-
Notifications
You must be signed in to change notification settings - Fork 26.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Transformers Agents #23214
Transformers Agents #23214
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hope the review helps a bit, mostly nits / suggestions
For demonstration purposes and so that this can be used with all setups, we have created remote executors for several | ||
of the default tools the agent has access to. These are created using | ||
[inference endpoints](https://huggingface.co/inference-endpoints). To see how to setup remote executors tools yourself, | ||
we recommend reading the custom tool guide [TODO LINK]. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Here is still a TODO - not 100% sure what should go there
…s into test_composition
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
…s into test_composition
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥
Super exciting PR!
🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥
I ignored the TODOs dotted around as I know it's still being iterated on. Just left a few nits here and there but overall a super impressive piece of work and one I can't way to start playing with!
TOOL_CONFIG_FILE = "tool_config.json" | ||
|
||
|
||
def get_repo_type(repo_id, repo_type=None, **hub_kwargs): |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Bit of a sneaky name as it's downloading in the background :D
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah but get_repo_type_by_downloading_tool_config
is a bit long ;-p
raise ValueError("This tool does not implement a default checkpoint, you need to pass one.") | ||
model = self.default_checkpoint | ||
if pre_processor is None: | ||
pre_processor = model |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Does this mean we do a double forward pass if there's no preprocessor?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Unless the user passed an instantiated model (which we could check in further development). Normally at this stage model
is a checkpoint name.
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
* Remove nestedness in tool config * Really do it * Use remote tools descriptions * Work * Clean up eval * Changes * Tools * Tools * tool * Fix everything * Use last result/assign for evaluation * Prompt * Remove hardcoded selection * Evaluation for chat agents * correct some spelling * Small fixes * Change summarization model (#23172) * Fix link displayed * Update description of the tool * Fixes in chat prompt * Custom tools, custom prompt * Tool clean up * save_pretrained and push_to_hub for tool * Fix init * Tests * Fix tests * Tool save/from_hub/push_to_hub and tool->load_tool * Clean push_to_hub and add app file * Custom inference API for endpoints too * Clean up * old remote tool and new remote tool * Make a requirements * return_code adds tool creation * Avoid redundancy between global variables * Remote tools can be loaded * Tests * Text summarization tests * Quality * Properly mark tests * Test the python interpreter * And the CI shall be green. * fix loading of additional tools * Work on RemoteTool and fix tests * General clean up * Guard imports * Fix tools * docs: Fix broken link in 'How to add a model...' (#23216) fix link * Get default endpoint from the Hub * Add guide * Simplify tool config * Docs * Some fixes * Docs * Docs * Docs * Fix code returned by agent * Try this * Match args with signature in remote tool * Should fix python interpreter for Python 3.8 * Fix push_to_hub for tools * Other fixes to push_to_hub * Add API doc page * Docs * Docs * Custom tools * Pin tensorflow-probability (#23220) * Pin tensorflow-probability * [all-test] * [all-test] Fix syntax for bash * PoC for some chaining API * Text to speech * J'ai pris des libertés * Rename * Basic python interpreter * Add agents * Quality * Add translation tool * temp * GenQA + LID + S2T * Quality + word missing in translation * Add open assistance, support f-strings in evaluate * captioning + s2t fixes * Style * Refactor descriptions and remove chain * Support errors and rename OpenAssistantAgent * Add setup * Deal with typos + example of inference API * Some rename + README * Fixes * Update prompt * Unwanted change * Make sure everyone has a default * One prompt to rule them all. * SD * Description * Clean up remote tools * More remote tools * Add option to return code and update doc * Image segmentation * ControlNet * Gradio demo * Diffusers protection * Lib protection * ControlNet description * Cleanup * Style * Remove accelerate and try to be reproducible * No randomness * Male Basic optional in token * Clean description * Better prompts * Fix args eval in interpreter * Add tool wrapper * Tool on the Hub * Style post-rebase * Big refactor of descriptions, batch generation and evaluation for agents * Make problems easier - interface to debug * More problems, add python primitives * Back to one prompt * Remove dict for translation * Be consistent * Add prompts * New version of the agent * Evaluate new agents * New endpoints agents * Make all tools a dict variable * Typo * Add problems * Add to big prompt * Harmonize * Add tools * New evaluation * Add more tools * Build prompt with tools descriptions * Tools on the Hub * Let's chat! * Cleanup * Temporary bs4 safeguard * Cache agents and clean up * Blank init * Fix evaluation for agents * New format for tools on the Hub * Add method to reset state * Remove nestedness in tool config * Really do it * Use remote tools descriptions * Work * Clean up eval * Changes * Tools * Tools * tool * Fix everything * Use last result/assign for evaluation * Prompt * Remove hardcoded selection * Evaluation for chat agents * correct some spelling * Small fixes * Change summarization model (#23172) * Fix link displayed * Update description of the tool * Fixes in chat prompt * Custom tools, custom prompt * Tool clean up * save_pretrained and push_to_hub for tool * Fix init * Tests * Fix tests * Tool save/from_hub/push_to_hub and tool->load_tool * Clean push_to_hub and add app file * Custom inference API for endpoints too * Clean up * old remote tool and new remote tool * Make a requirements * return_code adds tool creation * Avoid redundancy between global variables * Remote tools can be loaded * Tests * Text summarization tests * Quality * Properly mark tests * Test the python interpreter * And the CI shall be green. * Work on RemoteTool and fix tests * fix loading of additional tools * General clean up * Guard imports * Fix tools * Get default endpoint from the Hub * Simplify tool config * Add guide * Docs * Some fixes * Docs * Docs * Fix code returned by agent * Try this * Docs * Match args with signature in remote tool * Should fix python interpreter for Python 3.8 * Fix push_to_hub for tools * Other fixes to push_to_hub * Add API doc page * Fixes * Doc fixes * Docs * Fix audio * Custom tools * Audio fix * Improve custom tools docstring * Docstrings * Trigger CI * Mode docstrings * More docstrings * Improve custom tools * Fix for remote tools * Style * Fix repo consistency * Quality * Tip * Cleanup on doc * Cleanup toc * Add disclaimer for starcoder vs openai * Remove disclaimer * Small fixed in the prompts * 4.29 * Update src/transformers/tools/agents.py Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr> * Complete documentation * Small fixes * Agent evaluation * Note about gradio-tools & LC * Clean up agents and prompt * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Note about gradio-tools & LC * Add copyrights and address review comments * Quality * Add all language codes * Add remote tool tests * Move custom prompts to other docs * Apply suggestions from code review Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * TTS tests * Quality --------- Co-authored-by: Lysandre <hi@lyand.re> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Philipp Schmid <32632186+philschmid@users.noreply.github.com> Co-authored-by: Connor Henderson <connor.henderson@talkiatry.com> Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr> Co-authored-by: Lysandre <lysandre@huggingface.co> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
picture = agent.run("Draw me a picture of rivers and lakes") ==Code generated by the agent== ==Result== |
Ah yes we did that example with openAI. Will fine-tune the prompt so that example works before the release, thanks for the pointer! |
* Remove nestedness in tool config * Really do it * Use remote tools descriptions * Work * Clean up eval * Changes * Tools * Tools * tool * Fix everything * Use last result/assign for evaluation * Prompt * Remove hardcoded selection * Evaluation for chat agents * correct some spelling * Small fixes * Change summarization model (huggingface#23172) * Fix link displayed * Update description of the tool * Fixes in chat prompt * Custom tools, custom prompt * Tool clean up * save_pretrained and push_to_hub for tool * Fix init * Tests * Fix tests * Tool save/from_hub/push_to_hub and tool->load_tool * Clean push_to_hub and add app file * Custom inference API for endpoints too * Clean up * old remote tool and new remote tool * Make a requirements * return_code adds tool creation * Avoid redundancy between global variables * Remote tools can be loaded * Tests * Text summarization tests * Quality * Properly mark tests * Test the python interpreter * And the CI shall be green. * fix loading of additional tools * Work on RemoteTool and fix tests * General clean up * Guard imports * Fix tools * docs: Fix broken link in 'How to add a model...' (huggingface#23216) fix link * Get default endpoint from the Hub * Add guide * Simplify tool config * Docs * Some fixes * Docs * Docs * Docs * Fix code returned by agent * Try this * Match args with signature in remote tool * Should fix python interpreter for Python 3.8 * Fix push_to_hub for tools * Other fixes to push_to_hub * Add API doc page * Docs * Docs * Custom tools * Pin tensorflow-probability (huggingface#23220) * Pin tensorflow-probability * [all-test] * [all-test] Fix syntax for bash * PoC for some chaining API * Text to speech * J'ai pris des libertés * Rename * Basic python interpreter * Add agents * Quality * Add translation tool * temp * GenQA + LID + S2T * Quality + word missing in translation * Add open assistance, support f-strings in evaluate * captioning + s2t fixes * Style * Refactor descriptions and remove chain * Support errors and rename OpenAssistantAgent * Add setup * Deal with typos + example of inference API * Some rename + README * Fixes * Update prompt * Unwanted change * Make sure everyone has a default * One prompt to rule them all. * SD * Description * Clean up remote tools * More remote tools * Add option to return code and update doc * Image segmentation * ControlNet * Gradio demo * Diffusers protection * Lib protection * ControlNet description * Cleanup * Style * Remove accelerate and try to be reproducible * No randomness * Male Basic optional in token * Clean description * Better prompts * Fix args eval in interpreter * Add tool wrapper * Tool on the Hub * Style post-rebase * Big refactor of descriptions, batch generation and evaluation for agents * Make problems easier - interface to debug * More problems, add python primitives * Back to one prompt * Remove dict for translation * Be consistent * Add prompts * New version of the agent * Evaluate new agents * New endpoints agents * Make all tools a dict variable * Typo * Add problems * Add to big prompt * Harmonize * Add tools * New evaluation * Add more tools * Build prompt with tools descriptions * Tools on the Hub * Let's chat! * Cleanup * Temporary bs4 safeguard * Cache agents and clean up * Blank init * Fix evaluation for agents * New format for tools on the Hub * Add method to reset state * Remove nestedness in tool config * Really do it * Use remote tools descriptions * Work * Clean up eval * Changes * Tools * Tools * tool * Fix everything * Use last result/assign for evaluation * Prompt * Remove hardcoded selection * Evaluation for chat agents * correct some spelling * Small fixes * Change summarization model (huggingface#23172) * Fix link displayed * Update description of the tool * Fixes in chat prompt * Custom tools, custom prompt * Tool clean up * save_pretrained and push_to_hub for tool * Fix init * Tests * Fix tests * Tool save/from_hub/push_to_hub and tool->load_tool * Clean push_to_hub and add app file * Custom inference API for endpoints too * Clean up * old remote tool and new remote tool * Make a requirements * return_code adds tool creation * Avoid redundancy between global variables * Remote tools can be loaded * Tests * Text summarization tests * Quality * Properly mark tests * Test the python interpreter * And the CI shall be green. * Work on RemoteTool and fix tests * fix loading of additional tools * General clean up * Guard imports * Fix tools * Get default endpoint from the Hub * Simplify tool config * Add guide * Docs * Some fixes * Docs * Docs * Fix code returned by agent * Try this * Docs * Match args with signature in remote tool * Should fix python interpreter for Python 3.8 * Fix push_to_hub for tools * Other fixes to push_to_hub * Add API doc page * Fixes * Doc fixes * Docs * Fix audio * Custom tools * Audio fix * Improve custom tools docstring * Docstrings * Trigger CI * Mode docstrings * More docstrings * Improve custom tools * Fix for remote tools * Style * Fix repo consistency * Quality * Tip * Cleanup on doc * Cleanup toc * Add disclaimer for starcoder vs openai * Remove disclaimer * Small fixed in the prompts * 4.29 * Update src/transformers/tools/agents.py Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr> * Complete documentation * Small fixes * Agent evaluation * Note about gradio-tools & LC * Clean up agents and prompt * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Note about gradio-tools & LC * Add copyrights and address review comments * Quality * Add all language codes * Add remote tool tests * Move custom prompts to other docs * Apply suggestions from code review Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * TTS tests * Quality --------- Co-authored-by: Lysandre <hi@lyand.re> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Philipp Schmid <32632186+philschmid@users.noreply.github.com> Co-authored-by: Connor Henderson <connor.henderson@talkiatry.com> Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr> Co-authored-by: Lysandre <lysandre@huggingface.co> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
* Remove nestedness in tool config * Really do it * Use remote tools descriptions * Work * Clean up eval * Changes * Tools * Tools * tool * Fix everything * Use last result/assign for evaluation * Prompt * Remove hardcoded selection * Evaluation for chat agents * correct some spelling * Small fixes * Change summarization model (huggingface#23172) * Fix link displayed * Update description of the tool * Fixes in chat prompt * Custom tools, custom prompt * Tool clean up * save_pretrained and push_to_hub for tool * Fix init * Tests * Fix tests * Tool save/from_hub/push_to_hub and tool->load_tool * Clean push_to_hub and add app file * Custom inference API for endpoints too * Clean up * old remote tool and new remote tool * Make a requirements * return_code adds tool creation * Avoid redundancy between global variables * Remote tools can be loaded * Tests * Text summarization tests * Quality * Properly mark tests * Test the python interpreter * And the CI shall be green. * fix loading of additional tools * Work on RemoteTool and fix tests * General clean up * Guard imports * Fix tools * docs: Fix broken link in 'How to add a model...' (huggingface#23216) fix link * Get default endpoint from the Hub * Add guide * Simplify tool config * Docs * Some fixes * Docs * Docs * Docs * Fix code returned by agent * Try this * Match args with signature in remote tool * Should fix python interpreter for Python 3.8 * Fix push_to_hub for tools * Other fixes to