All recipes: Difference between revisions
From CMP Wiki
More actions
Wiki lane: site structure (templates, forms, help pages) |
Wiki lane: site structure (templates, forms, help pages) |
||
| Line 1: | Line 1: | ||
Every recipe on the wiki: a model running on CMP 170HX cards, with the settings and the numbers. Click a column to | {{#cmpcache:300}}Every recipe on the wiki: a model running on CMP 170HX cards, with the settings and the numbers. Click a column to | ||
sort; type in the search box to filter (for example <code>10 GB</code> or <code>vLLM</code>). | sort; type in the search box to filter (for example <code>10 GB</code> or <code>vLLM</code>). | ||
Revision as of 07:04, 30 October 2026
Every recipe on the wiki: a model running on CMP 170HX cards, with the settings and the numbers. Click a column to
sort; type in the search box to filter (for example 10 GB or vLLM).
Filter by card variant: GB 8 GB only · GB 10 GB only · recipes for both · Category:Recipes for 8 GB · Category:Recipes for 10 GB. Status is computed: Proven means someone else reproduced it (see Status taxonomy).
| Recipe | Model | Quant | Engine | Variant | Cards | Topology | Context | Decode tok/s | Prefill tok/s | VRAM per card | Power cap | Driver | Status | Tested by | Tested on |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Recipe | Model | Quant | Engine | Variant | Cards | Topology | Context | Decode tok/s | Prefill tok/s | VRAM per card | Power cap | Driver | Status | Tested by | Tested on |
| GLM 5.3 Flash on 4x 8 GB | GLM-5.3-Flash (320B MoE, 18B active) | W4A16 (INT4 weights, group size 128, FP16 activations): canada-quant/GLM-5.3-Flash-W4A16-MTP | Morrowmake/vllm-cmp170hx (vLLM fork, container image) | 8 GB | 4 | TP4 (default) or PP4 | 262,144 | 442.1 | 3,161.0 | 64 GiB exposed, --gpu-memory-utilization 0.95 | 180 W | NVIDIA 580 or newer; P2P needs the Morrowmake/cmpunlocker build | Proven | voidgiraffe | 2026-10-16 |
| Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2) | Qwen3.8-Flash-Next (125B MoE per the recipe) | W4A16: halt95/Qwen3.8-Flash-Next-W4A16-Merlin (AutoRound INT4 g128 experts, INT8 GDN projections, FP8 n-gram table) | Morrowmake/vllm-cmp170hx image (Qwen EXL3 release) with patch files mounted read-only | 8 GB | 2 | TP2 with whole-expert expert parallel | 262,144 | 349.7 | 4,696.0 | 64 GiB, GPU_UTIL 0.95 (fixed) | 180 W | 580 or newer (measured on 610.57.04); open modules with HMM for the table-on-disk mode | Proven | nathang6866 | 2026-10-23 |
| Qwen3.8 Flash Next on 1x 8 GB (ElectroHobo 3 bpw) | Qwen3.8-Flash-Next (125B MoE per the recipe) | AutoRound mixed 2/3-bit (klee100, 2.998 effective bpw, 47 GiB on card) plus halt95's FP8 n-gram table in host RAM | Morrowmake/vllm-cmp170hx image with patch files mounted read-only | 8 GB | 1 | single card | 262,144 | 296.4 | 4,243.0 | 64 GiB, --gpu-memory-utilization 0.95 (fixed) | 180 W | 580 or newer (measured on 610.57.04) | Reported | nathang6866 | 2026-10-23 |
