nemo_automodel.components.models.glm5_next

Native AutoModel support for zai-org/GLM-5.3-Flash.

Submodules

Package Contents

Classes

Name	Description
`Glm5NextConfig`	Top-level GLM-5.3 vision-language configuration.
`Glm5NextForConditionalGeneration`	Trainable GLM-5.3 VLM with EP and contiguous packed CP support.
`Glm5NextTextConfig`	Configuration for the GLM-5.3 hybrid KDA/KPool-DSA text backbone.
`Glm5NextVisionConfig`	Configuration for the GLM-5.3 image/video encoder.

Data

ModelClass

API

class nemo_automodel.components.models.glm5_next.config.Glm5NextConfig(
    text_config: dict[str, typing.Any] | nemo_automodel.components.models.glm5_next.config.Glm5NextTextConfig | None = None,
    vision_config: dict[str, typing.Any] | nemo_automodel.components.models.glm5_next.config.Glm5NextVisionConfig | None = None,
    image_token_id: int = 154854,
    video_token_id: int = 154855,
    image_start_token_id: int = 154830,
    image_end_token_id: int = 154831,
    video_start_token_id: int = 154832,
    video_end_token_id: int = 154833,
    tie_word_embeddings: bool = False,
    kwargs: typing.Any = {}
)

Bases: PretrainedConfig

Top-level GLM-5.3 vision-language configuration.

hidden_size

= text_config.hidden_size

keys_to_ignore_at_inference

= ['past_key_values']

max_position_embeddings

= text_config.max_position_embeddings

model_type

= 'glm5_next'

sub_configs

vocab_size

= text_config.vocab_size

nemo_automodel.components.models.glm5_next.config.Glm5NextConfig.get_text_config(
    decoder: bool = False
) -> nemo_automodel.components.models.glm5_next.config.Glm5NextTextConfig

Return the decoder config using the Transformers multimodal protocol.

nemo_automodel.components.models.glm5_next.config.Glm5NextConfig.to_dict() -> dict[str, typing.Any]

class nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration(
    config: nemo_automodel.components.models.glm5_next.config.Glm5NextConfig,
    moe_config: nemo_automodel.components.moe.config.MoEConfig | None = None,
    backend: nemo_automodel.components.models.common.BackendConfig | None = None,
    kwargs: typing.Any = {}
)

Bases: HFCheckpointingMixin, Module, MoEFSDPSyncMixin

Trainable GLM-5.3 VLM with EP and contiguous packed CP support.

_keep_in_fp32_modules_strict

_packed_cp_attn_backends

= ('sdpa', 'cudnn')

backend

language_model

Glm5NextTextModel

Expose the text module through the multimodal discovery protocol.

lm_head

model

state_dict_adapter

tie_word_embeddings_support

TieSupport = TieSupport.UNTIED_ONLY

vocab_size

= text_config.vocab_size

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration._embed_and_splice(
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor | None,
    image_grid_thw: torch.Tensor | None
) -> torch.Tensor

Embed the full sequence and replace image placeholder positions.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.forward(
    input_ids: torch.Tensor | None = None,
    attention_mask: torch.Tensor | None = None,
    position_ids: torch.Tensor | None = None,
    inputs_embeds: torch.Tensor | None = None,
    padding_mask: torch.Tensor | None = None,
    pixel_values: torch.Tensor | None = None,
    image_grid_thw: torch.Tensor | None = None,
    pixel_values_videos: torch.Tensor | None = None,
    logits_to_keep: int | torch.Tensor = 0,
    output_hidden_states: bool | None = None,
    kwargs: typing.Any = {}
) -> transformers.modeling_outputs.CausalLMOutputWithPast

Run image splice, contiguous CP slicing, text decoding and lm head.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.from_config(
    config: nemo_automodel.components.models.glm5_next.config.Glm5NextConfig,
    moe_config: nemo_automodel.components.moe.config.MoEConfig | None = None,
    backend: nemo_automodel.components.models.common.BackendConfig | None = None,
    kwargs: typing.Any = {}
) -> 'Glm5NextForConditionalGeneration'

classmethod

Construct from an already resolved native config.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.from_pretrained(
    pretrained_model_name_or_path: str,
    model_args: typing.Any = (),
    kwargs: typing.Any = {}
) -> 'Glm5NextForConditionalGeneration'

classmethod

Resolve the local config; checkpoint loading is owned by AutoModel.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.get_image_features(
    pixel_values: torch.Tensor,
    image_grid_thw: torch.Tensor
) -> nemo_automodel.components.models.glm5_next.vision.Glm5NextVisionOutput

Return raw and merged features for flattened image patches.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.get_input_embeddings() -> torch.nn.Module

Return the token embedding table.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.get_output_embeddings() -> torch.nn.Module

Return the untied language-model head.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.initialize_weights(
    buffer_device: torch.device | None = None,
    dtype: torch.dtype = torch.bfloat16
) -> None

