Brunobkr/diffusion-gemma-http.cpp

ΩFFΣLLIα — diffusion-gemma-http

A native, single-file, OpenAI-compatible HTTP server for DiffusionGemma block text-diffusion models (GGUF / llama.cpp).

diffusion-gemma-http.cpp makes a block text-diffusion model behave like a regular llama-server: load the GGUF once, listen on a port, answer /v1/chat/completions. Everything — tokenization, chat template, the block-diffusion denoising loop, and the HTTP API — runs in a single C++ process on top of the llama.cpp diffusion fork. No Python, no external tokenizer files, no per-step IPC.

Validated end-to-end on CPU-only consumer hardware (AMD Ryzen 5 5625U, 8 GB shared UMA, Kali Linux), serving a DiffusionGemma 26B-A4B MoE (NVFP4 GGUF) to a local chat UI over :8080.

---

Why this exists

Diffusion language models cannot be served by the standard llama-server: generation is not autoregressive token-by-token decoding, but iterative denoising of a fixed-length canvas ([prompt | canvas] bidirectional forwards, region-aware masks, self-conditioning). Until now the options were a raw logits server driven by an external Python loop, or a one-shot CLI. This tool closes the gap: a persistent HTTP server speaking the OpenAI Chat Completions protocol, with the entire diffusion decode loop in-process.

Features

• OpenAI-compatible API: POST /v1/chat/completions (streaming via SSE and non-streaming), GET /v1/models, GET /health, GET /
• Native tokenization from the GGUF vocabulary (control tokens parsed atomically; no tokenizer.json needed)
• Model chat template built in: <|turn>role\n … <turn|> turns with <|channel>thought … <channel|> reasoning channels; thinking can be disabled per request ("enable_thinking": false, the default) by pre-filling an empty thought channel, or enabled globally with --thinking
• Entropy-bound block-diffusion sampler (see below), with all parameters read from GGUF metadata
• Prompt-KV caching (DG_KVCACHE=1 / --kvcache): the prompt is prefilled once per block; each denoising step forwards only the canvas
• Self-conditioning across denoising steps (previous-step logits fed back from step 2 onward)
• Multi-block generation: when a block fills without an end-of-turn, it is appended to the prompt and a new canvas is denoised, until max_tokens or end of content
• Streaming per block with correct finish_reason (stop vs length)
• Memory-conscious: per-position statistics are computed in streaming passes over the logits — the probability matrix (canvas × 262k vocab) is never materialized

Requirements

1. A llama.cpp fork with the diffusion-gemma architecture (the DIFFUSION_GEMMA model class providing llama_diffusion_set_phase / llama_diffusion_set_sc).
2. A DiffusionGemma GGUF carrying the diffusion metadata:

GGUF key	Meaning
diffusion.canvas_length	Canvas size per block (required; the C++ graph splits [prompt | canvas] on it)
diffusion.eb_max_steps	Max denoising steps per block
diffusion.eb_t_min / diffusion.eb_t_max	Temperature schedule (linear, t_max → t_min)
diffusion.eb_entropy_bound	Per-position entropy bound for locking
diffusion.eb_stability_threshold	Consecutive stable-argmax steps required to lock
diffusion.eb_confidence_threshold	Reference-decoder parameter (read, reported, not used by this sampler — see Limitations)

3. The vendored headers already shipped with the llama.cpp tree: vendor/cpp-httplib/httplib.h (+ httplib.cpp) and vendor/nlohmann/json.hpp.

