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.

_with_global_batch_size

_nemotron_3_super_nvfp4_precision

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

_apply_nemotron_3_super_perf_defaults

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

_apply_nemotron_3_ultra_perf_defaults

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

_apply_nemotron_3_ultra_fsdp_hsdp

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

Data#

API#

bridge.perf_recipes.nemotronh.common._TE_QUANT_CFG_PATH#

‘with_name(…)’

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._with_global_batch_size(
cfg: megatron.bridge.training.config.ConfigContainer,
global_batch_size: int,
) megatron.bridge.training.config.ConfigContainer#
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._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._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_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.