bridge.perf_recipes.nemotronh.common

Common helpers for nemotronh performance recipes.

Module Contents

Functions

_apply_nemotron_3_nano_perf_defaults	Apply the canonical Nemotron 3 Nano performance workload defaults.
_nemotron_3_super_nvfp4_precision	Return the NVFP4 precision config used by Nemotron 3 Super perf recipes.
_nemotron_3_ultra_nvfp4_precision	Return the NVFP4 precision config used by Nemotron 3 Ultra perf recipes.
_apply_nemotron_3_super_perf_defaults	Apply shared Nemotron 3 Super perf defaults after recipe-specific overrides.
_enable_nemotron_3_super_full_iteration	Enable the HybridEP full-iteration stack for Nemotron 3 Super.
_enable_nemotron_3_5_lightning_full_iteration	Enable full-iteration capture with static MoE buffers for Lightning.
_apply_nemotron_3_ultra_perf_defaults	Apply shared Nemotron 3 Ultra perf defaults after recipe-specific overrides.
_apply_nemotron_3_ultra_gtp	Apply rack-local dense GTP settings for Nemotron 3 Ultra on GB300.
_apply_nemotron_3_ultra_fsdp_hsdp	Apply Megatron-FSDP (HSDP) settings for Nemotron 3 Ultra on GB300.

Data

_TE_QUANT_CFG_PATH
_GB300_NVLINK_DOMAIN_GPUS

API

bridge.perf_recipes.nemotronh.common._TE_QUANT_CFG_PATH

‘with_name(…)’

bridge.perf_recipes.nemotronh.common._GB300_NVLINK_DOMAIN_GPUS

64

bridge.perf_recipes.nemotronh.common._apply_nemotron_3_nano_perf_defaults(

cfg: megatron.bridge.training.config.ConfigContainer,

) → None

Apply the canonical Nemotron 3 Nano performance workload defaults.

bridge.perf_recipes.nemotronh.common._nemotron_3_super_nvfp4_precision() → megatron.bridge.training.mixed_precision.MixedPrecisionConfig

Return the NVFP4 precision config used by Nemotron 3 Super perf recipes.

bridge.perf_recipes.nemotronh.common._nemotron_3_ultra_nvfp4_precision() → megatron.bridge.training.mixed_precision.MixedPrecisionConfig

Return the NVFP4 precision config used by Nemotron 3 Ultra perf recipes.

bridge.perf_recipes.nemotronh.common._apply_nemotron_3_super_perf_defaults(

cfg: megatron.bridge.training.config.ConfigContainer,

) → None

Apply shared Nemotron 3 Super perf defaults after recipe-specific overrides.

bridge.perf_recipes.nemotronh.common._enable_nemotron_3_super_full_iteration(

cfg: megatron.bridge.training.config.ConfigContainer,

) → None

Enable the HybridEP full-iteration stack for Nemotron 3 Super.

bridge.perf_recipes.nemotronh.common._enable_nemotron_3_5_lightning_full_iteration(

cfg: megatron.bridge.training.config.ConfigContainer,

) → None

Enable full-iteration capture with static MoE buffers for Lightning.

bridge.perf_recipes.nemotronh.common._apply_nemotron_3_ultra_perf_defaults(

cfg: megatron.bridge.training.config.ConfigContainer,

) → None

Apply shared Nemotron 3 Ultra perf defaults after recipe-specific overrides.

bridge.perf_recipes.nemotronh.common._apply_nemotron_3_ultra_gtp(

cfg: megatron.bridge.training.config.ConfigContainer,

) → None

Apply rack-local dense GTP settings for Nemotron 3 Ultra on GB300.

Dense weights are sharded across one 64-GPU NVLink domain. Routed-expert weights remain unsharded by GTP because EP64 already spans that domain; this avoids expert-weight all-gathers across racks.

bridge.perf_recipes.nemotronh.common._apply_nemotron_3_ultra_fsdp_hsdp(

cfg: megatron.bridge.training.config.ConfigContainer,

num_gpus: int,

) → None

Apply Megatron-FSDP (HSDP) settings for Nemotron 3 Ultra on GB300.

Shards params/grads/optimizer within each NVLink domain and replicate (optimizer-sharded) across domains, with BF16 gradient comm, FP32 main params, and BF16 main grads. Applied last so it wins over the generic perf defaults.