core.models.hybrid.hybrid_layer_specs#

Module Contents#

Functions#

_get_gated_delta_product_mamba_layer_spec

_get_wide_residual_hybrid_stack_spec

Build a static HybridStack spec whose layers explicitly own wide connections.

Data#

API#

core.models.hybrid.hybrid_layer_specs.moe#

‘get_moe_module_spec(…)’

core.models.hybrid.hybrid_layer_specs.moe_inference#

‘get_inference_optimized_moe_spec(…)’

core.models.hybrid.hybrid_layer_specs._csa_compressor#

‘partial(…)’

core.models.hybrid.hybrid_layer_specs._csa_indexer#

‘partial(…)’

core.models.hybrid.hybrid_layer_specs._csa_qk_norm#

‘layer_norm(…)’

core.models.hybrid.hybrid_layer_specs._hybrid_mtp_block_spec#

‘ModuleSpec(…)’

core.models.hybrid.hybrid_layer_specs._get_gated_delta_product_mamba_layer_spec(in_proj, out_proj)#
core.models.hybrid.hybrid_layer_specs.hybrid_stack_spec#

‘ModuleSpec(…)’

core.models.hybrid.hybrid_layer_specs.gated_delta_product_stack_spec#

‘ModuleSpec(…)’

core.models.hybrid.hybrid_layer_specs.hybrid_inference_stack_spec#

‘ModuleSpec(…)’

core.models.hybrid.hybrid_layer_specs.gated_delta_product_inference_stack_spec#

‘ModuleSpec(…)’

core.models.hybrid.hybrid_layer_specs._get_wide_residual_hybrid_stack_spec(
stack_spec: megatron.core.transformer.spec_utils.ModuleSpec,
) → megatron.core.transformer.spec_utils.ModuleSpec#

Build a static HybridStack spec whose layers explicitly own wide connections.

core.models.hybrid.hybrid_layer_specs.wide_residual_hybrid_stack_spec#

‘_get_wide_residual_hybrid_stack_spec(…)’

core.models.hybrid.hybrid_layer_specs.wide_residual_gated_delta_product_stack_spec#

‘_get_wide_residual_hybrid_stack_spec(…)’

core.models.hybrid.hybrid_layer_specs.wide_residual_hybrid_inference_stack_spec#

‘_get_wide_residual_hybrid_stack_spec(…)’

core.models.hybrid.hybrid_layer_specs.wide_residual_gated_delta_product_inference_stack_spec#

‘_get_wide_residual_hybrid_stack_spec(…)’

core.models.hybrid.hybrid_layer_specs.mamba_stack_spec#

None

core.models.hybrid.hybrid_layer_specs.mamba_inference_stack_spec#

None

core.models.hybrid.hybrid_layer_specs.gdp_stack_spec#

None

core.models.hybrid.hybrid_layer_specs.gdp_inference_stack_spec#

None

core.models.hybrid.hybrid_layer_specs.hybrid_dsv4_stack_spec#

None