mojo-dllm 0.1.0

First release: CPU inference for two diffusion language models from GGUF.
Measured speed against llama.cpp and diffuse-cpp is in
docs/benchmarks.md; agreement with llama.cpp is in
docs/correctness.md.

Added

• LLaDA-8B and Dream-7B from Q4_K_M GGUF files (F32, Q4_K and Q6_K tensors).
• GGUF v2/v3 reader over a memory map, with mojo-dllm inspect.
• Int8-activation GEMM over weights repacked 8 rows per register, using
  AVX-VNNI when present, with a work-claiming thread pool for hybrid CPUs.
• Bidirectional attention with grouped-query heads; NORM and NEOX RoPE; Q/K/V
  bias; logits computed only for the positions a step samples.
• Byte-level BPE tokenizers (bailingmoe, qwen2) matching llama-tokenize
  id for id on a 36-case multilingual corpus each.
• LLaDA's block sampler (low-confidence or random remasking, Gumbel
  temperature) and Dream's timestep sampler (entropy, maskgit_plus,
  topk_margin).
• mojo-dllm run, tokenize, inspect and logits.
• Benchmark, parity and quality harnesses, with every published number
  rendered from their result files.
• Local-first CI (scripts/ci-local.sh, git hooks) mirrored in GitHub
  Actions; mdBook documentation.

Measured (LLaDA-8B-Instruct Q4_K_M, i5-13420H, 12 threads, 32 steps)

Runtime	ms / step	tokens / s	peak RSS
mojo-dllm	3,005	1.33	5.9 GB
llama.cpp llama-diffusion-cli	5,807	0.69	8.4 GB

Medians of 9 runs from bench/results/2026-10-03-cpu-llada.json; the full tables, Dream-7B and diffuse-cpp included, are in docs/benchmarks.md.

Requirements

Linux on x86-64 with AVX2, about 6 GB free memory, and pixi. Build from source: pixi install && pixi run build. No binaries are attached: the Mojo runtime libraries a build links against ship under Modular's own terms (see NOTICE).