Bump mlx-lm/mlx-vlm deps and add Qwen3.5-0.8B smoke test

pyproject.toml

@@ -28,9 +28,9 @@ classifiers = [
 
 dependencies = [
     # MLX - Required for Apple Silicon GPU acceleration
-    "mlx>=0.29.2; platform_system == 'Darwin' and platform_machine == 'arm64'",
-    "mlx-lm>=0.28.4; platform_system == 'Darwin' and platform_machine == 'arm64'",
-    "mlx-vlm>=0.3.0; platform_system == 'Darwin' and platform_machine == 'arm64'",  # Vision-language model support
+    "mlx>=0.31.0; platform_system == 'Darwin' and platform_machine == 'arm64'",
+    "mlx-lm>=0.31.0; platform_system == 'Darwin' and platform_machine == 'arm64'",
+    "mlx-vlm>=0.4.0; platform_system == 'Darwin' and platform_machine == 'arm64'",  # Vision-language model support
     # Model loading and weights
     "transformers>=5.0.0",
     "accelerate>=0.26.0",

tests/test_qwen35_smoke.py

@@ -0,0 +1,97 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Smoke test for Qwen3.5-0.8B: proves transformers 5.x model works end-to-end.
+
+Qwen3.5 uses the `qwen3_5` architecture which requires transformers>=5.0.0.
+This test verifies that the upgraded dependency stack (mlx-lm>=0.31.0,
+mlx-vlm>=0.4.0, transformers>=5.0.0) works correctly with vLLM on Metal.
+
+Golden token IDs were generated with greedy decoding (temperature=0) on
+Qwen/Qwen3.5-0.8B, one sequence at a time (max_num_seqs=1).
+
+Run:
+    VLLM_ENABLE_V1_MULTIPROCESSING=0 \
+        python -m pytest tests/test_qwen35_smoke.py -v -s
+"""
+
+from __future__ import annotations
+
+import os
+
+import pytest
+from vllm import LLM, SamplingParams
+
+MODEL_NAME = "Qwen/Qwen3.5-0.8B"
+MAX_TOKENS = 10
+
+PROMPTS = [
+    "The capital of France is",
+    "The weather today is not",
+    "One plus one equals",
+    "The largest planet in our solar system is",
+    "Water boils at a temperature of",
+]
+
+# fmt: off
+# Golden token IDs from MLX inline cache, greedy decoding (Qwen3.5-0.8B).
+GOLDEN_MLX = {
+    "The capital of France is":                     [11751, 13, 198, 760, 6511, 314, 9338, 369, 11751, 13],
+    "The weather today is not":                     [1603, 13, 198, 760, 8831, 3242, 369, 524, 1603, 13],
+    "One plus one equals":                          [1330, 13, 198, 3833, 5346, 799, 16327, 1330, 13, 198],
+    "The largest planet in our solar system is":    [279, 7806, 13, 271, 248068, 271, 248069, 271, 332, 2665],
+    "Water boils at a temperature of":              [220, 16, 15, 15, 29922, 13, 1368, 264, 1637, 369],
+}
+# fmt: on
+
+
+def _setenv_default(mp: pytest.MonkeyPatch, key: str, default: str) -> str:
+    """Set an env var only when absent and return the effective value."""
+    value = os.environ.get(key)
+    if value is None:
+        mp.setenv(key, default)
+        return default
+    return value
+
+
+@pytest.fixture(autouse=True, scope="module")
+def _set_env():
+    """Set default env vars for this test."""
+    with pytest.MonkeyPatch.context() as mp:
+        mp.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
+        _setenv_default(mp, "VLLM_METAL_MEMORY_FRACTION", "auto")
+        yield
+
+
+@pytest.fixture(scope="module")
+def vllm_outputs():
+    """Run vLLM offline inference once for all prompts."""
+    llm = LLM(model=MODEL_NAME, max_model_len=512, max_num_seqs=1)
+    sp = SamplingParams(temperature=0, max_tokens=MAX_TOKENS)
+    outputs = llm.generate(PROMPTS, sp)
+    return {o.prompt: o for o in outputs}
+
+
+class TestQwen35Smoke:
+    @pytest.mark.slow
+    @pytest.mark.parametrize("prompt", PROMPTS)
+    def test_generate_matches_golden(self, vllm_outputs, prompt):
+        output = vllm_outputs[prompt]
+        token_ids = list(output.outputs[0].token_ids)
+        text = output.outputs[0].text
+
+        expected = GOLDEN_MLX[prompt]
+        matched = token_ids == expected
+
+        print(f"\n  prompt: {prompt!r}")
+        print(f"  output: {text!r}")
+        print(f"  ids:    {token_ids}")
+        if matched:
+            print("  result: MATCHED golden")
+        else:
+            print("  result: NO MATCH")
+            print(f"  expected: {expected}")
+
+        assert matched, (
+            f"Output for {prompt!r} did not match golden set.\n"
+            f"Got:      {token_ids}\n"
+            f"Expected: {expected}"
+        )