{"meta":{"count":2,"page":1,"pages":1,"per_page":12,"applied":{"engine":null,"model":null,"quant":null,"storage":null,"hardware":["rtx-4090","dgx-station-gb300-x2"],"generation":null,"task":null,"tier":null,"tags":[],"vram":null,"search":null,"sort":"updated","featured":false},"origin":"https://localairecipes.com","openapi":"/openapi.json","updated_at":"2026-11-03T06:10:19+00:00"},"recipes":[{"slug":"qwen3-8-flash-next-iq3-s-mtp-on-1x-rtx-4090-llama-cpp-tandem-cpp","title":"Qwen3.8-Flash-Next IQ3_S + MTP on 1x RTX 4090 \u2014 llama.cpp + tandem.cpp MoE expert cache (ewhacc)","summary":"MoE expert-cache offload for llama.cpp via the tandem.cpp patch set: hot experts stay in a GPU expert cache, missed experts are split between CPU compute and an upload into the cache, sized from measured CPU and PCIe speeds; single-user decode is the target. Author's GSM8K first-200 decode numbers on the consumer PC (i5-12400F, 64 GB DDR4, RTX 4090 24 GB): Qwen3.8-Flash-Next IQ3_S + MTP 97.3 tok/s greedy and 96.1 sampled vs llama.cpp f5222f25c 29.8/29.4 (3.3x); gpt-oss-120b 51.5/48.0. REPRO CAVEAT: all numbers are author-self-reported on a repo <24h old with one commit day of history (initial commit 2026-11-02, no releases/tags); nothing reproduced on-box or by a third party - evidence is claimed, tier community. Requires CUDA 13.0-13.4 on Linux (sm_89/sm_120) and 64 GB host RAM for the 78 GiB IQ3_S weights; 1.4-2.1x gains on other MoE models that don't fit VRAM.","tier":"community","beta":false,"generation":"ada","decode_tok_s":73.7999999999999971578290569595992565155029296875,"prefill_tok_s":0,"updated_at":"2026-11-02 09:53:20.10132-07","commit_sha":"543df76dcb783c7c365b8ea88a2d81ed9d36de76","content_hash":"fd8e878ab3aa782deec1708f8b0913a0ee5ea7a7","votes":0,"credit":[{"name":"ewhacc - recipe, patch set and measurements","url":"https://github.com/ewhacc/tandem.cpp"},{"name":"llama.cpp / ggml (MIT)","url":"https://github.com/ggml-org/llama.cpp"},{"name":"MoE expert-split rule adapted from Strata (Niko1221)","url":"https://github.com/Niko1221/Strata"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-flash-next-iq3-s-mtp-on-1x-rtx-4090-llama-cpp-tandem-cpp.json","alternate":"/r/qwen3-8-flash-next-iq3-s-mtp-on-1x-rtx-4090-llama-cpp-tandem-cpp"}},{"slug":"deepseek-v4-1-flash-1017-fp8-tp1xpp2-dspark-2x-dgx-station-gb300","title":"DeepSeek V4.1 Flash 1017 FP8 TP1xPP2 DSpark 2x DGX Station GB300 RoCE vLLM 0.31 overlay","summary":"Pipeline-parallel TP1xPP2 + DSpark serving of DeepSeek V4.1 Flash (1017 official FP8 weights) across two DGX Station GB300 nodes over 400G CX8 RoCE. PP2 splits the 40 layers 20/20 and removes all 84 per-step all-reduces, replacing them with two point-to-point hops: single-stream temp-0 391 tok/s vs TP2 stock 290.5 (+34%), step 12.3 ms -> ~9.4 ms; aggregate soak 323/406/595/974 tok/s at C1/C2/C4/C8 (3.01x C1); GSM8K 97/100, 0 errors; 1M-token context, FP8 KV, Engram host table pinned in Grace RAM on rank 0. Ships the five-file vLLM v0.31.0 source-fix overlay (bind-mounted over stock vllm/vllm-openai:v0.31.0) that makes DSpark work under PP, plus preflight, launch scripts and soak/receipt verifiers. ALL perf numbers author-measured on their own pair (2026-10-25) - not reproduced locally; site has no GB300 hardware. Recipe by ebfio (port by Iris, a GLM-5.3 model, with Eduardo & Kestrel).","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":357,"prefill_tok_s":0,"updated_at":"2026-10-27 16:38:42.712204-06","commit_sha":"2acdf45851ff045dc48344004a9ec5fde3af3de4","content_hash":"739e619707a224aa4cd8782dc5ed8d43f68dce86","votes":0,"credit":[{"name":"ebfio - DeepSeek V4.1 Flash PP2+DSpark GB300 pair recipe (port by Iris on 4x RTX Pro 6000, with Eduardo & Kestrel)","url":"https://github.com/ebfio/deepseek-v41-flash-pp2-pair"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/deepseek-v4-1-flash-1017-fp8-tp1xpp2-dspark-2x-dgx-station-gb300.json","alternate":"/r/deepseek-v4-1-flash-1017-fp8-tp1xpp2-dspark-2x-dgx-station-gb300"}}]}