MLX error: Failed to load the default metallib. This library is using a deployment target (0x00020008) that is not supported on this OS

[mumaxiaozi]

mac book pro m4 48G
ProductName: macOS
ProductVersion: 15.3.1
BuildVersion: 24D70

vllm-swift version
vllm-swift 0.2.2
dylib: /opt/homebrew/Cellar/vllm-swift/0.2.2/lib/libVLLMBridge.dylib
vLLM: 0.19.1

startup & log:
 vllm-swift serve ~/.lmstudio/models/mlx-community/Qwen3.6-27B-4bit --max-model-len 32876 --served-model-name qwen3.6-27b --port 8001 INFO 04-27 19:00:15 [__init__.py:44] Available plugins for group vllm.platform_plugins: INFO 04-27 19:00:15 [__init__.py:46] - swift -> vllm_swift:register INFO 04-27 19:00:15 [__init__.py:49] All plugins in this group will be loaded. Set VLLM_PLUGINSto control which plugins to load. INFO 04-27 19:00:15 [__init__.py:239] Platform plugin swift is activated (APIServer pid=58692) INFO 04-27 19:00:16 [utils.py:299] (APIServer pid=58692) INFO 04-27 19:00:16 [utils.py:299] █ █ █▄ ▄█ (APIServer pid=58692) INFO 04-27 19:00:16 [utils.py:299] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.19.1 (APIServer pid=58692) INFO 04-27 19:00:16 [utils.py:299] █▄█▀ █ █ █ █ model Qwen/Qwen3-0.6B (APIServer pid=58692) INFO 04-27 19:00:16 [utils.py:299] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀ (APIServer pid=58692) INFO 04-27 19:00:16 [utils.py:299] (APIServer pid=58692) INFO 04-27 19:00:16 [utils.py:233] non-default args: {'model_tag': '/Users/kk/.lmstudio/models/mlx-community/Qwen3.6-27B-4bit', 'port': 8001, 'max_model_len': 32876, 'served_model_name': ['qwen3.6-27b']} (APIServer pid=58692) Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads. (APIServer pid=58692) INFO 04-27 19:00:19 [model.py:549] Resolved architecture: Qwen3ForCausalLM (APIServer pid=58692) INFO 04-27 19:00:19 [model.py:1678] Using max model len 32876 (APIServer pid=58692) INFO 04-27 19:00:19 [vllm.py:790] Asynchronous scheduling is enabled. (APIServer pid=58692) INFO 04-27 19:00:19 [importing.py:68] Triton not installed or not compatible; certain GPU-related functions will not be available. INFO 04-27 19:00:23 [__init__.py:44] Available plugins for group vllm.platform_plugins: INFO 04-27 19:00:23 [__init__.py:46] - swift -> vllm_swift:register INFO 04-27 19:00:23 [__init__.py:49] All plugins in this group will be loaded. SetVLLM_PLUGINS to control which plugins to load. INFO 04-27 19:00:23 [__init__.py:239] Platform plugin swift is activated (EngineCore pid=58708) INFO 04-27 19:00:24 [core.py:105] Initializing a V1 LLM engine (v0.19.1) with config: model='Qwen/Qwen3-0.6B', speculative_config=None, tokenizer='Qwen/Qwen3-0.6B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=32876, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, decode_context_parallel_size=1, dcp_comm_backend=ag_rs, disable_custom_all_reduce=True, quantization=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cpu, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=0, served_model_name=qwen3.6-27b, enable_prefix_caching=False, enable_chunked_prefill=False, pooler_config=None, compilation_config={'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'splitting_ops': ['vllm::unified_attention', 'vllm::unified_attention_with_output', 'vllm::unified_mla_attention', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::olmo_hybrid_gdn_full_forward', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::unified_kv_cache_update', 'vllm::unified_mla_kv_cache_update'], 'compile_mm_encoder': False, 'cudagraph_mm_encoder': False, 'encoder_cudagraph_token_budgets': [], 'encoder_cudagraph_max_images_per_batch': 0, 'compile_sizes': None, 'compile_ranges_endpoints': [32876], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'size_asserts': False, 'alignment_asserts': False, 'scalar_asserts': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.NONE: 0>, 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': None, 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': None, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} (EngineCore pid=58708) INFO 04-27 19:00:24 [parallel_state.py:1400] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://198.18.0.1:52723 backend=gloo (EngineCore pid=58708) INFO 04-27 19:00:24 [parallel_state.py:1716] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A MLX error: Failed to load the default metallib. This library is using a deployment target (0x00020008) that is not supported on this OS. library not found library not found library not found at /Users/tom/dev/vllm-swift/swift/.build/checkouts/mlx-swift/Source/Cmlx/mlx-c/mlx/c/memory.cpp:69 (APIServer pid=58692) Traceback (most recent call last): (APIServer pid=58692) File "<frozen runpy>", line 198, in _run_module_as_main (APIServer pid=58692) File "<frozen runpy>", line 88, in _run_code (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 712, in <module> (APIServer pid=58692) uvloop.run(run_server(args)) (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/uvloop/__init__.py", line 96, in run (APIServer pid=58692) return __asyncio.run( (APIServer pid=58692) ^^^^^^^^^^^^^^ (APIServer pid=58692) File "/opt/homebrew/Cellar/python@3.12/3.12.13_1/Frameworks/Python.framework/Versions/3.12/lib/python3.12/asyncio/runners.py", line 195, in run (APIServer pid=58692) return runner.run(main) (APIServer pid=58692) ^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/opt/homebrew/Cellar/python@3.12/3.12.13_1/Frameworks/Python.framework/Versions/3.12/lib/python3.12/asyncio/runners.py", line 118, in run (APIServer pid=58692) return self._loop.run_until_complete(task) (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/uvloop/__init__.py", line 48, in wrapper (APIServer pid=58692) return await main (APIServer pid=58692) ^^^^^^^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 672, in run_server (APIServer pid=58692) await run_server_worker(listen_address, sock, args, **uvicorn_kwargs) (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 686, in run_server_worker (APIServer pid=58692) async with build_async_engine_client( (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/opt/homebrew/Cellar/python@3.12/3.12.13_1/Frameworks/Python.framework/Versions/3.12/lib/python3.12/contextlib.py", line 210, in __aenter__ (APIServer pid=58692) return await anext(self.gen) (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 100, in build_async_engine_client (APIServer pid=58692) async with build_async_engine_client_from_engine_args( (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/opt/homebrew/Cellar/python@3.12/3.12.13_1/Frameworks/Python.framework/Versions/3.12/lib/python3.12/contextlib.py", line 210, in __aenter__ (APIServer pid=58692) return await anext(self.gen) (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/entrypoints/openai/api_server.py", line 136, in build_async_engine_client_from_engine_args (APIServer pid=58692) async_llm = AsyncLLM.from_vllm_config( (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/v1/engine/async_llm.py", line 225, in from_vllm_config (APIServer pid=58692) return cls( (APIServer pid=58692) ^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/v1/engine/async_llm.py", line 154, in __init__ (APIServer pid=58692) self.engine_core = EngineCoreClient.make_async_mp_client( (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper (APIServer pid=58692) return func(*args, **kwargs) (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/v1/engine/core_client.py", line 130, in make_async_mp_client (APIServer pid=58692) return AsyncMPClient(*client_args) (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/tracing/otel.py", line 178, in sync_wrapper (APIServer pid=58692) return func(*args, **kwargs) (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/v1/engine/core_client.py", line 887, in __init__ (APIServer pid=58692) super().__init__( (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/v1/engine/core_client.py", line 535, in __init__ (APIServer pid=58692) with launch_core_engines( (APIServer pid=58692) ^^^^^^^^^^^^^^^^^^^^ (APIServer pid=58692) File "/opt/homebrew/Cellar/python@3.12/3.12.13_1/Frameworks/Python.framework/Versions/3.12/lib/python3.12/contextlib.py", line 144, in __exit__ (APIServer pid=58692) next(self.gen) (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/v1/engine/utils.py", line 998, in launch_core_engines (APIServer pid=58692) wait_for_engine_startup( (APIServer pid=58692) File "/Users/kk/.vllm-swift/venv/lib/python3.12/site-packages/vllm/v1/engine/utils.py", line 1057, in wait_for_engine_startup (APIServer pid=58692) raise RuntimeError( (APIServer pid=58692) RuntimeError: Engine core initialization failed. See root cause above. Failed core proc(s): {'EngineCore': 255}

