forked from phoenix-oss/llama-stack-mirror
feat: support ClientTool output metadata (#1426)
# Summary: Client side change in https://github.com/meta-llama/llama-stack-client-python/pull/180 Changes the resume_turn API to accept `ToolResponse` instead of `ToolResponseMessage`: 1. `ToolResponse` contains `metadata` 2. `ToolResponseMessage` is a concept for model inputs. Here we are just submitting the outputs of tool execution. # Test Plan: Ran integration tests with newly added test using client tool with metadata LLAMA_STACK_CONFIG=fireworks pytest -s -v tests/integration/agents/test_agents.py --safety-shield meta-llama/Llama-Guard-3-8B --record-responses
This commit is contained in:
ehhuang
GitHub
parent
ac717f38dc
commit
6cf79437b3
10 changed files with 3984 additions and 2172 deletions
20
docs/_static/llama-stack-spec.html
vendored
20
docs/_static/llama-stack-spec.html
vendored
|
|
@ -9321,11 +9321,21 @@
|
|||
"type": "object",
|
||||
"properties": {
|
||||
"tool_responses": {
|
||||
"type": "array",
|
||||
"items": {
|
||||
"$ref": "#/components/schemas/ToolResponseMessage"
|
||||
},
|
||||
"description": "The tool call responses to resume the turn with."
|
||||
"oneOf": [
|
||||
{
|
||||
"type": "array",
|
||||
"items": {
|
||||
"$ref": "#/components/schemas/ToolResponse"
|
||||
}
|
||||
},
|
||||
{
|
||||
"type": "array",
|
||||
"items": {
|
||||
"$ref": "#/components/schemas/ToolResponseMessage"
|
||||
}
|
||||
}
|
||||
],
|
||||
"description": "The tool call responses to resume the turn with. NOTE: ToolResponseMessage will be deprecated. Use ToolResponse."
|
||||
},
|
||||
"stream": {
|
||||
"type": "boolean",
|
||||
|
|
|
|||
13
docs/_static/llama-stack-spec.yaml
vendored
13
docs/_static/llama-stack-spec.yaml
vendored
|
|
@ -6287,11 +6287,16 @@ components:
|
|||
type: object
|
||||
properties:
|
||||
tool_responses:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/ToolResponseMessage'
|
||||
oneOf:
|
||||
- type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/ToolResponse'
|
||||
- type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/ToolResponseMessage'
|
||||
description: >-
|
||||
The tool call responses to resume the turn with.
|
||||
The tool call responses to resume the turn with. NOTE: ToolResponseMessage
|
||||
will be deprecated. Use ToolResponse.
|
||||
stream:
|
||||
type: boolean
|
||||
description: Whether to stream the response.
|
||||
|
|
|
|||
|
|
@ -353,7 +353,7 @@ class AgentTurnResumeRequest(BaseModel):
|
|||
agent_id: str
|
||||
session_id: str
|
||||
turn_id: str
|
||||
tool_responses: List[ToolResponseMessage]
|
||||
tool_responses: Union[List[ToolResponse], List[ToolResponseMessage]]
|
||||
stream: Optional[bool] = False
|
||||
|
||||
|
||||
|
|
@ -432,7 +432,7 @@ class Agents(Protocol):
|
|||
agent_id: str,
|
||||
session_id: str,
|
||||
turn_id: str,
|
||||
tool_responses: List[ToolResponseMessage],
|
||||
tool_responses: Union[List[ToolResponse], List[ToolResponseMessage]],
|
||||
stream: Optional[bool] = False,
|
||||
) -> Union[Turn, AsyncIterator[AgentTurnResponseStreamChunk]]:
|
||||
"""Resume an agent turn with executed tool call responses.
|
||||
|
|
@ -443,6 +443,7 @@ class Agents(Protocol):
|
|||
:param session_id: The ID of the session to resume.
|
||||
:param turn_id: The ID of the turn to resume.
|
||||
:param tool_responses: The tool call responses to resume the turn with.
|
||||
NOTE: ToolResponseMessage will be deprecated. Use ToolResponse.
|
||||
:param stream: Whether to stream the response.
|
||||
:returns: A Turn object if stream is False, otherwise an AsyncIterator of AgentTurnResponseStreamChunk objects.
