{"meta":{"count":15,"page":1,"pages":2,"per_page":12,"applied":{"engine":null,"model":null,"quant":null,"storage":null,"hardware":["170hx-x3","dgx-spark","170hx"],"generation":null,"task":null,"tier":null,"tags":[],"vram":null,"search":null,"sort":"updated","featured":false},"origin":"https://localairecipes.com","openapi":"/openapi.json","updated_at":"2026-10-31T16:42:02+00:00"},"recipes":[{"slug":"qwen3-8-flash-next-w4a16-autoround-dgx-spark-vllm-mtp4-iter7g","title":"Qwen3.8-Flash-Next W4A16-AutoRound DGX Spark vLLM MTP4 Iter7g","summary":"Recipe by kyaky. Qwen3.8-Flash-Next W4A16-AutoRound hybrid (INT4 experts + FP8 side layers + INT8 head; 51B-token n-gram PLE table on NVMe mmap) on one DGX Spark (GB10, 121 GB unified memory) on a verified-deterministic vLLM floor (image qwen38-flash-dgx:iter7g-20261015 on the digest-pinned vllm/vllm-openai:qwen38-flash-next base): MoE align arrival-order shim, drafter-Marlin atomicAdd fix, pinned triton config, compile-cache mount. Measured line v19a-w5 = MTP=4 window-5 block verification: 76.8 tok/s single-stream, 261.2 tok/s aggregate at 8 streams (+9.2% N1 / +4.8% N8 over MTP=3, two boots per arm, same image). Promoted serving line v20-fp8: fp8 KV cache on a 24 GB pool at native 262,144 context, 1,291,120-token capacity. Chinese +31.3% via a 25,862-id (39.5%) Han draft-vocab rebalance (prose-zhen 34.8 to 45.7 tok/s). Stochastic-correctness gate 37/37 vs bit-exact fp64 oracle; per-round raw data + scripts ship in-repo; copy-agentic-1 failure disclosed.","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":61.2999999999999971578290569595992565155029296875,"prefill_tok_s":0,"updated_at":"2026-10-29 21:32:28.107205-06","commit_sha":"dc8982acabf5e6b6417f4871ad5149d61f01755d","content_hash":"e6a2cfdf4903c7ca1b80df3d0f496d3ca3f13840","votes":0,"credit":[{"name":"kyaky","url":"https://github.com/kyaky/qwen3.8-flash-next-gb10-ultrafast"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-flash-next-w4a16-autoround-dgx-spark-vllm-mtp4-iter7g.json","alternate":"/r/qwen3-8-flash-next-w4a16-autoround-dgx-spark-vllm-mtp4-iter7g"}},{"slug":"qwen3-8-flash-next-nvfp4-dgx-spark-gb10-sglang-nextn-mtp","title":"Qwen3.8-Flash-Next NVFP4 DGX Spark GB10 SGLang NEXTN MTP","summary":"Production single-node DGX Spark (Grace Blackwell GB10, SM121, 128GB unified memory) recipe for Qwen3.8-Flash-Next on SGLang: ModelOpt NVFP4 quantization with zero-copy NVMe mmap of the 47.7 GiB ple_table.bin via --ple-offload-embedding (eliminates 51 GiB of unified-memory footprint), custom Triton FP4 prefill kernels + trtllm_mha decode kernels on SM121, NEXTN MTP (3 verification steps / 4 draft tokens) lifting output to 110-152 tok/s, GDN+QSA hybrid (131,072-token served window, 262k native capacity), 96-slot mamba recurrence buffers, sharp froggeric v22.1 Jinja chat template (-18% response length), persistent Triton/Inductor SM121 JIT caches on NVMe, 8-worker parallel NVMe pre-warm (206 shards in 17s) + 16-core Grace OMP. Author-measured vs same-box 27B dense baseline: 110.4-152.8 tok/s decode w/ MTP (baseline 12.1 base / 21.5 MTP), single-stream CoT 25.1-28.8 tok/s (2.62x speculative gain), TTFT ~347ms radix hit / ~465ms cold, 113 GiB / 121 GiB RAM used (~8 GiB