Issues

• MLA/DSA/QSA/DSV4 paged-cache kernels still dereference block_table/cache_seqlens unguarded (follow-up to #436/#437)

  turboderp-org/exllamav3#438 · Filter by author ghazni101 opened on Oct 2, 2026

• Paged decode: one KV pass per kv head for speculative verify (patch ready, asking before a PR)

  turboderp-org/exllamav3#435 · Filter by author UHHHHHHHHHHHHHH opened on Oct 2, 2026

• Recurrent checkpoint workload retains ~1.5 GiB of host heap despite existing malloc_trim calls

  turboderp-org/exllamav3#432 · Filter by author szafranski opened on Oct 2, 2026
  2

• -rcs 0 (or any recurrent cache smaller than one checkpoint) crashes with KeyError: 'dictionary is empty' in RecurrentCache.put

  turboderp-org/exllamav3#431 · Filter by author szafranski opened on Oct 2, 2026

• Audio Support

  turboderp-org/exllamav3#420 · Filter by author TingTingin opened on Sep 28, 2026
  1

• MiMo-V2.6 gibberish output

  turboderp-org/exllamav3#418 · Filter by author HsYn35 opened on Sep 28, 2026
  1

• Turing (sm_75): prefill ×4–9 at 16–64K and decode ×1.6–2.2 on top of v1.5.3 — data + fork

  turboderp-org/exllamav3#417 · Filter by author rafatxf opened on Sep 28, 2026
  5

• exl3_moe: 3.5 bpw (half-integer) experts fall back to the runtime-switch kernel on passes over 8 rows

  turboderp-org/exllamav3#414 · Filter by author issi-hub opened on Sep 27, 2026
  2

• v1.5.2 Windows wheel fails to load: DLL load failed while importing exllamav3_ext (WinError 193); .nv_fatb section grew to 1.93 GiB

  turboderp-org/exllamav3#413 · Filter by author lslogis opened on Sep 27, 2026
  3

• Insufficient VRAM in latest release

  turboderp-org/exllamav3#409 · Filter by author rpvelloso opened on Sep 27, 2026
  4

• Dynamic draft sizing is inert under tensor parallel: MTP/DFlash drafters never export draft confidence

  turboderp-org/exllamav3#405 · Filter by author evaanp opened on Sep 25, 2026
  2

• Low DFlash2 draft acceptance (~40%) on Qwen3.8-27B EXL3 4bpw — no end-to-end win over MTP at 90–135k context

  turboderp-org/exllamav3#403 · Filter by author dfein38347g opened on Sep 23, 2026
  3

• DFlash2 draft loader rejects checkpoints with .weight-suffixed codebook keys

  turboderp-org/exllamav3#390 · Filter by author rktchip opened on Sep 21, 2026
  1

• Higher VRAM use on V1.5.1 vs V1.5.0 runing GLM5.3-Flash

  turboderp-org/exllamav3#388 · Filter by author stiggy2k16 opened on Sep 21, 2026
  25

• [Doc Request] Configuration for 1.5.0 test cases

  turboderp-org/exllamav3#385 · Filter by author xashr opened on Sep 20, 2026
  3

• SM89 / Ada prefill tuning: F16ACC and paged-attention benchmark results on RTX 4060 Ti

  turboderp-org/exllamav3#384 · Filter by author grimlee opened on Sep 20, 2026
  1

• [Feat Request] Numa support.

  turboderp-org/exllamav3#382 · Filter by author Ph0rk0z opened on Sep 18, 2026
  1

• [Feature Request] Chained Speculative Decoding

  turboderp-org/exllamav3#374 · Filter by author HermiG opened on Sep 14, 2026
  3

• [Feature Request] K2-Horizon-MoVA-36B-A4B Support

  turboderp-org/exllamav3#373 · Filter by author teee9 opened on Sep 14, 2026

• Pinned arena for CPU-MoE does not work with python from conda/conda forge

  turboderp-org/exllamav3#372 · Filter by author maximus2600 opened on Sep 14, 2026
  2

• Windows: hugepage promotion for the MoE CPU expert arena is a no-op (EXL3_MOE_ARENA_HUGEPAGE is inert)

  turboderp-org/exllamav3#368 · Filter by author sorasoras opened on Sep 12, 2026
  1

• Insufficient vram for model and cache ambiguity.

  turboderp-org/exllamav3#361 · Filter by author Ph0rk0z opened on Sep 12, 2026
  4

• Tensor parallelism for Qwen3.8 Flash Next

  turboderp-org/exllamav3#343 · Filter by author krutovvv opened on Sep 7, 2026
  3

• Feature request: Make the CPU page cache and recurrent checkpoint cache *injectable*, so third-party tiers can be plugged in?

  turboderp-org/exllamav3#336 · Filter by author niceblue88 opened on Sep 5, 2026
  2

• Two-GPU layer split decodes gibberish on a 2× RTX PRO 6000 (sm_120) box — root cause: platform P2P false positive (Torch reports peer access, copies return zeros); EXLLAMA_NO_P2P_COPY=1 confirmed. Originally reported as GLM-5.3-Flash-specific

  turboderp-org/exllamav3#333 · Filter by author ppickle1989 opened on Sep 4, 2026
  5