From e65113fb14b07eca6210a66d342ac3900ffc8a59 Mon Sep 17 00:00:00 2001 From: Malay Nagda Date: Mon, 9 Feb 2026 22:38:21 +0530 Subject: [PATCH 1/2] qwen gbs 2x Signed-off-by: Malay Nagda --- scripts/performance/configs/qwen/qwen3_workload_base_configs.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/scripts/performance/configs/qwen/qwen3_workload_base_configs.py b/scripts/performance/configs/qwen/qwen3_workload_base_configs.py index 0cd9f66c8b..33bc08a644 100644 --- a/scripts/performance/configs/qwen/qwen3_workload_base_configs.py +++ b/scripts/performance/configs/qwen/qwen3_workload_base_configs.py @@ -408,6 +408,7 @@ QWEN3_30B_A3B_PRETRAIN_CONFIG_H100_BF16_V1 = replace( BASE_QWEN3_30B_A3B_CONFIG, num_gpus=16, + global_batch_size=1024, pipeline_model_parallel_size=2, virtual_pipeline_model_parallel_size=12, moe_a2a_overlap=True, @@ -419,6 +420,7 @@ QWEN3_30B_A3B_PRETRAIN_CONFIG_H100_FP8_CS_V1 = replace( BASE_QWEN3_30B_A3B_CONFIG, num_gpus=16, + global_batch_size=1024, pipeline_model_parallel_size=2, virtual_pipeline_model_parallel_size=12, moe_a2a_overlap=True, From 141bf81345a47e1b433d96f1ca65fd76708fafa9 Mon Sep 17 00:00:00 2001 From: Malay Nagda Date: Thu, 12 Feb 2026 15:29:46 +0530 Subject: [PATCH 2/2] add valid flex dispatcher Signed-off-by: Malay Nagda --- scripts/performance/configs/qwen/qwen3_workload_base_configs.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/scripts/performance/configs/qwen/qwen3_workload_base_configs.py b/scripts/performance/configs/qwen/qwen3_workload_base_configs.py index 33bc08a644..b18f0f8b2b 100644 --- a/scripts/performance/configs/qwen/qwen3_workload_base_configs.py +++ b/scripts/performance/configs/qwen/qwen3_workload_base_configs.py @@ -414,6 +414,7 @@ moe_a2a_overlap=True, cuda_graph_impl="transformer_engine", cuda_graph_scope=["moe_router", "moe_preprocess"], + moe_flex_dispatcher_backend="deepep", ) @@ -424,6 +425,7 @@ pipeline_model_parallel_size=2, virtual_pipeline_model_parallel_size=12, moe_a2a_overlap=True, + moe_flex_dispatcher_backend="deepep", )