bridge.perf_recipes.qwen.h100.qwen3_moe

H100 performance recipes for Qwen3 MoE.

Module Contents

Functions

qwen3_30b_a3b_pretrain_16gpu_h100_bf16_config	Qwen3 30B-A3B pretrain: 16× H100, BF16, EP=16.
qwen3_30b_a3b_pretrain_16gpu_h100_fp8cs_config	Qwen3 30B-A3B pretrain: 16× H100, FP8 current-scaling, EP=16.
qwen3_235b_a22b_pretrain_256gpu_h100_bf16_config	Qwen3 235B-A22B pretrain: 256× H100, BF16, TP=2 PP=8 VP=4 EP=32.
qwen3_235b_a22b_pretrain_256gpu_h100_fp8cs_config	Qwen3 235B-A22B pretrain: 256× H100, FP8 current-scaling, TP=2 PP=8 VP=4 EP=32.
qwen3_235b_a22b_pretrain_256gpu_h100_fp8cs_large_scale_config	Qwen3 235B A22B pretrain: 256× H100, FP8-CS, large-scale proxy (GBS=512).
qwen3_30b_a3b_pretrain_64gpu_h100_bf16_config	Qwen3 30B-A3B pretrain: 64× H100, BF16, legacy-scaled GBS.
qwen3_30b_a3b_pretrain_64gpu_h100_fp8cs_config	Qwen3 30B-A3B pretrain: 64× H100, FP8 current-scaling, legacy-scaled GBS.
qwen3_next_80b_a3b_pretrain_128gpu_h100_bf16_config	Qwen3 Next 80B-A3B pretrain: 128× H100, BF16, EP=128, MBS=1.
qwen3_next_80b_a3b_pretrain_128gpu_h100_fp8cs_config	Qwen3 Next 80B-A3B pretrain: 128× H100, FP8-CS (same layout as BF16).

API

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_30b_a3b_pretrain_16gpu_h100_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 16× H100, BF16, EP=16.

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_30b_a3b_pretrain_16gpu_h100_fp8cs_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 16× H100, FP8 current-scaling, EP=16.

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_h100_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B-A22B pretrain: 256× H100, BF16, TP=2 PP=8 VP=4 EP=32.

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_h100_fp8cs_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B-A22B pretrain: 256× H100, FP8 current-scaling, TP=2 PP=8 VP=4 EP=32.

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_h100_fp8cs_large_scale_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 235B A22B pretrain: 256× H100, FP8-CS, large-scale proxy (GBS=512).

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_30b_a3b_pretrain_64gpu_h100_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 64× H100, BF16, legacy-scaled GBS.

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_30b_a3b_pretrain_64gpu_h100_fp8cs_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 30B-A3B pretrain: 64× H100, FP8 current-scaling, legacy-scaled GBS.

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_next_80b_a3b_pretrain_128gpu_h100_bf16_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 Next 80B-A3B pretrain: 128× H100, BF16, EP=128, MBS=1.

bridge.perf_recipes.qwen.h100.qwen3_moe.qwen3_next_80b_a3b_pretrain_128gpu_h100_fp8cs_config() → megatron.bridge.perf_recipes.qwen.common.ConfigContainer

Qwen3 Next 80B-A3B pretrain: 128× H100, FP8-CS (same layout as BF16).