OpenAI APIs - Vision

OpenAI APIs - Vision#

SGLang provides OpenAI-compatible APIs to enable a smooth transition from OpenAI services to self-hosted local models. A complete reference for the API is available in the OpenAI API Reference. This tutorial covers the vision APIs for vision language models.

SGLang supports vision language models such as Llama 3.2, LLaVA-OneVision, and QWen-VL2
- meta-llama/Llama-3.2-11B-Vision-Instruct
- lmms-lab/llava-onevision-qwen2-72b-ov-chat
- Qwen/Qwen2-VL-7B-Instruct

Launch A Server#

This code block is equivalent to executing

python3 -m sglang.launch_server --model-path meta-llama/Llama-3.2-11B-Vision-Instruct \
  --port 30010 --chat-template llama_3_vision

in your terminal and wait for the server to be ready.

Remember to add --chat-template llama_3_vision to specify the vision chat template, otherwise the server only supports text. We need to specify --chat-template for vision language models because the chat template provided in Hugging Face tokenizer only supports text.

[1]:

from sglang.utils import (
    execute_shell_command,
    wait_for_server,
    terminate_process,
    print_highlight,
)

embedding_process = execute_shell_command(
"""
python3 -m sglang.launch_server --model-path meta-llama/Llama-3.2-11B-Vision-Instruct \
    --port=30010 --chat-template=llama_3_vision
"""
)

wait_for_server("http://localhost:30010")

[2024-11-02 03:46:49] server_args=ServerArgs(model_path='meta-llama/Llama-3.2-11B-Vision-Instruct', tokenizer_path='meta-llama/Llama-3.2-11B-Vision-Instruct', tokenizer_mode='auto', skip_tokenizer_init=False, load_format='auto', trust_remote_code=False, dtype='auto', kv_cache_dtype='auto', quantization=None, context_length=None, device='cuda', served_model_name='meta-llama/Llama-3.2-11B-Vision-Instruct', chat_template='llama_3_vision', is_embedding=False, host='127.0.0.1', port=30010, mem_fraction_static=0.88, max_running_requests=None, max_total_tokens=None, chunked_prefill_size=8192, max_prefill_tokens=16384, schedule_policy='lpm', schedule_conservativeness=1.0, tp_size=1, stream_interval=1, random_seed=101706643, constrained_json_whitespace_pattern=None, decode_log_interval=40, log_level='info', log_level_http=None, log_requests=False, show_time_cost=False, api_key=None, file_storage_pth='SGLang_storage', enable_cache_report=False, watchdog_timeout=600, dp_size=1, load_balance_method='round_robin', dist_init_addr=None, nnodes=1, node_rank=0, json_model_override_args='{}', enable_double_sparsity=False, ds_channel_config_path=None, ds_heavy_channel_num=32, ds_heavy_token_num=256, ds_heavy_channel_type='qk', ds_sparse_decode_threshold=4096, lora_paths=None, max_loras_per_batch=8, attention_backend='flashinfer', sampling_backend='flashinfer', grammar_backend='outlines', disable_flashinfer=False, disable_flashinfer_sampling=False, disable_radix_cache=False, disable_regex_jump_forward=False, disable_cuda_graph=False, disable_cuda_graph_padding=False, disable_disk_cache=False, disable_custom_all_reduce=False, disable_mla=False, disable_penalizer=False, disable_nan_detection=False, enable_overlap_schedule=False, enable_mixed_chunk=False, enable_torch_compile=False, torch_compile_max_bs=32, cuda_graph_max_bs=160, torchao_config='', enable_p2p_check=False, triton_attention_reduce_in_fp32=False, num_continuous_decode_steps=1)
[2024-11-02 03:46:56] Use chat template for the OpenAI-compatible API server: llama_3_vision
[2024-11-02 03:47:06 TP0] Automatically turn off --chunked-prefill-size and adjust --mem-fraction-static for multimodal models.
[2024-11-02 03:47:06 TP0] Init torch distributed begin.
[2024-11-02 03:47:06 TP0] Load weight begin. avail mem=78.59 GB
[2024-11-02 03:47:06 TP0] lm_eval is not installed, GPTQ may not be usable
INFO 11-02 03:47:07 weight_utils.py:243] Using model weights format ['*.safetensors']
Loading safetensors checkpoint shards:   0% Completed | 0/5 [00:00<?, ?it/s]
Loading safetensors checkpoint shards:  20% Completed | 1/5 [00:00<00:03,  1.07it/s]
Loading safetensors checkpoint shards:  40% Completed | 2/5 [00:01<00:02,  1.01it/s]
Loading safetensors checkpoint shards:  60% Completed | 3/5 [00:02<00:01,  1.00it/s]
Loading safetensors checkpoint shards:  80% Completed | 4/5 [00:04<00:01,  1.02s/it]
Loading safetensors checkpoint shards: 100% Completed | 5/5 [00:04<00:00,  1.22it/s]
Loading safetensors checkpoint shards: 100% Completed | 5/5 [00:04<00:00,  1.11it/s]

