amlllm-openai-server 0.1.1

Open AI API compatiable LLM server for Amlogic Device

AML LLM OpenAI Server

An OpenAI-compatible API server for Amlogic LLM devices, with built-in tool-calling evaluation and Agent capability benchmarking tools.

• Server: OpenAI-compatible /v1/chat/completions endpoint (ADLA & llama.cpp backends)
• Tools Evaluation: Measure tool-calling accuracy of local vs cloud models
• Agent Benchmark: Evaluate LLM agent capabilities on MMLU-Pro, TAU2-Bench, and BFCL

---

Prerequisites

This project is using uv please use following command or follow the official document to install uv

curl -LsSf https://astral.sh/uv/install.sh | sh

Install the wheel

uv venv --python 3.11
uv pip install amlllm-openai-server -i <index url>

Bootstrap

Use a bootstrap tool

You need Internet connection to use bootstrap scripts to download the models or you can manually put files in folder

1. List the available model

uv run -m amlllm_openai_server.bootstrap list

You will see the available models as following table

#	Name	Type	Thinking	Multimodal	Context	Tools
0	Qwen3-1.7B	Qwen3	Yes	No	8192	Yes
1	Qwen3-4B-Instruct-2507	Qwen3	No	No	8192	Yes
2	Qwen3-4B	Qwen3	Yes	No	4096	Yes

2. Download the model

Replace the model name as your expectation. Here use Qwen3-4B-Instruct-2507 for example.

uv run -m amlllm_openai_server.bootstrap download Qwen3-4B-Instruct-2507

---

Server Configuration

config/server.yaml

server:
  host: "0.0.0.0"
  port: 8000
  api_key: ""          # optional, set for authentication
  log_level: "info"
  title: "AMLLLM OneAPI Proxy"
  version: "0.2.0"

models:
  root_dir: "../models"
  enabled:
    - default           # model directories under root_dir

# Optional: upstream proxy to a cloud API
# upstream:
#   base_url: "https://api.openai.com"
#   api_key: "sk-..."
#   model: "gpt-4o"
#   force_upstream: false
#   skill_injection: false

models/default/model.json

Example for Qwen3 Seriers:

{
  "id": "amlogic-default",
  "weights": "Qwen3-1.7B-Q4AM_PB32.adla",
  "model_type": "qwen",
  "sampling_mode": "top_p",
  "top_k": 40,
  "top_p": 0.85,
  "temperature": 0.7,
  "repeat_penalty": 1.15,
  "presence_penalty": 10,
  "system_prompt": "You are a helpful assistant.",
  "prompt_prefix": "",
  "prompt_postfix": "",
  "retain_history": true,
  "loglevel": "ERROR",
  "metadata": {
    "family": "default",
    "device": "amlogic"
  },
  "token_pair": { "151657": 151658 },
  "chat_format": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n\\n' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if message.content is string %}\n        {%- set content = message.content %}\n    {%- else %}\n        {%- set content = '' %}\n    {%- endif %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n        {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is string %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in content %}\n                {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n                {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {%- if loop.last or (not loop.last and reasoning_content) %}\n                {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<|im_start|>' + message.role + '\\n' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n    {%- if enable_thinking is defined and enable_thinking is false %}\n        {{- '<think>\\n\\n</think>\\n\\n' }}\n    {%- endif %}\n{%- endif %}"
}

Model Configuration Parameters

Parameter	Type	Required	Default	Description
id	string	Yes	—	Unique model identifier, used as model in OpenAI API requests
weights	string	Yes	—	Path to the model weights file (.adla for ADLA, .gguf for llama.cpp)
model_type	string	Yes	—	Model architecture type: "qwen", "llama", "gemma", etc. Determines chat template and tokenization
sampling_mode	string	No	"top_p"	Sampling strategy: "top_p", "top_k", or "greedy"
top_k	int	No	40	Top-K sampling: keep only the K most likely tokens
top_p	float	No	0.85	Nucleus sampling: keep tokens with cumulative probability ≥ p
temperature	float	No	0.7	Controls randomness; higher = more creative, lower = more deterministic
repeat_penalty	float	No	1.15	Penalizes repeated tokens; > 1.0 discourages repetition
presence_penalty	float	No	0	ADLA-specific presence penalty for token repetition
system_prompt	string	No	""	Default system prompt prepended to every conversation
prompt_prefix	string	No	""	Custom text prepended before the prompt
prompt_postfix	string	No	""	Custom text appended after the prompt
retain_history	bool	No	true	Whether to keep conversation history across turns
loglevel	string	No	"ERROR"	ADLA runtime log level: "DEBUG", "INFO", "WARNING", "ERROR"
metadata	dict	No	{}	Arbitrary key-value metadata exposed in /v1/models response
token_pair	dict	No	{}	Special token ID mapping, e.g. {"151657": 151658} maps <tool_call> → </tool_call>, the token MUST appear in pair, as 151657 ->xxx -> 151658, otherwise, the generation will be stopped.
chat_format	string	No	""	Jinja2 chat template for formatting messages (see Jinja2 Templates) or Path to tokenizer_config.json
chat_template_kwargs	dict	No	{}	Extra parameters passed to chat templates

Jinja2 Chat Templates

The chat_format field uses Jinja2 syntax or tokenizer config to define how messages are formatted into the model's native prompt structure.

1. Prepare the model tokenizer_config.json
2. For Qwen series, it is possible to put the Jinja template string directly
3. For other series, put the path to tokenizer_config.json

Backend values:

• "adla" — Amlogic ADLA runtime (AARCH64 only)

---

Running the Server

# Default: reads config/server.yaml
uv run python -m amlllm_openai_server

# Custom config path
uv run python -m amlllm_openai_server --config /path/to/server.yaml

# Override host/port/log-level
uv run python -m amlllm_openai_server --host 0.0.0.0 --port 8000 --log-level debug

Test the endpoint:

curl http://localhost:8000/v1/models
curl http://localhost:8000/healthz

If you want to install PicoClaw

dpkg -i picoclaw_modified_aarch64.deb

Run demo with PicoClaw

1. Start the AMLLLM server (see above)
2. Start PicoClaw:

   picoclaw-launcher &
   

3. Configure PicoClaw model:
   • Provider: openai
   • Model identifier: amlogic-default
   • API Base URL: http://localhost:8000/v1

Tips: The agent work folder is under $HOME/.picoclaw/workspace. If you are using local models with context limitations, it's better to remove AGENTS.md and SOUL.md.