diff --git a/scripts/performance/utils/overrides.py b/scripts/performance/utils/overrides.py
index 903c63d026..b14323609e 100644
--- a/scripts/performance/utils/overrides.py
+++ b/scripts/performance/utils/overrides.py
@@ -67,7 +67,7 @@ def _set_common_perf_overrides(recipe: ConfigContainer) -> ConfigContainer:
     if hasattr(recipe.model, "apply_rope_fusion"):
         recipe.model.apply_rope_fusion = True
     if hasattr(recipe.model, "cross_entropy_fusion_impl"):
-        recipe.model.cross_entropy_fusion_impl = "te"
+        recipe.model.cross_entropy_fusion_impl = "native"
 
     # TODO: This needs to be adjusted when overlapping HybridEP with computation or
     # the number of SMs for HybridEP is reduced.
diff --git a/src/megatron/bridge/diffusion/recipes/nemotron_labs_diffusion/ar_to_dlm.py b/src/megatron/bridge/diffusion/recipes/nemotron_labs_diffusion/ar_to_dlm.py
index 7d99ab8c89..6247779334 100644
--- a/src/megatron/bridge/diffusion/recipes/nemotron_labs_diffusion/ar_to_dlm.py
+++ b/src/megatron/bridge/diffusion/recipes/nemotron_labs_diffusion/ar_to_dlm.py
@@ -229,7 +229,7 @@ def _nemotron_labs_diffusion_common(
         model_cfg.recompute_num_layers = 1
 
     model_cfg.cross_entropy_loss_fusion = True
-    model_cfg.cross_entropy_fusion_impl = "te"
+    model_cfg.cross_entropy_fusion_impl = "native"
 