[2024-11-02 03:47:11 TP0] Load weight end. type=MllamaForConditionalGeneration, dtype=torch.bfloat16, avail mem=58.43 GB
[2024-11-02 03:47:11 TP0] Memory pool end. avail mem=11.80 GB
[2024-11-02 03:47:12 TP0] Capture cuda graph begin. This can take up to several minutes.
[2024-11-02 03:47:21 TP0] max_total_num_tokens=298440, max_prefill_tokens=16384, max_running_requests=2049, context_len=131072
[2024-11-02 03:47:21] INFO:     Started server process [3455368]
[2024-11-02 03:47:21] INFO:     Waiting for application startup.
[2024-11-02 03:47:21] INFO:     Application startup complete.
[2024-11-02 03:47:21] INFO:     Uvicorn running on http://127.0.0.1:30010 (Press CTRL+C to quit)
[2024-11-02 03:47:22] INFO:     127.0.0.1:36556 - "GET /v1/models HTTP/1.1" 200 OK
[2024-11-02 03:47:22] INFO:     127.0.0.1:36562 - "GET /get_model_info HTTP/1.1" 200 OK
[2024-11-02 03:47:22 TP0] Prefill batch. #new-seq: 1, #new-token: 7, #cached-token: 0, cache hit rate: 0.00%, token usage: 0.00, #running-req: 0, #queue-req: 0
[2024-11-02 03:47:22] INFO:     127.0.0.1:36564 - "POST /generate HTTP/1.1" 200 OK
[2024-11-02 03:47:22] The server is fired up and ready to roll!


NOTE: Typically, the server runs in a separate terminal.
In this notebook, we run the server and notebook code together, so their outputs are combined.
To improve clarity, the server logs are displayed in the original black color, while the notebook outputs are highlighted in blue.

Using cURL#

Once the server is up, you can send test requests using curl.

[2]:

import subprocess

curl_command = """
curl http://localhost:30010/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer None" \
  -d '{
    "model": "meta-llama/Llama-3.2-11B-Vision-Instruct",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "What’s in this image?"
          },
          {
            "type": "image_url",
            "image_url": {
              "url": "https://github.com/sgl-project/sglang/blob/main/test/lang/example_image.png?raw=true"
            }
          }
        ]
      }
    ],
    "max_tokens": 300
  }'
"""

response = subprocess.check_output(curl_command, shell=True).decode()
print_highlight(response)

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100   485    0     0  100   485      0    151  0:00:03  0:00:03 --:--:--   151

/actions-runner/_work/_tool/Python/3.9.20/x64/lib/python3.9/site-packages/torch/storage.py:414: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  return torch.load(io.BytesIO(b))
[2024-11-02 03:47:31 TP0] Prefill batch. #new-seq: 1, #new-token: 6463, #cached-token: 0, cache hit rate: 0.00%, token usage: 0.00, #running-req: 0, #queue-req: 0

100   485    0     0  100   485      0    115  0:00:04  0:00:04 --:--:--   115

[2024-11-02 03:47:32] INFO:     127.0.0.1:36616 - "POST /v1/chat/completions HTTP/1.1" 200 OK

100   930  100   445  100   485     98    107  0:00:04  0:00:04 --:--:--   134
{"id":"5bc9c6df24be44d396003a0f892bfe4a","object":"chat.completion","created":1730519252,"model":"meta-llama/Llama-3.2-11B-Vision-Instruct","choices":[{"index":0,"message":{"role":"assistant","content":"The image depicts a man ironing clothes on the back of a yellow taxi cab."},"logprobs":null,"finish_reason":"stop","matched_stop":128009}],"usage":{"prompt_tokens":6463,"total_tokens":6481,"completion_tokens":18,"prompt_tokens_details":null}}

Using OpenAI Python Client#

You can use the OpenAI Python API library to send requests.

