bridge.perf_recipes.deepseek.gb300.deepseek_v3

GB300 performance recipes for DeepSeek V3.

Module Contents

Functions

_build_deepseek_v3_gb300_bf16	Shared HybridEP BF16 base for the DeepSeek V3 256-GPU GB300 recipe and its VRNVL72 alias.
deepseek_v3_pretrain_256gpu_gb300_bf16_config	DeepSeek V3 pretrain: 256× GB300, BF16, NCCL EP.
deepseek_v3_pretrain_256gpu_gb300_fp8cs_config	DeepSeek V3 pretrain: 256× GB300, FP8 current-scaling.
_build_deepseek_v3_gb300_fp8mx	Shared HybridEP MXFP8 base for the DeepSeek V3 256-GPU GB300 recipe and its VRNVL72 alias.
deepseek_v3_pretrain_256gpu_gb300_fp8mx_config	DeepSeek V3 pretrain: 256× GB300, MXFP8, NCCL EP.
deepseek_v3_pretrain_256gpu_gb300_fp8mx_deterministic_config	DeepSeek V3 pretrain: 256× GB300, MXFP8, deterministic.
_build_deepseek_v3_gb300_nvfp4	Shared HybridEP NVFP4 base for the DeepSeek V3 256-GPU GB300 recipe and its VRNVL72 alias.
deepseek_v3_pretrain_256gpu_gb300_nvfp4_config	DeepSeek V3 pretrain: 256× GB300, NVFP4, NCCL EP.
deepseek_v3_pretrain_64gpu_gb300_bf16_fsdp_config	DeepSeek V3 pretrain: 64× GB300, BF16, Megatron FSDP.
deepseek_v3_pretrain_64gpu_gb300_fp8mx_fsdp_config	DeepSeek V3 pretrain: 64× GB300, MXFP8, Megatron FSDP.
deepseek_v3_pretrain_256gpu_gb300_fp8mx_hsdp_config	DeepSeek V3 pretrain: 256× GB300, MXFP8, Megatron Hybrid FSDP (HSDP).
deepseek_v3_pretrain_256gpu_gb300_fp8mx_large_scale_config	DeepSeek V3 pretrain: 256× GB300, MXFP8, large-scale proxy (PP=4, VP=4, EP=64).
deepseek_v3_pretrain_256gpu_gb300_fp8mx_partial_cg_dev_config	DeepSeek V3 pretrain: 256× GB300, MXFP8 and scoped CUDA graphs.

API

bridge.perf_recipes.deepseek.gb300.deepseek_v3._build_deepseek_v3_gb300_bf16() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

Shared HybridEP BF16 base for the DeepSeek V3 256-GPU GB300 recipe and its VRNVL72 alias.

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_256gpu_gb300_bf16_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 256× GB300, BF16, NCCL EP.

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_256gpu_gb300_fp8cs_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 256× GB300, FP8 current-scaling.

bridge.perf_recipes.deepseek.gb300.deepseek_v3._build_deepseek_v3_gb300_fp8mx() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

Shared HybridEP MXFP8 base for the DeepSeek V3 256-GPU GB300 recipe and its VRNVL72 alias.

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_256gpu_gb300_fp8mx_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 256× GB300, MXFP8, NCCL EP.

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_256gpu_gb300_fp8mx_deterministic_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 256× GB300, MXFP8, deterministic.

Wraps :func:deepseek_v3_pretrain_256gpu_gb300_fp8mx_config and applies

Func:

~megatron.bridge.recipes.utils.determinism_utils.apply_determinism_overrides.

bridge.perf_recipes.deepseek.gb300.deepseek_v3._build_deepseek_v3_gb300_nvfp4() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

Shared HybridEP NVFP4 base for the DeepSeek V3 256-GPU GB300 recipe and its VRNVL72 alias.

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_256gpu_gb300_nvfp4_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 256× GB300, NVFP4, NCCL EP.

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_64gpu_gb300_bf16_fsdp_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 64× GB300, BF16, Megatron FSDP.

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_64gpu_gb300_fp8mx_fsdp_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 64× GB300, MXFP8, Megatron FSDP.

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_256gpu_gb300_fp8mx_hsdp_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 256× GB300, MXFP8, Megatron Hybrid FSDP (HSDP).

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_256gpu_gb300_fp8mx_large_scale_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 256× GB300, MXFP8, large-scale proxy (PP=4, VP=4, EP=64).

bridge.perf_recipes.deepseek.gb300.deepseek_v3.deepseek_v3_pretrain_256gpu_gb300_fp8mx_partial_cg_dev_config() → megatron.bridge.perf_recipes.deepseek.common.ConfigContainer

DeepSeek V3 pretrain: 256× GB300, MXFP8 and scoped CUDA graphs.