     if lr_decay_style == "WSD":
         opt_cfg, scheduler_cfg = distributed_fused_adam_with_cosine_annealing_dllm(
diff --git a/src/megatron/bridge/models/deepseek/deepseek_v2_bridge.py b/src/megatron/bridge/models/deepseek/deepseek_v2_bridge.py
index 0a900ca287..8f09f54cf7 100644
--- a/src/megatron/bridge/models/deepseek/deepseek_v2_bridge.py
+++ b/src/megatron/bridge/models/deepseek/deepseek_v2_bridge.py
@@ -65,7 +65,7 @@ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> MLAModelProvider
         provider.gradient_accumulation_fusion = True
         provider.bias_activation_fusion = True
         provider.bias_dropout_fusion = True
-        provider.cross_entropy_fusion_impl = "te"
+        provider.cross_entropy_fusion_impl = "native"
         provider.cross_entropy_loss_fusion = True
         provider.masked_softmax_fusion = True
         provider.persist_layer_norm = True
diff --git a/src/megatron/bridge/models/deepseek/deepseek_v3_bridge.py b/src/megatron/bridge/models/deepseek/deepseek_v3_bridge.py
index 1333aee9d2..ed61cd7f5e 100644
--- a/src/megatron/bridge/models/deepseek/deepseek_v3_bridge.py
+++ b/src/megatron/bridge/models/deepseek/deepseek_v3_bridge.py
@@ -79,7 +79,7 @@ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> MLAModelProvider
         provider.gradient_accumulation_fusion = True
         provider.bias_activation_fusion = True
         provider.bias_dropout_fusion = True
-        provider.cross_entropy_fusion_impl = "te"
+        provider.cross_entropy_fusion_impl = "native"
         provider.cross_entropy_loss_fusion = True
         provider.masked_softmax_fusion = True
         provider.persist_layer_norm = True
diff --git a/src/megatron/bridge/models/deepseek/deepseek_v4_bridge.py b/src/megatron/bridge/models/deepseek/deepseek_v4_bridge.py
index 9fbd657f65..8011d7efbb 100644
--- a/src/megatron/bridge/models/deepseek/deepseek_v4_bridge.py
+++ b/src/megatron/bridge/models/deepseek/deepseek_v4_bridge.py
@@ -441,7 +441,7 @@ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> MLAModelProvider
         provider.share_embeddings_and_output_weights = bool(hf_config.tie_word_embeddings)
         provider.gradient_accumulation_fusion = True
         provider.bias_dropout_fusion = True
-        provider.cross_entropy_fusion_impl = "te"
+        provider.cross_entropy_fusion_impl = "native"
         provider.cross_entropy_loss_fusion = True
         provider.masked_softmax_fusion = True
         provider.persist_layer_norm = True
diff --git a/src/megatron/bridge/models/kimi/kimi_bridge.py b/src/megatron/bridge/models/kimi/kimi_bridge.py
index 058a026f16..c019bbccea 100755
--- a/src/megatron/bridge/models/kimi/kimi_bridge.py
+++ b/src/megatron/bridge/models/kimi/kimi_bridge.py
@@ -67,7 +67,7 @@ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> MLAModelProvider
         provider.gradient_accumulation_fusion = True
         provider.bias_activation_fusion = True
         provider.bias_dropout_fusion = True
-        provider.cross_entropy_fusion_impl = "te"
+        provider.cross_entropy_fusion_impl = "native"
         provider.cross_entropy_loss_fusion = True
         provider.masked_softmax_fusion = True
         provider.persist_layer_norm = True
diff --git a/src/megatron/bridge/models/kimi_vl/kimi_k25_vl_bridge.py b/src/megatron/bridge/models/kimi_vl/kimi_k25_vl_bridge.py
index 6087fbbba6..cdff57e555 100644
--- a/src/megatron/bridge/models/kimi_vl/kimi_k25_vl_bridge.py
+++ b/src/megatron/bridge/models/kimi_vl/kimi_k25_vl_bridge.py
@@ -105,7 +105,7 @@ def provider_bridge(self, hf_pretrained: PreTrainedVLM) -> KimiK25VLModelProvide
         provider.apply_rope_fusion = False
         provider.bias_activation_fusion = True
         provider.bias_dropout_fusion = True
-        provider.cross_entropy_fusion_impl = "te"
+        provider.cross_entropy_fusion_impl = "native"
         provider.cross_entropy_loss_fusion = True
         provider.masked_softmax_fusion = True
         provider.persist_layer_norm = True
diff --git a/src/megatron/bridge/models/sarvam/sarvam_provider.py b/src/megatron/bridge/models/sarvam/sarvam_provider.py
index 0057e6e001..c0986914ac 100644
--- a/src/megatron/bridge/models/sarvam/sarvam_provider.py
+++ b/src/megatron/bridge/models/sarvam/sarvam_provider.py
@@ -107,7 +107,7 @@ class SarvamMoEModelProvider(GPTModelProvider):
     attention_softmax_in_fp32: bool = True
     persist_layer_norm: bool = True
 
-    cross_entropy_fusion_impl: str = "te"
+    cross_entropy_fusion_impl: str = "native"
     cp_comm_type: str = "p2p"
     recompute_granularity: str = "selective"
     recompute_modules: List[str] = field(default_factory=lambda: ["layernorm", "shared_experts", "mlp", "moe_act"])
@@ -192,7 +192,7 @@ class SarvamMLAModelProvider(MLATransformerConfig, GPTModelProvider):
     attention_softmax_in_fp32: bool = True
     persist_layer_norm: bool = True
 
-    cross_entropy_fusion_impl: str = "te"
+    cross_entropy_fusion_impl: str = "native"
     cp_comm_type: str = "p2p"
     recompute_granularity: str = "selective"
     recompute_modules: List[str] = field(default_factory=lambda: ["moe"])
diff --git a/src/megatron/bridge/recipes/deepseek/deepseek_v2.py b/src/megatron/bridge/recipes/deepseek/deepseek_v2.py
index b00fe7c5b5..6af056fc09 100644
--- a/src/megatron/bridge/recipes/deepseek/deepseek_v2.py
+++ b/src/megatron/bridge/recipes/deepseek/deepseek_v2.py
@@ -84,7 +84,7 @@ def deepseek_v2_lite_pretrain_config() -> ConfigContainer:
         True  # MoE-specific: Use grouped GEMM for experts (default from DeepSeekModelProvider)
     )
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"  # Default from DeepSeekModelProvider
+    cfg.model.cross_entropy_fusion_impl = "native"  # Default from DeepSeekModelProvider
 
