Issues

• [Question] Has a HRX backend instead of HIP been considered?

  gpu

  Low-prioFilter by Priority: Low-prio

  High-effortFilter by Effort: High-effort
  Feature

  gufo-org/gufo#652 · Filter by author Mushoz opened on Oct 18, 2026
  1

• inference_backend_gpu_test fails on main: the min checkpoint step from #348 skips the root it restores

  Qwen3.8 27bFilter by Model: Qwen3.8 27b

  High-prioFilter by Priority: High-prio

  Low-effortFilter by Effort: Low-effort
  Bug

  gufo-org/gufo#648 · Filter by author headbouyJB opened on Oct 18, 2026
  1

• gfx1150 (Strix Point) field report: build support + six performance observations

  gufo-org/gufo#634 · Filter by author winaes13-ux opened on Oct 17, 2026
  3

• serve: on-demand reproducible hard wedge on 0.17.0 — long request frozen in component-cache refill (processed=61384 byte-identical twice, or processed=0 right after reuse=ram restore) while the concurrent short request's decode dies at dec=1; no device_lost; SIGTERM honored

  cacheserving

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  High-prioFilter by Priority: High-prio

  Medium-effortFilter by Effort: Medium-effort
  Bug

  gufo-org/gufo#631 · Filter by author titonio opened on Oct 17, 2026
  2

  

• DeepSeek dspark: greedy peer in batched-w2 skips EOS and repeats its answer next to a stop-sequence request (intermittent)

  DeepSeek v4 flashFilter by Model: DeepSeek v4 flash

  Low-prioFilter by Priority: Low-prio

  Medium-effortFilter by Effort: Medium-effort
  Bug

  gufo-org/gufo#612 · Filter by author francescobozzo opened on Oct 15, 2026

• docs: Add community Flash-Next Uncensored quant

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  Low-prioFilter by Priority: Low-prio

  Low-effortFilter by Effort: Low-effort
  Enhancement

  gufo-org/gufo#604 · Filter by author iwr-redmond opened on Oct 15, 2026

• Make loader concurrency (kReaders) configurable at runtime

  serving

  Low-prioFilter by Priority: Low-prio

  Low-effortFilter by Effort: Low-effort
  Feature

  gufo-org/gufo#601 · Filter by author ablindman44 opened on Oct 14, 2026

• Feature request: optional post-load VRAM read-back verification

  serving

  Low-prioFilter by Priority: Low-prio

  Low-effortFilter by Effort: Low-effort
  Feature

  gufo-org/gufo#600 · Filter by author ablindman44 opened on Oct 14, 2026

• Intermittent data-fabric sync-flood reset during model load on Strix Halo (gfx1151), Qwen3.8-Flash-Next

  memory

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  Medium-prioFilter by Priority: Medium-prio

  Medium-effortFilter by Effort: Medium-effort
  Bug

  gufo-org/gufo#599 · Filter by author ablindman44 opened on Oct 14, 2026
  4

  

• Flash-Next component cache: disk persistence silently stops writing, then a GB/s background chunk-read loop wedges decode (<1 t/s) and stop ignores SIGTERM

  cache

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  Medium-prioFilter by Priority: Medium-prio

  Medium-effortFilter by Effort: Medium-effort
  Bug

  gufo-org/gufo#597 · Filter by author crawlingweapon opened on Oct 14, 2026
  

• serve: zero-token replies (finish=stop, ttft 0, plan=serial-fallback) for all novel prompts after sustained concurrent load — 0.13.0 and 0.15.0

  cache

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  Medium-prioFilter by Priority: Medium-prio
  Bug

  gufo-org/gufo#595 · Filter by author JohannesBertens opened on Oct 14, 2026
  1

  

• Speed up and fix the Flash-Next functional suites

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  DeepSeek v4 flashFilter by Model: DeepSeek v4 flash

  Qwen3.8 27bFilter by Model: Qwen3.8 27b

  Low-prioFilter by Priority: Low-prio

  +1Show 1 more
  Enhancement

  gufo-org/gufo#593 · Filter by author fedeizzo opened on Oct 14, 2026
  8

  

• perf(npu): overlap deferred prefill projections on XDNA2 (evaluate the mikealanni/gufo fork)

  npu

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  Low-prioFilter by Priority: Low-prio

  High-effortFilter by Effort: High-effort
  Enhancement

  gufo-org/gufo#558 · Filter by author fedeizzo opened on Oct 11, 2026
  6

• serve: Claude Code system entries are hoisted, so every /v1/messages turn re-prefills the history

  cache

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  DeepSeek v4 flashFilter by Model: DeepSeek v4 flash

