NInfer-all is maintained by iamwavecut. It grew from
Neroued/ninfer through the NInfer-3090 line that
Don-Chad started and
ashalliants/ninfer-3090 continued (v0.12.0 and the
multi-GPU pipeline stages, most of both by Warlax), and it carries
work from the forks and pull requests listed below. Ported commits keep their authors and name
their source commit, so git log is the complete record; this page keeps the credit visible.
| Contributor | Contribution | Pull request |
|---|---|---|
| airtonix | Linux and Docker build and release support | #1 |
| justinlime | NixOS build support | #5 |
| sry9681 | Device-wide GPU-memory startup fix | #6 |
| ColeWheatley | SM86 runtime/GDN fixes, ECC diagnostics, and GeForce-safe Docker support | #7 |
| iamwavecut | swscale destination-alignment JPEG safety fix | #11 |
| nasedkinpv | Tool-call parser crash fix | #12 |
| wmehanna | In-place system-turn rendering for Claude Code prefix reuse | #13 |
| mgscreativa | Thinking-budget boundary fix; premature KV-loan race diagnosis and gate fix behind a vision std::bad_alloc |
#156, #138 (kept in #145) |
| Source | Contribution |
|---|---|
| TertiumOrganum1/ninfer-3090 | the rk4v4-e8 KV cache, the 128x64 int8 tile for ternary prefill, tool-call recovery, shared-capture surplus handling, sm_120a on the mma.sync path |
| UDPSendToFailed/ninfer-4090 (UDPSendToFailed, Daniel Parker, Matt Anderson, keylimesoda, Sergiusz Michalik) | the E8-lattice KV codecs and rk2v4-e8, the whole-program CUDA build, shared-memory scale reads in INT8 attention, keys past 262,144 and four times the native window with YaRN, SFU sigmoid/SiLU/softplus, TCP_NODELAY, the WDDM evictable budget and DirectStorage restores on Windows, MTP windows up to 15, /metrics, /slots and /props, the compiled-in WebUI, split compilation |
| IMGillusion/ninfer-disk-kv | the disk KV tier and --context-cache-policy rolling |
| kido5217/frinfer (Fedor Suchkov) | token log probabilities on Chat Completions and Responses |
| MirkoCovizzi/ninfer-rtx5090-mobile | adaptive MTP |
| Wallawalla47/ninfer-custom (Ian Ranson, Duncan Betts) | the fast INT8 and NVFP4 prompt-attention kernels, the hybrid prefix cache, n-gram copy drafting (with remesis), recency eviction, Device KV lease growth, the Host cache budget, automatic long anchors, --assistant-prefill, grouped --help, levelled logs, programmatic dependent launches, split-KV attention for short prefill steps, ModelOpt and Quasar NVFP4/FP8 conversion, a native Windows build |
| gzenz/ninfer (Gideon Zenz, David Oelfke) | thorough admission search, value-aware demotion, concurrent prefill, worker out-of-memory recovery, the MTP attention window, post-thinking sampling, GET /stats and the tools/monitor dashboard, request-log rotation, derived session keys, CPU Vision, RoPE position interpolation, branch anchors, PackGQA |
| tmark00/ninfer | table-decoded E8 roots, the SM-count RMSNorm cutoff, the WebUI MCP proxy |
| pelebel, natpate | openable server URLs and CORS preflight echoes |
| Matt Anderson | wide Q5 split tiles and an idle Engine that fails only the request it cannot place |
| hyz615, lededev, NikolaFC | schema assertions in constrained decoding; the disk tier persisted on an orderly stop and MSVC-safe test builtins; Q8_0 ssm_alpha/ssm_beta in GGUF conversion |
| PR | Contribution | Author |
|---|---|---|
| #54 | the exception that terminates the process is named | Aleksandr Iakimov |
| #97, #297 | incremental native compilation in container builds; layout state survives an overflow | Duncan Betts |
| #162, #163 | llama.cpp model metadata on /v1/models; timings during prompt progress |
Héctor Ramón Jiménez |
| #197 | ignore_eos documented and pinned in the schema test |
Thireus |
| #199, #221, #264, #268, #273, #286-#290 | sparse-MoE Q4 quads, MTP topology classes, the NVFP4 SwiGLU tail tile, the gated small-T attention epilogue, NVFP4 expert banks and the qwen3_6_35b_a3b_nvfp4 recipe |
Mykhailo Dementii |
| #282 | GGUF as a conversion source | giveen |
| #284 | the qwen3_8_27b_q6 recipe |
bingchengcc |
| #294 | structured output through xgrammar | Andrey Shvartsman |
| #295 | reasoning.summary and encrypted reasoning content |
Mac |
| #299 | a repeated tool-call parameter keeps its last value | adubkov |
| #300 | the agent-harness XML tool-call forms | pkochubey, Ian Ranson |
| #304, #305, #307 | the speed-of-light estimator, fused attention RMSNorm with NVFP4 quantization, single-pass target log probabilities | Duncan Betts |
| #309 | quoted reasoning closes and tool-call markers | Fedor Suchkov |
| #311 | pipelined Q4/Q5 groups for the attention input | MOVIBALE |
| #316 | Copilot tool shapes and longer tool names | Damian Sromek |
Upstream master changes ported here keep their authors: the unified Linear templates, the
two-stage GDN prefill, parallel query tiles, MX FP8 MMA and Minnnn's column-band routing.
- The Ternary Bonsai 2 27B artifact combines PrismML's ternary weights, ProCreations' MTP head and DFlash2 adapter, and Qwen's checkpoint, all Apache-2.0.
- The Qwen3.8-Flash-Next artifacts import ISTA-DASLab's GSQ-RCO GGUF releases and Unsloth's MTP block, under the Qwen Community License 1.0 of the model.
- Qwen3.8-Flash-Next support was designed against the transformers, vLLM and llama.cpp implementations of the model and Strata (MIT, copyright 2026 Niko1221 and contributors), whose transcription served as a reference for the n-gram hash, the expert cache and its measured pitfalls.
- llama.cpp's GGUF block kernels are vendored unmodified in
third_party/ggml-quantsunder their MIT license.