Fix genai_config.json provider names to use short format

src/mobius/integrations/ort_genai/auto_export_test.py

@@ -796,7 +796,7 @@ def test_ep_cuda_passes_through(self, tmp_path):
             data = json.load(f)
         provider_opts = data["model"]["decoder"]["session_options"]["provider_options"]
         assert len(provider_opts) == 1
-        assert "CUDAExecutionProvider" in provider_opts[0]
+        assert "cuda" in provider_opts[0]
 
     def test_raises_when_pkg_config_is_none(self, tmp_path):
         """ValueError is raised when pkg.config is None."""

src/mobius/integrations/ort_genai/ep_config.py

@@ -17,12 +17,14 @@
 
 from mobius._execution_providers import ep_registry
 
-# ORT GenAI provider name mapping (internal name → ORT GenAI provider string)
+# ORT GenAI provider name mapping (internal name → ORT GenAI provider string).
+# GenAI expects short lowercase names (e.g. "cuda", "dml"), not the full
+# ORT EP class names (e.g. "CUDAExecutionProvider").
 _ORT_PROVIDER_NAMES: dict[str, str] = {
-    "cpu": "CPUExecutionProvider",
-    "cuda": "CUDAExecutionProvider",
-    "dml": "DmlExecutionProvider",
-    "webgpu": "WebGpuExecutionProvider",
+    "cpu": "cpu",
+    "cuda": "cuda",
+    "dml": "dml",
+    "webgpu": "webgpu",
     "trt-rtx": "NvTensorRtRtx",
 }
 

src/mobius/integrations/ort_genai/ep_config_test.py

@@ -20,26 +20,26 @@ def test_cpu_returns_empty(self):
     def test_cuda_default(self):
         result = make_provider_options("cuda")
         assert len(result) == 1
-        assert "CUDAExecutionProvider" in result[0]
-        assert result[0]["CUDAExecutionProvider"]["enable_cuda_graph"] == "0"
+        assert "cuda" in result[0]
+        assert result[0]["cuda"]["enable_cuda_graph"] == "0"
 
     def test_cuda_with_graph(self):
         result = make_provider_options("cuda", enable_cuda_graph=True)
-        assert result[0]["CUDAExecutionProvider"]["enable_cuda_graph"] == "1"
+        assert result[0]["cuda"]["enable_cuda_graph"] == "1"
 
     def test_dml(self):
         result = make_provider_options("dml")
         assert len(result) == 1
-        assert "DmlExecutionProvider" in result[0]
+        assert "dml" in result[0]
 
     def test_webgpu_default(self):
         result = make_provider_options("webgpu")
-        assert result[0]["WebGpuExecutionProvider"]["enableGraphCapture"] == "0"
-        assert result[0]["WebGpuExecutionProvider"]["validationMode"] == "basic"
+        assert result[0]["webgpu"]["enableGraphCapture"] == "0"
+        assert result[0]["webgpu"]["validationMode"] == "basic"
 
     def test_webgpu_with_graph(self):
         result = make_provider_options("webgpu", enable_webgpu_graph=True)
-        opts = result[0]["WebGpuExecutionProvider"]
+        opts = result[0]["webgpu"]
         assert opts["enableGraphCapture"] == "1"
         assert opts["validationMode"] == "disabled"
 

src/mobius/integrations/ort_genai/genai_config_test.py

@@ -692,22 +692,22 @@ def test_cpu_has_empty_provider_options(self):
         assert opts["provider_options"] == []
 
     def test_cuda_has_cuda_provider_options(self):
-        """CUDA EP produces a provider_options entry for CUDAExecutionProvider."""
+        """CUDA EP produces a provider_options entry for cuda."""
         from mobius.integrations.ort_genai.genai_config import _make_session_options
 
         opts = _make_session_options("cuda")
         assert opts["log_id"] == "onnxruntime-genai"
         assert len(opts["provider_options"]) == 1
-        assert "CUDAExecutionProvider" in opts["provider_options"][0]
+        assert "cuda" in opts["provider_options"][0]
 
     def test_dml_has_dml_provider_options(self):
-        """DML EP produces a provider_options entry for DmlExecutionProvider."""
+        """DML EP produces a provider_options entry for dml."""
         from mobius.integrations.ort_genai.genai_config import _make_session_options
 
         opts = _make_session_options("dml")
         assert opts["log_id"] == "onnxruntime-genai"
         assert len(opts["provider_options"]) == 1
-        assert "DmlExecutionProvider" in opts["provider_options"][0]
+        assert "dml" in opts["provider_options"][0]
 
 
 class TestGenaiConfigGeneratorEp:
@@ -735,7 +735,7 @@ def test_cuda_ep_decoder_has_cuda_provider_options(self):
         config = self._gen("cuda").generate()
         opts = config["model"]["decoder"]["session_options"]["provider_options"]
         assert len(opts) == 1
-        assert "CUDAExecutionProvider" in opts[0]
+        assert "cuda" in opts[0]
 
     def test_cuda_ep_all_blocks_have_cuda_session_options(self):
         """CUDA EP applied to all 4 session blocks (decoder, vision, embedding, audio)."""
@@ -761,6 +761,4 @@ def test_cuda_ep_all_blocks_have_cuda_session_options(self):
             session_opts = config["model"][block]["session_options"]
             provider_options = session_opts["provider_options"]
             assert len(provider_options) == 1, f"{block} missing CUDA provider options"
-            assert "CUDAExecutionProvider" in provider_options[0], (
-                f"{block} has wrong EP in provider_options"
-            )
+            assert "cuda" in provider_options[0], f"{block} has wrong EP in provider_options"