Initialize all tensors for checkpoint-free construction.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.prepare_model_inputs_for_cp(
    batch: dict[str, typing.Any],
    num_chunks: int = 1
) -> dict[str, typing.Any]

Install GLM’s contiguous packed sharder while leaving media and ids global.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.set_input_embeddings(
    value: torch.nn.Module
) -> None

Replace the token embedding table.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.set_output_embeddings(
    value: torch.nn.Module
) -> None

Replace the language-model head.

nemo_automodel.components.models.glm5_next.model.Glm5NextForConditionalGeneration.update_moe_gate_bias() -> None

Update no-aux-loss router correction biases after an optimizer step.

class nemo_automodel.components.models.glm5_next.config.Glm5NextTextConfig(
    vocab_size: int = 154880,
    hidden_size: int = 4096,
    intermediate_size: int = 12288,
    moe_intermediate_size: int = 2048,
    num_hidden_layers: int = 45,
    num_attention_heads: int = 64,
    num_key_value_heads: int = 64,
    n_shared_experts: int = 1,
    n_routed_experts: int = 288,
    routed_scaling_factor: float = 2.5,
    kv_lora_rank: int = 512,
    q_lora_rank: int = 1536,
    qk_rope_head_dim: int = 0,
    qk_nope_head_dim: int = 256,
    v_head_dim: int = 256,
    n_group: int = 1,
    topk_group: int = 1,
    num_experts_per_tok: int = 8,
    norm_topk_prob: bool = True,
    mlp_layer_types: list[str] | None = None,
    layer_types: list[str] | None = None,
    indexer_types: list[str] | None = None,
    index_topk_pattern: str | list[str] | None = None,
    index_topk_freq: int = 1,
    index_skip_topk_offset: int = 2,
    index_topk: int = 2048,
    index_head_dim: int = 128,
    index_n_heads: int = 32,
    index_kpool: int = 16,
    index_kpool_always_select_tail: bool = True,
    hidden_act: str = 'silu',
    swiglu_limit: float = 10.0,
    linear_head_dim: int = 128,
    linear_num_heads: int = 64,
    linear_conv_kernel_dim: int = 4,
    linear_lower_bound: float | None = -5.0,
    linear_attn_config: dict[str, typing.Any] | None = None,
    hc_mult: int = 4,
    hc_eps: float = 1e-06,
    hc_sinkhorn_iters: int = 20,
    max_position_embeddings: int = 1048576,
    initializer_range: float = 0.02,
    rms_norm_eps: float = 1e-05,
    use_cache: bool = False,
    attention_bias: bool = False,
    attention_dropout: float = 0.0,
    output_router_logits: bool = False,
    router_aux_loss_coef: float = 0.001,
    num_nextn_predict_layers: int = 1,
    pad_token_id: int | None = 154820,
    bos_token_id: int | None = None,
    eos_token_id: int | list[int] | None = None,
    tie_word_embeddings: bool = False,
    kwargs: typing.Any = {}
)

Bases: PretrainedConfig

Configuration for the GLM-5.3 hybrid KDA/KPool-DSA text backbone.

attribute_map

= {'num_local_experts': 'n_routed_experts'}

indexer_types

= list(indexer_types)

keys_to_ignore_at_inference

= ['past_key_values']

layer_types

= list(layer_types)

linear_attn_config

mlp_layer_types

= list(mlp_layer_types)

model_type

= 'glm5_next_text'

num_local_experts

int

Alias used by Hugging Face expert implementations.

qk_head_dim

= qk_nope_head_dim + qk_rope_head_dim

nemo_automodel.components.models.glm5_next.config.Glm5NextTextConfig.to_dict() -> dict[str, typing.Any]

class nemo_automodel.components.models.glm5_next.config.Glm5NextVisionConfig(
    depth: int = 24,
    hidden_size: int = 1024,
    hidden_act: str = 'silu',
    attention_bias: bool = True,
    attention_dropout: float = 0.0,
    num_heads: int = 16,
    in_channels: int = 3,
    image_size: int = 448,
    patch_size: int = 14,
    rms_norm_eps: float = 1e-05,
    spatial_merge_size: int = 2,
    temporal_patch_size: int = 2,
    out_hidden_size: int = 4096,
    intermediate_size: int = 4096,
    projection_intermediate_size: int = 10240,
    initializer_range: float = 0.02,
    swiglu_limit: float = 10.0,
    kwargs: typing.Any = {}
)

Bases: PretrainedConfig

Configuration for the GLM-5.3 image/video encoder.

model_type

= 'glm5_next_vision'

nemo_automodel.components.models.glm5_next.config.Glm5NextVisionConfig.to_dict() -> dict[str, typing.Any]

nemo_automodel.components.models.glm5_next.ModelClass = Glm5NextForConditionalGeneration