     # Memory saving (recompute & offloading) - ENABLED for V2-Lite
     cfg.model.recompute_granularity = "full"
@@ -202,7 +202,7 @@ def deepseek_v2_pretrain_config() -> ConfigContainer:
         True  # MoE-specific: Use grouped GEMM for experts (default from DeepSeekModelProvider)
     )
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"  # Default from DeepSeekModelProvider
+    cfg.model.cross_entropy_fusion_impl = "native"  # Default from DeepSeekModelProvider
 
     # Memory saving (recompute & offloading) - ENABLED for V2
     cfg.model.recompute_granularity = "full"
diff --git a/src/megatron/bridge/recipes/deepseek/deepseek_v3.py b/src/megatron/bridge/recipes/deepseek/deepseek_v3.py
index 8520d992a2..e112a2e098 100644
--- a/src/megatron/bridge/recipes/deepseek/deepseek_v3.py
+++ b/src/megatron/bridge/recipes/deepseek/deepseek_v3.py
@@ -170,7 +170,7 @@ def deepseek_v3_pretrain_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True  # MoE-specific: Fuse permute operations
     cfg.model.moe_grouped_gemm = True  # MoE-specific: Use grouped GEMM for experts
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"  # Default from DeepSeekModelProvider
+    cfg.model.cross_entropy_fusion_impl = "native"  # Default from DeepSeekModelProvider
 
