bridge.perf_recipes.nemotronh.vrnvl72.nemotronh

VRNVL72 performance recipes for Nemotron 3 and 3.5.

Module Contents

Functions

nemotron_3_nano_pretrain_8gpu_vrnvl72_bf16_config	Nemotron 3 Nano pretrain: 8× VRNVL72, BF16 (alias of GB300).
nemotron_3_nano_pretrain_8gpu_vrnvl72_fp8mx_config	Nemotron 3 Nano pretrain: 8× VRNVL72, FP8-MX (alias of GB300).
nemotron_3_nano_pretrain_8gpu_vrnvl72_nvfp4_config	Nemotron 3 Nano pretrain: 8× VRNVL72, NVFP4 (alias of GB300).
nemotron_3_super_pretrain_64gpu_vrnvl72_bf16_config	Nemotron 3 Super pretrain: 64× VRNVL72, BF16 (alias of GB300).
nemotron_3_super_pretrain_64gpu_vrnvl72_fp8mx_config	Nemotron 3 Super pretrain: 64× VRNVL72, MXFP8 with full-iteration CUDA graph.
nemotron_3_super_pretrain_64gpu_vrnvl72_nvfp4_config	Nemotron 3 Super pretrain: 64× VRNVL72, NVFP4 with full-iteration CUDA graph.
nemotron_3_ultra_pretrain_256gpu_vrnvl72_fp8mx_config	Nemotron 3 Ultra pretrain: 256× VRNVL72, MXFP8 (alias of GB300).
nemotron_3_ultra_pretrain_256gpu_vrnvl72_nvfp4_config	Nemotron 3 Ultra pretrain: 256× VRNVL72, NVFP4 (alias of GB300).
nemotron_3_5_lightning_pretrain_8gpu_vrnvl72_bf16_config	Nemotron 3.5 Lightning pretrain: 8× VRNVL72, BF16 (alias of GB300).
nemotron_3_5_lightning_pretrain_8gpu_vrnvl72_fp8mx_config	Nemotron 3.5 Lightning pretrain: 8× VRNVL72, MXFP8 with full-iteration CUDA graph.

API

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_nano_pretrain_8gpu_vrnvl72_bf16_config() → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3 Nano pretrain: 8× VRNVL72, BF16 (alias of GB300).

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_nano_pretrain_8gpu_vrnvl72_fp8mx_config() → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3 Nano pretrain: 8× VRNVL72, FP8-MX (alias of GB300).

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_nano_pretrain_8gpu_vrnvl72_nvfp4_config() → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3 Nano pretrain: 8× VRNVL72, NVFP4 (alias of GB300).

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_super_pretrain_64gpu_vrnvl72_bf16_config() → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3 Super pretrain: 64× VRNVL72, BF16 (alias of GB300).

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_super_pretrain_64gpu_vrnvl72_fp8mx_config() → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3 Super pretrain: 64× VRNVL72, MXFP8 with full-iteration CUDA graph.

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_super_pretrain_64gpu_vrnvl72_nvfp4_config() → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3 Super pretrain: 64× VRNVL72, NVFP4 with full-iteration CUDA graph.

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_ultra_pretrain_256gpu_vrnvl72_fp8mx_config(

*,

num_gpus: int = 256,

) → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3 Ultra pretrain: 256× VRNVL72, MXFP8 (alias of GB300).

Parameters:

num_gpus – Launch GPU count used to size HSDP optimizer instances. Defaults to the canonical 256-GPU benchmark allocation.

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_ultra_pretrain_256gpu_vrnvl72_nvfp4_config(

*,

num_gpus: int = 256,

) → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3 Ultra pretrain: 256× VRNVL72, NVFP4 (alias of GB300).

Parameters:

num_gpus – Retained for compatibility; sharding is rack-local and the distributed-optimizer instance count is fixed at one.

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_5_lightning_pretrain_8gpu_vrnvl72_bf16_config() → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3.5 Lightning pretrain: 8× VRNVL72, BF16 (alias of GB300).

bridge.perf_recipes.nemotronh.vrnvl72.nemotronh.nemotron_3_5_lightning_pretrain_8gpu_vrnvl72_fp8mx_config() → megatron.bridge.perf_recipes.nemotronh.common.ConfigContainer

Nemotron 3.5 Lightning pretrain: 8× VRNVL72, MXFP8 with full-iteration CUDA graph.