#!/usr/bin/env bash
set -Eeuo pipefail

RESULT_DIR=/data/wxy/sskj-h3/throughput/sglang-base/results/subblock-sm120-probe
SGLANG_BIN=/root/.miniconda3/envs/sglang-lora/bin/sglang
MEDIA_BIN_DIR=/root/.miniconda3/envs/deploy/bin

mkdir -p "$RESULT_DIR"
env -u SGLANG_CACHE_DIT_ENABLED -u SGLANG_CACHE_DIT_FN -u SGLANG_CACHE_DIT_BN \
  -u SGLANG_CACHE_DIT_RDT -u SGLANG_CACHE_DIT_MC -u SGLANG_CACHE_DIT_WARMUP \
  -u SGLANG_CACHE_DIT_TAYLORSEER -u SGLANG_CACHE_DIT_SCM_PRESET \
  -u SGLANG_CACHE_DIT_SCM_POLICY \
  CUDA_VISIBLE_DEVICES=0,1 \
  PATH="$MEDIA_BIN_DIR:$PATH" \
  PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  PYTHONUNBUFFERED=1 TOKENIZERS_PARALLELISM=false SGLANG_USE_RUNAI_MODEL_STREAMER=false \
  "$SGLANG_BIN" serve \
    --model-path /data/hf_models/MiniMax-H3 \
    --model-variant Ref2VA \
    --backend sglang \
    --performance-mode speed \
    --num-gpus 2 \
    --tp-size 2 \
    --ulysses-degree 1 \
    --use-fsdp-inference false \
    --enable-torch-compile false \
    --batching-max-size 1 \
    --batching-delay-ms 0 \
    --host 127.0.0.1 \
    --port 34910 \
    --master-port 35910 \
    --scheduler-port 36910 \
    --output-path "$RESULT_DIR/outputs" \
    --attention-backend subblock_sparse_attn \
    --component-attention-backends text_encoder=fa \
    --attention-backend-config /data/wxy/sskj-h3/throughput/sglang-base/configs/subblock_sparse_075.json