| DeepSeek V4.1 Flash on 4x 8 GB (Morrowmake) | DeepSeek-V4.1-Flash-1017 | Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16; KV BF16 | Morrowmake/vllm-cmp170hx (vLLM fork, container image) | 8 GB | 4 | TP4 with expert parallel (96 whole experts per card) | 262,144 | 287.0 | 3,818.0 | 64 GiB exposed, --gpu-memory-utilization 0.95 | 180 W | NVIDIA 580 or newer; P2P needs the Morrowmake/cmpunlocker build | Reported | voidgiraffe | 2026-10-24 |
| MiMo V2.6 Flash on 4x 8 GB | MiMo-V2.6-Flash-RL (309B MoE, 15B active; text, image, video and audio input) | NVFP4 (ModelOpt mixed precision): ProCreations/MiMo-V2.6-Flash-RL-NVFP4, run as 4-bit weights with BF16 activations through Marlin | Morrowmake/vllm-cmp170hx image (tag qwen38-recipe-1.0.0) plus one mounted file (vLLM #58235 vision-sink fix) | 8 GB | 4 | TP4 (P2P on); expert parallel off | 1,048,576 | 274.9 | 2,722.0 | 64 GiB; weights 45.5 GiB plus 10 GiB KV per card; GPU_UTIL 0.94 (fixed) | 180 W | 580 or newer (measured on 610.57.04) | Reported | nathang6866 | 2026-10-18 |
| DeepSeek V4.1 Flash on 4x 8 GB (GNmac) | DeepSeek-V4.1-Flash-1017 | Hybrid: routed experts EXL3 (mostly 3 bpw, 2 bpw on layers 18-22) from LibertAI's 2.9 bpw pack; attention, shared experts, indexer and DSpark drafter official FP8; KV fp8_ds_mla | vLLM with a pinned sm_80 patch stack (built locally from source) | 8 GB | 4 | TP4 (default) or PP4 for x1/x4 links | 262,144 (up to 1,048,576) | 258.0 | 2,416.0 | 64 GiB, GPU_MEMORY_UTILIZATION 0.95 | 180 W (nvidia-smi -pl 180) | CUDA 13 capable; measured on 610.43.02 | Proven | greencandle9753 | 2026-10-26 |
| Qwen3.8 Flash Next on 2x 8 GB (Plean PP2) | Qwen3.8-Flash-Next | Mixed INT4 AutoRound (Minachist/Qwen3.8-Flash-Next-INT4-Mixed-AutoRound, about 65 GB) with the full BF16 PLE table on NVMe | vLLM 0.29.1rc1 nightly plus the checkpoint author's patches | 8 GB | 2 | PP2, MTP 3, no P2P | 262,144 (256K in the post) | 197.0 | 5,800.0 | 64 GB | Reported | plean. | 2026-10-07 | ||
| Qwen3.8 Flash Next on 2x 8 GB (mannydelrio PP2) | Qwen3.8-Flash-Next-Uncensored (refusal-removed, from orcarouter) | AWQ W4A16 g32 (leoncca/Qwen3.8-Flash-Next-Uncensored-AWQ-g32); PLE table converted to INT4 and served from SSD | wtdcode/vllm-backport 3142ded1 plus a two-file patch and two tinybeast PLE plugins (venv, no container) | 8 GB | 2 | PP2 (layer partition 25,23), no P2P | 262,144 | 177.8 | 7,006.0 | 64 GB, --gpu-memory-utilization 0.94, TB_TORCH_MEM_CAP_GIB=61.5 | 180 W | nvidia-open 610.43.02 + cmpunlocker v0.5 or newer | Reported | mrmannyatx | 2026-10-15 |
| Qwen3.8 27B on 1x 8 GB (W4A16 vLLM) | Qwen3.8-27B (dense) | W4A16 (INT4 weights, Marlin kernels) | vLLM | 8 GB | 1 | single card | 262,144 native (YaRN to about 500k in the recipe's long variant) | 170.0 | 64 GB | Reported | kellertreppe123 | 2026-10-05 | |||
