qqtang/sskj
1
2
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[Feat] use nightly SGLang FlashInfer MoE

Browse Source
This commit is contained in:
Quantong Qiu 2026-07-21 15:22:47 +08:00
parent d0319504e4
commit 8c737b840b
10 changed files with 16 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
README.md
View File

@ -1,5 +1,8 @@
# sskj — 多平台大模型推理性能基准测试项目
> **更新（2026-07-21 14:12:30 +0800）**
> - RTX 6000D / DeepSeek-V4-Flash 的 SGLang 全量与 tiny 脚本切换到 `lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45`，MoE backend 保持 `auto`，由 nightly 自动选择可用的 FlashInfer MoE 实现；同时使用独立 nightly CUDA 13 cache，避免复用旧 SM120 修补镜像的编译缓存。
> **更新（2026-07-21 14:00:30 +0800）**
> - adaptive 并发搜索新增可配置的 `C=1 OOM` 配置级短路：同一 TP/DP 在最小并发仍 OOM 时，直接跳过该组更长输入，避免把已经确定不可部署的长度继续压测；6000D DSV4-Flash SGLang tiny 已启用，并支持从已有结果恢复时识别此前的该边界。

6
experiments/pro6000/dsv4_pro6000_sglang_tiny_1k_output/config.env
View File

@ -34,7 +34,7 @@ declare -a PARALLEL_CONFIGS=(
"8 1"
)
# SGLang server settings, verified for this model and machine.
# SGLang nightly server settings for this model and machine.
MEM_FRACTION_STATIC="${MEM_FRACTION_STATIC:-0.90}"
MOE_RUNNER_BACKEND="${MOE_RUNNER_BACKEND:-auto}"
CONTEXT_LENGTH="${CONTEXT_LENGTH:-131072}"
@ -44,13 +44,13 @@ SKIP_CONFIG_ON_MIN_CONCURRENCY_OOM="${SKIP_CONFIG_ON_MIN_CONCURRENCY_OOM:-1}"
# Deployment switch. 0 = native sglang venv, 1 = Docker.
USE_DOCKER="${USE_DOCKER:-1}"
DOCKER_IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
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:-${ROOT_DIR}/sglang_sm120_cache}"
CACHE_DIR="${CACHE_DIR:-${ROOT_DIR}/sglang_nightly_cu13_cache}"
# Matrix contains ISL=1K..128K with fixed OSL=1K.
MATRIX_FILE="${MATRIX_FILE:-${SCRIPT_DIR:-.}/matrix.json}"

2
experiments/pro6000/dsv4_pro6000_sglang_tiny_1k_output/run_adaptive_concurrency_add16.sh
View File

@ -27,7 +27,7 @@ else
PYTHON="$(command -v python3)"
fi
DOCKER_IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
DOCKER_IMAGE="${DOCKER_IMAGE:-lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45}"
engine_is_healthy() {
curl --fail --silent --show-error --max-time 5 \

2
experiments/pro6000/dsv4_pro6000_sglang_tiny_1k_output/start_sglang_docker.sh
View File

@ -16,7 +16,7 @@ source "${SCRIPT_DIR}/config.env"
RUNTIME_BASE="${RUNTIME_BASE:-${SCRIPT_DIR}/runtime}"
mkdir -p "${RUNTIME_BASE}/logs" "${RUNTIME_BASE}/tmp" "$CACHE_DIR"
IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
IMAGE="${DOCKER_IMAGE:-lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45}"
PORT="${SGLANG_PORT:-30031}"
NAME="${EXPERIMENT}_sglang_tp${TP}_dp${DP}"
PID_FILE="${RUNTIME_BASE}/${EXPERIMENT}_sglang_tp${TP}_dp${DP}.pid"

6
experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/config.env
View File

@ -33,7 +33,7 @@ declare -a PARALLEL_CONFIGS=(
"8 1"
)
# SGLang server settings, verified for RTX 6000D / DeepSeek-V4-Flash.
# 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="${CONTEXT_LENGTH:-131072}"
@ -41,13 +41,13 @@ MAX_RUNNING_REQUESTS="${MAX_RUNNING_REQUESTS:-64}"
# Deployment switch. 0 = native sglang venv, 1 = Docker.
USE_DOCKER="${USE_DOCKER:-1}"
DOCKER_IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
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:-${ROOT_DIR}/sglang_sm120_cache}"
CACHE_DIR="${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}"

2
experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/run_adaptive_concurrency.sh
View File

@ -27,7 +27,7 @@ else
PYTHON="$(command -v python3)"
fi
DOCKER_IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
DOCKER_IMAGE="${DOCKER_IMAGE:-lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45}"
engine_is_healthy() {
curl --fail --silent --show-error --max-time 5 \

2
experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh
View File

@ -27,7 +27,7 @@ else
PYTHON="$(command -v python3)"
fi
DOCKER_IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
DOCKER_IMAGE="${DOCKER_IMAGE:-lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45}"
engine_is_healthy() {
curl --fail --silent --show-error --max-time 5 \

2
experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/run_bench.sh
View File

@ -30,7 +30,7 @@ else
PYTHON="$(command -v python3)"
fi
DOCKER_IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
DOCKER_IMAGE="${DOCKER_IMAGE:-lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45}"
log_dir_global="${RESULT_BASE}/${RUN_ID}/logs"
mkdir -p "$log_dir_global"

2
experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/start_sglang_container.sh
View File

@ -13,7 +13,7 @@ source "${SCRIPT_DIR}/config.env"
RUNTIME_BASE="${RUNTIME_BASE:-${SCRIPT_DIR}/runtime}"
mkdir -p "${RUNTIME_BASE}/logs" "${RUNTIME_BASE}/tmp"
IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
IMAGE="${DOCKER_IMAGE:-lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45}"
PORT="${SGLANG_PORT:-30031}"
NAME="${EXPERIMENT}_sglang_container"
PID_FILE="${RUNTIME_BASE}/${EXPERIMENT}_container.pid"

2
experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/start_sglang_docker.sh
View File

@ -16,7 +16,7 @@ source "${SCRIPT_DIR}/config.env"
RUNTIME_BASE="${RUNTIME_BASE:-${SCRIPT_DIR}/runtime}"
mkdir -p "${RUNTIME_BASE}/logs" "${RUNTIME_BASE}/tmp" "$CACHE_DIR"
IMAGE="${DOCKER_IMAGE:-sglang-sm120-dsv4:0.5.15.post1-fi0.6.14-sm120fix1}"
IMAGE="${DOCKER_IMAGE:-lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45}"
PORT="${SGLANG_PORT:-30031}"
NAME="${EXPERIMENT}_sglang_tp${TP}_dp${DP}"
PID_FILE="${RUNTIME_BASE}/${EXPERIMENT}_sglang_tp${TP}_dp${DP}.pid"