bridge.perf_recipes.deepseek.b200.deepseek_v3#
B200 performance recipes for DeepSeek V3.
Module Contents#
Functions#
DeepSeek V3 pretrain: 256× B200, BF16. |
|
DeepSeek V3 pretrain: 256× B200, FP8 current-scaling. |
|
DeepSeek V3 pretrain: 256× B200, MXFP8. |
|
DeepSeek V3 pretrain: 256× B200, NVFP4 (same layout as BF16). |
|
DeepSeek V3 pretrain: 256× B200, MXFP8, large-scale proxy (GBS=256). |
API#
- bridge.perf_recipes.deepseek.b200.deepseek_v3.deepseek_v3_pretrain_256gpu_b200_bf16_config() megatron.bridge.perf_recipes.deepseek.common.ConfigContainer#
DeepSeek V3 pretrain: 256× B200, BF16.
- bridge.perf_recipes.deepseek.b200.deepseek_v3.deepseek_v3_pretrain_256gpu_b200_fp8cs_config() megatron.bridge.perf_recipes.deepseek.common.ConfigContainer#
DeepSeek V3 pretrain: 256× B200, FP8 current-scaling.
- bridge.perf_recipes.deepseek.b200.deepseek_v3.deepseek_v3_pretrain_256gpu_b200_fp8mx_config() megatron.bridge.perf_recipes.deepseek.common.ConfigContainer#
DeepSeek V3 pretrain: 256× B200, MXFP8.
- bridge.perf_recipes.deepseek.b200.deepseek_v3.deepseek_v3_pretrain_256gpu_b200_nvfp4_config() megatron.bridge.perf_recipes.deepseek.common.ConfigContainer#
DeepSeek V3 pretrain: 256× B200, NVFP4 (same layout as BF16).
- bridge.perf_recipes.deepseek.b200.deepseek_v3.deepseek_v3_pretrain_256gpu_b200_fp8mx_large_scale_config() megatron.bridge.perf_recipes.deepseek.common.ConfigContainer#
DeepSeek V3 pretrain: 256× B200, MXFP8, large-scale proxy (GBS=256).