| DeepSeek V4.1 Flash on 6x 8 GB (Deepinfer-SM80 TP2 x PP3) | DeepSeek-V4.1-Flash-1017 | Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 | vLLM PR #56201 head (79a7108d) plus the two Deepinfer-SM80 patches and 30 drop-in modules | 8 GB | 6 | TP2 x PP3, partition 14,14,12 | 1,048,576 | 160.8 | 4,037.0 | 64 GiB | Reported | HalfVulpes | 2026-10-28 | ||
| DeepSeek V4.1 Flash heretic on 4x 8 GB | DeepSeek-V4.1-Flash-1017 heretic (pqhaz abliteration) | Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16 with the 38 heretic attn.wo_b tensors | Morrowmake/vllm-cmp170hx (vLLM fork, container image) | 8 GB | 4 | TP4 | 262,144 | 158.0 | 2,453.0 | 64 GiB, --gpu-memory-utilization 0.92 (miner co-resident) | Reported | darkmaniac7 | 2026-10-29 | ||
| DeepSeek V4.1 Flash on 8x 8 GB (Deepinfer-SM80 TP4 x PP2) | DeepSeek-V4.1-Flash-1017 | Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 | vLLM PR #56201 head (79a7108d) plus the two Deepinfer-SM80 patches and 30 drop-in modules | 8 GB | 8 | TP4 x PP2, partition 20,20 | 1,048,576 | 153.35 | 3,463.0 | 64 GiB, --gpu-memory-utilization 0.95 | Reported | HalfVulpes | 2026-10-28 | ||
| Qwen3.8 Flash Next on 2x 8 GB (Morrowmake EXL3) | Qwen3.8-Flash-Next | EXL3 5.05 bpw; PLE (n-gram) table read from SSD | Morrowmake/vllm-cmp170hx (EXL3 path) | 8 GB | 2 | TP2 with expert parallel | 262,144 | 152.6 | 2,651.0 | 64 GiB, --gpu-memory-utilization 0.95 | Reported | voidgiraffe | 2026-10-16 | ||
| DeepSeek V4.1 Flash on 4x 8 GB (CurtisAccelerate TC-MoE) | DeepSeek-V4.1-Flash-1017 | Hybrid: LibertAI EXL3 2.9 bpw routed experts (layers 0-39), official FP8/FP4 for attention, shared experts, head, draft and Engram; KV fp8_ds_mla | Morrowmake/vllm-cmp170hx (bind-mounted sm80 package) plus the author's tensor-core MoE kernels | 8 GB | 4 | TP4, peer access off (NCCL_P2P_DISABLE=1), host all-reduce | 262,144 (weights address 1,048,576) | 144.2 | 1,523.0 | about 62.5 GB used of 64 GB, --gpu-memory-utilization 0.97 | 140 W | Reported | CurtisAccelerate | 2026-10-22 | |
| Qwen3.8 Flash Next on 2x 8 GB (DUFRelic PP2) | Qwen3.8-Flash-Next | W4A16 INT4 g128 with the PLE table repacked INT4 g32: zankich/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV; KV BF16 | vLLM v0.30.0 + zankich INT4-PLE plugin + six hash-guarded patch layers | 8 GB | 2 | PP2, MTP depth 3 | 262,144 | 133.39 | 6,213.66 | 64 GiB, --gpu-memory-utilization 0.90 | 140 W (200 W observed-draw abort guard) | NVIDIA 610.43.02 | Reported | dufrelic | 2026-10-03 |
| DeepSeek V4.1 Flash on 8x 8 GB (Zanooda PP8) | DeepSeek-V4.1-Flash-1017 | Official checkpoint as downloaded: MXFP4 routed experts as Marlin W4A16, MXFP8 dense as Marlin W8A16; KV fp8 | vLLM PR #56201 head (79a7108d) plus a 20-patch sm_80 series; image zanooda/vllm-sm80-ds41f:v41-sm80 | 8 GB | 8 | PP8, partition 5,5,5,5,5,5,5,5 | 1,048,576 | 117.0 | 6,066.0 | 64 GB; weights 35.9 GB per rank (ranks 1-6), 37.6 GB rank 0, 46.3 GB rank 7 | 180 W (one card unstable at 250 W) | Reported | Zanooda | 2026-10-19 | |
