Neroued/ninfer

Issues somebody may already be working on are listed last. List everything by date

• [Bug] Anthropic-style thinking={type:disabled} combined with output_config.effort is rejected with HTTP 400 "conflicting template option"

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 3/5 1-2 days Newbie friendliness 57/100

  Neroued/ninfer#441 · Opened Nov 5, 2026

  Maintainers usually reply within 9 days

• [Bug] Qwen3.8 chat template only accepts xhigh/medium/low reasoning_effort — serve-layer 7-tier effort gets HTTP 400 "invalid prompt"

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 2/5 1-3 hours Newbie friendliness 67/100

  Neroued/ninfer#440 · Opened Nov 5, 2026

  Maintainers usually reply within 9 days

• Startup crash on RTX 5090D (sm_120a) at HEAD a754fcf3: decode_graph.cpp:19 CUDA_CHECK(cudaGraphAddChildGraphNode) failed: cudaErrorInvalidValue

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 18/100

  Neroued/ninfer#439 · 2 comments · Opened Nov 3, 2026

  Maintainers usually reply within 9 days

• Support A100 gpus

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#438 · Opened Nov 3, 2026

  Maintainers usually reply within 9 days

• [Perf] NVFP4 Compact64-T6: reduce shared-memory wavefronts and investigate short-context decode consistency

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 8/100

  Neroued/ninfer#437 · Opened Nov 2, 2026

  Maintainers usually reply within 9 days

• unsloth Dynamic V3.1 DS4.1 only needs 64G RAM

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 8/100

  Neroued/ninfer#435 · Opened Nov 1, 2026

  Maintainers usually reply within 9 days

• [Bug] Requests stall and never make progress

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 12/100

  Neroued/ninfer#434 · 6 comments · 1 reaction · Opened Oct 31, 2026

  Maintainers usually reply within 9 days

• [Feature] Qwen3.5-9B geometry support on master (verified implementation, branch available)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 38/100

  Neroued/ninfer#433 · Opened Oct 31, 2026

  Maintainers usually reply within 9 days

• [Feature Request] Add a configurable GPU utilization ceiling to leave headroom for desktop/office use

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#431 · Opened Oct 30, 2026

  Maintainers usually reply within 9 days

• [Feature Request] Support nullable string parameters in strict tool calling (OpenAI compatibility)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  Neroued/ninfer#425 · Opened Oct 26, 2026

  Maintainers usually reply within 9 days

• serve prefill ~22% below CLI at identical engine config (RTX 5090, nvfp4)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 43/100

  Neroued/ninfer#424 · 1 comment · Opened Oct 26, 2026

  Maintainers usually reply within 9 days

• Optional bounded idle keep-warm for interactive single-GPU serving

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#423 · Opened Oct 25, 2026

  Maintainers usually reply within 9 days

• KV history growth under-reserved when a replay write crosses a page boundary

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 1-2 days Newbie friendliness 54/100

  Neroued/ninfer#419 · Opened Oct 22, 2026

  Maintainers usually reply within 9 days

• Anyone know if Ninfer rotates activations?

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 3/5 1-2 days Newbie friendliness 25/100

  Neroued/ninfer#418 · 2 comments · Opened Oct 22, 2026

  Maintainers usually reply within 9 days

• Long image chats get stuck with 400 media_budget_exceeded — the server should drop the oldest images automatically

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  Neroued/ninfer#417 · Opened Oct 22, 2026

  Maintainers usually reply within 9 days

• Concurrent requests sharing a prefix crash the engine: “KV write exceeds its page reservation” (regression after 4de9e425)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  Neroued/ninfer#414 · 1 comment · Opened Oct 21, 2026

  Maintainers usually reply within 9 days

• Support EAGLE 4 Drafter

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 35/100

  Neroued/ninfer#413 · 1 comment · Opened Oct 21, 2026

  Maintainers usually reply within 9 days

• Report whether the loaded template supports tool calls

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  Neroued/ninfer#400 · Opened Oct 15, 2026

  Maintainers usually reply within 9 days

• Grammar violation on one request kills the whole engine (`selected token violates the request grammar`) — 503 until process restart

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#399 · Opened Oct 15, 2026

  Maintainers usually reply within 9 days

• Context cache: once full, a plain multi-turn chat resumes only from the engine-discovered system prefix

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 42/100

  Neroued/ninfer#394 · 1 comment · Opened Oct 11, 2026

  Maintainers usually reply within 9 days

• Host memory duplication after system suspend/resume (CUDA context re-initialization leaks pinned RAM)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  Neroued/ninfer#393 · Opened Oct 10, 2026