ai analyse：
MLX error: Failed to load the default metallib. This library is using a deployment target (0x00020008) that is not supported on this OS.

根本原因：vllm-swift 0.2.2 的 MLX metallib 库与你的 macOS 版本不兼容。

具体分析：

1. MLX Metal 库部署目标不匹配 — 0x00020008 对应 macOS 26.0+ (即 macOS Tahoe/下一代)，而你当前系统是 macOS 15.3.1 (Sequoia)。vllm-swift 0.2.2 编译时使用了过高的 macOS deployment
   target，导致 Metal shader metallib 无法在当前系统加载。
2. 连锁反应 — MLX 初始化失败后，vLLM EngineCore 无法通过 Swift/Metal 后端运行模型，随后整个引擎启动流程崩溃，抛出 RuntimeError: Engine core initialization failed。
3. 另外注意 — 日志显示 Resolved architecture: Qwen3ForCausalLM 和 model='Qwen/Qwen3-0.6B'，说明 vLLM 把这个 Qwen3.5 hybrid (含 linear_attention 层) 模型错误地解析为
   Qwen3ForCausalLM（纯文本 Qwen3），而非 Qwen3_5ForConditionalGeneration。即使 MLX 加载成功，这个架构映射错误也会导致问题。

[TheTom]

@mumaxiaozi this is fixed in v0.3.0. the metallib in v0.2.2 was built on macOS 26 (Tahoe) without an explicit -mmacosx-version-min, so it inherited the SDK default and rejected on Sequoia. rebuilt in commit d54d0a0 with -mmacosx-version-min=15.0.

upgrade:

vllm-swift update
# or
brew upgrade vllm-swift

then re-run your serve command. metallib should load on 15.3.1.

separate gotcha visible in your log: model='Qwen/Qwen3-0.6B' is wrong (issue #11). vLLM 0.19.1's argparse maps a stray positional to the deprecated model_tag slot and falls back to that placeholder. fix is in PR #12 but not bottled yet. workaround until v0.3.1 ships:

vllm-swift serve --model ~/.lmstudio/models/mlx-community/Qwen3.6-27B-4bit \
  --max-model-len 32876 --served-model-name qwen3.6-27b --port 8001

(note the explicit --model instead of positional path.)

let me know if upgrade resolves the metallib error and i'll close.

[handicraftsman]

Happening on macOS 26.6 Tahoe as well