| Qwen3.8 Flash Next on 4x 8 GB | Qwen3.8-Flash-Next | AWQ W4A16: wtdcode/Qwen3.8-Flash-Next-AWQ-W4A16 | wtdcode/vllm-backport (image lazymio/vllm-backport:latest-sm80) | 8 GB | 4 | TP4 with expert parallel, MTP 3 | 1,000,000 (YaRN x4 from 262,144) | 116.49 | 2,101.4 | 64 GB | Reported | matstatman | 2026-10-16 | ||
| Qwen3.8 Flash Next on 1x 8 GB (iIIusi0n SSD PLE) | Qwen3.8-Flash-Next | AutoRound 3 bpw + MTP: klee100/Qwen3.8-Flash-Next-AutoRound-3bpw-MTP; 95.37 GiB BF16 PLE table on SSD | vLLM a5a30471 plus a 13-file downstream patch (native Linux AIO PLE reader) | 8 GB | 1 | single card | 262,144 advertised (tested through 32K) | 110.99 | 2,636.0 | 64 GiB, --gpu-memory-utilization 0.96 | 180 W | 610.43.03 | Proven | iIIusi0n | 2026-09-19 |
| GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) | GLM-5.3-Flash (320B MoE, 18B active) | EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3; DFlash2 drafter converted to EXL3 6 bpw | ExLlamaV3 (patched, pinned) | 8 GB | 2 | TP2 over NCCL (GPU P2P), DFlash2 K7 plus hybrid drafting | 392,960 per request; 655,360-token Q8 KV pool shared by two requests | 107.67 | 1,533.0 | about 62.9 / 62.3 GiB idle, peaks about 64.4 / 63.0 GiB with vision | 180 W (author's cards); about 170 W with a modest memory OC in his Discord numbers | 610.57.04 | Reported | nestaa51 | 2026-10-13 |
| DeepSeek V4.1 Flash on 4x 8 GB (alanoo fidelity build) | DeepSeek-V4.1-Flash-1017 | Hybrid: all 40 expert layers at 3 bits (LibertAI 2.9 bpw pack with layers 18-22 grafted from the 3.5 bpw pack), official FP8 dense tensors | GNmac-lab recipe image 1.2.0 (vLLM with sm_80 patches) plus three patches | 8 GB | 4 | TP4, P2P on | 262,144 | 105.0 | 2,514.0 | 64 GiB (64,911 MiB usable), GPU_MEMORY_UTILIZATION 0.98 | 200 W | NVIDIA open 610.43.02 from a cmpunlocker fork with a P2P patch | Reported | alanoo | 2026-10-27 |
| DeepSeek V4.1 Flash on 4x 8 GB plus RTX PRO 6000 (ExLlamaV3 TP5) | DeepSeek-V4.1-Flash-1017 | EXL3 3.5 bpw (LibertAIDAI; routed experts 4 bpw on layers 0-9 and 30-39, 3 bpw on 10-29; attention and shared experts 6 bpw) plus the official Engram shards | ExLlamaV3 fork (anoane/exllamav3-anemone) + TabbyAPI fork (anoane/tabbyAPI) | 8 GB | 4 | TP5 across four CMP 170HX and one RTX PRO 6000 Blackwell (native backend, reduces through host memory); MTP draft on the PRO | 1,048,576 FP16 | 99.9 | 2,116.0 | gpu_split 58 GB per CMP, 70 GB on the PRO | 150 W per CMP (PRO 400 W) | NVIDIA open 610.57.04 patched with Morrowmake/cmpunlocker 1f2c498 (profile 8gb) | Reported | anoane | 2026-10-20 |
| DeepSeek V4.1 Flash on 6x 8 GB (TP6) | DeepSeek-V4.1-Flash-1017 | Official checkpoint: dense MXFP8 through an SM80 Marlin W8A16 path, routed experts MXFP4 through W4A16; KV FP8 | Custom SM80 vLLM (patched from public commit 79a7108d9) | 8 GB | 6 | TP6 (virtual-head padding), PP1 | 1,048,576 | 78.55 | 64 GB, --gpu-memory-utilization 0.95 | Reported | wgkbfcp | 2026-10-19 | |||