headroom), 103.5 GiB","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":110.400000000000005684341886080801486968994140625,"prefill_tok_s":0,"updated_at":"2026-10-29 21:32:27.869274-06","commit_sha":"2c5220aa85b5379d7373abd0aa14bb13b1ec993f","content_hash":"bf0c02f95b99494d6720a515e2f8f438f126bc87","votes":0,"credit":[{"name":"Michael Bemler (Bemler Labs)","url":"https://github.com/bemlerlabs/qwen3.8-flash-next-dgx-spark-sglang"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-flash-next-nvfp4-dgx-spark-gb10-sglang-nextn-mtp.json","alternate":"/r/qwen3-8-flash-next-nvfp4-dgx-spark-gb10-sglang-nextn-mtp"}},{"slug":"qwen3-8-flash-next-fp4-local-1x-dgx-spark-vllm-b12x","title":"Qwen3.8-Flash-Next FP4-Local 1x DGX Spark vLLM b12x","summary":"Recipe by sojufx. Serves NVIDIA's nvidia/Qwen3.8-Flash-Next-FP4-Local (96 GiB NVFP4 quantization-aware-distilled checkpoint in a new compressed NVFP4 scale format) on one DGX Spark (GB10) via the digest-pinned eugr/spark-vllm-b12x community container (reported vLLM 0.1.dev21830+g8244d1c44.d20261015) - the model card requires community b12x containers until the format lands upstream. FP8 KV cache, exact MTP K=3 with standard rejection sampling, 262,144-token context, 4 scheduler slots, 4096-token batched prefill, disk-backed io_uring PLE, prefix cache with Mamba align, vision plus qwen3_xml tool calling, thinking off by default, localhost-only 127.0.0.1:8001 OpenAI-compatible backend with a deliberate no-replace/no-port-take guard. Measured 2026-10-15/16 on the author's Spark: 66.2 tok/s single-stream sparkDash counting decode, 166.4 tok/s at C4, 41-54 tok/s single-stream on the four-prompt wall-clock suite - different measurement bases, not interchangeable.","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":66.2000000000000028421709430404007434844970703125,"prefill_tok_s":0,"updated_at":"2026-10-29 21:32:27.177524-06","commit_sha":"5fc0d101d8e9b1c6a1419f3d465e21649f4ecbad","content_hash":"be42677d0cc5ca7f71fe72210e756f1622fe3604","votes":0,"credit":[{"name":"sojufx","url":"https://github.com/sojufx/Qwen3.8-Flash-Next-FP4-Local-DGX-Spark-Recipe"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-flash-next-fp4-local-1x-dgx-spark-vllm-b12x.json","alternate":"/r/qwen3-8-flash-next-fp4-local-1x-dgx-spark-vllm-b12x"}},{"slug":"qwen3-8-flash-next-3bpw-fp8-ple-1x-cmp-170hx-dflash2-recoverssm","title":"Qwen3.8-Flash-Next 3bpw FP8-PLE 1x CMP 170HX DFlash2 RecoverSSM","summary":"Qwen3.8-Flash-Next (125B MoE) 3 bpw AutoRound experts on ONE CMP 170HX card with the full 262,144-token context: 197.8/296.4/135.5 tok/s one user (structured/code/prose), 371 tok/s editing code, 799.5/791.3/586.4 across eight, up to 48 requests at once - 1,384 tok/s total with the default DFlash2 drafter (1,403 with the checkpoint's MTP head), 4,243 tok/s cold prefill, 292,784-token KV pool. Serves klee100's 3 bpw AutoRound checkpoint (3-bit g128 / 2-bit g64 experts, 47 GiB on card) with halt95's FP8 n-gram table in HOST RAM (48 GiB, hard-linked at build) on Morrowmake's vllm-cmp170hx fork @ 18305948acba (stock release image, digest-pinned) so the whole decode step is one CUDA graph on a 2016 dual-Xeon host. inco.ai's DFlash2 drafter (7/5/3 by