For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python module

max.pipelines.architectures.wan

Wan diffusion architecture for video generation.

BlockLevelModel

class max.pipelines.architectures.wan.BlockLevelModel(pre, post, *, combined_blocks)

source

Bases: object

Executes transformer forward pass as pre -> combined blocks -> post.

All transformer blocks are compiled into a single Model graph, so the runtime allocates one shared workspace.

Parameters:

• pre (Model)
• post (Model)
• combined_blocks (Model)

WanArchConfig

class max.pipelines.architectures.wan.WanArchConfig(*, pipeline_config, quantization_encoding=None)

source

Bases: ArchConfig

Pipeline-level config for Wan (implements ArchConfig; no KV cache).

Parameters:

• pipeline_config (PipelineConfig)
• quantization_encoding (SupportedEncoding | None)

DEFAULT_ENCODING

DEFAULT_ENCODING: ClassVar[SupportedEncoding] = 'bfloat16'

source

SUPPORTED_ENCODINGS

SUPPORTED_ENCODINGS: ClassVar[set[SupportedEncoding]] = {'bfloat16', 'float32', 'float8_e4m3fn'}

source

get_max_seq_len()

get_max_seq_len()

source

Returns the default maximum sequence length for the model.

Subclasses should determine whether this value can be overridden by setting the --max-length (pipeline_config.model.max_length) flag.

Return type:

int

initialize()

classmethod initialize(pipeline_config, model_config=None)

source

Initialize the config from a PipelineConfig.

Parameters:

• pipeline_config (PipelineConfig) – The pipeline configuration.
• model_config (MAXModelConfig | None) – The model configuration to read from. When None (the default), pipeline_config.model is used. Pass an explicit config (e.g. pipeline_config.draft_model) to initialize the arch config for a different model.

Return type:

Self

pipeline_config

pipeline_config: PipelineConfig

source

quantization_encoding

quantization_encoding: SupportedEncoding | None = None

source

WanConfig

class max.pipelines.architectures.wan.WanConfig(*, config_file=None, section_name=None, patch_size=(1, 2, 2), num_attention_heads=40, attention_head_dim=128, in_channels=16, out_channels=16, text_dim=4096, freq_dim=256, ffn_dim=13824, num_layers=40, cross_attn_norm=True, qk_norm='rms_norm_across_heads', eps=1e-06, image_dim=None, added_kv_proj_dim=None, rope_max_seq_len=1024, pos_embed_seq_len=None, dtype=bfloat16, device=<factory>, quant_config=None)

source

Bases: WanConfigBase

Parameters:

• config_file (str | None)
• section_name (str | None)
• patch_size (tuple[int, int, int])
• num_attention_heads (int)
• attention_head_dim (int)
• in_channels (int)
• out_channels (int)
• text_dim (int)
• freq_dim (int)
• ffn_dim (int)
• num_layers (int)
• cross_attn_norm (bool)
• qk_norm (str | None)
• eps (float)
• image_dim (int | None)
• added_kv_proj_dim (int | None)
• rope_max_seq_len (int)
• pos_embed_seq_len (int | None)
• dtype (DType)
• device (DeviceRef)
• quant_config (QuantConfig | None)

generate()

static generate(config_dict, encoding, devices)

source

Parameters:

• config_dict (dict[str, Any])
• encoding (Literal['float32', 'float16', 'bfloat16', 'q4_k', 'q4_0', 'q6_k', 'float8_e4m3fn', 'float4_e2m1fnx2', 'gptq'])
• devices (list[Device])

Return type:

WanConfig

model_config

model_config: ClassVar[ConfigDict] = {'arbitrary_types_allowed': True, 'extra': 'forbid', 'strict': False}

source

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

WanConfigBase

class max.pipelines.architectures.wan.WanConfigBase(*, config_file=None, section_name=None, patch_size=(1, 2, 2), num_attention_heads=40, attention_head_dim=128, in_channels=16, out_channels=16, text_dim=4096, freq_dim=256, ffn_dim=13824, num_layers=40, cross_attn_norm=True, qk_norm='rms_norm_across_heads', eps=1e-06, image_dim=None, added_kv_proj_dim=None, rope_max_seq_len=1024, pos_embed_seq_len=None, dtype=bfloat16, device=<factory>, quant_config=None)

source

Bases: MAXModelConfigBase

Parameters:

• config_file (str | None)
• section_name (str | None)
• patch_size (tuple[int, int, int])
• num_attention_heads (int)
• attention_head_dim (int)
• in_channels (int)
• out_channels (int)
• text_dim (int)
• freq_dim (int)
• ffn_dim (int)
• num_layers (int)
• cross_attn_norm (bool)
• qk_norm (str | None)
• eps (float)
• image_dim (int | None)
• added_kv_proj_dim (int | None)
• rope_max_seq_len (int)
• pos_embed_seq_len (int | None)
• dtype (DType)
• device (DeviceRef)
• quant_config (QuantConfig | None)

added_kv_proj_dim

added_kv_proj_dim: int | None

source

attention_head_dim

attention_head_dim: int

source

cross_attn_norm

cross_attn_norm: bool

source

device

device: DeviceRef

source

dtype

dtype: DType

source

eps

eps: float

source

ffn_dim

ffn_dim: int

source

freq_dim

freq_dim: int

source

image_dim

image_dim: int | None

source

in_channels

in_channels: int

source

model_config

model_config: ClassVar[ConfigDict] = {'arbitrary_types_allowed': True, 'extra': 'forbid', 'strict': False}

source

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

num_attention_heads

num_attention_heads: int

source

num_layers

num_layers: int

source

out_channels

out_channels: int

source

patch_size

patch_size: tuple[int, int, int]

source

pos_embed_seq_len

pos_embed_seq_len: int | None

source

qk_norm

qk_norm: str | None

source

quant_config

quant_config: QuantConfig | None

source

Static per-tensor FP8 quantization config, populated when the transformer encoding is float8_e4m3fn. None for bfloat16.

rope_max_seq_len

rope_max_seq_len: int

source

text_dim

text_dim: int

source