bridge.models.kimi.kimi_k3_provider#
Model provider for the Kimi K3 language backbone.
Module Contents#
Classes#
Megatron configuration and provider for Kimi K3. |
API#
- class bridge.models.kimi.kimi_k3_provider.KimiK3ModelProvider#
Bases:
megatron.bridge.models.mla_provider.MLAModelProviderMegatron configuration and provider for Kimi K3.
- variable_seq_lengths: bool#
True
- kimi_kda_layers: tuple[int, ...]#
()
- kimi_linear_num_heads: int#
96
- kimi_linear_head_dim: int#
128
- kimi_linear_conv_kernel_size: int#
4
- kimi_kda_gate_lower_bound: float#
None
- kimi_attn_res_block_size: int#
12