no non det algo (#6282)

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

Author: malay-nagda

src/megatron/bridge/perf_recipes/deepseek/vr200/deepseek_v3.py

@@ -328,7 +328,7 @@ def deepseek_v3_pretrain_256gpu_vr200_fp8mx_config() -> ConfigContainer:
         "NVTE_NORM_BWD_USE_CUDNN": 1,
         "NVTE_NORM_FWD_USE_CUDNN": 1,
         # Keep DeepSeek kernel selection aligned with the measured baseline.
-        "NVTE_ALLOW_NONDETERMINISTIC_ALGO": 1,
+        "NVTE_ALLOW_NONDETERMINISTIC_ALGO": 0,
     }
     return cfg
 
@@ -369,7 +369,7 @@ def deepseek_v3_pretrain_64gpu_vr200_fp8mx_proxy_config() -> ConfigContainer:
         "NVTE_FWD_LAYERNORM_SM_MARGIN": 20,
         "NVTE_NORM_BWD_USE_CUDNN": 1,
         "NVTE_NORM_FWD_USE_CUDNN": 1,
-        "NVTE_ALLOW_NONDETERMINISTIC_ALGO": 1,
+        "NVTE_ALLOW_NONDETERMINISTIC_ALGO": 0,
     }
     return cfg
 

src/megatron/bridge/perf_recipes/deepseek/vr200/deepseek_v4.py

@@ -47,7 +47,6 @@ def deepseek_v4_flash_pretrain_128gpu_vr200_fp8mx_config() -> ConfigContainer:
         "NVTE_FWD_LAYERNORM_SM_MARGIN": 20,
         "NVTE_NORM_BWD_USE_CUDNN": 1,
         "NVTE_NORM_FWD_USE_CUDNN": 1,
-        # This benchmark-only recipe permits Scaled-SwiGLU dprob atomics in the fused path.
-        "NVTE_ALLOW_NONDETERMINISTIC_ALGO": 1,
+        "NVTE_ALLOW_NONDETERMINISTIC_ALGO": 0,
     }
     return cfg

tests/unit_tests/recipes/test_deepseek_v4_flash_perf_recipes.py

@@ -144,13 +144,12 @@ def test_deepseek_v4_flash_128gpu_vr200_fp8mx_disables_clamp_for_cutedsl() -> No
     assert vr200_cfg.model.use_transformer_engine_op_fuser is True
     assert vr200_cfg.model.moe_use_grouped_tensor is True
     assert vr200_cfg.env_vars["NVTE_CUTEDSL_FUSED_GROUPED_MLP"] == 1
-    assert vr200_cfg.env_vars["NVTE_ALLOW_NONDETERMINISTIC_ALGO"] == 1
+    assert vr200_cfg.env_vars["NVTE_ALLOW_NONDETERMINISTIC_ALGO"] == 0
     assert is_full_iteration_cuda_graph(vr200_cfg.model)
 
     gb300_cfg.model.activation_func_clamp_value = None
     gb300_cfg.model.use_transformer_engine_op_fuser = True
     gb300_cfg.model.moe_use_grouped_tensor = True
-    gb300_cfg.env_vars["NVTE_ALLOW_NONDETERMINISTIC_ALGO"] = 1
     assert vars(vr200_cfg.model) == vars(gb300_cfg.model)
     for field in fields(vr200_cfg):
         if field.name != "model":

tests/unit_tests/recipes/test_perf_recipe_environment.py

@@ -223,16 +223,7 @@ def test_explicit_environment_invariants_across_all_flat_recipes():
                 # so these recipes must not pin NVTE_ALLOW_NONDETERMINISTIC_ALGO=0.
                 assert "NVTE_ALLOW_NONDETERMINISTIC_ALGO" not in environment
             else:
-                # These VR200 MXFP8 recipes explicitly opt into this path.
-                expected_nondeterminism = int(
-                    function_name
-                    in {
-                        "deepseek_v3_pretrain_256gpu_vr200_fp8mx_config",
-                        "deepseek_v3_pretrain_64gpu_vr200_fp8mx_proxy_config",
-                        "deepseek_v4_flash_pretrain_128gpu_vr200_fp8mx_config",
-                    }
-                )
-                assert environment["NVTE_ALLOW_NONDETERMINISTIC_ALGO"] == expected_nondeterminism
+                assert environment["NVTE_ALLOW_NONDETERMINISTIC_ALGO"] == 0
             if "NCCL_EP_HT_EM_PULL_PUSH" in environment:
                 # NCCL EP recipes drop the HybridEP tuning that nothing reads on that backend.
                 assert not hybrid_ep_names