# OOM probe for 256k input + 4k output on DeepSeek-V4-Flash (H200).
# Tests whether SGLang and vLLM can serve this context length without OOM
# at concurrency=2 before adding it to the full long-context matrix.

EXPERIMENT="dsv4_h200_256k_4k_probe"
MODEL_NAME="DeepSeek-V4-Flash"
MODEL_PATH="/data/models/DeepSeek-V4-Flash"
SERVED_MODEL_NAME="deepseek-v4-flash"

SGLANG_PORT="${SGLANG_PORT:-30006}"
VLLM_PORT="${VLLM_PORT:-30005}"

VENV_SGLANG="${VENV_SGLANG:-/data/user1/yy/envs/sglang}"
VENV_VLLM="${VENV_VLLM:-/data/user1/yy/envs/vllm}"

export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
TP=8

# 256k input + 4k output + padding -> use 270000.
MAX_MODEL_LEN=270000
MAX_NUM_SEQS=8
MAX_RUNNING=8

# "concurrency input_len output_len num_prompts"
declare -a SCENARIOS=(
  "2 262144 4096 4"
)

VENV_CLIENT="${VENV_CLIENT:-$VENV_SGLANG}"

SGLANG_START_SCRIPT="${SCRIPT_DIR:-.}/start_sglang.sh"
VLLM_START_SCRIPT="${SCRIPT_DIR:-.}/start_vllm.sh"