Qwen3.8-Flash-Next on 3x CMP170HX (SM80): NVMe-streamed PLE, FP8 KV, 1M context (with NVMe offload)
-
Updated
Oct 11, 2026 - Jinja
Qwen3.8-Flash-Next on 3x CMP170HX (SM80): NVMe-streamed PLE, FP8 KV, 1M context (with NVMe offload)
CMP 170HX PCIe Gen2 x4 under Proxmox: experimental findings, source patch, benchmarks and recovery notes.
DeepSeek-V4-Flash-Vision-Exp on NVIDIA CMP 170HX/SM80 with vLLM, PP DSpark and 1M context
Qwen3.8 Flash Next FP8 on 4x CMP 170HX: PP3 transformer stages with the 51B PLE table on GPU 3
CMP 170HX 10GB to 40GB on Proxmox VFIO: pinned Johnny fix, Gen2 x4, 150W, Chinese guide and temperature-monitored gpu-burn evidence.
GLM-5.3 (753B MoE) on 2x RTX PRO 6000 Blackwell + 8x CMP 170HX: a two-host TP2 x PP5 layout and three overlays (pipeline hand-off fix, structured output with drafts, probabilistic drafts) for cebeuq's glm53-cmp170hx-recipe
Bench tools and firmware patches for the Pico voltage-glitch method on CMP 170HX 8 GB cards. Glitch and read only; nothing here burns fuses.
Independent Qwen3.8-Flash-Next reproduction on 2x CMP 170HX Gen2 x4, with raw benchmarks and quality harnesses.
CMP 170HX(GA100/sm80) 上 Qwen3.8-27B W8A16+MTP 的 vLLM 0.32.0 部署配方与四臂实测:split-KV verify attention 让 128K–256K 单流解码 2.3–4.8×(权重不变)
To associate your repository with the cmp170hx topic, visit your repo's landing page and select "manage topics."