Issues

• multi-slot: every thinking request (incl. the no-fields default) 400s on templates that let the model open <think> itself

  warpfront/hipfire#800 · Filter by author ghazni101 opened on Oct 2, 2026

• Tracking: 0.4.1 (Tue 10/06) — Flash-Next prefill, bugfixes, in-flight work

  warpfront/hipfire#799 · Filter by author Kaden-Schutt opened on Oct 1, 2026
  10

• Tracking: 0.4.0 (Thu/Fri), 0.4.1 (Tue 10/06), 0.4.2 (Sun 10/11) — cadence, scope, preliminary release notes

  warpfront/hipfire#798 · Filter by author Kaden-Schutt opened on Sep 30, 2026
  1

• Dense TP: wire block verify on the mesh so n-gram/DFlash/DSpark can run at tp>1

  warpfront/hipfire#796 · Filter by author alpineQ opened on Sep 30, 2026

• gfx1100: registry Qwen3.6-27B MQ3 fails in decode with unsupported gemv.swiglu_residual

  warpfront/hipfire#789 · Filter by author HUSRCF opened on Sep 29, 2026

• size mismatch when downloading qwen3.8 mq4 xt mpdel

  warpfront/hipfire#777 · Filter by author igagis opened on Sep 26, 2026
  3

• Benchmark: RDNA1 - gfx1010 - RX 5600 XT

  warpfront/hipfire#773 · Filter by author okamoto9060 opened on Sep 23, 2026

• gfx1030 - qwen3.8 + dflash issue

  warpfront/hipfire#771 · Filter by author zues75 opened on Sep 19, 2026
  2

• dense TP=2 decode crashes HSA_STATUS_ERROR_INVALID_ALLOCATION on tied/aliased-lm_head dense Qwen (qwen3.5-4b); separate-lm_head (3.8-27b) works

  warpfront/hipfire#770 · Filter by author sklarsa opened on Sep 17, 2026

• Dflash hurts perfomance with Qwen 3.8 on Strix Halo

  warpfront/hipfire#754 · Filter by author alucryd opened on Sep 11, 2026
  3

• feat: DeepSeek-V4.1-Flash implementation plan — MQ3/MQ2 weights and SSD-streamed Engram

  warpfront/hipfire#743 · Filter by author fivetide opened on Sep 10, 2026

• serve: nondeterministic model-load hang — CPU spin at random layer, GPU idle, frozen IO counters (gfx1100, qwen3.8:27b mq4r, 59669d9)

  warpfront/hipfire#739 · Filter by author nadaboulaares opened on Sep 9, 2026

• Create a new release version

  warpfront/hipfire#733 · Filter by author HUSRCF opened on Sep 8, 2026

• Qwen3.8-27B native MTP crashes when Q8 batched GEMM exceeds MAX_BATCH=64.

  warpfront/hipfire#732 · Filter by author Sergey7210 opened on Sep 7, 2026
  3

• Memory preflight OOM guard is unconditionally enabled — refuses valid loads on discrete-GPU dev boxes

  warpfront/hipfire#696 · Filter by author ghazni101 opened on Sep 4, 2026

• DFlash decode collapses below AR past ~10k ctx: verify-block target attention costs B×L per cycle

  warpfront/hipfire#693 · Filter by author Kaden-Schutt opened on Sep 3, 2026
  6

• JetSpec: Parallel Tree Drafting

  warpfront/hipfire#684 · Filter by author SlobCoder opened on Sep 2, 2026
  1

• EP MoE generate fails after a successful tp=4 load: 'EP arch mismatch (expected dense Qwen TP)' (ornith-1.5:35b-a3b-mq4r, master)

  warpfront/hipfire#683 · Filter by author Kaden-Schutt opened on Sep 2, 2026
  2

• Publish Gemma 4 E2B-it typed MQ4/MQ6 + PLE-Q8 artifact

  warpfront/hipfire#672 · Filter by author HUSRCF opened on Sep 1, 2026

• hipfire diag and serve always use GPU 0

  warpfront/hipfire#669 · Filter by author nwalk7800 opened on Aug 31, 2026
  5

• Tracking: upstream-native device mesh port replacing PR #527

  warpfront/hipfire#666 · Filter by author fivetide opened on Aug 31, 2026
  16

• hipfire bug reports — RX 7900 XTX / Windows

  warpfront/hipfire#655 · Filter by author drsohailimran opened on Aug 27, 2026
  5

• quantize/runtime: support PARO group_size=64 for SmolLM2-360M

  area:quantizekind:feat
  warpfront/hipfire#651 · Filter by author Kaden-Schutt opened on Aug 27, 2026

• kv: investigate FWHT residual QJL transform

  area:kernelskind:research
  warpfront/hipfire#650 · Filter by author Kaden-Schutt opened on Aug 27, 2026

• redline: screen next ordinary-AR retained-replay levers

  area:redlinekind:research
  warpfront/hipfire#649 · Filter by author Kaden-Schutt opened on Aug 27, 2026
  2