push_to_hub * Add API doc page * Docs * Docs * Custom tools * Pin tensorflow-probability (huggingface#23220) * Pin tensorflow-probability * [all-test] * [all-test] Fix syntax for bash * PoC for some chaining API * Text to speech * J'ai pris des libertés * Rename * Basic python interpreter * Add agents * Quality * Add translation tool * temp * GenQA + LID + S2T * Quality + word missing in translation * Add open assistance, support f-strings in evaluate * captioning + s2t fixes * Style * Refactor descriptions and remove chain * Support errors and rename OpenAssistantAgent * Add setup * Deal with typos + example of inference API * Some rename + README * Fixes * Update prompt * Unwanted change * Make sure everyone has a default * One prompt to rule them all. * SD * Description * Clean up remote tools * More remote tools * Add option to return code and update doc * Image segmentation * ControlNet * Gradio demo * Diffusers protection * Lib protection * ControlNet description * Cleanup * Style * Remove accelerate and try to be reproducible * No randomness * Male Basic optional in token * Clean description * Better prompts * Fix args eval in interpreter * Add tool wrapper * Tool on the Hub * Style post-rebase * Big refactor of descriptions, batch generation and evaluation for agents * Make problems easier - interface to debug * More problems, add python primitives * Back to one prompt * Remove dict for translation * Be consistent * Add prompts * New version of the agent * Evaluate new agents * New endpoints agents * Make all tools a dict variable * Typo * Add problems * Add to big prompt * Harmonize * Add tools * New evaluation * Add more tools * Build prompt with tools descriptions * Tools on the Hub * Let's chat! * Cleanup * Temporary bs4 safeguard * Cache agents and clean up * Blank init * Fix evaluation for agents * New format for tools on the Hub * Add method to reset state * Remove nestedness in tool config * Really do it * Use remote tools descriptions * Work * Clean up eval * Changes * Tools * Tools * tool * Fix everything * Use last result/assign for evaluation * Prompt * Remove hardcoded selection * Evaluation for chat agents * correct some spelling * Small fixes * Change summarization model (huggingface#23172) * Fix link displayed * Update description of the tool * Fixes in chat prompt * Custom tools, custom prompt * Tool clean up * save_pretrained and push_to_hub for tool * Fix init * Tests * Fix tests * Tool save/from_hub/push_to_hub and tool->load_tool * Clean push_to_hub and add app file * Custom inference API for endpoints too * Clean up * old remote tool and new remote tool * Make a requirements * return_code adds tool creation * Avoid redundancy between global variables * Remote tools can be loaded * Tests * Text summarization tests * Quality * Properly mark tests * Test the python interpreter * And the CI shall be green. * Work on RemoteTool and fix tests * fix loading of additional tools * General clean up * Guard imports * Fix tools * Get default endpoint from the Hub * Simplify tool config * Add guide * Docs * Some fixes * Docs * Docs * Fix code returned by agent * Try this * Docs * Match args with signature in remote tool * Should fix python interpreter for Python 3.8 * Fix push_to_hub for tools * Other fixes to push_to_hub * Add API doc page * Fixes * Doc fixes * Docs * Fix audio * Custom tools * Audio fix * Improve custom tools docstring * Docstrings * Trigger CI * Mode docstrings * More docstrings * Improve custom tools * Fix for remote tools * Style * Fix repo consistency * Quality * Tip * Cleanup on doc * Cleanup toc * Add disclaimer for starcoder vs openai * Remove disclaimer * Small fixed in the prompts * 4.29 * Update src/transformers/tools/agents.py Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr> * Complete documentation * Small fixes * Agent evaluation * Note about gradio-tools & LC * Clean up agents and prompt * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Note about gradio-tools & LC * Add copyrights and address review comments * Quality * Add all language codes * Add remote tool tests * Move custom prompts to other docs * Apply suggestions from code review Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * TTS tests * Quality --------- Co-authored-by: Lysandre <hi@lyand.re> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Philipp Schmid <32632186+philschmid@users.noreply.github.com> Co-authored-by: Connor Henderson <connor.henderson@talkiatry.com> Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr> Co-authored-by: Lysandre <lysandre@huggingface.co> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
Introducing Transformers Agents
This PR adds a new API called Transformers Agents. Agents allow you to use Transformers with zero code experience, directly talking to Transformers or Diffusers via natural language. It is based on
Agent
s andTool
s. The agent is an LLM prompted to generate code using the tools, which are simple functions performing a single task.Tools can live in Transformers or on the Hub, this PR introduces both. You can read more about this in the added documentation but here is an example:
Define an agent using the starcoder model:
Use the command
run
to execute a given problem:Use the command
chat
to chat with the agent and execute instructions one after the other: