Models

Diffulex model support is defined by three choices:

• model_name: selects the registered model implementation and sampler factory;

• decoding_strategy: selects request state, scheduler, KV cache manager, and model runner;

• sampling_mode: selects standard or edit-style sampler behavior.

The config validator normalizes some model/strategy combinations and rejects known invalid combinations.

Supported Model Families

Model family	model_name values	Typical strategy	Notes
Dream / D2F-Dream	dream	d2f	D2F-style full-prefix block decoding.
DiffuCoder / D2F-DiffuCoder	diffucoder	d2f	Uses shifted sampler behavior.
Dream reasoner	dream_reasoner	multi_bd	Block-causal MultiBD path.
Stable-DiffCoder	stable_diffcoder	multi_bd	Block-causal MultiBD path.
LLaDA / D2F-LLaDA	llada	d2f	Use D2F LoRA-style configs when applicable.
Fast-dLLM-v2	fast_dllm_v2	multi_bd or fast_dllm_v2	Dual-Cache sub-block refinement; fast_dllm_v2 strategy enables the native FDv2 decoding path.
SDAR	sdar	multi_bd	Dense SDAR path.
SDAR-MoE	sdar_moe	multi_bd	MoE path; keep expert parallel at 1 unless extending the runtime.
LLaDA2 family	llada2, llada2_mini, llada2_moe, llada2dot1_mini, llada2_mini_dmax	multi_bd, dmax, or fast_dllm_v2	llada2_mini_dmax enables DMax sampler defaults.
DiffusionGemma	diffusion_gemma	diffusion_gemma	Native 256-token canvas/block decoder.

The original full-attention inference implementations from some upstream dLLM projects are not the target runtime. Diffulex adds support model by model through block-wise adapters, samplers, and strategy registrations.

Strategy Compatibility

Strategy	Use it for	Important behavior
d2f	D2F-style LLaDA, Dream, and DiffuCoder paths	Forces full-prefix multi-block behavior and disables prefix caching.
multi_bd	LLaDA2, SDAR, stable DiffuCoder/Dream reasoner paths	Implements Multi-Block Diffusion with block-causal visibility and prefix caching.
fast_dllm_v2	Fast-dLLM-v2 native dual-cache decoding	3-mode sub-block refinement FSM with dedicated CUDA graphs per mode.
dmax	Supported LLaDA2 edit-sampling experiments	Token merge + edit sampling; requires sampling_mode="edit".
diffusion_gemma	DiffusionGemma	Uses DiffusionGemma request, sampler, block/page size, and canvas defaults.

Sampling Modes

Sampling mode	Use it for
naive	Standard confidence-based diffusion sampling. This is the default for most supported models.
edit	LLaDA2-family edit/remask sampling. Required by DMax-style decoding.

Model Path Requirements

Model and tokenizer paths should point to local directories. During startup, Diffulex loads tokenizer metadata, Hugging Face config, and then the registered Diffulex model implementation.

Before a full benchmark, verify:

Requirement	Check
Checkpoint path	The model directory exists and contains the expected config and weights.
Tokenizer	The tokenizer loads from tokenizer_path or the model path.
model_name	The model name is listed in the table above and registered under diffulex/model/.
Strategy	The strategy is compatible with the model family.
Mask token	mask_token_id matches the tokenizer for the selected model.
Page/block size	block_size <= page_size; DiffusionGemma uses 256/256.

Maintained Benchmark Configs

Common starting points live under diffulex_bench/configs/:

Config	Purpose
llada2_mini_gsm8k.yml	LLaDA2-mini, multi_bd, GSM8K.
llada2_mini_dmax_gsm8k.yml	LLaDA2-mini DMax/edit sampling.
diffusion_gemma_gsm8k.yml	DiffusionGemma native Diffulex benchmark.
diffucoder_instruct_gsm8k.yml	DiffuCoder D2F-style benchmark.
dream_base_gsm8k.yml	Dream D2F-style benchmark.
fast_dllm_v2_gsm8k.yml	Fast-dLLM-v2 multi-block benchmark.
sdar_chat_gsm8k.yml	SDAR dense benchmark.
sdar_moe_chat_gsm8k.yml	SDAR-MoE benchmark.
fast_dllm_v2_multibd_gsm8k.yml	Fast-dLLM-v2 training-free MultiBD variant (blksz=4, bufsz=6).
llada_instruct_gsm8k.yml	LLaDA D2F-style benchmark.
llada2_mini_gsm8k_tp1_limit200_maxnfe2048.yml	Parity/evaluation variant with TP=1 and 200-sample limit.

Start with --dataset-limit before running a full dataset.