Wire up muse-glimmer CI

.github/workflows/mlx.yml

@@ -14,6 +14,7 @@ on:
       - extension/llm/export/**
       - extension/audio/**
       - examples/models/gemma4_31b/**
+      - examples/models/muse-glimmer/**
       - examples/models/parakeet/**
       - examples/models/voxtral_realtime/**
       - examples/models/qwen3_5_moe/**
@@ -90,6 +91,7 @@ jobs:
           backends/mlx/test/test_slot_recycling.py \
           backends/mlx/test/test_sample.py \
           examples/models/gemma4_31b/tests/test_mlx_pipeline.py \
+          examples/models/muse-glimmer/tests/test_mlx_pipeline.py \
           -v
         echo "::endgroup::"
 

examples/models/muse-glimmer/tests/test_cuda_pipeline.py

@@ -21,9 +21,22 @@
 import unittest
 from dataclasses import replace
 
-import executorch.backends.cuda.quantize_op_dispatch as _quantize_op_dispatch  # noqa: F401
 import torch
-from executorch.backends.cuda.coalesced_int4_tensor import CudaCoalescedInt4Tensor
+
+try:
+    import executorch.backends.cuda.quantize_op_dispatch as _quantize_op_dispatch  # noqa: F401
+    from executorch.backends.cuda.coalesced_int4_tensor import CudaCoalescedInt4Tensor
+    from executorch.examples.models.muse_glimmer.source_transformations.cuda import (
+        add_dflash_hidden_tapping,
+        add_on_device_sampler,
+        cuda_source_transformations,
+    )
+except ImportError as e:
+    # The CUDA backend pulls in triton, which is absent on non-CUDA builds.
+    # Every test here is CUDA-specific, so skip the module rather than fail
+    # collection for anyone running the suite without CUDA.
+    raise unittest.SkipTest(f"CUDA backend unavailable: {e}")
+
 from executorch.examples.models.muse_glimmer.export.common import (
     mutable_buffer_metadata,
 )
@@ -38,11 +51,6 @@
 )
 from executorch.examples.models.muse_glimmer.loaders.checkpoint_loader import _finalize
 from executorch.examples.models.muse_glimmer.model.model import FlatKVCache
-from executorch.examples.models.muse_glimmer.source_transformations.cuda import (
-    add_dflash_hidden_tapping,
-    add_on_device_sampler,
-    cuda_source_transformations,
-)
 from executorch.examples.models.muse_glimmer.tests.test_pipeline import (
     build_random_tiny_model,
     DEFAULT_RECIPE,

examples/models/muse-glimmer/tests/test_ondevice_serving.py

@@ -10,8 +10,15 @@
 
 import pytest
 
-from executorch.examples.llm_server.python.chat_template import ChatTemplate
-from executorch.examples.llm_server.python.protocol import ChatMessage
+# The serving stack is built on pydantic, which ships with the llm_server
+# extras (examples/llm_server/python/requirements.txt) rather than core
+# ExecuTorch, so skip instead of failing collection when it is absent.
+pytest.importorskip("pydantic", reason="requires llm_server serving dependencies")
+
+from executorch.examples.llm_server.python.chat_template import (  # noqa: E402
+    ChatTemplate,
+)
+from executorch.examples.llm_server.python.protocol import ChatMessage  # noqa: E402
 
 _SERVER = os.environ.get("MUSE_GLIMMER_SERVER_URL")
 _HF_DIR = os.environ.get("MUSE_GLIMMER_HF_DIR")

examples/models/muse-glimmer/tests/test_serve.py

@@ -13,11 +13,20 @@
 
 import pytest
 
-from executorch.examples.llm_server.python import chat_template
-from executorch.examples.llm_server.python.errors import APIError
-from executorch.examples.llm_server.python.protocol import ChatCompletionRequest
-from executorch.examples.llm_server.python.session_runtime import PromptInput
-from executorch.examples.models.muse_glimmer.serving import serve
+# The serving stack is built on pydantic, which ships with the llm_server
+# extras (examples/llm_server/python/requirements.txt) rather than core
+# ExecuTorch, so skip instead of failing collection when it is absent.
+pytest.importorskip("pydantic", reason="requires llm_server serving dependencies")
+
+from executorch.examples.llm_server.python import chat_template  # noqa: E402
+from executorch.examples.llm_server.python.errors import APIError  # noqa: E402
+from executorch.examples.llm_server.python.protocol import (  # noqa: E402
+    ChatCompletionRequest,
+)
+from executorch.examples.llm_server.python.session_runtime import (  # noqa: E402
+    PromptInput,
+)
+from executorch.examples.models.muse_glimmer.serving import serve  # noqa: E402
 
 _HERE = pathlib.Path(serve.__file__).resolve().parent
 _LLM_SERVER_PYTHON = pathlib.Path(chat_template.__file__).resolve().parent

pytest.ini

@@ -98,6 +98,11 @@ testpaths =
     examples/models/llama3_2_vision/text_decoder/test
     examples/models/llava/test
 
+    # Backend-agnostic Muse Glimmer tests. The MLX pipeline test skips itself
+    # when the backend is absent and is also run by mlx.yml; the CUDA one is
+    # ignored above.
+    examples/models/muse-glimmer/tests
+
     # exir
     exir/