Repository Issue Activity (beta)

#456 Flash-Next: active_bytes creeps past the Metal limit under deep-session churn (103.1 GB > 96 GiB with bank=0, cache=0, live sessions 7.8 GB) — nothing sheddable, only restart recovers

#447 prefill_admission_shed refuses a warm +23K turn (507) while 12.6 GB of evictable bank is resident, and cache_cleared makes retries unrecoverable

#446 Flash-Next: incremental deep session loses its cached prefix at turn 4-5 (block-prefix remnant of 14336/18432 tokens), forcing a 70-100K re-prefill — 5/5 chains on 2.11.1

#440 UPDATE: Root cause is concurrency, NOT unsafe-force-unverified (this param doesn't exist in 2.11.0)

#420 [QSACache] ar_batch mode fails when batching requests with history

#477 `_consolidate_system_messages` hoists late `system` turns into msg0, preventing prefix reuse from advancing

#476 Re: Issue #440 Test Results - v2.11.2 Confirmed Fixed

#468 Bare-name first line keeps `_InitialOrphanToolControlStreamGuard` in `hold` for the entire stream

#438 Memory leak in 2.10.2?

#472 `mtplx connect opencode` hardcodes modalities.input ["text"], so vision-capable builds lose image input in OpenCode / T3 Code

#451 Add optional aria2c backend for resumable Hugging Face model downloads

#455 Lastest update Qwen3.8 27b Optimised-quality Bug

#449 Flash-Next n-gram sidecar: the table cannot be replaced by a function; the miss path pays three page-ins per row

#460 Hermes integration silently writes ~/.hermes/profiles/mtplx with hardcoded terminal.backend: local — overrides user's Docker sandbox, re-stamped every launch

#463 One-shot mtplx run/chat crashes on Flash-Next (qwen4_exp): generic gdn_capture paths assume legacy Qwen3.5 layer layout

#465 AR SessionBank restore fails with policy_mismatch for Hermes; mtp_history_policy metadata disagrees with fingerprint

#450 qwen4 fixed-M4 lane gate counts idle session-bank bytes as live and never asks the bank to yield (128 GB Mac, 2.11.1)

#469 error: [metal::Device] Unable to build metal library from source | API only | after latest 2.11 app update

#461 MTPLX 2.11.1 Unable to build metal library (header file not found) on macOS 15.7

#467 Qwen 3.8 semantic corruption even in MTPLX chat window with no agents starting from 2.11.1

#464 Qwen3.8 + Hermes produces semantic corruption in MTP depth 1/3 while identical cold request works in AR

#459 [MTPLX + OpenCode] - Qwen 3.8 27B MTP3 collapsing into hallucinatory loops on agnetic tasks

#321 Verify kernels: bits != 4 early return leaves 8-bit shapes unbenchmarked

#445 Huggingface Hub Cache Support

#403 Flash-Next: engine deadlock when ssd-session-cache is enabled and an incremental session crosses ~131K tokens (2.10.1)