlmserve/sampler.py — extended confidence_metric enum to support Dream's alg strategies (top1_prob, margin, entropy, random) and alg_temp ordering noise per ADR 006.
    • dlmserve/api/protocol.py — added top_p and alg_temp params to ChatCompletionRequest.
  • What we used: Discrete masked-diffusion inference contract from paper §3-4; the four alg strategies for unmasking-order selection; alg_temp noise injection; LLaMA-derived tokenizer auto-discovery for mask token.
  • Modifications in dlmserve's reimplementation: integrated into the existing absorb-and-resample loop (denoise_loop.py); per-model LocalLeap thresholds calibrated separately for Dream-7B; mask token auto-discovered from tokenizer rather than hard-coded (Dream doesn't have a fixed mask ID across versions like LLaDA does).

LocalLeap — Local-Aware Anchor Propagation for Diffusion LLM Inference

  • Paper: Kong et al., "LocalLeap: Local-Aware Anchor Propagation for Diffusion Language Model Inference Acceleration", arXiv:2510.07081 (Oct 2025)
  • Source repo: https://github.com/friedrichor/LocalLeap
  • License: Apache-2.0 (Klear Team / Kuaishou Technology)
  • Adapted in:
    • dlmserve/sampler.py — commit_with_local_leap() implements §3 anchor-propagation algorithm
    • dlmserve/denoise_loop.py — LocalLeap commit path gated behind SamplingParams.use_local_leap=False (opt-in default-off, matching industry precedent for lossy accelerations — vLLM/SGLang speculative decoding follow the same pattern)
    • dlmserve/scheduler.py — use_local_leap is part of the batch group key so LocalLeap and non-LocalLeap requests never share a batch
  • What we used: Anchor token identification (top-k by confidence with anchor threshold κ) and local neighborhood propagation (neighbors within radius W committed if their confidence ≥ τ). Paper §3.2–3.3. Paper reports up to 6.94× speedup with HumanEval/GSM8K/MBPP/IFEval scores preserved within 1-2pp at LLaDA-8B-Instruct's recommended κ=0.9, τ=0.75, W=4.
  • Modifications: Integrated into dlmserve's batch DenoiseState and the existing top-k commit path; thresholds exposed via SamplingParams; no changes to the absorb-and-resample math contract.

No code is adapted without a corresponding entry here and an inline citation in the source file.