diff --git a/scripts/performance/configs/deepseek/deepseek_llm_pretrain.py b/scripts/performance/configs/deepseek/deepseek_llm_pretrain.py
index d30c9b2f14..4c4235c7d2 100644
--- a/scripts/performance/configs/deepseek/deepseek_llm_pretrain.py
+++ b/scripts/performance/configs/deepseek/deepseek_llm_pretrain.py
@@ -210,7 +210,11 @@ def deepseek_v3_pretrain_config_h100(
     cfg.model.pipeline_model_parallel_size = base_cfg.pipeline_model_parallel_size
     cfg.model.virtual_pipeline_model_parallel_size = base_cfg.virtual_pipeline_model_parallel_size
     cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
-    cfg.model.pipeline_model_parallel_layout = "Et|(tt|)*30mL"
+    if base_cfg.pp_layout:
+        cfg.model.pipeline_model_parallel_layout = base_cfg.pp_layout
+    else:
+        # Recompute layout based on updated PP/VP sizes
+        set_deepseek_v3_pipeline_model_parallel_layout(cfg.model)
 
     set_deepseek_v3_common_configs(cfg)
     set_workload_base_configs(cfg, base_cfg)
diff --git a/scripts/performance/configs/deepseek/deepseek_workload_base_configs.py b/scripts/performance/configs/deepseek/deepseek_workload_base_configs.py
index 6de05c2eba..ca06801718 100644
--- a/scripts/performance/configs/deepseek/deepseek_workload_base_configs.py
+++ b/scripts/performance/configs/deepseek/deepseek_workload_base_configs.py
@@ -120,7 +120,7 @@
 DEEPSEEK_V3_PRETRAIN_CONFIG_H100_V1 = replace(
     BASE_DEEPSEEK_V3_CONFIG,
     num_gpus=1024,
-    tensor_model_parallel_size=4,  # TODO: TP=2 is OOM. Resolve it and revert it to recover perf
+    tensor_model_parallel_size=2,
     pipeline_model_parallel_size=8,
     virtual_pipeline_model_parallel_size=4,
     expert_model_parallel_size=64,
@@ -128,6 +128,7 @@
     recompute_modules=["mla_up_proj", "mlp"],
     moe_flex_dispatcher_backend="hybridep",
     moe_a2a_overlap=False,
+    pp_layout="Et|(tt|)*30mL",
 )
 DEEPSEEK_V3_PRETRAIN_CONFIG_H100_BF16_V1 = DEEPSEEK_V3_PRETRAIN_CONFIG_H100_V1
 DEEPSEEK_V3_PRETRAIN_CONFIG_H100_FP8_CS_V1 = DEEPSEEK_V3_PRETRAIN_CONFIG_H100_V1
@@ -189,7 +190,11 @@
 )
 DEEPSEEK_V3_PRETRAIN_CONFIG_H100_BF16_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_H100_V2
 DEEPSEEK_V3_PRETRAIN_CONFIG_H100_FP8_CS_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_H100_V2
-DEEPSEEK_V3_PRETRAIN_CONFIG_H100_FP8_SC_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_H100_FP8_CS_V2
+DEEPSEEK_V3_PRETRAIN_CONFIG_H100_FP8_SC_V2 = replace(
+    DEEPSEEK_V3_PRETRAIN_CONFIG_H100_FP8_CS_V2,
+    virtual_pipeline_model_parallel_size=2,
+    pp_layout=None,
+)
 
 
 # =============================================================================
diff --git a/scripts/performance/perf_plugins.py b/scripts/performance/perf_plugins.py
index 40b3701419..84ffe3a970 100644
--- a/scripts/performance/perf_plugins.py
+++ b/scripts/performance/perf_plugins.py
@@ -267,6 +267,14 @@ def _set_model_specific_environment_variables(
         ):
             if compute_dtype in ["fp8_cs", "fp8_mx"]:
                 executor.env_vars["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+        elif (
+            model_family_name in ["deepseek"]
+            and model_recipe_name in ["deepseek_v3"]
+            and train_task == "pretrain"
+            and gpu in ["h100"]
+        ):
+            executor.env_vars["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+
         del_cudnn_ln = True
         if gpu in ["h100"]:
             if model_family_name == "llama" and model_recipe_name == "llama3_8b" and train_task == "pretrain":