request count) + the recipe's RecoverSSM + BF16 GDN-state patches (9 files mounted over the image) fit 48 concurrent requests on one card. Measured 2026-10-23 under the MiaAI-Lab protocol.","tier":"community","beta":false,"generation":"ampere","decode_tok_s":197.80000000000001136868377216160297393798828125,"prefill_tok_s":0,"updated_at":"2026-10-29 21:32:26.955646-06","commit_sha":"af7d7dde12376551476c8b9e36cd72c68dc32408","content_hash":"c82e80ffa9955d905b61bda381cf7d89db1350ba","votes":0,"credit":[{"name":"ElectroHobo","url":"https://github.com/ElectroHobo/qwen38-flash-next-3bpw-1x-cmp170hx-recipe"},{"name":"Morrowmake (vllm-cmp170hx engine + image + Qwen EXL3 recipe scripts this descends from, MIT)","url":"https://github.com/Morrowmake/vllm-cmp170hx"},{"name":"klee100 (3 bpw AutoRound checkpoint with its MTP head, made for a 64 GiB card; not involved in this recipe)","url":"https://huggingface.co/klee100/Qwen3.8-Flash-Next-AutoRound-3bpw-MTP"},{"name":"halt95 (W4A16-Merlin checkpoint whose FP8 n-gram table this recipe serves; not involved in this recipe)","url":"https://huggingface.co/halt95/Qwen3.8-Flash-Next-W4A16-Merlin"},{"name":"inco.ai (DFlash2 drafter, Apache-2.0)","url":"https://huggingface.co/incoai/Qwen3.8-Flash-Next-DFlash2"},{"name":"Intel (AutoRound quantisation) + RadixArk (FP8 n-gram table quantisation)","url":"https://huggingface.co/Intel/Qwen3.8-Flash-Next-W4A16-AutoRound"},{"name":"Qwen (Qwen3.8-Flash-Next base model, Qwen Community License 1.0)","url":"https://huggingface.co/Qwen/Qwen3.8-Flash-Next"},{"name":"MiaAI-Lab (benchmark protocol + category prompts, used as data)","url":"https://huggingface.co/Mia-AiLab"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-flash-next-3bpw-fp8-ple-1x-cmp-170hx-dflash2-recoverssm.json","alternate":"/r/qwen3-8-flash-next-3bpw-fp8-ple-1x-cmp-170hx-dflash2-recoverssm"}},{"slug":"nemotron-3-5-bolt-74b-a9b-1x-dgx-spark-vllm-b12x-lilicorr-262k","title":"Nemotron 3.5 Bolt 74B-A9B 1x DGX Spark vLLM b12x LiLiCorr 262K","summary":"NVIDIA Nemotron 3.5 Bolt 74B-A9B-24GB (hybrid Mamba-2/Transformer latent-MoE, 74B with 9B active, IQ2_XXS 2-bit experts, NVFP4 Mamba layers, Q8_0 rest; text+image+video in, 262,144-token window, OpenMDW-1.1) on one DGX Spark (GB10/SM121) with vLLM nightly-77154ca6 (base image vllm/vllm-openai@sha256:1ee85ca8) + b12x 1.5.1 kernels (--linear-backend b12x --moe-backend b12x), FP8 KV cache, gpu-memory-utilization 0.70, max-num-seqs 16, and the NVIDIA LiLiCorr MTP drafter at 7 speculative tokens. Pinned revisions: model c6cc2e8f, drafter 4c6a5613. Measured on the reference box: ~140 tok/s code rewrite / 63 summarise / 50 prose single-stream with the drafter (mean acceptance 4.0 tokens/pass), 31 tok/s without; 74.8/113.4/163.6/201.2 tok/s at 1/4/8/16 streams; KV cache ~6.52M tokens at the 0.70 share. On a 64 GB Spark the lane takes its own memory share 0.60: 667,874 KV tokens (2.55 windows), 110.8 tok/s greedy, 7.1 GiB left at the peak (2026-10-24).","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":63,"prefill_tok_s":0,"updated_at":"2026-10-29 