  Qwen3.8 27bFilter by Model: Qwen3.8 27b

  Low-prioFilter by Priority: Low-prio

  +1Show 1 more
  Enhancement

  gufo-org/gufo#522 · Filter by author jabata opened on Oct 10, 2026
  

• serve: resuming captured decoders delays concurrent peers' prefill

  cacheserving

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  DeepSeek v4 flashFilter by Model: DeepSeek v4 flash

  Qwen3.8 27bFilter by Model: Qwen3.8 27b

  Medium-prioFilter by Priority: Medium-prio

  +1Show 1 more
  Enhancement

  gufo-org/gufo#455 · Filter by author fedeizzo opened on Oct 6, 2026
  3

• Flash-Next: support Unsloth's smaller quants (UD-Q3_K_XL / UD-Q2_K_XL) so 64 GB and 96 GB Strix Halo owners can run it

  quantization

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  Low-prioFilter by Priority: Low-prio

  High-effortFilter by Effort: High-effort
  Feature

  gufo-org/gufo#454 · Filter by author sebt3 opened on Oct 6, 2026

• Implement DeepSeek V4 Flash Vision-Exp image input

  vision

  DeepSeek v4 flashFilter by Model: DeepSeek v4 flash

  Low-prioFilter by Priority: Low-prio

  Medium-effortFilter by Effort: Medium-effort
  Feature

  gufo-org/gufo#433 · Filter by author spencergilbert opened on Oct 4, 2026
  4

• serve: memory-bounded expert streaming from SSD (--expert-streaming + --expert-cache), model-specific

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  DeepSeek v4 flashFilter by Model: DeepSeek v4 flash

  Low-prioFilter by Priority: Low-prio

  Medium-effortFilter by Effort: Medium-effort
  Enhancement

  gufo-org/gufo#427 · Filter by author Farenheith opened on Oct 4, 2026
  8

• serve: /v1/messages needs streaming and tools for Claude Code sessions

  serving

  Low-prioFilter by Priority: Low-prio

  Low-effortFilter by Effort: Low-effort
  Enhancement

  gufo-org/gufo#417 · Filter by author fedeizzo opened on Oct 4, 2026
  1

• Auto --cache-ram-bytes (up to 32 GiB) exhausts the KFD memory limit on unified-memory APUs when other models are co-resident

  cache

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  DeepSeek v4 flashFilter by Model: DeepSeek v4 flash

  Qwen3.8 27bFilter by Model: Qwen3.8 27b

  Qwen TTSFilter by Model: Qwen TTS

  +3Show 3 more
  Enhancement

  gufo-org/gufo#387 · Filter by author Nawart opened on Oct 2, 2026
  8

• Qwen3.8-Flash-Next: opt-in YaRN context extension past 262,144 (scope check)

  memory

  Qwen3.8 Flash NextFilter by Model: Qwen3.8 Flash Next

  Low-prioFilter by Priority: Low-prio

  High-effortFilter by Effort: High-effort
  Feature

  gufo-org/gufo#287 · Filter by author kh0pper opened on Sep 26, 2026
  1

  

• Feature request: Support tensor parallelism with RDMA

  Low-prioFilter by Priority: Low-prio

  High-effortFilter by Effort: High-effort
  Feature

  gufo-org/gufo#258 · Filter by author neuhaus opened on Sep 24, 2026
  24

• native windows support

  Low-prioFilter by Priority: Low-prio

  High-effortFilter by Effort: High-effort
  Feature

  gufo-org/gufo#254 · Filter by author ElderOrb opened on Sep 24, 2026
  7

• Track official MiniMax H3 sparse-attention support

  gpuminimax-h3roadmap-placeholdervideo

  Minimax H3Filter by Model: Minimax H3

  Low-prioFilter by Priority: Low-prio

  High-effortFilter by Effort: High-effort
  Enhancement

  gufo-org/gufo#179 · Filter by author francescobozzo opened on Aug 21, 2026 · 13 - Video Generation
  1

• Add MiniMax H3 ordered Ref2VA image, video, and audio references

  audiogpuminimax-h3roadmap-placeholdervideovision

  Minimax H3Filter by Model: Minimax H3

  Low-prioFilter by Priority: Low-prio

  High-effortFilter by Effort: High-effort
  Enhancement

  gufo-org/gufo#178 · Filter by author francescobozzo opened on Aug 21, 2026 · 13 - Video Generation