# Remote vLLM inference config for Llama 3.2 11B Vision Instruct.
#
# Requirements:
#   - Run `pip install oumi[gpu]`
#   - Log into HF: `hf auth login`
#   - Request access to Llama 3.2: https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct
#
# Usage:
#   oumi infer -i -c configs/recipes/vision/llama3_2_vision/inference/11b_rvllm_infer.yaml \
#     --image "tests/testdata/images/the_great_wave_off_kanagawa.jpg"
#
# Sample python command to start vLLM server:
#
# Note, for models that vLLM supports LoRA-adapters adapt below line
# export USE_LORA='--enable-lora --lora-modules your_lora_adapter=path/to/your/lora/adapter'

# python -u -m vllm.entrypoints.openai.api_server \
#   --port 6864 \
#   --model meta-llama/Llama-3.2-11B-Vision-Instruct  \
#   --trust-remote-code \
#   --dtype=bfloat16 \
#   --device=cuda \
#   --max-model-len 256 \
#   --enforce-eager \
#   --disable-custom-all-reduce \
#   --disable-log-requests \
#   --max_num_seqs=2 \
#   --enable-chunked-prefill=false \
#   --gpu-memory-utilization=0.95 \
#   --tensor-parallel-size 1 \
#   ${USE_LORA}
#
# See Also:
#   - Documentation: https://oumi.ai/docs/en/latest/user_guides/infer/infer.html
#   - Config class: oumi.core.configs.InferenceConfig
#   - Config source: https://github.com/oumi-ai/oumi/blob/main/src/oumi/core/configs/inference_config.py
#   - Other inference configs: configs/**/inference/

model:
  model_name: "meta-llama/Llama-3.2-11B-Vision-Instruct"
  torch_dtype_str: "bfloat16"
  model_max_length: 1024
  chat_template: "llama3-instruct"
  trust_remote_code: True

generation:
  max_new_tokens: 32

remote_params:
  # This address is just an example (it's what you may get when you start vllm sever locally).
  # For details, see https://platform.openai.com/docs/api-reference/chat/create
  api_url: "http://localhost:6864/v1/chat/completions"

engine: REMOTE_VLLM