21:32:26.717952-06","commit_sha":"2cdb5f11ffa1fb1af8676496fdd631976c3d536d","content_hash":"28887fb2cf3bab1914c803e792898fff92b9afe0","votes":0,"credit":[{"name":"hasso5703","url":"https://github.com/hasso5703"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/nemotron-3-5-bolt-74b-a9b-1x-dgx-spark-vllm-b12x-lilicorr-262k.json","alternate":"/r/nemotron-3-5-bolt-74b-a9b-1x-dgx-spark-vllm-b12x-lilicorr-262k"}},{"slug":"glm-5-3-flash-exlr8-exl3-experts-on-dgx-spark-tp1-4","title":"GLM-5.3-Flash exlr8 EXL3 experts on DGX Spark TP1-4","summary":"GLM-5.3-Flash exlr8: MoE layers are EXL3-trellis units decoded by the custom exlr8 MoE kernel; dense linears stay FP8/NVFP4. One release carries three expert widths (k2/k3/k4 = 2/3/4 bpw), chosen at boot via VLLM_EXLR8_WIDTHS. Reference topology TP=4 on four GB10 boxes; TP=1 serves the whole model (~87 GiB) on one 128 GB DGX Spark / ASUS GX10 via tp1.yaml plus a distinct MENTAT_GROUP. TP=2 and TP=3 also measured. 2026-10-19, clocks TP-locked at 2200 MHz: decode code 128.0 tok/s, prose 72.5 at TP=4; 53.0 / 29.5 at TP=1; prefill 4,393 tok/s at 8K, 4,190 at 32K. Trade vs the shipping NVFP4 build: decode +3-8%, NLL equal or better, KV pool ~+15%, prefill 7.7% lower at Second 1x-Spark GLM-5.3-Flash lane: not the VQ Q4_K GGUF lane (/r/glm-5-3-flash-vq-q4-k-1x-dgx-spark-1m-ctx-jlfn), and not a fork of the 2x ATLAS build (/r/glm-5-3-flash-nvfp4-on-2x-dgx-spark-atlas-rc4-2-tp2-dflash2-1m) whose NVFP4 Flash base this stack overlays.","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":65.2000000000000028421709430404007434844970703125,"prefill_tok_s":3438.5,"updated_at":"2026-10-29 21:32:25.385656-06","commit_sha":"2ddd4d836d8efa003f7203ba30597c20dc21e495","content_hash":"79db03010b10026e9788b7c749df25daf5fb38be","votes":0,"credit":[{"name":"kindlingai","url":"https://github.com/kindlingai/glm-5.3-flash-exlr8-gx10"},{"name":"nvidia","url":"https://huggingface.co/nvidia/GLM-5.3-Flash-NVFP4"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/glm-5-3-flash-exlr8-exl3-experts-on-dgx-spark-tp1-4.json","alternate":"/r/glm-5-3-flash-exlr8-exl3-experts-on-dgx-spark-tp1-4"}},{"slug":"ds4-flash-3x170hx-pp3-dspark-k5","title":"DeepSeek V4 Flash-0731 abliterated on 3x CMP 170HX \u2014 vLLM patched PP3 DSpark K5","summary":"The proven 3-card pipeline-parallel recipe on CMP 170HX (sm_80, 64GB unlocked): PP=3 with manual layer partition 15,16,12 at util 0.98, FP8 KV, block-size 256, DSpark block drafting at exactly 5 speculative tokens, async scheduling, CPU pinning (dedicated physical cores, placeholder in launch args), hard memory cap 64g with swap disallowed. Structured decode 181 tok/s at 99% acceptance. Upstream reference repo: allover326/deepseek-v4-cmp170hx @ 3dd2d88 (Aug 13 2026) \u2014 documents 15,15,13 3-card split, retracted 4-card requirement, INT4 524288 ctx w/ DSV4_LOGITS_ROW_CHUNK=64, +7% decode via c3046d1 rebase patches (full source build required). First-party 15,16,12 numbers on this page are ours.","tier":"curated","beta":false,"generation":"ampere","decode_tok_s":129.900000000000005684341886080801486968994140625,"prefill_tok_s":2350,"updated_at":"2026-10-29 