Issues

• Memory explosion / leak when serving Qwen3.6-27B-ConfigI-MLX with TurboQuant kv_scheme=turbo4

  TheTom/vllm-swift#21 · Filter by author heykb opened on May 14, 2026

• Homebrew bottle install: ModuleNotFoundError: No module named 'vllm_swift'

  TheTom/vllm-swift#20 · Filter by author heykb opened on May 11, 2026

• stream_options.include_usage not honored in chat completion streaming responses

  TheTom/vllm-swift#19 · Filter by author yaanfpv opened on May 9, 2026

• Long-context prefill (40k+ on Qwen3.5-9B at 131k ctx) blows up memory; chunked prefill is force-disabled in platform plugin

  TheTom/vllm-swift#18 · Filter by author yaanfpv opened on May 9, 2026

• TurboQuant unavailable on Qwen3.5 (head_dim=256): kernel turbo_dequant_rotated_4_256_bf16 missing in v0.6.0 metallib bottle

  TheTom/vllm-swift#17 · Filter by author yaanfpv opened on May 9, 2026

• Qwen3.5-VL vision returns hallucinated output for every image (mm_features + load order + prefill shape)

  TheTom/vllm-swift#16 · Filter by author yaanfpv opened on May 9, 2026

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

  TheTom/vllm-swift#10 · Filter by author mumaxiaozi opened on Apr 27, 2026
  2