revert dsv3, qw 30b_a3b

Signed-off-by: Malay Nagda <malayn@nvidia.com>

Author: malay-nagda

scripts/performance/configs/deepseek/deepseek_workload_base_configs.py

@@ -94,7 +94,7 @@
     pipeline_model_parallel_size=16,
     expert_model_parallel_size=8,
     global_batch_size=2048,
-    recompute_modules=["moe_act"],
+    recompute_modules=["mla_up_proj"],
     moe_flex_dispatcher_backend="hybridep",
     moe_a2a_overlap=False,
 )
@@ -173,7 +173,7 @@
     DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V2,
     pipeline_model_parallel_size=8,
     virtual_pipeline_model_parallel_size=None,
-    recompute_modules=["moe_act"],
+    recompute_modules=["mla_up_proj"],
 )
 DEEPSEEK_V3_PRETRAIN_CONFIG_B300_FP8_CS_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V2
 DEEPSEEK_V3_PRETRAIN_CONFIG_B300_FP8_MX_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_B300_FP8_CS_V2

scripts/performance/configs/qwen/qwen3_workload_base_configs.py

@@ -91,7 +91,7 @@
     global_batch_size=1024,
     moe_flex_dispatcher_backend="hybridep",
     cuda_graph_impl="transformer_engine",
-    cuda_graph_scope=["attn", "moe_router"],
+    cuda_graph_scope=["attn", "moe_router", "moe_preprocess"],
 )
 
 
@@ -103,7 +103,7 @@
     global_batch_size=1024,
     moe_flex_dispatcher_backend="hybridep",
     cuda_graph_impl="transformer_engine",
-    cuda_graph_scope=["attn", "moe_router"],
+    cuda_graph_scope=["attn", "moe_router", "moe_preprocess"],
 )