21:32:23.860633-06","commit_sha":"24a9ac90156459e4efb49c951590edbea775df38","content_hash":"02d56ece2249f2b3e2b1ee2ffce158e841e20d26","votes":0,"credit":[{"name":"allover326","url":"https://github.com/allover326/deepseek-v4-cmp170hx"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/ds4-flash-3x170hx-pp3-dspark-k5.json","alternate":"/r/ds4-flash-3x170hx-pp3-dspark-k5"}},{"slug":"qwen3-8-27b-sglang-nvfp4-dflash2-on-1x-dgx-spark-deterministic","title":"Qwen3.8-27B SGLang NVFP4 DFlash2 on 1x DGX Spark (deterministic)","summary":"First dense Qwen3.8-27B recipe on 1x DGX Spark (GB10, 128 GB unified): SGLang v0.5.22 release image + NVFP4 target + DFlash2 speculative decoding (NVFP4 calibrated 5-layer draft, depth 16), served by a boot-persistent systemd unit from a one-command installer. Author-measured 68.3 tok/s greedy median single stream (bench.sh, thinking on; v1.39.0 A/B 2026-10-23), 81.7 tok/s at 4 streams (serving-bench), 135-148 tok/s at 8 streams and 258 at 32. Fully pinned (image digest, checkpoint revisions, draft depth) and reproducible to the decimal across boots via --disable-flashinfer-autotune (\"The boot lottery\", BENCHMARKS.md); 1M-token YaRN context by default, native 262144 as fallback; OpenAI- and Anthropic-compatible API on :30001 through the box's keepalive proxy (the engine itself answers on 127.0.0.1:30000, on the box only), with opencode wired out of the box.","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":68.2999999999999971578290569595992565155029296875,"prefill_tok_s":0,"updated_at":"2026-10-28 16:49:58.824279-06","commit_sha":"307fa2334d111e38258e27fa9e690c44e96b441b","content_hash":"7d8b3e291e1b8abfa9db4ded3e90cd601b8bba78","votes":0,"credit":[{"name":"hasso5703","url":"https://github.com/hasso5703"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-27b-sglang-nvfp4-dflash2-on-1x-dgx-spark-deterministic.json","alternate":"/r/qwen3-8-27b-sglang-nvfp4-dflash2-on-1x-dgx-spark-deterministic"}},{"slug":"qwen3-8-flash-next-nvfp4-1x-dgx-spark-vllm-71-3-c1-salient-data","title":"Qwen3.8-Flash-Next NVFP4 1x DGX Spark vLLM 71.3 c1 (salient-data)","summary":"Qwen3.8-Flash-Next (~180B multimodal MoE, NVFP4) on one NVIDIA DGX Spark (GB10, 121 GiB unified memory) with a pinned vLLM nightly image + 14 provenance-tracked patches (patch --fuzz=0) + one docker run, fitting the 51.2 GB PLE n-gram table by never making it resident (rows read per forward-pass step from the checkpoint file). 69.3 tok/s single-stream median decode (runs 66.8/69.5/69.3), 41.0 prose median, 91.9 tok/s 6-stream aggregate (1 run), GSM8K 99/100 on a fixed 100-problem subset (temp 0.6 top-p 0.95 thinking on), 12/12 arithmetic + 131K needle pass + prefix-cache canary pass, prefix caching ON. Measured with tonyd2wild's bench_categories.py @ 6ad1c8f unchanged on build 2026-10-22-8e2beaa3; clean first-time run 2026-10-25 re-measured 69.8/41.5/82.6. Prod/dev: prod image digest-pinned + cosign-signed + SPDX SBOM; dev rebuilds the same serving code from published patches with an AST-equivalence proof. Outside the model card's support statement (B200/B300 listed, GB10 unlisted).","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":69.2999999999999971578290569595992565155029296875,"prefill_tok_s":0,"updated_at":"2026-10-28 15:59:39.152527-06","commit_sha":"85e2e4120ef69829bff82736c39745ac579c9a6b","content_hash":"63f86642d07e8833929005e2087044adaecc1ca4","votes":0,"credit":[{"name":"Jeremy