|
||||
"""
|
||||
|
|
|
|||
|
|
@ -216,13 +216,25 @@ class ChatAgent(ShieldRunnerMixin):
|
|||
steps = []
|
||||
messages = await self.get_messages_from_turns(turns)
|
||||
if is_resume:
|
||||
messages.extend(request.tool_responses)
|
||||
if isinstance(request.tool_responses[0], ToolResponseMessage):
|
||||
tool_response_messages = request.tool_responses
|
||||
tool_responses = [
|
||||
ToolResponse(call_id=x.call_id, tool_name=x.tool_name, content=x.content)
|
||||
for x in request.tool_responses
|
||||
]
|
||||
else:
|
||||
tool_response_messages = [
|
||||
ToolResponseMessage(call_id=x.call_id, tool_name=x.tool_name, content=x.content)
|
||||
for x in request.tool_responses
|
||||
]
|
||||
tool_responses = request.tool_responses
|
||||
messages.extend(tool_response_messages)
|
||||
last_turn = turns[-1]
|
||||
last_turn_messages = self.turn_to_messages(last_turn)
|
||||
last_turn_messages = [
|
||||
x for x in last_turn_messages if isinstance(x, UserMessage) or isinstance(x, ToolResponseMessage)
|
||||
]
|
||||
last_turn_messages.extend(request.tool_responses)
|
||||
last_turn_messages.extend(tool_response_messages)
|
||||
|
||||
# get steps from the turn
|
||||
steps = last_turn.steps
|
||||
|
|
@ -238,14 +250,7 @@ class ChatAgent(ShieldRunnerMixin):
|
|||
step_id=(in_progress_tool_call_step.step_id if in_progress_tool_call_step else str(uuid.uuid4())),
|
||||
turn_id=request.turn_id,
|
||||
tool_calls=(in_progress_tool_call_step.tool_calls if in_progress_tool_call_step else []),
|
||||
tool_responses=[
|
||||
ToolResponse(
|
||||
call_id=x.call_id,
|
||||
tool_name=x.tool_name,
|
||||
content=x.content,
|
||||
)
|
||||
for x in request.tool_responses
|
||||
],
|
||||
tool_responses=tool_responses,
|
||||
completed_at=now,
|
||||
started_at=(in_progress_tool_call_step.started_at if in_progress_tool_call_step else now),
|
||||
)
|
||||
|
|
|
|||
|
|
@ -27,6 +27,7 @@ from llama_stack.apis.agents import (
|
|||
from llama_stack.apis.inference import (
|
||||
Inference,
|
||||
ToolConfig,
|
||||
ToolResponse,
|
||||
ToolResponseMessage,
|
||||
UserMessage,
|
||||
)
|
||||
|
|
@ -168,7 +169,7 @@ class MetaReferenceAgentsImpl(Agents):
|
|||
agent_id: str,
|
||||
session_id: str,
|
||||
turn_id: str,
|
||||
tool_responses: List[ToolResponseMessage],
|
||||
tool_responses: Union[List[ToolResponse], List[ToolResponseMessage]],
|
||||
stream: Optional[bool] = False,
|
||||
) -> AsyncGenerator:
|
||||
request = AgentTurnResumeRequest(
|
||||
|
|
|
|||
|
|
@ -4,6 +4,7 @@
|
|||
# This source code is licensed under the terms described in the LICENSE file in
|
||||
# the root directory of this source tree.
|
||||
|
||||
from typing import Any, Dict
|
||||
from uuid import uuid4
|
||||
|
||||
import pytest
|
||||
|
|
@ -40,6 +41,25 @@ def get_boiling_point(liquid_name: str, celcius: bool = True) -> int:
|
|||
return -1
|
||||
|
||||
|
||||
@client_tool
|
||||
def get_boiling_point_with_metadata(liquid_name: str, celcius: bool = True) -> Dict[str, Any]:
|
||||
"""
|
||||
Returns the boiling point of a liquid in Celcius or Fahrenheit
|
||||
|
||||
:param liquid_name: The name of the liquid
|
||||
:param celcius: Whether to return the boiling point in Celcius
|
||||
:return: The boiling point of the liquid in Celcius or Fahrenheit
|
||||
"""
|
||||
if liquid_name.lower() == "polyjuice":
|
||||
if celcius:
|
||||
temp = -100
|
||||
else:
|
||||
temp = -212
|
||||
else:
|
||||
temp = -1
|
||||
return {"content": temp, "metadata": {"source": "https://www.google.com"}}
|
||||
|
||||
|
||||
@pytest.fixture(scope="session")
|
||||
def agent_config(llama_stack_client_with_mocked_inference, text_model_id):
|
||||
available_shields = [shield.identifier for shield in llama_stack_client_with_mocked_inference.shields.list()]
|
||||
|
|
@ -551,8 +571,9 @@ def test_rag_and_code_agent(llama_stack_client_with_mocked_inference, agent_conf
|
|||
assert expected_kw in response.output_message.content.lower()
|
||||
|
||||
|
||||
def test_create_turn_response(llama_stack_client_with_mocked_inference, agent_config):
|
||||
client_tool = get_boiling_point
|
||||
