fix(perf): select supported DPA recipe for DeepSeek proxy

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

Author: malay-nagda

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

@@ -284,6 +284,8 @@ def deepseek_v3_pretrain_8gpu_gb300_nvfp4_config() -> ConfigContainer:
         "NVTE_ALLOW_NONDETERMINISTIC_ALGO": 0,
         # NVFP4 fast-math path.
         "NVTE_USE_FAST_MATH": 1,
+        # Use a supported FP8 attention recipe under the NVFP4 linear-layer recipe.
+        "NVTE_DPA_FP8_RECIPE": "Float8CurrentScaling",
         "NVTE_CUTEDSL_FUSED_GROUPED_MLP": 1,
     }
     return cfg

tests/unit_tests/recipes/test_r050_perf_recipe_ports.py

@@ -161,6 +161,13 @@ def test_deepseek_v3_8gpu_gb300_nvfp4_proxy_has_valid_topology() -> None:
     assert cfg.env_vars["NUM_OF_HYBRID_EP_RANKS_PER_NVLINK_DOMAIN"] == 4
 
 
+def test_deepseek_v3_8gpu_gb300_nvfp4_uses_fp8_current_scaling_for_dpa() -> None:
+    cfg = deepseek_v3_pretrain_8gpu_gb300_nvfp4_config()
+
+    assert cfg.mixed_precision.fp8_dot_product_attention is True
+    assert cfg.env_vars["NVTE_DPA_FP8_RECIPE"] == "Float8CurrentScaling"
+
+
 def test_deepseek_v3_b300_mxfp8_preserves_r050_hybridep_settings() -> None:
     cfg = deepseek_v3_pretrain_256gpu_b300_fp8mx_config()