bridge.perf_recipes.qwen.gb200.qwen3_moe

GB200 performance recipes for Qwen3 MoE.

Module Contents

Functions

qwen3_235b_a22b_pretrain_64gpu_gb200_bf16_config	Qwen3 235B-A22B pretrain: 64× GB200, BF16, PP=8 EP=8.
qwen3_235b_a22b_pretrain_64gpu_gb200_fp8cs_config	Qwen3 235B-A22B pretrain: 64× GB200, FP8 current-scaling, PP=8 EP=8.
qwen3_235b_a22b_pretrain_64gpu_gb200_fp8mx_config	Qwen3 235B-A22B pretrain: 64× GB200, MXFP8, PP=8 EP=8.
qwen3_30b_a3b_pretrain_8gpu_gb200_bf16_config	Qwen3 30B-A3B pretrain: 8× GB200, BF16, EP=8.
qwen3_30b_a3b_pretrain_8gpu_gb200_fp8cs_config	Qwen3 30B-A3B pretrain: 8× GB200, FP8 current-scaling, EP=8.
qwen3_30b_a3b_pretrain_8gpu_gb200_fp8mx_config	Qwen3 30B-A3B pretrain: 8× GB200, MXFP8, EP=8.
qwen3_235b_a22b_pretrain_256gpu_gb200_bf16_config	Qwen3 235B-A22B pretrain: 256× GB200, BF16, PP=8 EP=8.
qwen3_235b_a22b_pretrain_256gpu_gb200_fp8cs_config	Qwen3 235B-A22B pretrain: 256× GB200, FP8 current-scaling, PP=8 EP=32.
qwen3_235b_a22b_pretrain_256gpu_gb200_fp8mx_config	Qwen3 235B-A22B pretrain: 256× GB200, MXFP8, PP=8 EP=32.
qwen3_235b_a22b_pretrain_256gpu_gb200_fp8mx_large_scale_config	Qwen3 235B A22B pretrain: 256× GB200, FP8-MX, large-scale proxy (GBS=512).
qwen3_235b_a22b_pretrain_64gpu_gb200_nvfp4_config	Qwen3 235B A22B pretrain: 64× GB200, NVFP4 (same layout as FP8-CS).
qwen3_235b_a22b_pretrain_256gpu_gb200_nvfp4_config	Qwen3 235B A22B pretrain: 256× GB200, NVFP4 (same layout as FP8-CS).
qwen3_30b_a3b_pretrain_32gpu_gb200_bf16_config	Qwen3 30B-A3B pretrain: 32× GB200, BF16, legacy-scaled GBS.
qwen3_30b_a3b_pretrain_32gpu_gb200_fp8cs_config	Qwen3 30B-A3B pretrain: 32× GB200, FP8 current-scaling, legacy-scaled GBS.
qwen3_next_80b_a3b_pretrain_64gpu_gb200_bf16_config	Qwen3 Next 80B-A3B pretrain: 64× GB200, BF16 (GB300 layout, MBS=2).
qwen3_next_80b_a3b_pretrain_64gpu_gb200_fp8mx_config	Qwen3 Next 80B-A3B pretrain: 64× GB200, MXFP8 (same layout as BF16).

API

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_64gpu_gb200_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B-A22B pretrain: 64× GB200, BF16, PP=8 EP=8.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_64gpu_gb200_fp8cs_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B-A22B pretrain: 64× GB200, FP8 current-scaling, PP=8 EP=8.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_64gpu_gb200_fp8mx_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B-A22B pretrain: 64× GB200, MXFP8, PP=8 EP=8.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_30b_a3b_pretrain_8gpu_gb200_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 8× GB200, BF16, EP=8.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_30b_a3b_pretrain_8gpu_gb200_fp8cs_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 8× GB200, FP8 current-scaling, EP=8.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_30b_a3b_pretrain_8gpu_gb200_fp8mx_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 8× GB200, MXFP8, EP=8.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_gb200_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B-A22B pretrain: 256× GB200, BF16, PP=8 EP=8.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_gb200_fp8cs_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B-A22B pretrain: 256× GB200, FP8 current-scaling, PP=8 EP=32.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_gb200_fp8mx_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B-A22B pretrain: 256× GB200, MXFP8, PP=8 EP=32.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_gb200_fp8mx_large_scale_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B A22B pretrain: 256× GB200, FP8-MX, large-scale proxy (GBS=512).

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_64gpu_gb200_nvfp4_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B A22B pretrain: 64× GB200, NVFP4 (same layout as FP8-CS).

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_gb200_nvfp4_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B A22B pretrain: 256× GB200, NVFP4 (same layout as FP8-CS).

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_30b_a3b_pretrain_32gpu_gb200_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 32× GB200, BF16, legacy-scaled GBS.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_30b_a3b_pretrain_32gpu_gb200_fp8cs_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 32× GB200, FP8 current-scaling, legacy-scaled GBS.

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_next_80b_a3b_pretrain_64gpu_gb200_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 Next 80B-A3B pretrain: 64× GB200, BF16 (GB300 layout, MBS=2).

bridge.perf_recipes.qwen.gb200.qwen3_moe.qwen3_next_80b_a3b_pretrain_64gpu_gb200_fp8mx_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 Next 80B-A3B pretrain: 64× GB200, MXFP8 (same layout as BF16).