     # Memory saving (recompute & offloading) — no recompute by default.
     # Setting granularity="selective" with modules=None would cause MCore's
@@ -326,7 +326,7 @@ def deepseek_v3_pretrain_config_32nodes() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving - FULL recompute for 32 nodes (memory efficiency)
     cfg.model.recompute_granularity = "full"
diff --git a/src/megatron/bridge/recipes/deepseek/deepseek_v4.py b/src/megatron/bridge/recipes/deepseek/deepseek_v4.py
index a8a04963ed..18956d43c0 100644
--- a/src/megatron/bridge/recipes/deepseek/deepseek_v4.py
+++ b/src/megatron/bridge/recipes/deepseek/deepseek_v4.py
@@ -121,7 +121,7 @@ def deepseek_v4_flash_pretrain_config() -> ConfigContainer:
     cfg.model.moe_aux_loss_coeff = 0.0
     cfg.model.moe_router_force_load_balancing = False
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = "selective"
     cfg.model.recompute_modules = ["moe_act", "mhc"]
diff --git a/src/megatron/bridge/recipes/kimi/kimi_k2.py b/src/megatron/bridge/recipes/kimi/kimi_k2.py
index 62fe3a7784..8aa2397598 100644
--- a/src/megatron/bridge/recipes/kimi/kimi_k2.py
+++ b/src/megatron/bridge/recipes/kimi/kimi_k2.py
@@ -146,7 +146,7 @@ def kimi_k2_pretrain_config(optimizer_type: str = "muon") -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading) - already set in model provider
     # cfg.model.recompute_granularity = "selective"
diff --git a/src/megatron/bridge/recipes/kimi_vl/kimi_k25_vl.py b/src/megatron/bridge/recipes/kimi_vl/kimi_k25_vl.py
index 69ded93599..744a3d19c0 100644
--- a/src/megatron/bridge/recipes/kimi_vl/kimi_k25_vl.py
+++ b/src/megatron/bridge/recipes/kimi_vl/kimi_k25_vl.py
@@ -135,7 +135,7 @@ def kimi_k25_vl_sft_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving
     cfg.model.fine_grained_activation_offloading = False
diff --git a/src/megatron/bridge/recipes/llama/llama3.py b/src/megatron/bridge/recipes/llama/llama3.py
index 3ca001a5ad..03d623e4a3 100644
--- a/src/megatron/bridge/recipes/llama/llama3.py
+++ b/src/megatron/bridge/recipes/llama/llama3.py
@@ -95,7 +95,7 @@ def llama32_1b_pretrain_config() -> ConfigContainer:
     # Kernel selections
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -182,7 +182,7 @@ def llama32_3b_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -268,7 +268,7 @@ def llama3_8b_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -349,7 +349,7 @@ def llama3_8b_16k_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -430,7 +430,7 @@ def llama3_8b_64k_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -511,7 +511,7 @@ def llama3_8b_128k_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -614,7 +614,7 @@ def llama3_8b_low_precision_pretrain_config(mixed_precision_recipe: str) -> Conf
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -701,7 +701,7 @@ def llama3_70b_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -804,7 +804,7 @@ def llama3_70b_16k_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -893,7 +893,7 @@ def llama3_70b_64k_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -987,7 +987,7 @@ def llama31_8b_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -1070,7 +1070,7 @@ def llama31_70b_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
@@ -1165,7 +1165,7 @@ def llama31_405b_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     cfg.model.recompute_granularity = None
     cfg.model.recompute_modules = None
diff --git a/src/megatron/bridge/recipes/moonlight/moonlight_16b.py b/src/megatron/bridge/recipes/moonlight/moonlight_16b.py
index 6a1530769f..cddbd10a12 100644
--- a/src/megatron/bridge/recipes/moonlight/moonlight_16b.py
+++ b/src/megatron/bridge/recipes/moonlight/moonlight_16b.py
@@ -125,7 +125,7 @@ def moonlight_16b_pretrain_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = "selective"
@@ -307,7 +307,7 @@ def moonlight_16b_sft_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     # recompute_granularity already set in model provider
@@ -520,7 +520,7 @@ def moonlight_16b_peft_config(
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving
     cfg.model.fine_grained_activation_offloading = False
diff --git a/src/megatron/bridge/recipes/nemotronh/nemotron_3_super.py b/src/megatron/bridge/recipes/nemotronh/nemotron_3_super.py
index 74cb6957c0..8098ed414f 100644
--- a/src/megatron/bridge/recipes/nemotronh/nemotron_3_super.py
+++ b/src/megatron/bridge/recipes/nemotronh/nemotron_3_super.py
@@ -86,7 +86,7 @@ def nemotron_3_super_pretrain_config() -> ConfigContainer:
 
     # Kernel Selections
     cfg.model.attention_backend = "fused"
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
     cfg.model.use_te_rng_tracker = True
 
     # MTP Settings (HF config has num_nextn_predict_layers=1 for the shared block;
diff --git a/src/megatron/bridge/recipes/qwen/qwen3.py b/src/megatron/bridge/recipes/qwen/qwen3.py
index fb5a49ea06..befed2a929 100644
--- a/src/megatron/bridge/recipes/qwen/qwen3.py
+++ b/src/megatron/bridge/recipes/qwen/qwen3.py
@@ -67,7 +67,7 @@ def qwen3_600m_pretrain_config() -> ConfigContainer:
     # Kernel selections
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -148,7 +148,7 @@ def qwen3_1p7b_pretrain_config() -> ConfigContainer:
     # Kernel selections
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -229,7 +229,7 @@ def qwen3_4b_pretrain_config() -> ConfigContainer:
     # Kernel selections
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -310,7 +310,7 @@ def qwen3_8b_pretrain_config() -> ConfigContainer:
     # Kernel selections
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -391,7 +391,7 @@ def qwen3_14b_pretrain_config() -> ConfigContainer:
     # Kernel selections
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -464,7 +464,7 @@ def qwen3_32b_pretrain_config() -> ConfigContainer:
 
     cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading) - ENABLED for 32B
     cfg.model.recompute_granularity = "full"
diff --git a/src/megatron/bridge/recipes/qwen/qwen3_moe.py b/src/megatron/bridge/recipes/qwen/qwen3_moe.py
index 58f1fd2ca6..7a978abab1 100644
--- a/src/megatron/bridge/recipes/qwen/qwen3_moe.py
+++ b/src/megatron/bridge/recipes/qwen/qwen3_moe.py
@@ -77,7 +77,7 @@ def qwen3_30b_a3b_pretrain_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading) - ENABLED for 30B MoE
     cfg.model.recompute_granularity = "full"
@@ -191,7 +191,7 @@ def qwen3_235b_a22b_pretrain_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     # Enable if needed for memory optimization
@@ -321,7 +321,7 @@ def qwen3_30b_a3b_sft_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -454,7 +454,7 @@ def qwen3_235b_a22b_sft_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -601,7 +601,7 @@ def qwen3_30b_a3b_peft_config(peft_scheme: str | PEFT = "lora") -> ConfigContain
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
@@ -748,7 +748,7 @@ def qwen3_235b_a22b_peft_config(peft_scheme: str | PEFT = "lora") -> ConfigConta
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = None
diff --git a/src/megatron/bridge/recipes/stepfun/step35.py b/src/megatron/bridge/recipes/stepfun/step35.py
index d4b17d87d6..95133ff7db 100644
--- a/src/megatron/bridge/recipes/stepfun/step35.py
+++ b/src/megatron/bridge/recipes/stepfun/step35.py
@@ -72,7 +72,7 @@ def step35_196b_a11b_pretrain_config() -> ConfigContainer:
     cfg.model.moe_permute_fusion = True
     cfg.model.moe_grouped_gemm = True
     cfg.model.cross_entropy_loss_fusion = True
-    cfg.model.cross_entropy_fusion_impl = "te"
+    cfg.model.cross_entropy_fusion_impl = "native"
 
     # Memory saving (recompute & offloading)
     cfg.model.recompute_granularity = "full"
diff --git a/tests/unit_tests/recipes/gemma/test_gemma2_recipes.py b/tests/unit_tests/recipes/gemma/test_gemma2_recipes.py
index 754dc15260..6d784b5257 100644
--- a/tests/unit_tests/recipes/gemma/test_gemma2_recipes.py
+++ b/tests/unit_tests/recipes/gemma/test_gemma2_recipes.py
@@ -61,7 +61,7 @@ def _safe_overrides_for(name: str) -> dict:
 
 class _FakeModelCfg:
     def __init__(self):
-        self.cross_entropy_fusion_impl = "te"
+        self.cross_entropy_fusion_impl = "native"
         self.vocab_size = 256000
         self.context_parallel_size = 1
 
diff --git a/tests/unit_tests/recipes/kimi/test_kimi_k2.py b/tests/unit_tests/recipes/kimi/test_kimi_k2.py
index 7778ff1478..c21a2fca99 100644
--- a/tests/unit_tests/recipes/kimi/test_kimi_k2.py
+++ b/tests/unit_tests/recipes/kimi/test_kimi_k2.py
@@ -225,7 +225,7 @@ def test_pretrain_config_kernel_selections(self):
 
         assert cfg.model.attention_backend is None
         assert cfg.model.cross_entropy_loss_fusion is True
-        assert cfg.model.cross_entropy_fusion_impl == "te"
+        assert cfg.model.cross_entropy_fusion_impl == "native"
 
