Issues

• Feature Request: Make automatic KV capacity headroom configurable

  Neroued/ninfer#376 · Filter by author DuncanBetts opened on Oct 6, 2026

• KV precision tail (--kv-tail-tokens): port results and measured data (fork link, no code proposed)

  Neroued/ninfer#375 · Filter by author qwased opened on Oct 6, 2026

• Fused residual rounding in NVFP4 A4 / FP8 A8 linear_add depends on tile occupancy (FullTokens contracts to an FMA, tails do not)

  Neroued/ninfer#374 · Filter by author ValerioDolci opened on Oct 5, 2026

• request log: model_thinking_tokens stays 0 unless a thinking budget is set

  Neroued/ninfer#373 · Filter by author ValerioDolci opened on Oct 5, 2026

• Materializer assumes direct-I/O staging slots are 4096-byte aligned (small pinned allocations are not)

  Neroued/ninfer#372 · Filter by author ValerioDolci opened on Oct 5, 2026

• A cancelled follow-up discards the consumed continuation; a different follow-up (or an edited last message) re-prefills the whole conversation

  Neroued/ninfer#371 · Filter by author ValerioDolci opened on Oct 5, 2026

• Performance: allow parallel CPU workers during weight conversion

  Neroued/ninfer#370 · Filter by author Xtravaganz opened on Oct 5, 2026

• Proposal: land the #142 automatic shared-prefix candidate (rebase of #152 onto current master)

  Neroued/ninfer#369 · Filter by author kaushikvira opened on Oct 5, 2026

• Nix Flake

  Neroued/ninfer#367 · Filter by author oakshortbow opened on Oct 4, 2026

• Context cache rewrite and request preemption are now on master

  Neroued/ninfer#366 · Filter by author Neroued opened on Oct 4, 2026
  13

• NVFP4 W4A4 linear_add [5120,17408]: M32N64 launches 80 CTAs at verify widths, half a 5090 idle; M32N32 (4 stages) is -8..-13% at T=8..32, bit-identical

  Neroued/ninfer#364 · Filter by author CaptainArni opened on Oct 4, 2026
  1

• Opt-in active retrieval KV windows with Program-owned Host history

  Neroued/ninfer#362 · Filter by author qzshch opened on Oct 3, 2026

• Encoded DFlash2 component composition for compatible v3 artifacts

  Neroued/ninfer#361 · Filter by author qzshch opened on Oct 3, 2026

• StateImage transfer byte counters accumulate image counts instead of payload bytes

  Neroued/ninfer#360 · Filter by author qzshch opened on Oct 3, 2026

• Engine crash (0xC0000409 / 0xC0000005) in hybrid Host-tier restore path...

  Neroued/ninfer#357 · Filter by author WormholeFusion opened on Oct 2, 2026

• NVFP4 tiled prefill under-occupies the device

  Neroued/ninfer#356 · Filter by author DuncanBetts opened on Oct 1, 2026

• Proposal: L3 disk KV-cache tier for multi-turn serving (content-addressed pages, proactive per-turn checkpoints, seam anchors) - production-validated on RTX 5090

  Neroued/ninfer#350 · Filter by author IMGillusion opened on Oct 1, 2026

• OpenAI Codex CLI against /v1/responses: three hard rejections (web_search tool, reasoning.summary, include) — gateway workaround + feature request

  Neroued/ninfer#348 · Filter by author alexveli1 opened on Sep 30, 2026
  2

• HTTP listener sets only SO_REUSEPORT, so SO_REUSEADDR servers cannot bind the port for ~1 min after NInfer stops (and vice versa)

  Neroued/ninfer#341 · Filter by author DuncanBetts opened on Sep 30, 2026
  1

• serve wedges with 503 for all requests after Qwen3.5 resource subtraction underflow (checked_resource_difference, context_work.cpp:222) under --max-concurrency 2

  Neroued/ninfer#339 · Filter by author jwin-ling opened on Sep 30, 2026

• Use Zluda for AMD GPUs

  Neroued/ninfer#336 · Filter by author fayakun-it-consulting opened on Sep 30, 2026

• Proposal: replace the checkpoint-catalog prefix cache with a content-addressed KV-block and state-snapshot cache

  Neroued/ninfer#334 · Filter by author Wallawalla47 opened on Sep 30, 2026
  1

• cudaErrorLaunchTimeout crash during DFlash2 prefill immediately after ~98%-cache-replay prefill (RTX 5090)

  Neroued/ninfer#333 · Filter by author na8r opened on Sep 29, 2026

• Suggestion: Port the web panel from strata

  Neroued/ninfer#331 · Filter by author leonbollerup opened on Sep 28, 2026
  2

• NVFP4 A16 sliced-K at T=3..16 is MIO-throttled on 70-SM GPUs: two schedule knobs give −8…−10 % on RTX 5070 Ti, bit-identical

  Neroued/ninfer#330 · Filter by author ValerioDolci opened on Sep 28, 2026