#!/usr/bin/env bash
# TP×DP matrix experiment for DeepSeek-V4-Flash on RTX 6000D (8 GPUs) using SGLang.
# TP=2/DP=4 is excluded because Marlin weight loading OOMs before KV Cache
# initialization on this machine.
# Tests SGLang with two parallel configurations:
#   TP=4, DP=2  -> 4 GPUs per replica, 2 replicas
#   TP=8, DP=1  -> 8 GPUs, no data parallelism

EXPERIMENT="dsv4_pro6000_sglang_tp_dp_matrix"
MODEL_NAME="DeepSeek-V4-Flash"
MODEL_PATH="/data/6000D/DeepSeek-V4-Flash"
SERVED_MODEL_NAME="deepseek-v4-flash"

SGLANG_PORT="${SGLANG_PORT:-30031}"

# Python interpreter for orchestration scripts (parse_backend.py, compare.py, etc.)
# and the benchmark client. Defaults to the system python3 if the sglang venv
# does not exist on the host.
VENV_CLIENT="${VENV_CLIENT:-/root/.miniconda3/envs/sglang}"

# Run the benchmark client natively (0) or inside Docker (1).
USE_DOCKER_CLIENT="${USE_DOCKER_CLIENT:-1}"

export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"

# Runtime working directory for logs, pid files, and tmp. Defaults to a local
# directory under this experiment so the benchmark is self-contained.
RUNTIME_BASE="${RUNTIME_BASE:-${SCRIPT_DIR}/runtime}"

# Parallel configurations to test. Format: "TP DP"
declare -a PARALLEL_CONFIGS=(
  "4 2"
  "8 1"
)

# SGLang nightly server settings for RTX 6000D / DeepSeek-V4-Flash.
MEM_FRACTION_STATIC="${MEM_FRACTION_STATIC:-0.90}"
MOE_RUNNER_BACKEND="${MOE_RUNNER_BACKEND:-auto}"
CONTEXT_LENGTH="${CONTEXT_LENGTH:-131072}"
MAX_RUNNING_REQUESTS="${MAX_RUNNING_REQUESTS:-64}"

# Deployment switch. 0 = native sglang venv, 1 = Docker.
USE_DOCKER="${USE_DOCKER:-1}"
DOCKER_IMAGE="${DOCKER_IMAGE:-lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45}"

# To use ShareGPT, set BENCH_DATASET_NAME=random and DATASET_PATH explicitly.
BENCH_DATASET_NAME="${BENCH_DATASET_NAME:-random}"
DATASET_PATH="${DATASET_PATH:-${ROOT_DIR}/dataset/ShareGPT_V3_unfiltered_cleaned_split.json}"
SGLANG_BENCH_MODULE="${SGLANG_BENCH_MODULE:-sglang.benchmark.serving}"
CACHE_DIR="${CACHE_DIR:-${ROOT_DIR}/sglang_nightly_cu13_cache}"

# Matrix and concurrency rules are defined in matrix.json by default.
MATRIX_FILE="${MATRIX_FILE:-${SCRIPT_DIR:-.}/matrix.json}"
MATRIX_MODE="${MATRIX_MODE:-Y}"

# Sampling density for concurrency.
# 0 = use the default heuristic in generate_scenarios.py (6-8 points).
# 2 = only test the low and high endpoints.
export CONCURRENCY_SAMPLES="${CONCURRENCY_SAMPLES:-2}"

# Per-scenario timeout to avoid hangs (seconds).
SCENARIO_TIMEOUT_S="${SCENARIO_TIMEOUT_S:-1800}"

# GPU memory sampling interval (seconds).
GPU_MEM_SAMPLE_INTERVAL_S="${GPU_MEM_SAMPLE_INTERVAL_S:-1}"

# Dry-run mode: if 1, only log the server args and scenario plan without starting
# any server or sending requests.
DRY_RUN="${DRY_RUN:-0}"

# Per-config scenario limit for quick smoke tests. 0 = run all generated scenarios.
GRID_LIMIT="${GRID_LIMIT:-0}"

# PyTorch CUDA allocator setting for the SGLang server. expandable_segments
# reduces fragmentation from the GiB-scale indexer temporaries that OOM the
# SM120 torch fallback (fp8_paged_mqa_logits_torch_sm120) at ISL >= 4096.
PYTORCH_CUDA_ALLOC_CONF="${PYTORCH_CUDA_ALLOC_CONF:-expandable_segments:True}"