vllm_omni.diffusion.data ¶

default: AttentionSpec | None = None

per_role: dict[str, AttentionSpec] = field(
    default_factory=dict
)

backend: str

extra: dict[str, Any] = field(default_factory=dict)

Bn_compute_blocks: int = 0

Fn_compute_blocks: int = 1

coefficients: list[float] | None = None

enable_taylorseer: bool = False

force_refresh_step_hint: int | None = None

force_refresh_step_policy: str = 'once'

mag_calibrate: bool = False

mag_max_skip_steps: int = 5

mag_ratios: list[float] | None = None

mag_retention_ratio: float = 0.1

mag_threshold: float = 0.24

max_cached_steps: int = -1

max_continuous_cached_steps: int = 3

max_warmup_steps: int = 4

num_inference_steps: int | None = None

rel_l1_thresh: float = 0.2

residual_diff_threshold: float = 0.24

scm_steps_mask_policy: str | None = None

scm_steps_policy: str = 'dynamic'

taylorseer_order: int = 1

abort_message: str | None = None

aborted: bool = False

custom_output: dict[str, Any] = field(default_factory=dict)

error: str | None = None

output: Tensor | tuple[Any, ...] | dict[str, Any] | None = (
    None
)

peak_memory_mb: float = 0.0

post_process_func: Callable[..., Any] | None = None

stage_durations: dict[str, float] = field(
    default_factory=dict
)

to_cpu: bool = False

trajectory_decoded: list[Image] | None = None

trajectory_latents: Tensor | dict[str, Any] | None = None

trajectory_log_probs: Tensor | dict[str, Any] | None = None

trajectory_timesteps: Tensor | dict[str, Any] | None = None

cfg_parallel_size: int = 1

data_parallel_size: int = 1

enable_expert_parallel: bool = False

hsdp_replicate_size: int = 1

hsdp_shard_size: int = -1

pipeline_parallel_size: int = 1

ring_degree: int = 1

sequence_parallel_size: int | None = None

tensor_parallel_size: int = 1

ulysses_degree: int = 1

ulysses_mode: str = 'strict'

use_hsdp: bool = False

vae_patch_parallel_size: int = 1

error_msg: str | None = None

freed_bytes: int = 0

metadata: dict[str, Any] = field(default_factory=dict)

rank: int | None = None

stage_id: int | None = None

status: str

task_id: str

VSA_sparsity: float = 0.0

additional_config: dict[str, Any] = field(
    default_factory=dict
)

boundary_ratio: float | None = None

cache_backend: str = 'none'

cache_config: DiffusionCacheConfig | dict[str, Any] = field(
    default_factory=dict
)

cache_strategy: str = 'none'

cfg_kv_collect_func: Any | None = None

custom_pipeline_args: dict[str, Any] | None = None

diffusers_call_kwargs: dict[str, Any] = field(
    default_factory=dict
)

diffusers_load_kwargs: dict[str, Any] = field(
    default_factory=dict
)

diffusers_pipeline_cls: type[DiffusionPipeline] | None = (
    None
)

diffusion_attention_config: AttentionConfig = field(
    default_factory=lambda: AttentionConfig()
)

diffusion_kv_cache_dtype: str | None = None

diffusion_kv_cache_skip_layer_indices: set[int] | None = (
    None
)

diffusion_kv_cache_skip_layers: str | None = None

diffusion_kv_cache_skip_step_indices: set[int] | None = None

diffusion_kv_cache_skip_steps: str | None = None

diffusion_load_format: str = 'default'

disable_autocast: bool = False

dist_timeout: int | None = None

distributed_executor_backend: str = 'mp'

dtype: dtype = bfloat16

enable_cache_dit_summary: bool = False

enable_cpu_offload: bool = False

enable_diffusion_pipeline_profiler: bool = False

enable_layerwise_offload: bool = False

enable_multithread_weight_load: bool = True

enable_prompt_embed_cache: bool = False

enable_sleep_mode: bool = False

enable_stage_verification: bool = True

enforce_eager: bool = False

extras: dict[str, Any] = Field(default_factory=dict)

flow_shift: float | None = None

force_cutlass_fp8: bool = False

host: str | None = None

is_moe: bool

log_level: str = 'info'

lora_path: str | None = None

lora_scale: float = 1.0

mask_strategy_file_path: str | None = None

master_port: int | None = None

max_cpu_loras: int | None = None

max_multimodal_image_inputs: int | None = None

max_num_seqs: int = 1

moba_config_path: str | None = None

model: str | None = None

model_class_name: str | None = None

model_config: dict[str, Any] = field(default_factory=dict)

model_loaded: dict[str, bool] = field(
    default_factory=lambda: {
        "transformer": True,
        "vae": True,
    }
)

model_paths: dict[str, str] = field(default_factory=dict)

nccl_port: int | None = None

num_gpus: int | None = None

num_weight_load_threads: int = 4

omni_kv_config: dict[str, Any] = field(default_factory=dict)

output_type: str = 'pil'

override_transformer_cls_name: str | None = None

parallel_config: DiffusionParallelConfig = field(
    default_factory=DiffusionParallelConfig
)

pin_cpu_memory: bool = True

port: int | None = None

profiler_config: ProfilerConfig | dict[str, Any] | None = (
    None
)

prompt_embed_cache_size: int = 32

prompt_file_path: str | None = None

quantization_config: (
    str | QuantizationConfig | dict[str, Any] | None
) = None

revision: str | None = None

scheduler_port: int = 5555

skip_time_steps: int = 15

stage_id: int = 0

step_execution: bool = False

supports_multimodal_inputs: bool = False

tf_model_config: TransformerConfig = field(
    default_factory=TransformerConfig
)

trust_remote_code: bool = False

vae_use_slicing: bool = False

vae_use_tiling: bool = False

worker_extension_cls: str | None = None

level: int = 2

metadata: dict[str, Any] = field(default_factory=dict)

task_id: str

tags: list[str] | None = None

task_id: str

params: dict[str, Any] = field(default_factory=dict)

quant_config: QuantizationConfig | None = None

quant_method: str | None = None