Search results

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

Status: Open.

#21 In TheTom/vllm-swift;

· heykb opened on May 14, 2026

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

Status: Open.

#20 In TheTom/vllm-swift;

· heykb opened on May 11, 2026

stream_options.include_usage not honored in chat completion streaming responses

Status: Open.

#19 In TheTom/vllm-swift;

· 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

Status: Open.

#18 In TheTom/vllm-swift;

· 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

Status: Open.

#17 In TheTom/vllm-swift;

· yaanfpv opened on May 9, 2026

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

Status: Open.

#16 In TheTom/vllm-swift;

· 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

Status: Open.

#10 In TheTom/vllm-swift;

· mumaxiaozi opened on Apr 27, 2026