`bridge.models.sarvam.sarvam_provider`#

Module Contents#

Classes#

`SarvamMoEModelProvider`	Sarvam 30B model provider.
`SarvamMLAModelProvider`	Sarvam 105B model provider.

Data#

logger

API#

bridge.models.sarvam.sarvam_provider.logger#: ‘getLogger(…)’

class bridge.models.sarvam.sarvam_provider.SarvamMoEModelProvider#

Bases: megatron.bridge.models.gpt_provider.GPTModelProvider

Sarvam 30B model provider.

transformer_layer_spec: Union[megatron.core.transformer.ModuleSpec, Callable[[megatron.bridge.models.gpt_provider.GPTModelProvider], megatron.core.transformer.ModuleSpec]]#: ‘partial(…)’

normalization: str#: ‘RMSNorm’

activation_func: Callable#: None

gated_linear_unit: bool#: True

position_embedding_type: str#: ‘rope’

add_bias_linear: bool#: False

share_embeddings_and_output_weights: bool#: False

make_vocab_size_divisible_by: int#: 128

add_qkv_bias: bool#: False

qk_layernorm: bool#: True

init_method_std: float#: 0.006

hidden_dropout: float#: 0.0

attention_dropout: float#: 0.0

layernorm_epsilon: float#: 1e-06

moe_aux_loss_coeff: float#: 0

moe_router_pre_softmax: bool#: True

moe_router_enable_expert_bias: bool#: True

moe_router_bias_update_rate: float#: 0.001

moe_grouped_gemm: bool#: True

moe_permute_fusion: bool#: True

moe_router_topk_scaling_factor: float#: 2.5

moe_shared_expert_overlap: bool#: False

moe_router_dtype: Optional[str]#: ‘fp32’

moe_router_score_function: str#: ‘sigmoid’

moe_token_dispatcher_type: str#: ‘alltoall’

attention_softmax_in_fp32: bool#: True

persist_layer_norm: bool#: True

cross_entropy_fusion_impl: str#: ‘te’

cp_comm_type: str#: ‘p2p’

recompute_granularity: str#: ‘selective’

recompute_modules: List[str]#: ‘field(…)’

kv_channels: Optional[int]#: 64

seq_length: int#: 131072

rotary_base: float#: 8000000.0

vocab_size: int#: 262144

num_moe_experts: int#: 128

moe_router_topk: int#: 6

num_layers: int#: 19

hidden_size: int#: 4096

num_attention_heads: int#: 64

ffn_hidden_size: int#: 8192

moe_ffn_hidden_size: int#: 1024

moe_shared_expert_intermediate_size: int#: 1024

moe_layer_freq: Union[int, List[int]]#: ‘field(…)’

bf16: bool#: True

num_query_groups: int#: 4

class bridge.models.sarvam.sarvam_provider.SarvamMLAModelProvider#

Bases: megatron.bridge.models.transformer_config.MLATransformerConfig, megatron.bridge.models.gpt_provider.GPTModelProvider

Sarvam 105B model provider.

transformer_layer_spec: Union[megatron.core.transformer.ModuleSpec, Callable[[megatron.bridge.models.gpt_provider.GPTModelProvider], megatron.core.transformer.ModuleSpec]]#: ‘partial(…)’

normalization: str#: ‘RMSNorm’

activation_func: Callable#: None

gated_linear_unit: bool#: True

position_embedding_type: str#: ‘rope’

add_bias_linear: bool#: False

share_embeddings_and_output_weights: bool#: False

make_vocab_size_divisible_by: int#: 128

add_qkv_bias: bool#: False

qk_layernorm: bool#: True

init_method_std: float#: 0.006

hidden_dropout: float#: 0.0

attention_dropout: float#: 0.0

layernorm_epsilon: float#: 1e-06

moe_aux_loss_coeff: float#: 0

moe_router_pre_softmax: bool#: True

moe_router_enable_expert_bias: bool#: True

moe_router_bias_update_rate: float#: 0.001

moe_grouped_gemm: bool#: True

moe_permute_fusion: bool#: True

moe_router_topk_scaling_factor: float#: 2.5

moe_shared_expert_overlap: bool#: False

moe_router_dtype: Optional[str]#: ‘fp32’

moe_router_score_function: str#: ‘sigmoid’

moe_token_dispatcher_type: str#: ‘alltoall’

attention_softmax_in_fp32: bool#: True

persist_layer_norm: bool#: True

cross_entropy_fusion_impl: str#: ‘te’

cp_comm_type: str#: ‘p2p’

recompute_granularity: str#: ‘selective’

recompute_modules: List[str]#: ‘field(…)’

multi_latent_attention: bool#: True

rope_type: str#: ‘yarn’

rotary_scaling_factor: float#: 40

original_max_position_embeddings: int#: 4096

beta_fast: float#: 32.0

beta_slow: float#: 1.0

mscale: float#: 1.0

mscale_all_dim: float#: 1.0

kv_channels: Optional[int]#: 64

seq_length: int#: 131072

rotary_base: float#: 10000.0

vocab_size: int#: 262144

num_moe_experts: int#: 128

moe_router_topk: int#: 8

num_layers: int#: 32

hidden_size: int#: 4096

num_attention_heads: int#: 64

ffn_hidden_size: int#: 16384

moe_ffn_hidden_size: int#: 2048

moe_shared_expert_intermediate_size: int#: 2048

moe_layer_freq: Union[int, List[int]]#: ‘field(…)’

bf16: bool#: True

q_lora_rank: Optional[int]#: None

kv_lora_rank: int#: 512

qk_head_dim: int#: 128

qk_pos_emb_head_dim: int#: 64

v_head_dim: int#: 128

bridge.models.sarvam.sarvam_provider#

Module Contents#

Classes#

Data#

API#

`bridge.models.sarvam.sarvam_provider`#