EXPERIMENT_BASE=kimi3_pro6000_pd_dflash_attribution 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 # The P deployment is the established PP8 TTFT optimum. D is TP32/PP1. P_TP=4 P_PP=8 P_EP=4 D_TP=32 D_PP=1 D_EP=4 P_MEM=0.84 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 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 SPEED_BENCH_DATASET=/data/speedbench16k_complete.jsonl BENCH_PROMPTS=40 PROFILE_PROMPTS=8 OUTPUT_LEN=512 CONCURRENCY=8 # Per-node temporary telemetry/profile output; copied back to the 601 result tree. REMOTE_WORK_BASE=/data/hzy/kimi3_pd_dflash_attribution TELEMETRY_DURATION_S=900 TELEMETRY_INTERVAL_S=1 P_PROFILE_STEPS=4 D_PROFILE_STEPS=20