Build

Place the file at tools/diffusion-gemma-http/diffusion-gemma-http.cpp inside the fork, then:

# tools/diffusion-gemma-http/CMakeLists.txt
set(TARGET llama-diffusion-gemma-http)
add_executable(${TARGET}
    diffusion-gemma-http.cpp
    ${CMAKE_SOURCE_DIR}/vendor/cpp-httplib/httplib.cpp)
target_include_directories(${TARGET} PRIVATE
    ${CMAKE_SOURCE_DIR}/vendor/cpp-httplib
    ${CMAKE_SOURCE_DIR}/vendor)
target_link_libraries(${TARGET} PRIVATE llama Threads::Threads)
target_compile_features(${TARGET} PRIVATE cxx_std_17)
install(TARGETS ${TARGET} RUNTIME)

echo 'add_subdirectory(diffusion-gemma-http)' >> tools/CMakeLists.txt
cmake -B build
cmake --build build --target llama-diffusion-gemma-http -j4

Usage

DG_KVCACHE=1 ./build/bin/llama-diffusion-gemma-http \
  -m model.gguf --port 8080 [--host 0.0.0.0] [-ngl N] [-c MAXTOK] [--thinking]

Flag / env	Default	Description
-m, --model	—	GGUF path (positional also accepted)
--port / --host	8080 / 0.0.0.0	Bind address
-c, --ctx / MAXTOK	2304	Context budget = prompt + accumulated blocks. Non-causal forwards require the whole sequence in one ubatch, so the compute buffer scales with this — raise gradually on small-RAM machines
-ngl / NGL	0	GPU layers
--kvcache / DG_KVCACHE=1	off	Prompt-KV caching (strongly recommended on CPU)
--thinking / DG_THINKING=1	off	Enable the reasoning channel by default
DG_MASK_ID	auto (<mask>)	Override the canvas mask token id
FA=1	off	Flash attention

API

curl -s http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "messages": [{"role": "user", "content": "Explique em duas frases o que é um número primo."}],
  "max_tokens": 200,
  "stream": false
}'

Any OpenAI-compatible client or front-end pointed at http://host:8080 works unchanged. Per-request fields: messages (system merged into the first user turn), max_tokens / max_completion_tokens, stream, enable_thinking.

The sampler

The decode loop implements an entropy-bound block-diffusion sampler:

1. The canvas starts fully masked. Each step runs one bidirectional forward and computes, per position, the best real token (mask excluded), its confidence, and the entropy of the mask-excluded distribution — in streaming passes, without materializing probabilities.
2. A position locks when its argmax has been stable for stability_threshold consecutive steps and its entropy is below entropy_bound.
3. A prediction of <mask> never locks: it means "not ready yet / end of content". Positions still masked when the model proposes nothing new for 2 consecutive steps signal end of content (the model's native length control).
4. Minimum progress per step is guaranteed by confidence ranking; adjacent positions never lock in the same step, and an anti-echo guard defers locking a token identical to an already-locked neighbor (legitimate repetition persists and passes; denoising echoes dissolve).
5. Temperature follows a linear t_max → t_min schedule; self-conditioning on the previous step's logits is active from step 2.

Honest limitations

• This sampler is a validated approximation, not a byte-exact port of the reference entropy-bound decoder: in particular, diffusion.eb_confidence_threshold is read and reported but plays no role in the locking rule, whose reference semantics differ from a naive confidence cutoff. Residual artifacts of parallel unmasking (rare token echoes) are mitigated by the guards above but not formally eliminated.
• Single-flight inference: concurrent requests are serialized by a mutex.
• Diffusion on CPU is compute-heavy: every denoising step is a dense forward over the canvas (plus the prompt without KV caching). Expect minutes, not seconds, for long answers on laptop-class CPUs.

Provenance

Developed iteratively against a live DiffusionGemma 26B-A4B (MoE, 30 layers, 262k vocab, canvas 256, Harmony-style <|turn>/<|channel> template) quantized to NVFP4 GGUF, debugged end-to-end from raw logits to a working chat UI. Part of the ΩFFΣLLIα local-first, zero-telemetry tooling line.

• Author: Bruno Becker — huggingface.co/Brunobkr
• Research: doi.org/10.5281/zenodo.20026837
• Built on llama.cpp (MIT) with a diffusion-gemma architecture fork; HTTP via cpp-httplib, JSON via nlohmann/json.

License

MIT, following the llama.cpp ecosystem it extends.