a4e38b9e33
Move all experiments under hardware-specific folders: - experiments/h200/ : H200 GPU experiments (15 dirs) - experiments/h20/ : H20 GPU experiments (2 dirs) - experiments/p800/ : Kunlun P800 experiments (3 dirs) - experiments/pro6000/ : RTX 6000D experiments (2 dirs) This improves discoverability and keeps hardware-specific configs isolated from each other.
36 lines
1.0 KiB
Bash
36 lines
1.0 KiB
Bash
# 64k-context focused comparison for SGLang vs vLLM on H200.
|
|
# Scenarios requested:
|
|
# 1) input=65536, output=256, concurrency=25
|
|
# 2) input=65536, output=1024, concurrency=8, num_prompts=100
|
|
|
|
EXPERIMENT="dsv4_h200_64k_sglang_vs_vllm"
|
|
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
|
|
|
|
# 64k + 1024 output + padding -> use 70000 to keep both backends identical.
|
|
MAX_MODEL_LEN=70000
|
|
MAX_NUM_SEQS=64
|
|
MAX_RUNNING=64
|
|
|
|
# "concurrency input_len output_len num_prompts"
|
|
# Following docs/EXPERIMENT_GUIDE.md: num_prompts = concurrency * 5.
|
|
declare -a SCENARIOS=(
|
|
"25 65536 256 125"
|
|
"8 65536 1024 40"
|
|
)
|
|
|
|
VENV_CLIENT="${VENV_CLIENT:-$VENV_SGLANG}"
|
|
|
|
SGLANG_START_SCRIPT="${SCRIPT_DIR:-.}/start_sglang.sh"
|
|
VLLM_START_SCRIPT="${SCRIPT_DIR:-.}/start_vllm.sh"
|