@pytest.mark.parametrize("client_tools", [(get_boiling_point, False), (get_boiling_point_with_metadata, True)])
|
||||
def test_create_turn_response(llama_stack_client_with_mocked_inference, agent_config, client_tools):
|
||||
client_tool, expectes_metadata = client_tools
|
||||
agent_config = {
|
||||
**agent_config,
|
||||
"input_shields": [],
|
||||
|
|
@ -577,7 +598,9 @@ def test_create_turn_response(llama_stack_client_with_mocked_inference, agent_co
|
|||
assert len(steps) == 3
|
||||
assert steps[0].step_type == "inference"
|
||||
assert steps[1].step_type == "tool_execution"
|
||||
assert steps[1].tool_calls[0].tool_name == "get_boiling_point"
|
||||
assert steps[1].tool_calls[0].tool_name.startswith("get_boiling_point")
|
||||
if expectes_metadata:
|
||||
assert steps[1].tool_responses[0].metadata["source"] == "https://www.google.com"
|
||||
assert steps[2].step_type == "inference"
|
||||
|
||||
last_step_completed_at = None
|
||||
|
|
|
|||
File diff suppressed because one or more lines are too long
Binary file not shown.
|
|
@ -1,4 +1,13 @@
|
|||
{
|
||||
"()_[('kwargs', {'session_id': '<UUID>', 'code': \"import pandas as pd\\nimport matplotlib.pyplot as plt\\n\\n# Load data\\ndf = pd.read_csv('inflation.csv')\\n\\n# Convert 'date' column to datetime\\ndf['date'] = pd.to_datetime(df['date'])\\n\\n# Group by year and calculate average inflation\\naverage_inflation = df.groupby(df['date'].dt.year)['inflation'].mean()\\n\\n# Plot the time series\\nplt.figure(figsize=(10,6))\\nplt.plot(average_inflation.index, average_inflation.values, marker='o')\\nplt.title('Average Yearly Inflation')\\nplt.xlabel('Year')\\nplt.ylabel('Average Inflation')\\nplt.grid(True)\\nplt.show()\"}), ('tool_name', 'code_interpreter')]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
"content": "completed\n[stderr]\nTraceback (most recent call last):\n line 5, in <module>\n from bwrap.core import main\nModuleNotFoundError: No module named 'bwrap.core'\n[/stderr]",
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": null
|
||||
}
|
||||
},
|
||||
"()_[('kwargs', {'session_id': '<UUID>', 'code': \"import pandas as pd\\nimport matplotlib.pyplot as plt\\n\\n# Load data\\ndf = pd.read_csv('inflation.csv')\\n\\n# Convert date column to datetime\\ndf['date'] = pd.to_datetime(df['date'])\\n\\n# Group by year and calculate average inflation\\naverage_inflation = df.groupby(df['date'].dt.year)['inflation'].mean()\\n\\n# Plot time series\\nplt.figure(figsize=(10,6))\\nplt.plot(average_inflation.index, average_inflation.values, marker='o')\\nplt.title('Average Yearly Inflation')\\nplt.xlabel('Year')\\nplt.ylabel('Average Inflation')\\nplt.grid(True)\\nplt.show()\"}), ('tool_name', 'code_interpreter')]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
|
|
@ -80,6 +89,15 @@
|
|||
"metadata": null
|
||||
}
|
||||
},
|
||||
"()_[('kwargs', {'session_id': '<UUID>', 'code': 'import pandas as pd\\nimport code_interpreter\\n\\n# Load the CSV file\\ndf = pd.read_csv(\"<TEMP_FILE>\")\\n\\n# Print the first few rows of the dataframe\\nprint(df.head())\\n\\n# Print the data types of each column\\nprint(df.dtypes)\\n\\n# Print the summary statistics of the dataframe\\nprint(df.describe())'}), ('tool_name', 'code_interpreter')]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
"content": "completed\n[stderr]\nTraceback (most recent call last):\n line 5, in <module>\n from bwrap.core import main\nModuleNotFoundError: No module named 'bwrap.core'\n[/stderr]",
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": null
|
||||
}
|
||||
},
|
||||
"()_[('kwargs', {'session_id': '<UUID>', 'code': 'import pandas as pd\\nimport matplotlib.pyplot as plt\\n\\n# Load the CSV file\\ndf = pd.read_csv(\"<TEMP_FILE>\")\\n\\n# Convert the \\'Year\\' column to datetime\\ndf[\\'Year\\'] = pd.to_datetime(df[\\'Year\\'], format=\\'%Y\\')\\n\\n# Group by \\'Year\\' and calculate the average inflation\\ndf_avg_inflation = df.groupby(\\'Year\\')[\\'Inflation\\'].mean().reset_index()\\n\\n# Plot the average inflation as a time series\\nplt.figure(figsize=(10,6))\\nplt.plot(df_avg_inflation[\\'Year\\'], df_avg_inflation[\\'Inflation\\'], marker=\\'o\\')\\nplt.title(\\'Average Yearly Inflation\\')\\nplt.xlabel(\\'Year\\')\\nplt.ylabel(\\'Inflation\\')\\nplt.grid(True)\\nplt.show()'}), ('tool_name', 'code_interpreter')]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
