phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-10-04 12:07:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

use templates for generating system prompts

Browse source
This commit is contained in:
Hardik Shah 2024-08-23 14:21:12 -07:00
parent 68855ed218
commit ab8193c88c
8 changed files with 410 additions and 218 deletions
Download patch file Download diff file Expand all files Collapse all files

6
llama_toolchain/inference/meta_reference/inference.py
View file

@ -22,7 +22,7 @@ from llama_toolchain.inference.api import (
ToolCallDelta,
ToolCallParseStatus,
)
from llama_toolchain.inference.prepare_messages import prepare_messages_for_tools
from llama_toolchain.inference.prepare_messages import prepare_messages
from .config import MetaReferenceImplConfig
from .model_parallel import LlamaModelParallelGenerator
@ -67,7 +67,7 @@ class MetaReferenceInferenceImpl(Inference):
) -> AsyncIterator[
Union[ChatCompletionResponseStreamChunk, ChatCompletionResponse]
]:
request = prepare_messages_for_tools(request)
messages = prepare_messages(request)
model = resolve_model(request.model)
if model is None:
raise RuntimeError(
@ -99,7 +99,7 @@ class MetaReferenceInferenceImpl(Inference):
ipython = False
for token_result in self.generator.chat_completion(
messages=request.messages,
messages=messages,
temperature=request.sampling_params.temperature,
top_p=request.sampling_params.top_p,
max_gen_len=request.sampling_params.max_tokens,

8
llama_toolchain/inference/ollama/ollama.py
View file

@ -32,7 +32,7 @@ from llama_toolchain.inference.api import (
ToolCallDelta,
ToolCallParseStatus,
)
from llama_toolchain.inference.prepare_messages import prepare_messages_for_tools
from llama_toolchain.inference.prepare_messages import prepare_messages
from .config import OllamaImplConfig
# TODO: Eventually this will move to the llama cli model list command
@ -111,7 +111,7 @@ class OllamaInference(Inference):
return options
async def chat_completion(self, request: ChatCompletionRequest) -> AsyncGenerator:
request = prepare_messages_for_tools(request)
messages = prepare_messages(request)
# accumulate sampling params and other options to pass to ollama
options = self.get_ollama_chat_options(request)
ollama_model = self.resolve_ollama_model(request.model)
@ -133,7 +133,7 @@ class OllamaInference(Inference):
if not request.stream:
r = await self.client.chat(
model=ollama_model,
messages=self._messages_to_ollama_messages(request.messages),
messages=self._messages_to_ollama_messages(messages),
stream=False,
options=options,
)
@ -161,7 +161,7 @@ class OllamaInference(Inference):
)
stream = await self.client.chat(
model=ollama_model,
messages=self._messages_to_ollama_messages(request.messages),
messages=self._messages_to_ollama_messages(messages),
stream=True,
options=options,
)