Newhouse / salient-data (recipe, patches, measurements)","url":"https://github.com/salient-data/qwen38-next-dgx-recipe"},{"name":"tonyd2wild (bench_categories.py @ 6ad1c8f + comparator stack/numbers)","url":"https://github.com/tonyd2wild/Qwen3.8-Flash-Next-NVFP4-DGX-Spark"},{"name":"NVIDIA (Qwen3.8-Flash-Next-NVFP4 checkpoint; HF token gated)","url":"https://huggingface.co/nvidia/Qwen3.8-Flash-Next-NVFP4"},{"name":"vLLM project (nightly base image 8e2beaa3)","url":"https://github.com/vllm-project/vllm"},{"name":"Sigstore/cosign (image signing + SBOM timestamping)","url":"https://docs.sigstore.dev"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-flash-next-nvfp4-1x-dgx-spark-vllm-71-3-c1-salient-data.json","alternate":"/r/qwen3-8-flash-next-nvfp4-1x-dgx-spark-vllm-71-3-c1-salient-data"}},{"slug":"glm-5-3-flash-vq-q4-k-1x-dgx-spark-1m-ctx-jlfn","title":"GLM-5.3-Flash VQ Q4_K 1x DGX Spark 1M ctx (JLFN)","summary":"Recipe by JLFN (dgx-recipe, verified on-box 2026-10-17). Baekpica/GLM-5.3-Flash-VQ-GGUF (72 GiB GVQ1 release directory: routed experts vector-quantized with codebook, 316 nonexpert matrices Q4_K, rest source precision; 56 GGUF parts, embedded MTP predictor) served by the ds4-dfm-rs Rust host (ds4-server, CUDA) on one DGX Spark at the model's native 1048576-token context with two banks. Measured at temperature 0: 24.2-24.3 tok/s one stream, 12.1/12.0 each with two streams (aggregate flat - decode is bytes-per-token against the weight stream). Disclosed quality trade: prose PPL 4.8555 vs 8.0797, code-completion 10.5 vs 18.4%, Korean meeting-summary degeneration under greedy.","tier":"community","beta":false,"generation":"blackwell","decode_tok_s":24.300000000000000710542735760100185871124267578125,"prefill_tok_s":0,"updated_at":"2026-10-20 21:52:36.476254-06","commit_sha":"f5a53e8b06f37327380e8affb1c9ea298e270456","content_hash":"2d63af307f133a15a2baacaee8d45c16f33ff129","votes":0,"credit":[{"name":"JLFN","url":"https://github.com/JLFN/dgx-recipe"},{"name":"Baekpica","url":"https://huggingface.co/Baekpica"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/glm-5-3-flash-vq-q4-k-1x-dgx-spark-1m-ctx-jlfn.json","alternate":"/r/glm-5-3-flash-vq-q4-k-1x-dgx-spark-1m-ctx-jlfn"}},{"slug":"qwen3-8-flash-next-w4a16-3x-cmp-170hx-pp3-nvme-ple-fp8-kv-1m-cle","title":"Qwen3.8-Flash-Next W4A16 3x CMP 170HX PP3 NVMe PLE fp8 KV 1M (clee)","summary":"Working PP3 recipe for Qwen3.8-Flash-Next (Qwen4Exp 180B-A6B MoE, Intel INT4 W4A16 AutoRound) on 3x CMP 170HX (SM80, unlocked 64 GiB each): the 95.4 GiB BF16 PLE n-gram table streams live from NVMe (shard-direct preadv into a fixed GPU staging buffer, never fully loaded), fp8 e4m3 KV software-decoded on SM80, mamba SSM cache bf16, and 1,048,576-token context via YaRN x4 over the 262144-native checkpoint. KV pool 5.97M tokens (~5.7 concurrent 1M sessions). W4A16 experts run GPTQ-Marlin bit-exact; no GPU P2P on this card pairing, so PP stages through host RAM with NCCL_P2P_DISABLE=1 + --disable-custom-all-reduce. Measured cold-cache temp=0: prefill 2389 tok/s @4k rising to 4690 @1M (QSA sparse crossover past ~64k), decode 73.2 -> 59.8 tok/s @1M, TTFT 223 s @1M; exact needle recall at all depths through 1M. Fork