EXPERIMENT=kimi3_pro6000_pd_dflash_validation PD_IMAGE=local/sglang:kimi-k3-pp-dflash-33863-fi0618-situ4460-kvbounds1 MODEL_PATH=/data/hf_models/Kimi-K3 DRAFT_MODEL_PATH=/data/hf_models/Kimi-K3-DFlash P_NODES=(174.1.60.1 174.1.60.2 174.1.60.3 174.1.60.4) D_NODES=(174.1.60.5 174.1.60.6 174.1.60.7 174.1.60.8) SSH_USER=user SSH_OPTIONS=(-o BatchMode=yes -o ConnectTimeout=10) PORT=30000 DIST_PORT=20000 BOOTSTRAP_PORT=28800 ROUTER_PORT=31000 P_TP=4 P_PP=8 P_EP=4 D_TP=32 D_PP=1 D_EP=4 P_MEM=0.88 D_MEM=0.86 P_MAMBA_RATIO=0.36 D_MAMBA_RATIO=0.21 CHUNK=8192 PAGE_SIZE=64 KV_DTYPE=bfloat16 DRAFT_TOKENS=16 # GSM8K C1/C8 only; cap capture and scheduling to the measured concurrency. MAX_RUNNING=8 GRAPH_BS=8 IB_DEVICES=mlx5_0,mlx5_1,mlx5_2,mlx5_3 JIT_CACHE=/data/hzy/cache/kimi-dflash-fi0618-situ4460 HEALTH_WAIT_S=2400 GSM8K="${SCRIPT_DIR}/../../../datasets/gsm8k/test.jsonl" GSM8K_SHA256=3730d312f6e3440559ace48831e51066acaca737f6eabec99bccb9e4b3c39d14 # Reuse the established client and its prompt/telemetry definitions. BENCH_CLIENT="${SCRIPT_DIR}/bench_gsm8k_acceptance.py"