|
|
@ -98,6 +116,52 @@
|
|||
"metadata": null
|
||||
}
|
||||
},
|
||||
"()_[('kwargs', {'session_id': '<UUID>', 'query': 'How to use LoRA in Torchtune', 'vector_db_ids': ['vector_db_<UUID>']}), ('tool_name', 'knowledge_search')]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
"content": [
|
||||
{
|
||||
"text": "knowledge_search tool found 5 chunks:\nBEGIN of knowledge_search tool results.\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 1:\nDocument_id:cc646\nContent: .. _lora_finetune_label:\n\n============================\nFine-Tuning Llama2 with LoRA\n============================\n\nThis guide will teach you about `LoRA <https://arxiv.org/abs/2106.09685>`_, a parameter-efficient finetuning technique,\nand show you how you can use torchtune to finetune a Llama2 model with LoRA.\nIf you already know what LoRA is and want to get straight to running\nyour own LoRA finetune in torchtune, you can jump to :ref:`LoRA finetuning recipe in torchtune<lora_recipe_label>`.\n\n.. grid:: 2\n\n .. grid-item-card:: :octicon:`mortar-board;1em;` What you will learn\n\n * What LoRA is and how it saves memory during finetuning\n * An overview of LoRA components in torchtune\n * How to run a LoRA finetune using torchtune\n * How to experiment with different LoRA configurations\n\n .. grid-item-card:: :octicon:`list-unordered;1em;` Prerequisites\n\n * Be familiar with :ref:`torchtune<overview_label>`\n * Make sure to :ref:`install torchtune<install_label>`\n * Make sure you have downloaded the :ref:`Llama2-7B model weights<download_llama_label>`\n\nWhat is LoRA?\n-------------\n\n`LoRA <https://arxiv.org/abs/2106.09685>`_ is an adapter-based method for\nparameter-efficient finetuning that adds trainable low-rank decomposition matrices to different layers of a neural network,\nthen freezes the network's remaining parameters. LoRA is most commonly applied to\ntransformer models, in which case it is common to add the low-rank matrices\nto some of the linear projections in each transformer layer's self-attention.\n\n.. note::\n\n If you're unfamiliar, check out these references for the `definition of rank <https://en.wikipedia.org/wiki/Rank_(linear_algebra)>`_\n and discussion of `low-rank approximations <https://en.wikipedia.org/wiki/Low-rank_approximation>`_.\n\nBy finetuning with LoRA (as opposed to finetuning all model parameters),\nyou can expect to see memory savings due to a substantial reduction in the\nnumber of parameters with gradients. When using an optimizer with momentum,\nlike `AdamW <https://py\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 2:\nDocument_id:cc646\nContent: LoRA to Llama2 models\n------------------------------\n\nWith torchtune, we can easily apply LoRA to Llama2 with a variety of different configurations.\nLet's take a look at how to construct Llama2 models in torchtune with and without LoRA.\n\n.. code-block:: python\n\n from torchtune.models.llama2 import llama2_7b, lora_llama2_7b\n\n # Build Llama2 without any LoRA layers\n base_model = llama2_7b()\n\n # The default settings for lora_llama2_7b will match those for llama2_7b\n # We just need to define which layers we want LoRA applied to.\n # Within each self-attention, we can choose from [\"q_proj\", \"k_proj\", \"v_proj\", and \"output_proj\"].\n # We can also set apply_lora_to_mlp=True or apply_lora_to_output=True to apply LoRA to other linear\n # layers outside of the self-attention.\n lora_model = lora_llama2_7b(lora_attn_modules=[\"q_proj\", \"v_proj\"])\n\n.. note::\n\n Calling :func:`lora_llama_2_7b <torchtune.models.llama2.lora_llama2_7b>` alone will not handle the definition of which parameters are trainable.\n See :ref:`below<setting_trainable_params>` for how to do this.\n\nLet's inspect each of these models a bit more closely.