clee/vllm-sm80 on a PINNED nightly; newer nightlies not validated.","tier":"community","beta":false,"generation":"ampere","decode_tok_s":73.2000000000000028421709430404007434844970703125,"prefill_tok_s":3563,"updated_at":"2026-10-20 15:34:25.555323-06","commit_sha":"fb3da33b3299da70eb24ce68a2c772de0dc1a356","content_hash":"0dea5f8b482a14a47e6cef76e85ad322d56f9821","votes":1,"credit":[{"name":"clee (recipe, fork patches, bench + needle harness)","url":"https://github.com/clee/qwen3.8-flash-next-3x170hx"},{"name":"clee (vllm-sm80 fork: units-fix, fp8-kv-sm80, ple-pp3, ple-nvme-staged)","url":"https://github.com/clee/vllm-sm80"},{"name":"Kai / 2Wild (Qwen38-Flash-Next-4x3090 \u2014 staged-PLE architecture inspiration)","url":"https://github.com/tonyd2wild/Qwen38-Flash-Next-4x3090"},{"name":"Intel (INT4 W4A16 AutoRound quant)","url":"https://huggingface.co/Intel/Qwen3.8-Flash-Next-W4A16-AutoRound"},{"name":"froggeric (bugfixed Qwen3.8 chat templates)","url":"https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-flash-next-w4a16-3x-cmp-170hx-pp3-nvme-ple-fp8-kv-1m-cle.json","alternate":"/r/qwen3-8-flash-next-w4a16-3x-cmp-170hx-pp3-nvme-ple-fp8-kv-1m-cle"}},{"slug":"qwen3-8-27b-int8-mtp3-prefix-cache-on-1x-cmp-170hx-hynix","title":"Qwen3.8-27B INT8 MTP3 prefix-cache on 1x CMP 170HX Hynix","summary":"Qwen3.8-27B at 65-67 tok/s decode (74 structured) with MTP k=3 speculative decoding AND prefix caching simultaneously on a single NVIDIA CMP 170HX 64GB (Hynix HBM2e, GA100/sm80), 262,144-token context on fp8_e4m3 KV, 24/7 production on pinned upstream vLLM nightly 0.26.1rc1.dev1024+g1fe3a1571 \u2014 the first build where prefix caching and MTP spec decode coexist (GDN mamba-align). Cached TTFT on a 7K shared prefix drops ~4.0 s -> 1.44 s (33.5% hit rate) while decode gains 7-9% over the 0.27.1 fork fallback \u2014 the 'MTP vs prefix cache, pick one' tradeoff is gone. Load-bearing trio --async-scheduling + --max-num-batched-tokens 8192 + --max-num-scheduled-tokens 8192 (default chunked-prefill batch = 0.0% hit rate while looking healthy); --prefix-match-unit 400 halves cached TTFT further (must divide the 1600-token GDN block); batch budget 32000 OOMs and hard-wedges the card. Samsung-40GB side recipe (int8 + DFlash2, 97.8 tok/s) with an sm80 int8-repack CPU-fallback patch.","tier":"community","beta":false,"generation":"ampere","decode_tok_s":68.7000000000000028421709430404007434844970703125,"prefill_tok_s":0,"updated_at":"2026-10-20 15:34:22.585629-06","commit_sha":"ad3f55dab02cedfb2b187f7dcfc3f1647e0aaa0c","content_hash":"68f95048b88c132d1ddbffc8ceceaf5c38be5140","votes":0,"credit":[{"name":"ahnguyen17 (recipe, patches, probes, A/B benchmarks)","url":"https://github.com/ahnguyen17/cmp-170hx-vllm"},{"name":"lued (Qwen3.8-27B-INT8-W8A16-MTP checkpoint)","url":"https://huggingface.co/lued/Qwen3.8-27B-INT8-W8A16-MTP"},{"name":"vllm-project (vLLM upstream; #50021 GDN spec bounds, #45238 align-mode cache retention)","url":"https://github.com/vllm-project/vllm"}],"uploaded_by":"localai-recipes","measured":null,"links":{"self":"/api/v1/recipes/qwen3-8-27b-int8-mtp3-prefix-cache-on-1x-cmp-170hx-hynix.json","alternate":"/r/qwen3-8-27b-int8-mtp3-prefix-cache-on-1x-cmp-170hx-hynix"}}]}