Repository Issue Activity (beta)

#2229 0.7.0 @ b5379f69 sweep: Llama-3.2-11B-Vision (mllama) mask off-by-one crash, InternVL 3.5 'internvl' model_type unsupported, aya_vision leaks <|START_RESPONSE|> markers

#2228 mlx_vlm.generate CLI and generate() ignore the checkpoint's generation_config.json sampling defaults that the server honours since #1337

#2220 GLM-4.6V-nvfp4 (`glm4v_moe`) still returns `<|begin_of_box|>` / `<|end_of_box|>` on 0.7.0: the #2170 cleanup was applied to `glm4v` only

#2227 Mage-VL-OptiQ-4bit (`mage_vl`) fails to load on 0.7.0: flattened OptiQ `model.*` keys are not remapped by `sanitize` (904 parameters not in model)

#2221 20260911 -- Diagnostic run comparing 0.7.0rc0 → 0.7.0: like-for-like sweep of 33 cached VLMs (mostly "nothing to see")

#2201 Qwen assistant reasoning_content is dropped during prompt normalization

#2084 glm5_next: indexer incremental `_pool` crashes with `[argmax] Cannot argmax reduce zero size array` after any cache rewind (stale `t_prev`)

#2210 Exact APC hit path (0.6.17): single-row warm cache merged into Batch* caches, Qwen3.5 shortcut copies the whole cache per decode token (6.5 -> 9.9 tok/s at 33k; patch)

#2154 `TurboQuantKVCache` advertises `bits` but not `group_size`, so `mlx_lm`'s quantized SDPA path raises on it

#2033 **MTP drafter + concurrent streaming requests: infinite decode loop (repeated token_id=0 / '!') and prefill starvation**

#2217 GLM-5.3-Flash 4-bit + MTP drafter (split_mtp, mxfp8): mlx_vlm.server is SIGKILLed (rc 137) ~60 s into load on a 256 GB M3 Ultra; target alone loads and serves fine

#2216 glm5_next: sanitize() only remaps lm_head.weight — quantized community checkpoints (lm_head.scales/biases at top level) fail the strict load

#2215 Kernel support request: which shapes a packed 4-bit KV path has to accept to be reachable from the server

#2211 Chat Completions moves historical image markers to the latest user turn

#284 Issue with Finetuning Mistral (Mistral-Small-3.1-24B-Instruct-2503-4bit)

#1222 mlx_vlm.convert with mtp?

#824 LoRA training broken for Qwen3.5 VLM: 3 bugs causing corrupted generation

#1346 Server batching: `[concatenate] No arrays provided for concatenation` in `get_rope_index` (Qwen3.5/3.6) with mixed-length vision requests — one failure takes down all in-flight requests

#1718 sanitize() idempotency is an unstated invariant since #1498 — make it explicit and tested

#1619 Qwen3-Omni: decode steps lose the multimodal RoPE delta — generation degenerates instantly ("the the the…")

#1726 LoRA/SFT training crashes on multi-image records: image_grid_thw collated to (1, N, 3) instead of (N, 3) (Qwen3-VL / Qwen3.5)

#2041 qwen4_exp: Qwen3.8-Flash-Next MLX 4-bit checkpoint loads cleanly on main but generates garbage

#2179 Distributed LoRA/SFT training crashes with Metal GPU Timeout Error, even after applying the mlx-examples#1185 CPU-stream fix

#2183 Parakeet (and other NeMo-alignment STT models) never return "segments" via /v1/audio/transcriptions - AlignedResult exposes .sentences, not .segments

#2180 Video-Depth-Anything: two concrete speed levers found while optimizing a sibling Depth-Anything-V2 MLX port