Repository Issue Activity (beta)

#2781 Qwen ANE Prompt Processing on M3 Max: 16K +13%, 32K +57% prefill — plus GDN/packed-bank caveats

#2900 Quantization with MTP head merge for gemma 4 unified (fix inside, PR needed please)

#1631 How to use Native MTP?

#1233 specprefill pp-tps boost silently lost when dflash is also enabled

#2940 ANE Program Inference error at mlp_fraction 0.40 on main; runtime failure latch does not recover — all requests fail

#2938 [Regression] v0.6.3rc1: ANE Prefill Completely Disabled for oQ4e-mtp — TG Drops ~35%

#2763 Cluster: SSH issue. Coordinator key is refused by worker node. Worker node seems to work OK.

#2899 ANE tuner v2 can recommend a gdn_fraction below the structural floor, silently compiling 0 GDN procedures

#2924 settings.json written non-atomically: a torn file silently reverts the server to defaults, including disabled API-key auth

#2725 [BUG] MemoryMonitor unnecessary rejections: estimate uses total layer count instead of KVCache layer count (4x over-estimate on Qwen3.5/3.6/3.8 hybrids)

#2841 ANE prefill triggers hard memory-pressure abort + model eviction (Qwen3.8-27B)

#2914 [0.6.3rc1] ANE Tuner crashes: "Loaded engine does not expose its model" when the model is loaded via the DFlash engine

#2908 [Bug] ANE Tuner V2 (v0.6.3rc1) OOMs at Step 11/13 (_VERIFY_SLOT): calibration program banks and base engine retained during verification reload

#2839 [Bug] ANE prefill not executing on oQ4e-mtp (v0.6.2): configured_layers=0, ANE duty=0, TG drops 57%

#2824 qwen35_ane_prefill_enabled is a silent no-op on non-4bit checkpoints: no warning, no status field, only a load-time delta reveals it

#2813 version 0.6.1 use OsaurusAI/Qwen3.8-27B-MXFP8 can't give any answer，and there's some err in log

#2934 Bug Report: Higgs TTS v3 Outputs Silence (NaN) on OMLX Summary

#2689 Bug: Qwen3.8-27B-oQ8e-mtp decode collapses at 16K context (4-5.5 tok/s); 32K is faster than 16K — reproducible twice on M3 Max

#2874 [Final Report] Optimal Configuration for Qwen3.8-27B on M3 Max 128GB

#2931 [0.6.3rc1] Batched engine loses primed context after late-join handoff — request continues generating with `primed=0` and returns out-of-context output to the wrong conversation

#2928 memory_guard_custom_ceiling_gb silently ignored unless tier=custom — warn at startup

#1691 Orphaned transient prefill memory (Metal allocations) not cleared after stream failure / MidStreamFallbackError, permanently blinding Memory Guard

#2779 ANE prefill regresses on M5 (NAX) hardware; qwen35_ane_prefill doesn't consult is_nax_available()

#2119 Output corruption at batch>=3 for Qwen3.5 hybrid linear-attention (GatedDeltaNet) models; batch 2 clean

#2927 OMLX_WITH_CUSTOM_KERNEL=1: extensions compile but are not installed (wheel-cache hit); surfaces as 'Private ANE runtime unavailable'