  Maintainers usually reply within 9 days

• Ninfer process crashes on the second GPU in multi-ninfer, multi-GPU setup

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 42/100

  Neroued/ninfer#391 · 2 comments · Opened Oct 10, 2026

  Maintainers usually reply within 9 days

• Experimental DGX Spark / GB10 (sm_121a) build and inference support

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#390 · 2 comments · Opened Oct 10, 2026

  Maintainers usually reply within 9 days

• Host context arena: first-fit extents shared by three allocation sizes fragment — make it a block pool

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#379 · Opened Oct 8, 2026

  Maintainers usually reply within 9 days

• A complete tool call emitted inside an unclosed thinking span is never recovered

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  Neroued/ninfer#377 · 1 comment · Opened Oct 7, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 15/100

  Neroued/ninfer#375 · 1 reaction · Opened Oct 6, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 43/100

  Neroued/ninfer#374 · Opened Oct 5, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#371 · Opened Oct 5, 2026

  Maintainers usually reply within 9 days

• Performance: allow parallel CPU workers during weight conversion

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 38/100

  Neroued/ninfer#370 · Opened Oct 5, 2026

  Maintainers usually reply within 9 days

• Nix Flake

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 45/100

  Neroued/ninfer#367 · Opened Oct 4, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 15/100

  Neroued/ninfer#362 · Opened Oct 3, 2026

  Maintainers usually reply within 9 days

• Encoded DFlash2 component composition for compatible v3 artifacts

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 28/100

  Neroued/ninfer#361 · Opened Oct 3, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  Neroued/ninfer#357 · Opened Oct 2, 2026

  Maintainers usually reply within 9 days

• NVFP4 tiled prefill under-occupies the device

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 35/100

  Neroued/ninfer#356 · Opened Oct 1, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 18/100

  Neroued/ninfer#350 · Opened Oct 1, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 52/100

  Neroued/ninfer#348 · 2 comments · Opened Sep 30, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  Neroued/ninfer#339 · Opened Sep 30, 2026

  Maintainers usually reply within 9 days

• Use Zluda for AMD GPUs

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 20/100

  Neroued/ninfer#336 · Opened Sep 30, 2026

  Maintainers usually reply within 9 days

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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 42/100

  Neroued/ninfer#333 · Opened Sep 29, 2026

  Maintainers usually reply within 9 days

• Suggestion: Port the web panel from strata

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#331 · 2 comments · Opened Sep 28, 2026

  Maintainers usually reply within 9 days

• 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

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  Neroued/ninfer#330 · Opened Sep 28, 2026

  Maintainers usually reply within 9 days

• perf(ops): reciprocal-multiply in NVFP4 activation quantize on linear MMA route

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 38/100

  Neroued/ninfer#327 · Opened Sep 27, 2026

  Maintainers usually reply within 9 days

• Proposal: Gemma 4 31B (dense) as a contributed second model family. Would you accept it, and on what terms?

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 20/100

  Neroued/ninfer#326 · 2 comments · 2 reactions · Opened Sep 27, 2026

  Maintainers usually reply within 9 days

• v3 jinja chat template rejects 'tool' role messages, breaking OpenAI-compatible agents

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  Neroued/ninfer#325 · 2 comments · 1 reaction · Opened Sep 26, 2026

  Maintainers usually reply within 9 days

• Release log & version tagging

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 35/100

  Neroued/ninfer#323 · 3 reactions · Opened Sep 26, 2026

  Maintainers usually reply within 9 days

• Fixes for prefix caching, tool calling and other issues

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 10/100

  Neroued/ninfer#322 · 1 comment · 2 reactions · Opened Sep 26, 2026

  Maintainers usually reply within 9 days

• MCP tool calls are output as plain text; this involves the Hermes and Claude agent frameworks.

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 35/100

  Neroued/ninfer#321 · Opened Sep 26, 2026

  Maintainers usually reply within 9 days

• Xid 79 "GPU has fallen off the bus" during prefill — `activate()` still publishes the KV block table on the legacy stream (8 occurrences in 4.2 days)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 4/5 3-5 days Newbie friendliness 25/100

  Neroued/ninfer#320 · 3 comments · 2 reactions · Opened Sep 26, 2026

  Maintainers usually reply within 9 days

• Two-GPU tensor parallelism for 16 GB boards — would you consider it, or parts of it?

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 5/5 Over a week Newbie friendliness 25/100

  Neroued/ninfer#313 · 1 comment · Opened Sep 24, 2026

  Maintainers usually reply within 9 days

• Model request: ThinkingCap Qwen 3.8 27B

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel

  Difficulty 3/5 1-2 days Newbie friendliness 48/100

  Neroued/ninfer#306 · 7 comments · Opened Sep 23, 2026

  Maintainers usually reply within 9 days