bridge.perf_recipes.qwen.vrnvl72.qwen3_moe#
VRNVL72 performance recipes for Qwen3 MoE.
Module Contents#
Functions#
Qwen3 235B A22B pretrain: 256× VRNVL72, BF16 (alias of GB300). |
|
Qwen3 235B A22B pretrain: 256× VRNVL72, FP8-MX (alias of GB300). |
|
Qwen3 235B A22B pretrain: 256× VRNVL72, NVFP4, full-iteration CG and A2A overlap. |
|
Qwen3 30B-A3B pretrain: 8× VRNVL72, BF16 (alias of GB300). |
|
Qwen3 30B-A3B pretrain: 8× VRNVL72, FP8-MX (alias of GB300). |
|
Qwen3 30B-A3B pretrain: 8× VRNVL72, NVFP4 (same layout as GB300). |
API#
- bridge.perf_recipes.qwen.vrnvl72.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_vrnvl72_bf16_config() megatron.bridge.perf_recipes.qwen.common.ConfigContainer#
Qwen3 235B A22B pretrain: 256× VRNVL72, BF16 (alias of GB300).
- bridge.perf_recipes.qwen.vrnvl72.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_vrnvl72_fp8mx_config() megatron.bridge.perf_recipes.qwen.common.ConfigContainer#
Qwen3 235B A22B pretrain: 256× VRNVL72, FP8-MX (alias of GB300).
- bridge.perf_recipes.qwen.vrnvl72.qwen3_moe.qwen3_235b_a22b_pretrain_256gpu_vrnvl72_nvfp4_config() megatron.bridge.perf_recipes.qwen.common.ConfigContainer#
Qwen3 235B A22B pretrain: 256× VRNVL72, NVFP4, full-iteration CG and A2A overlap.
- bridge.perf_recipes.qwen.vrnvl72.qwen3_moe.qwen3_30b_a3b_pretrain_8gpu_vrnvl72_bf16_config() megatron.bridge.perf_recipes.qwen.common.ConfigContainer#
Qwen3 30B-A3B pretrain: 8× VRNVL72, BF16 (alias of GB300).
- bridge.perf_recipes.qwen.vrnvl72.qwen3_moe.qwen3_30b_a3b_pretrain_8gpu_vrnvl72_fp8mx_config() megatron.bridge.perf_recipes.qwen.common.ConfigContainer#
Qwen3 30B-A3B pretrain: 8× VRNVL72, FP8-MX (alias of GB300).
- bridge.perf_recipes.qwen.vrnvl72.qwen3_moe.qwen3_30b_a3b_pretrain_8gpu_vrnvl72_nvfp4_config() megatron.bridge.perf_recipes.qwen.common.ConfigContainer#
Qwen3 30B-A3B pretrain: 8× VRNVL72, NVFP4 (same layout as GB300).
NVFP4’s fp4_param_gather path is incompatible with TP comm overlap, so it is disabled here.