Wire up muse-glimmer CI (#21797)

As titled

Author: metascroy

.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_benchmark_dflash_long_context.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_ondevice_serving.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

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

@@ -30,6 +30,12 @@ addopts =
     # Ignore failing llava tests (missing accelerate dependency)
     --ignore=examples/models/llava/test/test_llava.py
 
+    # Backend-specific Muse Glimmer pipelines. These run from mlx.yml / cuda.yml,
+    # which name them explicitly -- --ignore only suppresses them while
+    # collecting the directory, not when a path is passed on the command line.
+    --ignore=examples/models/muse-glimmer/tests/test_mlx_pipeline.py
+    --ignore=examples/models/muse-glimmer/tests/test_cuda_pipeline.py
+
     # Ignore tests with missing custom_ops_generated_lib dependencies
     --ignore=exir/backend/test/demos/test_delegate_aten_mode.py
     --ignore=exir/backend/test/demos/test_xnnpack_qnnpack.py
@@ -98,6 +104,10 @@ testpaths =
     examples/models/llama3_2_vision/text_decoder/test
     examples/models/llava/test
 
+    # Backend-agnostic Muse Glimmer tests. The MLX and CUDA pipeline tests are
+    # excluded above and run from their own backend workflows.
+    examples/models/muse-glimmer/tests
+
     # exir
     exir/