vllm.config.diffusion
¶
Configuration for discrete diffusion (dLLM) models.
Classes:
-
DiffusionConfig–Configuration for discrete diffusion language models (dLLMs).
-
DiffusionModelOptions–Base class for model-specific
DiffusionConfig.model_options.
DiffusionConfig
¶
Configuration for discrete diffusion language models (dLLMs).
dLLMs generate tokens via iterative denoising over a fixed-length canvas rather than left-to-right autoregressive decoding. They reuse the speculative-decoding data path (draft token ids, scheduled spec decode tokens) with overloaded semantics for block-based generation.
Attributes:
-
canvas_length(int) –Length of the denoising canvas (block). Also determines the number of
-
max_denoising_steps(int | None) –Maximum number of denoising iterations per canvas block.
-
model_options(dict[str, Any] | DiffusionModelOptions) –Model-specific options. The accepted keys depend on the model and are
Source code in vllm/config/diffusion.py
canvas_length = Field(default=None, gt=0)
class-attribute
instance-attribute
¶
Length of the denoising canvas (block). Also determines the number of speculative tokens scheduled per step.
max_denoising_steps = None
class-attribute
instance-attribute
¶
Maximum number of denoising iterations per canvas block. If not set, read from the model's generation_config.json.
model_options = Field(default_factory=dict)
class-attribute
instance-attribute
¶
Model-specific options. The accepted keys depend on the model and are
validated once the model is known, e.g. NemotronDiffusionOptions or
DiffusionGemmaDiffusionOptions in vllm.model_executor.models.config.