#!/usr/bin/env bash
# Candidate-only FP16-K prefill variant measured on 2026-08-01.
set -euo pipefail
BIN_DIR=${BIN_DIR:-/opt/llama.cpp-kimi-k3-candidate/build-flashkda-iq1s-v1/bin}
MODEL=${MODEL:-$PWD/Kimi-K3-UD-IQ1_S-00001-of-00014.gguf}
RPC_WORKERS=${RPC_WORKERS:-10.10.10.1:50053,10.10.10.2:50053,10.10.10.4:50053}
PORT=${PORT:-8210}
exec env GGML_RPC_SKIP_HASH=1 CUDA_SCALE_LAUNCH_QUEUES=4x LLAMA_MMAP_PREFETCH=0 LLAMA_PARALLEL_DEVICE_LOAD=1 "$BIN_DIR/llama-server" \
  --model "$MODEL" --alias kimi-k3-f16k-600k-u1024 \
  --host 0.0.0.0 --port "$PORT" --ctx-size 600000 --parallel 1 \
  --n-gpu-layers 999 --rpc "$RPC_WORKERS" --split-mode layer --tensor-split 1,1,1,1 \
  --flash-attn on --fit off --cache-type-k f16 --cache-type-v f16 \
  --batch-size 2048 --ubatch-size 1024 --threads 16 --threads-batch 20 \
  --cache-reuse 256 --reasoning auto --metrics