[3]:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30010/v1", api_key="None")

response = client.chat.completions.create(
    model="meta-llama/Llama-3.2-11B-Vision-Instruct",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "What is in this image?",
                },
                {
                    "type": "image_url",
                    "image_url": {"url": "https://github.com/sgl-project/sglang/blob/main/test/lang/example_image.png?raw=true"},
                },
            ],
        }
    ],
    max_tokens=300,
)

print_highlight(response.choices[0].message.content)

[2024-11-02 03:47:32 TP0] Prefill batch. #new-seq: 1, #new-token: 11, #cached-token: 6452, cache hit rate: 49.89%, token usage: 0.02, #running-req: 0, #queue-req: 0
[2024-11-02 03:47:32 TP0] Decode batch. #running-req: 1, #token: 6479, token usage: 0.02, gen throughput (token/s): 3.38, #queue-req: 0
[2024-11-02 03:47:33 TP0] Decode batch. #running-req: 1, #token: 6519, token usage: 0.02, gen throughput (token/s): 105.20, #queue-req: 0
[2024-11-02 03:47:33 TP0] Decode batch. #running-req: 1, #token: 6559, token usage: 0.02, gen throughput (token/s): 104.32, #queue-req: 0
[2024-11-02 03:47:34] INFO:     127.0.0.1:59640 - "POST /v1/chat/completions HTTP/1.1" 200 OK
The image depicts a man ironing clothes on the back of a taxi cab as it drives down the street. The man is standing on a chair, which is placed on the back of the taxi, and is ironing a blue shirt. The taxi is yellow with white trim and has a white license plate that reads "TAXICAB" in black letters. The man is wearing a yellow jacket and has a white shirt on the ironing board. In the background, there are several buildings and trees, and another yellow taxi can be seen driving down the street. The image appears to be humorous, as it shows a man doing an unusual task in a public place.

Multiple-Image Inputs#

The server also supports multiple images and interleaved text and images if the model supports it.

[4]:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30010/v1", api_key="None")

response = client.chat.completions.create(
    model="meta-llama/Llama-3.2-11B-Vision-Instruct",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://github.com/sgl-project/sglang/blob/main/test/lang/example_image.png?raw=true",
                    },
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://raw.githubusercontent.com/sgl-project/sglang/main/assets/logo.png",
                    },
                },
                {
                    "type": "text",
                    "text": "I have two very different images. They are not related at all. "
                            "Please describe the first image in one sentence, and then describe the second image in another sentence.",
                },
            ],
        }
    ],
    temperature=0,
)

print_highlight(response.choices[0].message.content)

[2024-11-02 03:47:34 TP0] Prefill batch. #new-seq: 1, #new-token: 12895, #cached-token: 0, cache hit rate: 24.98%, token usage: 0.00, #running-req: 0, #queue-req: 0
[2024-11-02 03:47:35 TP0] Decode batch. #running-req: 1, #token: 12896, token usage: 0.04, gen throughput (token/s): 28.66, #queue-req: 0
[2024-11-02 03:47:35 TP0] Decode batch. #running-req: 1, #token: 12936, token usage: 0.04, gen throughput (token/s): 105.95, #queue-req: 0
[2024-11-02 03:47:35] INFO:     127.0.0.1:59644 - "POST /v1/chat/completions HTTP/1.1" 200 OK
The first image shows a man in a yellow shirt ironing a shirt on the back of a yellow taxi cab, with a small icon of a computer code snippet in the bottom left corner. The second image shows a large orange "S" and "G" and "L" on a white background.
[5]:

terminate_process(embedding_process)

Chat Template#

As mentioned before, if you do not specify a vision model’s --chat-template, the server uses Hugging Face’s default template, which only supports text.

We list popular vision models with their chat templates: