Drilldown: Recipe

DeepSeek V4.1 Flash on 5x 8 GB (PP5 full residency) (1) · GLM 5.3 on 8x 8 GB (vllm-170hx-glm5) (1) · Qwen3.8 Flash Next on 3x or 4x 8 GB (zebgop-ops FP8 PP) (1)

DeepSeek-V4.1-Flash (1) · GLM-5.3 (full 78-layer DSA MoE, not Flash) (1) · Qwen/Qwen3.8-Flash-Next-FP8 (revision bcd9f01d) (1)

GLM-5.3-Int4-Int8Mix-AWQ-g64; drafter incoai/GLM-5.3-DFlash2 (1) · Official block FP8, about 126 GiB including the about 50 GiB FP8 n-gram (PLE) table, which is served from host RAM; KV cache bf16 (the QSA layers refuse fp8) (1) · Official checkpoint (510 GB: MXFP4 routed experts, FP8 dense layers, 203 GB FP8 Engram table); KV cache fp8_ds_mla (1)

Morrowmake/vllm-cmp170hx branch ampere-glm53 plus fail-closed source patches (1) · vLLM: 344303947's fork of wtdcode/vllm-backport (shadow source layer, staged Marlin repack, exact-size Engram pinning) (1) · vLLM: the vllm/vllm-openai:qwen38-flash-next image with zebgop-ops patches 0001 to 0016 (never 0008-WITHDRAWN) and three ported files (1)

0.1.dev20073+g8e685d198; repo commit 2f551bb (patch 0016, 2026-10-02) (1) · v0.13.1.dev6, tag v41-pp5-1m-pinmem-20260916 (1) · vLLM 0.30.1rc1, fork commit 378c37b0 (1)

8 GB (3)

4 (1) · 5 (1) · 8 (1)

PP4, VLLM_PP_LAYER_PARTITION=12,12,12,12, MTP 3, FULL_AND_PIECEWISE CUDA graphs; PCIe Gen2 x4 per card, no P2P. Three cards: PP3 with 16,17,15 or 16,16,16 (1) · PP5 with an even 8,8,8,8,8 layer split; the repo's cards are on PCIe Gen2 x16 with no NVLink (1) · PP8 (also PP9 and PP10), TP1 (1)

1,048,576 (KV pool 1,632,827 tokens at utilisation 0.97; an 896K-token prompt passed) (1) · 1,048,576 per request (1) · 262,144 native; 524,288 with YaRN 2.0; 1,048,576 with YaRN 4.0 (needle verified on PP3) (1)

None (1) · 48.7 (1) · 89 (1)

1,833 (1) · 4,600 (1) · 11,100 (1)

64 GB, full residency (no expert offload), --gpu-memory-utilization 0.97 (1) · 64 GiB, GPU_UTIL 0.96 at PP8 (1) · 64 GiB; --gpu-memory-utilization 0.94 plus per-rank KV budgets; measured 34.0-34.3 GiB per 12-layer rank, 37.9 GiB on the drafter rank (1)

None (1) · 170 W (author's 10-card run) (1) · 200 W (1)

None · 580 or newer (measured on 610.57.04) · NVIDIA 580 or newer; P2P needs the Morrowmake/cmpunlocker build · 580 or newer (measured on 610.57.04); open modules with HMM for the table-on-disk mode · 610.43.02 (nvidia-driver-610-open) · 610.43.03 · 610.43.03 (open modules) · 610.57.04 · CUDA 13 capable; measured on 610.43.02 · NVIDIA 610.43.02 · NVIDIA open 610.43.02 from a cmpunlocker fork with a P2P patch · NVIDIA open 610.43.03, CUDA 13.0 · NVIDIA open 610.57.04 patched with Morrowmake/cmpunlocker 1f2c498 (profile 8gb) · nvidia-open 610.43.02 + cmpunlocker v0.5 or newer · Other

None (2) · cmpunlocker VRAM unlock (64 GiB per card) (1)

Proven (1) · Reported (2)

344303947 (1) · flipnburner (1) · reala7096 (1)

September 2, 2026 (1) · September 22, 2026 (1) · September 29, 2026 (1)

https://github.com/344303947/dsv41-flash-pp5-170hx (1) · https://github.com/arvkmr/vllm-170hx-glm5 (1) · https://github.com/zebgop-ops/qwen38-flashnext-pp (1)