215
llama_toolchain/inference/prepare_messages.py
View file

@ -1,203 +1,66 @@
import json
import os
import textwrap
from datetime import datetime
from llama_toolchain.inference.api import * # noqa: F403
from llama_toolchain.tools.builtin import (
BraveSearchTool,
CodeInterpreterTool,
PhotogenTool,
WolframAlphaTool,
from llama_toolchain.common.prompt_templates.system_prompts import (
BuiltinToolGenerator,
FunctionTagCustomToolGenerator,
JsonCustomToolGenerator,
SystemDefaultGenerator,
)
def tool_breakdown(tools: List[ToolDefinition]) -> str:
builtin_tools, custom_tools = [], []
for dfn in tools:
if isinstance(dfn.tool_name, BuiltinTool):
builtin_tools.append(dfn)
else:
custom_tools.append(dfn)
def prepare_messages(request: ChatCompletionRequest) -> List[Message]:
return builtin_tools, custom_tools
def prepare_messages_for_tools(request: ChatCompletionRequest) -> ChatCompletionRequest:
"""This functions takes a ChatCompletionRequest and returns an augmented request.
The request's messages are augmented to update the system message
corresponding to the tool definitions provided in the request.
"""
assert request.tool_choice == ToolChoice.auto, "Only `ToolChoice.auto` supported"
existing_messages = request.messages
existing_system_message = None
if existing_messages[0].role == Role.system.value:
existing_system_message = existing_messages.pop(0)
builtin_tools, custom_tools = tool_breakdown(request.tools)
assert (
existing_messages[0].role != Role.system.value
), "Should only have 1 system message"
messages = []
content = ""
if builtin_tools or custom_tools:
content += "Environment: ipython\n"
if builtin_tools:
tool_str = ", ".join(
[
t.tool_name.value
for t in builtin_tools
if t.tool_name != BuiltinTool.code_interpreter
]
)
if tool_str:
content += f"Tools: {tool_str}\n"
default_gen = SystemDefaultGenerator()
default_template = default_gen.gen()
current_date = datetime.now()
formatted_date = current_date.strftime("%d %B %Y")
date_str = textwrap.dedent(
f"""
Cutting Knowledge Date: December 2023
Today Date: {formatted_date}
"""
)
content += date_str.lstrip("\n")
sys_content = ""
tool_template = None
if request.tools:
tool_gen = BuiltinToolGenerator()
tool_template = tool_gen.gen(request.tools)
sys_content += tool_template.render()
sys_content += "\n"
sys_content += default_template.render()
if existing_system_message:
content += "\n"
content += existing_system_message.content
sys_content += "\n"
sys_content += existing_system_message.content
messages.append(SystemMessage(content=content))
messages.append(SystemMessage(content=sys_content))
if custom_tools:
if request.tool_prompt_format == ToolPromptFormat.function_tag:
text = prompt_for_function_tag(custom_tools)
messages.append(UserMessage(content=text))
elif request.tool_prompt_format == ToolPromptFormat.json:
text = prompt_for_json(custom_tools)
messages.append(UserMessage(content=text))
has_custom_tools = any(isinstance(dfn.tool_name, str) for dfn in request.tools)
if has_custom_tools:
if request.tool_prompt_format == ToolPromptFormat.json:
tool_gen = JsonCustomToolGenerator()
elif request.tool_prompt_format == ToolPromptFormat.function_tag:
tool_gen = FunctionTagCustomToolGenerator()
else:
raise NotImplementedError(
f"Tool prompt format {tool_prompt_format} is not supported"
raise ValueError(
f"Non supported ToolPromptFormat {request.tool_prompt_format}"
)
custom_tools = [t for t in request.tools if isinstance(t.tool_name, str)]
custom_template = tool_gen.gen(custom_tools)
messages.append(UserMessage(content=custom_template.render()))
# Add back existing messages from the request
messages += existing_messages
request.messages = messages
return request
def prompt_for_json(custom_tools: List[ToolDefinition]) -> str:
tool_defs = "\n".join(
translate_custom_tool_definition_to_json(t) for t in custom_tools
)
content = textwrap.dedent(
"""
Answer the user's question by making use of the following functions if needed.
If none of the function can be used, please say so.
Here is a list of functions in JSON format:
{tool_defs}
Return function calls in JSON format.
"""
)
content = content.lstrip("\n").format(tool_defs=tool_defs)
return content
def prompt_for_function_tag(custom_tools: List[ToolDefinition]) -> str:
custom_tool_params = ""
for t in custom_tools:
custom_tool_params += get_instruction_string(t) + "\n"
custom_tool_params += get_parameters_string(t) + "\n\n"
content = textwrap.dedent(
"""
You have access to the following functions:
{custom_tool_params}
Think very carefully before calling functions.
If you choose to call a function ONLY reply in the following format with no prefix or suffix:
<function=example_function_name>{{"example_name": "example_value"}}</function>
Reminder:
- If looking for real time information use relevant functions before falling back to brave_search
- Function calls MUST follow the specified format, start with <function= and end with </function>
- Required parameters MUST be specified
- Only call one function at a time
- Put the entire function call reply on one line
"""
)
return content.lstrip("\n").format(custom_tool_params=custom_tool_params)
def get_instruction_string(custom_tool_definition) -> str:
return f"Use the function '{custom_tool_definition.tool_name}' to '{custom_tool_definition.description}'"
def get_parameters_string(custom_tool_definition) -> str:
return json.dumps(
{
"name": custom_tool_definition.tool_name,
"description": custom_tool_definition.description,
"parameters": {
name: definition.__dict__
for name, definition in custom_tool_definition.parameters.items()
},
}
)
def translate_custom_tool_definition_to_json(tool_def):
"""Translates ToolDefinition to json as expected by model
eg. output for a function
{
"type": "function",
"function": {
"name": "conv_int",
"description": "Convert serialized fract24 integer into int value.",
"parameters": {
"type": "object",
"properties": [
{
"data": {
"type": "object",
"description": ""
}
}
],
"required": ["data"]
}
}
}
"""
assert isinstance(tool_def.tool_name, str)
func_def = {"type": "function", "function": {}}
func_def["function"]["name"] = tool_def.tool_name
func_def["function"]["description"] = tool_def.description or ""
if tool_def.parameters:
required = []
properties = []
for p_name, p_def in tool_def.parameters.items():
properties.append(
{
p_name: {
# TODO: see if this should not always be object
"type": "object",
"description": p_def.description or "",
}
}
)
if p_def.required:
required.append(p_name)
func_def["function"]["parameters"] = {
"type": "object",
"properties": properties,
"required": required,
}
else:
func_def["function"]["parameters"] = {}
return json.dumps(func_def, indent=4)
return messages