     def test_pretrain_config_comm_overlap(self):
         """Test communication overlap configuration."""
diff --git a/tests/unit_tests/recipes/kimi_vl/test_kimi_k25_vl.py b/tests/unit_tests/recipes/kimi_vl/test_kimi_k25_vl.py
index 54017acc48..9e5c4c3776 100644
--- a/tests/unit_tests/recipes/kimi_vl/test_kimi_k25_vl.py
+++ b/tests/unit_tests/recipes/kimi_vl/test_kimi_k25_vl.py
@@ -267,12 +267,12 @@ def test_sft_config_transformer_engine_and_cuda_graph(self):
         assert cfg.model.cuda_graph_warmup_steps == 3
 
     def test_sft_config_kernel_selections(self):
-        """Default attention backend is None; cross-entropy fusion uses TE."""
+        """Default attention backend is None; cross-entropy fusion uses native."""
         cfg = kimi_k25_vl_sft_config()
 
         assert cfg.model.attention_backend is None
         assert cfg.model.cross_entropy_loss_fusion is True
-        assert cfg.model.cross_entropy_fusion_impl == "te"
+        assert cfg.model.cross_entropy_fusion_impl == "native"
 
     def test_sft_config_comm_overlap(self):
         """Comm overlap is off (TP overlap, wgrad delay, MoE EP overlap)."""
diff --git a/tests/unit_tests/recipes/qwen/test_qwen2_recipes.py b/tests/unit_tests/recipes/qwen/test_qwen2_recipes.py
index 23737a0454..b8f9e4b7d2 100644
--- a/tests/unit_tests/recipes/qwen/test_qwen2_recipes.py
+++ b/tests/unit_tests/recipes/qwen/test_qwen2_recipes.py
@@ -66,7 +66,7 @@
 
 class _FakeModelCfg:
     def __init__(self):
-        self.cross_entropy_fusion_impl = "te"
+        self.cross_entropy_fusion_impl = "native"
         self.context_parallel_size = 1
 
     def finalize(self):
diff --git a/tests/unit_tests/recipes/test_cross_entropy_defaults.py b/tests/unit_tests/recipes/test_cross_entropy_defaults.py
new file mode 100644
index 0000000000..de64cd351e
--- /dev/null
+++ b/tests/unit_tests/recipes/test_cross_entropy_defaults.py
@@ -0,0 +1,76 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import ast
+from pathlib import Path
+
+import pytest
+
+
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+_SHIPPED_DEFAULT_ROOTS = (
+    Path("src/megatron/bridge/recipes"),
+    Path("src/megatron/bridge/diffusion/recipes"),
+    Path("src/megatron/bridge/models"),
+    Path("scripts/performance"),
+)
+
+
+def _iter_shipped_python_files() -> list[Path]:
+    files: list[Path] = []
+    for root in _SHIPPED_DEFAULT_ROOTS:
+        absolute_root = _REPO_ROOT / root
+        if absolute_root.exists():
+            files.extend(sorted(absolute_root.rglob("*.py")))
+    return files
+
+
+def _is_cross_entropy_impl_target(target: ast.expr) -> bool:
+    if isinstance(target, ast.Attribute):
+        return target.attr == "cross_entropy_fusion_impl"
+    return isinstance(target, ast.Name) and target.id == "cross_entropy_fusion_impl"
+
+
+def _is_te_literal(value: ast.expr | None) -> bool:
+    return isinstance(value, ast.Constant) and value.value == "te"
+
+
+def _find_te_cross_entropy_defaults(path: Path) -> list[str]:
+    tree = ast.parse(path.read_text(encoding="utf-8"), filename=str(path))
+    failures: list[str] = []
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Assign) and _is_te_literal(node.value):
+            for target in node.targets:
+                if _is_cross_entropy_impl_target(target):
+                    failures.append(f"{path.relative_to(_REPO_ROOT)}:{node.lineno}")
+        elif isinstance(node, ast.AnnAssign) and _is_te_literal(node.value):
+            if _is_cross_entropy_impl_target(node.target):
+                failures.append(f"{path.relative_to(_REPO_ROOT)}:{node.lineno}")
+        elif isinstance(node, ast.Call):
+            for keyword in node.keywords:
+                if keyword.arg == "cross_entropy_fusion_impl" and _is_te_literal(keyword.value):
+                    failures.append(f"{path.relative_to(_REPO_ROOT)}:{node.lineno}")
+    return failures
+
+
+@pytest.mark.unit
+def test_shipped_defaults_do_not_select_te_cross_entropy_fusion() -> None:
+    """Shipped defaults should prefer native CE fusion because MCore rejects the TE path."""
+    failures: list[str] = []
+    for path in _iter_shipped_python_files():
+        failures.extend(_find_te_cross_entropy_defaults(path))
+
+    assert not failures, "Shipped defaults must use cross_entropy_fusion_impl='native' instead of 'te':\n" + "\n".join(
+        failures
+    )
diff --git a/tests/unit_tests/recipes/test_llama_recipes.py b/tests/unit_tests/recipes/test_llama_recipes.py
index 698272e0cf..02f37b1da2 100644
--- a/tests/unit_tests/recipes/test_llama_recipes.py
+++ b/tests/unit_tests/recipes/test_llama_recipes.py
@@ -82,7 +82,7 @@ def _safe_overrides_for(name: str) -> dict:
 