| Qwen3.8 Flash Next on 2x 8 GB (vektorprime PP2) | Qwen3.8-Flash-Next | AWQ INT4: cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4; 95 GiB BF16 PLE table in host RAM (FP8 table optional) | vLLM nightly + PR #53899 (a5530b9) + the repo's patch scripts | 8 GB | 2 | PP2 with PLE CPU offload | 262,144 | 75.4 | 64 GiB; weights PP0 39.48 GiB, PP1 45.53 GiB; --gpu-memory-utilization 0.9 | Reported | vektor_prime | 2026-08-29 | |||
| Qwen3.8 Flash Next on 3x 8 GB (clee PP3) | Qwen3.8-Flash-Next | W4A16 AutoRound: Intel/Qwen3.8-Flash-Next-W4A16-AutoRound; 95.4 GiB BF16 PLE table streamed from NVMe; KV fp8 e4m3 (software-decoded on SM80) | vLLM nightly 7d0b4e57 plus the author's pure-Python fork clee/vllm-sm80 (branch q38fn/stack) | 8 GB | 3 | PP3, no P2P (works at PP2 on two cards) | 1,048,576 (YaRN x4) | 74.4 | 4,831.0 | 64 GiB, gpu-memory-utilization 0.92 | Proven | c133 | 2026-10-10 | ||
| DeepSeek V4.1 Flash on 2x 8 GB (Project Maya) | DeepSeek-V4.1-Flash-1017 (748B in the Maya README) | Maya GGUF (Maya-S 340 GB or Maya-XS 277 GB; Engram tables stay on SSD) | Project Maya (Strata and llama.cpp lineage) | 8 GB | 2 | engine-managed expert tiering across VRAM, host RAM and NVMe | 63.0 | 64 GB | Reported | pcsv. | 2026-10-25 | ||||
| GLM 5.3 Flash on 2x 8 GB (Project Maya) | GLM-5.3-Flash (321B in the Maya README) | Maya-M GGUF (116 GB), every expert in VRAM; Maya-S (96.5 GB) and Maya-S24 (94.7 GB) also reported | Project Maya (Strata and llama.cpp lineage, EXL3 loading, DFlash2 drafter) | 8 GB | 2 | engine-managed split with expert tiering across VRAM, RAM and NVMe | 61.3 | 514.0 | 64 GB | Reported | ZackO2o | ||||
| GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 layer split) | GLM-5.3-Flash (320B MoE, 18B active) | EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3 | ExLlamaV3 1.5.4 (built for SM80) with a small OpenAI-compatible server | 8 GB | 2 | layer split (no tensor parallel), no P2P | 392,960 (profile q8_384k) | 60.0 | about 58 GiB of 64 GiB (116.6 GiB resident across both) | Proven | 68x0 | 2026-10-20 | |||
| GLM 5.3 on 8x 8 GB (vllm-170hx-glm5) | GLM-5.3 (full 78-layer DSA MoE, not Flash) | GLM-5.3-Int4-Int8Mix-AWQ-g64; drafter incoai/GLM-5.3-DFlash2 | Morrowmake/vllm-cmp170hx branch ampere-glm53 plus fail-closed source patches | 8 GB | 8 | PP8 (also PP9 and PP10), TP1 | 1,048,576 per request | 48.7 | 1,833.0 | 64 GiB, GPU_UTIL 0.96 at PP8 | 170 W (author's 10-card run) | Proven | reala7096 | 2026-09-29 | |
| GLM 5.3 on 8x 8 GB (Int4-Int8Mix PP8) | GLM-5.3 (full 78-layer, not Flash) | GLM-5.3-Int4-Int8Mix (INT4 / INT8 mix); KV fp8_ds_mla | vLLM (build not stated) | 8 GB | 8 | PP8, TP1 | 1,048,576 | 25.0 | 64 GiB, --gpu-memory-utilization 0.97 | 200 W | Reported | 1a2b3c0967 | 2026-10-07 | ||
| GLM 5.3 Flash on 3x 8 GB | GLM-5.3-Flash (320B MoE, 18B active) | W4A16, about 4.3 bpw, 163 GiB on disk | Morrowmake/vllm-cmp170hx (via the glm53-flash-cmp170hx-recipe) | 8 GB | 3 | PP3 | about 1,000,000 (8-bit KV cache) | 64 GiB | Reported | shirogeek | 2026-10-19 |