[kernel] Tune LingBot-Video MoE TMA configs for H100 (#36543)

test/registered/unit/layers/moe/test_fused_moe_triton_config.py

@@ -22,6 +22,25 @@ def test_h200_bf16_config_is_available_for_current_triton_runtime():
     assert json.loads(config_path.read_text())["128"]["BLOCK_SIZE_M"] > 0
 
 
+def test_h100_lingbot_video_configs_enable_tma_only_for_the_tuned_shape():
+    config_root = (
+        Path(fused_moe_triton_config.__file__).parent / "configs" / "triton_3_7_1"
+    )
+
+    for suffix in ("", "_down"):
+        config_path = (
+            config_root / f"E=128,N=768,device_name=NVIDIA_H100_80GB_HBM3{suffix}.json"
+        )
+        configs = json.loads(config_path.read_text())
+
+        assert configs["4096"]["USE_TMA"] is True
+        assert all(
+            "USE_TMA" not in config
+            for num_tokens, config in configs.items()
+            if num_tokens != "4096"
+        )
+
+
 def test_down_moe_reuses_tuned_up_config_when_separate_config_is_absent(
     monkeypatch, tmp_path
 ):