`bridge.models.gpt_oss.gpt_oss_provider`#

Module Contents#

Classes#

`GPTOSSProvider`	Base config for GPT-OSS
`GPTOSSProvider120B`	Config for GPT-OSS 120B
`GPTOSSProvider20B`	Config for GPT-OSS 20B

Data#

logger

API#

bridge.models.gpt_oss.gpt_oss_provider.logger#: ‘getLogger(…)’

class bridge.models.gpt_oss.gpt_oss_provider.GPTOSSProvider#

Bases: megatron.bridge.models.gpt_provider.GPTModelProvider

Base config for GPT-OSS

hidden_size: int#: 2880

num_attention_heads: int#: 64

num_query_groups: int#: 8

ffn_hidden_size: int#: 2880

kv_channels: Optional[int]#: 64

normalization: str#: ‘RMSNorm’

gated_linear_unit: bool#: True

add_bias_linear: bool#: True

share_embeddings_and_output_weights: bool#: False

vocab_size: int#: 201088

hidden_dropout: float#: 0.0

attention_dropout: float#: 0.0

bf16: bool#: True

params_dtype: torch.dtype#: None

position_embedding_type: str#: ‘yarn’

rotary_base: int#: 150000

yarn_rotary_scaling_factor: float#: 32.0

yarn_original_max_position_embeddings: int#: 4096

yarn_beta_fast: float#: 32.0

yarn_beta_slow: float#: 1.0

yarn_correction_range_round_to_int: bool#: False

yarn_mscale: float#: 1.0

yarn_mscale_all_dim: float#: 1.0

moe_router_topk: int#: 4

moe_router_pre_softmax: bool#: False

moe_grouped_gemm: bool#: True

moe_token_dispatcher_type: str#: ‘alltoall’

moe_permute_fusion: bool#: True

moe_ffn_hidden_size: int#: 2880

moe_router_load_balancing_type: str#: ‘none’

seq_length: int#: 131072

window_size: Optional[Tuple[int, int]]#: (128, 0)

softmax_type: Literal[vanilla, off-by-one, learnable]#: ‘learnable’

activation_func: Callable#: None

glu_linear_offset: float#: 1.0

bias_activation_fusion: bool#: True

bias_dropout_fusion: bool#: False

window_attn_skip_freq: Optional[Union[int, List[int]]]#: 2

activation_func_clamp_value: Optional[float]#: 7.0

provide( pre_process=None, post_process=None, vp_stage=None, ) → megatron.core.models.gpt.GPTModel#

class bridge.models.gpt_oss.gpt_oss_provider.GPTOSSProvider120B#

Bases: bridge.models.gpt_oss.gpt_oss_provider.GPTOSSProvider

Config for GPT-OSS 120B

num_layers: int#: 36

num_moe_experts: int#: 128

class bridge.models.gpt_oss.gpt_oss_provider.GPTOSSProvider20B#