 class _FakeModelCfg:
     def __init__(self):
-        self.cross_entropy_fusion_impl = "te"
+        self.cross_entropy_fusion_impl = "native"
         self.context_parallel_size = 1
 
     def finalize(self):
@@ -186,11 +186,7 @@ def test_each_llama_recipe_builds_config(recipe_func: Callable, monkeypatch: pyt
     assert getattr(cfg.model, "pipeline_model_parallel_size", 1) >= 1
 
     if "llama3" in recipe_func.__name__.lower():
-        # Pretrain configs use "te", SFT/PEFT configs use "native"
-        expected_impl = (
-            "native" if ("sft" in recipe_func.__name__.lower() or "peft" in recipe_func.__name__.lower()) else "te"
-        )
-        assert cfg.model.cross_entropy_fusion_impl == expected_impl
+        assert cfg.model.cross_entropy_fusion_impl == "native"
 
 
 @pytest.mark.parametrize("recipe_func", _LLAMA3_SFT_FUNCS)
diff --git a/tests/unit_tests/recipes/test_qwen_recipes.py b/tests/unit_tests/recipes/test_qwen_recipes.py
index 540091f23c..ada66149ff 100644
--- a/tests/unit_tests/recipes/test_qwen_recipes.py
+++ b/tests/unit_tests/recipes/test_qwen_recipes.py
@@ -136,7 +136,7 @@ def test_each_qwen_recipe_builds_config(recipe_func: Callable, monkeypatch: pyte
     assert getattr(cfg.model, "pipeline_model_parallel_size", 1) >= 1
 
     if "qwen3" in recipe_name and "pretrain" in recipe_name and "next" not in recipe_name:
-        assert cfg.model.cross_entropy_fusion_impl == "te"
+        assert cfg.model.cross_entropy_fusion_impl == "native"
 
     # SFT and PEFT-specific assertions
     if is_sft_or_peft:
diff --git a/tests/unit_tests/recipes/test_run_plugins.py b/tests/unit_tests/recipes/test_run_plugins.py
index eaf62269e4..b471614132 100644
--- a/tests/unit_tests/recipes/test_run_plugins.py
+++ b/tests/unit_tests/recipes/test_run_plugins.py
@@ -105,7 +105,7 @@ def create_test_config(**kwargs):
         hidden_size=4096,
         ffn_hidden_size=14336,
         num_attention_heads=32,
-        cross_entropy_fusion_impl="te",
+        cross_entropy_fusion_impl="native",
         tensor_model_parallel_size=tensor_model_parallel_size,
         pipeline_model_parallel_size=pipeline_model_parallel_size,
         pipeline_dtype=pipeline_dtype,