fix: max_running_requests=256, cuda_graph_max_bs=256 to match max concurrency 128
This commit is contained in:
parent
da1d4d3d78
commit
cd362c5ca0
@ -32,9 +32,11 @@ NCCL_SOCKET_IFNAME="${NCCL_SOCKET_IFNAME:-eth0}"
|
|||||||
NCCL_DEBUG="${NCCL_DEBUG:-WARN}"
|
NCCL_DEBUG="${NCCL_DEBUG:-WARN}"
|
||||||
|
|
||||||
# ---- SGLang memory / perf ----
|
# ---- SGLang memory / perf ----
|
||||||
|
# max_running_requests must >= max concurrency tested (128) to avoid queueing
|
||||||
|
# cuda_graph_max_bs_decode must >= max concurrent batch size for accurate decode perf
|
||||||
MEM_FRACTION_STATIC="${MEM_FRACTION_STATIC:-0.9}"
|
MEM_FRACTION_STATIC="${MEM_FRACTION_STATIC:-0.9}"
|
||||||
CUDA_GRAPH_MAX_BS_DECODE="${CUDA_GRAPH_MAX_BS_DECODE:-8}"
|
CUDA_GRAPH_MAX_BS_DECODE="${CUDA_GRAPH_MAX_BS_DECODE:-256}"
|
||||||
MAX_RUNNING_REQUESTS="${MAX_RUNNING_REQUESTS:-32}"
|
MAX_RUNNING_REQUESTS="${MAX_RUNNING_REQUESTS:-256}"
|
||||||
|
|
||||||
# ---- Benchmark ----
|
# ---- Benchmark ----
|
||||||
BENCH_DATASET_NAME="${BENCH_DATASET_NAME:-random}"
|
BENCH_DATASET_NAME="${BENCH_DATASET_NAME:-random}"
|
||||||
@ -44,10 +46,6 @@ RANDOM_RANGE_RATIO="${RANDOM_RANGE_RATIO:-1.0}"
|
|||||||
WARMUP_REQUESTS="${WARMUP_REQUESTS:-16}"
|
WARMUP_REQUESTS="${WARMUP_REQUESTS:-16}"
|
||||||
NUM_REPETITIONS="${NUM_REPETITIONS:-3}"
|
NUM_REPETITIONS="${NUM_REPETITIONS:-3}"
|
||||||
|
|
||||||
# ---- SLO ----
|
|
||||||
TTFT_SLO_MS="${TTFT_SLO_MS:-4000}"
|
|
||||||
ENABLE_TTFT_SLO_STOP="${ENABLE_TTFT_SLO_STOP:-1}"
|
|
||||||
|
|
||||||
# ---- Timeouts ----
|
# ---- Timeouts ----
|
||||||
SCENARIO_TIMEOUT_S="${SCENARIO_TIMEOUT_S:-1800}"
|
SCENARIO_TIMEOUT_S="${SCENARIO_TIMEOUT_S:-1800}"
|
||||||
HEALTH_CHECK_RETRIES="${HEALTH_CHECK_RETRIES:-600}"
|
HEALTH_CHECK_RETRIES="${HEALTH_CHECK_RETRIES:-600}"
|
||||||
|
|||||||
Loading…
x
Reference in New Issue
Block a user