Profile Large Language Models with GenAI-Perf

Profile Large Language Models with GenAI-Perf#

Profile GPT2 running on Triton + TensorRT-LLM #

Run GPT2 on Triton Inference Server using TensorRT-LLM#

See instructions

Run Triton Inference Server with TensorRT-LLM backend container:

export RELEASE="yy.mm" # e.g. export RELEASE="24.06"

docker run -it --net=host --gpus=all --shm-size=2g --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/tritonserver:${RELEASE}-trtllm-python-py3

# Install Triton CLI (~5 min):
pip install "git+https://github.com/triton-inference-server/triton_cli@0.0.8"

# Download model:
triton import -m gpt2 --backend tensorrtllm

# Run server:
triton start

Run GenAI-Perf#

Run GenAI-Perf from Triton Inference Server SDK container:

export RELEASE="yy.mm" # e.g. export RELEASE="24.06"

docker run -it --net=host --gpus=all nvcr.io/nvidia/tritonserver:${RELEASE}-py3-sdk

# Run GenAI-Perf in the container:
genai-perf profile \
  -m gpt2 \
  --service-kind triton \
  --backend tensorrtllm \
  --num-prompts 100 \
  --random-seed 123 \
  --synthetic-input-tokens-mean 200 \
  --synthetic-input-tokens-stddev 0 \
  --streaming \
  --output-tokens-mean 100 \
  --output-tokens-stddev 0 \
  --output-tokens-mean-deterministic \
  --tokenizer hf-internal-testing/llama-tokenizer \
  --concurrency 1 \
  --measurement-interval 4000 \
  --profile-export-file my_profile_export.json \
  --url localhost:8001

Example output:

                                                  LLM Metrics
┏━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃                Statistic ┃         avg ┃         min ┃         max ┃         p99 ┃         p90 ┃         p75 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│ Time to first token (ns) │  13,266,974 │  11,818,732 │  18,351,779 │  16,513,479 │  13,741,986 │  13,544,376 │
│ Inter token latency (ns) │   2,069,766 │      42,023 │  15,307,799 │   3,256,375 │   3,020,580 │   2,090,930 │
│     Request latency (ns) │ 223,532,625 │ 219,123,330 │ 241,004,192 │ 238,198,306 │ 229,676,183 │ 224,715,918 │
│   Output sequence length │         104 │         100 │         129 │         128 │         109 │         105 │
│    Input sequence length │         199 │         199 │         199 │         199 │         199 │         199 │
└──────────────────────────┴─────────────┴─────────────┴─────────────┴─────────────┴─────────────┴─────────────┘
Output token throughput (per sec): 460.42
Request throughput (per sec): 4.44

Profile GPT2 running on Triton + vLLM #

Run GPT2 on Triton Inference Server using vLLM#

See instructions

Run Triton Inference Server with vLLM backend container:

export RELEASE="yy.mm" # e.g. export RELEASE="24.06"


docker run -it --net=host --gpus=1 --shm-size=2g --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/tritonserver:${RELEASE}-vllm-python-py3

# Install Triton CLI (~5 min):
pip install "git+https://github.com/triton-inference-server/triton_cli@0.0.8"

# Download model:
triton import -m gpt2 --backend vllm

# Run server:
triton start

Run GenAI-Perf#

Run GenAI-Perf from Triton Inference Server SDK container:

export RELEASE="yy.mm" # e.g. export RELEASE="24.06"

docker run -it --net=host --gpus=1 nvcr.io/nvidia/tritonserver:${RELEASE}-py3-sdk

# Run GenAI-Perf in the container:
genai-perf profile \
  -m gpt2 \
  --service-kind triton \
  --backend vllm \
  --num-prompts 100 \
  --random-seed 123 \
  --synthetic-input-tokens-mean 200 \
  --synthetic-input-tokens-stddev 0 \
  --streaming \
  --output-tokens-mean 100 \
  --output-tokens-stddev 0 \
  --output-tokens-mean-deterministic \
  --tokenizer hf-internal-testing/llama-tokenizer \
  --concurrency 1 \
  --measurement-interval 4000 \
  --profile-export-file my_profile_export.json \
  --url localhost:8001

Example output:

                                                  LLM Metrics
┏━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃                Statistic ┃         avg ┃         min ┃         max ┃         p99 ┃         p90 ┃         p75 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│ Time to first token (ns) │  15,786,560 │  11,437,189 │  49,550,549 │  40,129,652 │  21,248,091 │  17,824,695 │
│ Inter token latency (ns) │   3,543,380 │     591,898 │  10,013,690 │   6,152,260 │   5,039,278 │   4,060,982 │
│     Request latency (ns) │ 388,415,721 │ 312,552,612 │ 528,229,817 │ 518,189,390 │ 484,281,365 │ 459,417,637 │
│   Output sequence length │         113 │         105 │         123 │         122 │         119 │         115 │
│    Input sequence length │         199 │         199 │         199 │         199 │         199 │         199 │
└──────────────────────────┴─────────────┴─────────────┴─────────────┴─────────────┴─────────────┴─────────────┘
Output token throughput (per sec): 290.24
Request throughput (per sec): 2.57

Profile GPT2 running on OpenAI Chat API-Compatible Server #

Run GPT2 on OpenAI Chat Completions API-compatible server#

See instructions

Run the vLLM inference server:

docker run -it --net=host --gpus=all vllm/vllm-openai:latest --model gpt2 --dtype float16 --max-model-len 1024

Run GenAI-Perf#

Run GenAI-Perf from Triton Inference Server SDK container:

export RELEASE="yy.mm" # e.g. export RELEASE="24.06"

docker run -it --net=host --gpus=all nvcr.io/nvidia/tritonserver:${RELEASE}-py3-sdk

# Run GenAI-Perf in the container:
genai-perf profile \
  -m gpt2 \
  --service-kind openai \
  --endpoint v1/chat/completions \
  --endpoint-type chat \
  --num-prompts 100 \
  --random-seed 123 \
  --synthetic-input-tokens-mean 200 \
  --synthetic-input-tokens-stddev 0 \
  --streaming \
  --output-tokens-mean 100 \
  --output-tokens-stddev 0 \
  --tokenizer hf-internal-testing/llama-tokenizer \
  --concurrency 1 \
  --measurement-interval 4000 \
  --profile-export-file my_profile_export.json \
  --url localhost:8000

Example output:

                                                  LLM Metrics
┏━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃                Statistic ┃         avg ┃         min ┃         max ┃         p99 ┃         p90 ┃         p75 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│ Time to first token (ns) │  13,546,815 │   9,821,658 │  48,317,756 │  34,361,913 │  16,541,625 │  14,612,026 │
│ Inter token latency (ns) │   2,560,813 │     457,703 │   6,507,334 │   3,754,617 │   3,059,158 │   2,953,540 │
│     Request latency (ns) │ 283,597,027 │ 240,098,890 │ 361,730,568 │ 349,164,037 │ 323,279,761 │ 306,507,562 │
│   Output sequence length │         114 │         103 │         142 │         136 │         122 │         119 │
│    Input sequence length │         199 │         199 │         199 │         199 │         199 │         199 │
└──────────────────────────┴─────────────┴─────────────┴─────────────┴─────────────┴─────────────┴─────────────┘
Output token throughput (per sec): 401.62
Request throughput (per sec): 3.52

Profile GPT2 running on OpenAI Completions API-Compatible Server #

Running GPT2 on OpenAI Completions API-compatible server#

See instructions

Run the vLLM inference server:

docker run -it --net=host --gpus=all vllm/vllm-openai:latest --model gpt2 --dtype float16 --max-model-len 1024

Run GenAI-Perf#

Run GenAI-Perf from Triton Inference Server SDK container:

export RELEASE="yy.mm" # e.g. export RELEASE="24.06"

docker run -it --net=host --gpus=all nvcr.io/nvidia/tritonserver:${RELEASE}-py3-sdk


# Run GenAI-Perf in the container:
genai-perf profile \
  -m gpt2 \
  --service-kind openai \
  --endpoint v1/completions \
  --endpoint-type completions \
  --num-prompts 100 \
  --random-seed 123 \
  --synthetic-input-tokens-mean 200 \
  --synthetic-input-tokens-stddev 0 \
  --output-tokens-mean 100 \
  --output-tokens-stddev 0 \
  --tokenizer hf-internal-testing/llama-tokenizer \
  --concurrency 1 \
  --measurement-interval 4000 \
  --profile-export-file my_profile_export.json \
  --url localhost:8000

Example output:

                                                LLM Metrics
┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃              Statistic ┃         avg ┃        min ┃         max ┃         p99 ┃         p90 ┃         p75 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│   Request latency (ns) │ 296,990,497 │ 43,312,449 │ 332,788,242 │ 327,475,292 │ 317,392,767 │ 310,343,333 │
│ Output sequence length │         109 │         11 │         158 │         142 │         118 │         113 │
│  Input sequence length │           1 │          1 │           1 │           1 │           1 │           1 │
└────────────────────────┴─────────────┴────────────┴─────────────┴─────────────┴─────────────┴─────────────┘
Output token throughput (per sec): 366.78
Request throughput (per sec): 3.37
Contents