Initial tests for Kimi-K2.5 via KTransformers+SGLang, on a hybrid 4x RTX Pro 6000 Blackwell + 640GB/1.5TB CPU memory offload. Compute provided by Lium pods: - 19.97 output tok/s @ 10 concurrent requests - Mean TTFT: ~120s - Median TTFT: ~102s Need to play with the KT flags to further optimize this setup, which is heavily dependent on the overall system's CPU core count & available RAM. GPU <-> PCIe <-> RAM interconnectivity is the most obvious bottleneck Experts per MoE Layer on GPU: --kt-num-gpu-experts=128 CPU cores dedicated to MoE inference: --kt-cpuinfer=104 CPU experts work overlapping GPU work: --kt-max-deferred-experts-per-token=2 Max tokens per prefill chunk: --chunked-prefill-size=32658 CUDA graph capture disabled: --disable-cuda-graph

KTransformers+SGLang flags to reproduce work: ========== export CUDA_VISIBLE_DEVICES=0,1,2,3 export OMP_NUM_THREADS=1 export MKL_NUM_THREADS=1 export OPENBLAS_NUM_THREADS=1 export NUMEXPR_NUM_THREADS=1 export VECLIB_MAXIMUM_THREADS=1 python -m sglang.launch_server \ --model-path <HF_PATH>/models--moonshotai--Kimi-K2.5/snapshots/3367c8d1c68584429fab7faf845a32d5195b6ac1 \ --kt-weight-path <HF_PATH>/models--moonshotai--Kimi-K2.5/snapshots/3367c8d1c68584429fab7faf845a32d5195b6ac1 \ --kt-cpuinfer 104 \ --kt-threadpool-count 2 \ --kt-num-gpu-experts 128 \ --kt-max-deferred-experts-per-token 2 \ --kt-method RAWINT4 \ --kt-gpu-prefill-token-threshold 400 \ --kt-expert-placement-strategy uniform \ --trust-remote-code \ --mem-fraction-static 0.90 \ --served-model-name kimi_k2 \ --tool-call-parser kimi_k2 \ --reasoning-parser kimi_k2 \ --disable-radix-cache \ --disable-chunked-prefix-cache \ --enable-mixed-chunk \ --tensor-parallel-size 4 \ --enable-p2p-check \ --disable-shared-experts-fusion \ --chunked-prefill-size 32658 \ --max-total-tokens 120000 \ --attention-backend flashinfer \ --disable-cuda-graph \ --host 0.0.0.0 \ --port 8000