bridge.models.glm_moe_dsa.glm5_provider

Hybrid model provider for the GLM-5 DSA family.

Module Contents

Classes

GLM5ModelProvider

Construct GLM as DSA/dense/MoE Hybrid layers with MLA configuration.

Functions

glm_hybrid_pattern	Build an architecture-preserving pattern from HF block counts.
split_glm_pattern	Place PP boundaries without splitting attention/MLP pairs.
_partition_layer_groups	Split ordered layer groups into contiguous stages that minimize the largest stage.

Data

__all__

API

bridge.models.glm_moe_dsa.glm5_provider.glm_hybrid_pattern(

*,

num_layers: int,

first_k_dense_replace: int,

) → str

Build an architecture-preserving pattern from HF block counts.

bridge.models.glm_moe_dsa.glm5_provider.split_glm_pattern(pattern: str, block_counts: list[int]) → str

Place PP boundaries without splitting attention/MLP pairs.

bridge.models.glm_moe_dsa.glm5_provider._partition_layer_groups(

group_sizes: list[int],

num_stages: int,

) → list[int]

Split ordered layer groups into contiguous stages that minimize the largest stage.

Parameters:

• group_sizes – Number of layers in each group, in layer order.

• num_stages – Number of stages; must not exceed the number of groups.

Returns:

Number of layers in each stage. Every stage receives at least one whole group.

class bridge.models.glm_moe_dsa.glm5_provider.GLM5ModelProvider

Bases: megatron.bridge.models.hybrid.hybrid_provider.HybridModelProvider, megatron.bridge.models.transformer_config.MLATransformerConfig

Construct GLM as DSA/dense/MoE Hybrid layers with MLA configuration.

_resolve_hybrid_stack_spec() → megatron.core.transformer.spec_utils.ModuleSpec

finalize() → None

Validate physical layout and preserve DSA sharing within each stage.

bridge.models.glm_moe_dsa.glm5_provider.__all__

[‘GLM5ModelProvider’]