fix(recipe): set TORCH_CPP_LOG_LEVEL=ERROR in DeepSeek V3 and Qwen3 235B bf16 recipes

src/megatron/bridge/perf_recipes/deepseek/gb200/deepseek_v3.py

@@ -61,6 +61,10 @@ def deepseek_v3_pretrain_256gpu_gb200_bf16_config() -> ConfigContainer:
         "NCCL_GRAPH_REGISTER": 0,
         "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
         "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # PyTorch 2.14 prints a C++ warning on every deprecated CUDAGraph.register_generator_state() call.
+        # TE layer-graph capture with PP>1 makes thousands per rank, and the stderr flood can stall ranks
+        # past the NCCL timeout, so keep only C++ errors.
+        "TORCH_CPP_LOG_LEVEL": "ERROR",
         # NCCL user-buffer and launch settings.
         "NCCL_NVLS_ENABLE": 0,
         # HybridEP topology for the target system.

src/megatron/bridge/perf_recipes/deepseek/gb300/deepseek_v3.py

@@ -73,6 +73,10 @@ def deepseek_v3_pretrain_256gpu_gb300_bf16_config() -> ConfigContainer:
         "NCCL_GRAPH_REGISTER": 0,
         "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
         "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # PyTorch 2.14 prints a C++ warning on every deprecated CUDAGraph.register_generator_state() call.
+        # TE layer-graph capture with PP>1 makes thousands per rank, and the stderr flood can stall ranks
+        # past the NCCL timeout, so keep only C++ errors.
+        "TORCH_CPP_LOG_LEVEL": "ERROR",
         # NCCL user-buffer and launch settings.
         "NCCL_NVLS_ENABLE": 0,
         # NCCL EP dispatcher mode and one GPU per rank.

src/megatron/bridge/perf_recipes/qwen/gb200/qwen3_moe.py

@@ -379,6 +379,10 @@ def qwen3_235b_a22b_pretrain_256gpu_gb200_bf16_config() -> ConfigContainer:
         "NCCL_GRAPH_REGISTER": 0,
         "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
         "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # PyTorch 2.14 prints a C++ warning on every deprecated CUDAGraph.register_generator_state() call.
+        # TE layer-graph capture with PP>1 makes thousands per rank, and the stderr flood can stall ranks
+        # past the NCCL timeout, so keep only C++ errors.
+        "TORCH_CPP_LOG_LEVEL": "ERROR",
         # NCCL user-buffer and launch settings.
         "NCCL_NVLS_ENABLE": 0,
         # HybridEP topology for the target system.

src/megatron/bridge/perf_recipes/qwen/gb300/qwen3_moe.py

@@ -418,6 +418,10 @@ def qwen3_235b_a22b_pretrain_256gpu_gb300_bf16_config() -> ConfigContainer:
         "NCCL_GRAPH_REGISTER": 0,
         "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
         "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # PyTorch 2.14 prints a C++ warning on every deprecated CUDAGraph.register_generator_state() call.
+        # TE layer-graph capture with PP>1 makes thousands per rank, and the stderr flood can stall ranks
+        # past the NCCL timeout, so keep only C++ errors.
+        "TORCH_CPP_LOG_LEVEL": "ERROR",
         # NCCL user-buffer and launch settings.
         "NCCL_NVLS_ENABLE": 0,
         # NCCL EP dispatcher mode and one GPU per rank.

tests/unit_tests/recipes/test_perf_recipe_environment.py

@@ -274,3 +274,19 @@ def test_representative_recipe_specific_environment_is_visible(relative_path, fu
     environment = _explicit_environment(_RECIPE_ROOT / relative_path, function_name)
 
     assert environment.items() >= expected.items()
+
+
+@pytest.mark.parametrize(
+    ("relative_path", "function_name"),
+    [
+        ("deepseek/gb200/deepseek_v3.py", "deepseek_v3_pretrain_256gpu_gb200_bf16_config"),
+        ("deepseek/gb300/deepseek_v3.py", "deepseek_v3_pretrain_256gpu_gb300_bf16_config"),
+        ("qwen/gb200/qwen3_moe.py", "qwen3_235b_a22b_pretrain_256gpu_gb200_bf16_config"),
+        ("qwen/gb300/qwen3_moe.py", "qwen3_235b_a22b_pretrain_256gpu_gb300_bf16_config"),
+    ],
+)
+def test_te_layer_graph_pipeline_recipes_keep_only_cpp_errors(relative_path, function_name):
+    """TE layer-graph capture with PP>1 floods stderr with PyTorch 2.14 C++ deprecation warnings."""
+    environment = _explicit_environment(_RECIPE_ROOT / relative_path, function_name)
+
+    assert environment["TORCH_CPP_LOG_LEVEL"] == "ERROR"