A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 12:15 UTC
[Benchmarks] Remove invalid option under V1 engine (#17651)
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Russell Bryant committed
May 5, 2025 at 16:30 UTC
d3efde817619d9f573f7571ed49ba5e8dd2be9b4
2 files changed
+3
-20
benchmarks/benchmark_serving_structured_output.py
-12
@@ -414,7 +414,6 @@ async def benchmark(
414
ignore_eos: bool,
415
max_concurrency: Optional[int],
416
structured_output_ratio: float,
417
- structured_output_backend: str,
417
goodput_config_dict: Optional[dict[str, float]] = None,
418
):
419
if backend in ASYNC_REQUEST_FUNCS:
@@ -426,8 +425,6 @@ async def benchmark(
425
extra_body = {}
426
# Add the schema to the extra_body
427
extra_body[request.structure_type] = request.schema
429
- # Add the specific structured_output_backend
430
- extra_body["guided_decoding_backend"] = structured_output_backend
428
return extra_body
429
430
print("Starting initial single prompt test run...")
@@ -785,7 +782,6 @@ def main(args: argparse.Namespace):
782
ignore_eos=args.ignore_eos,
783
max_concurrency=args.max_concurrency,
784
structured_output_ratio=args.structured_output_ratio,
788
- structured_output_backend=args.structured_output_backend,
785
goodput_config_dict=goodput_config_dict,
786
))
787
@@ -1000,14 +996,6 @@ if __name__ == "__main__":
996
type=float,
997
default=1.0,
998
help="Ratio of Structured Outputs requests")
1003
- parser.add_argument("--structured-output-backend",
1004
- type=str,
1005
- choices=[
1006
- "outlines", "lm-format-enforcer", "xgrammar",
1007
- "guidance", "auto"
1008
- ],
1009
- default="auto",
1010
- help="Backend to use for structured outputs")
999
1000
args = parser.parse_args()
1001
main(args)
benchmarks/run_structured_output_benchmark.sh
+3
-8
@@ -9,13 +9,10 @@ BACKEND=${2:-"vllm"}
9
# Define the dataset to use
10
DATASET=${3:-"xgrammar_bench"}
11
12
-# Define the guided decoding backend
13
-GUIDED_BACKEND=${4:-"xgrammar"}
14
-
12
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
16
-OUTPUT_DIR=${5:-"$SCRIPT_DIR/structured_output_benchmark_results"}
13
+OUTPUT_DIR=${4:-"$SCRIPT_DIR/structured_output_benchmark_results"}
14
18
-GUIDED_RATIO=${6:-0.5}
15
+GUIDED_RATIO=${5:-0.5}
16
17
# Create output directory if it doesn't exist
18
mkdir -p "$OUTPUT_DIR"
@@ -27,7 +24,6 @@ QPS_VALUES=(70 60 50 25 20 15 10)
24
COMMON_PARAMS="--backend $BACKEND \
25
--model $MODEL \
26
--dataset $DATASET \
30
- --structured-output-backend $GUIDED_BACKEND \
27
--structured-output-ratio $GUIDED_RATIO \
28
--save-results \
29
--result-dir $OUTPUT_DIR"
@@ -35,7 +31,6 @@ COMMON_PARAMS="--backend $BACKEND \
31
echo "Starting structured output benchmark with model: $MODEL"
32
echo "Backend: $BACKEND"
33
echo "Dataset: $DATASET"
38
-echo "Structured output backend: $GUIDED_BACKEND"
34
echo "Results will be saved to: $OUTPUT_DIR"
35
echo "----------------------------------------"
36
@@ -48,7 +43,7 @@ for qps in "${QPS_VALUES[@]}"; do
43
GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown")
44
45
# Construct filename for this run
51
- FILENAME="${GUIDED_BACKEND}_${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json"
46
+ FILENAME="${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json"
47
48
# Run the benchmark
49
python "$SCRIPT_DIR/benchmark_serving_structured_output.py" $COMMON_PARAMS \