\n\n.. code-block:: bash\n\n # Print the first layer's self-attention in the usual Llama2 model\n >>> print(base_model.layers[0].attn)\n MultiHeadAttention(\n (q_proj): Linear(in_features=4096, out_features=4096, bias=False)\n (k_proj): Linear(in_features=4096, out_features=4096, bias=False)\n (v_proj): Linear(in_features=4096, out_features=4096, bias=False)\n (output_proj): Linear(in_features=4096, out_features=4096, bias=False)\n (pos_embeddings): RotaryPositionalEmbeddings()\n )\n\n # Print the same for Llama2 with LoRA weights\n >>> print(lora_model.layers[0].attn)\n MultiHeadAttention(\n (q_proj): LoRALinear(\n (dropout): Dropout(p=0.0, inplace=False)\n \n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 3:\nDocument_id:cc646\nContent: 06% of all params are trainable.\n\n.. note::\n If you are directly using the LoRA recipe (as detailed :ref:`here<lora_recipe_label>`), you need only pass the\n relevant checkpoint path. Loading model weights and setting trainable parameters will be taken care\n of in the recipe.\n\n\n.. _lora_recipe_label:\n\nLoRA finetuning recipe in torchtune\n-----------------------------------\n\nFinally, we can put it all together and finetune a model using torchtune's `LoRA recipe <https://github.com/pytorch/torchtune/blob/48626d19d2108f92c749411fbd5f0ff140023a25/recipes/lora_finetune.py>`_.\nMake sure that you have first downloaded the Llama2 weights and tokenizer by following :ref:`these instructions<download_llama_label>`.\nYou can then run the following command to perform a LoRA finetune of Llama2-7B with two GPUs (each having VRAM of at least 16GB):\n\n.. code-block:: bash\n\n tune run --nnodes 1 --nproc_per_node 2 lora_finetune_distributed --config llama2/7B_lora\n\n.. note::\n Make sure to point to the location of your Llama2 weights and tokenizer. This can be done\n either by adding :code:`checkpointer.checkpoint_files=[my_model_checkpoint_path] tokenizer_checkpoint=my_tokenizer_checkpoint_path`\n or by directly modifying the :code:`7B_lora.yaml` file. See our \"\":ref:`config_tutorial_label`\" recipe\n for more details on how you can easily clone and modify torchtune configs.\n\n.. note::\n You can modify the value of :code:`nproc_per_node` depending on (a) the number of GPUs you have available,\n and (b) the memory constraints of your hardware.\n\nThe preceding command will run a LoRA finetune with torchtune's factory settings, but we may want to experiment a bit.\nLet's take a closer look at some of the :code:`lora_finetune_distributed` config.\n\n.. code-block:: yaml\n\n # Model Arguments\n model:\n _component_: lora_llama2_7b\n lora_attn_modules: ['q_proj', 'v_proj']\n lora_rank: 8\n lora_alpha: 16\n ...\n\nWe see that the\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 4:\nDocument_id:cc646\nContent: from our Llama2\nmodel without any wrappers or custom checkpoint conversion logic.\n\n.. code-block:: python\n\n # Assuming that base_model already has the pretrained Llama2 weights,\n # this will directly load them into your LoRA model without any conversion necessary.\n lora_model.load_state_dict(base_model.state_dict(), strict=False)\n\n.. note::\n Whenever loading weights with :code:`strict=False`, you should verify that any missing or extra keys in\n the loaded :code:`state_dict` are as expected. torchtune's LoRA recipes do this by default via\n :func:`validate_missing_and_unexpected_for_lora() <torchtune.modules.peft.validate_missing_and_unexpected_for_lora>`.\n\nOnce we've loaded the base model weights, we also want to set only LoRA parameters to trainable.\n\n.. _setting_trainable_params:\n\n.. code-block:: python\n\n from torchtune.modules.peft.peft_utils import get_adapter_params, set_trainable_params\n\n # Fetch all params from the model that are associated with LoRA.\n lora_params = get_adapter_params(lora_model)\n\n # Set requires_grad=True on lora_params, and requires_grad=False on all others.\n set_trainable_params(lora_model, lora_params)\n\n # Print the total number of parameters\n total_params = sum([p.numel() for p in lora_model.parameters()])\n trainable_params = sum([p.numel() for p in lora_model.parameters() if p.requires_grad])\n print(\n f\"\"\"\n {total_params} total params,\n {trainable_params}\" trainable params,\n {(100.0 * trainable_params / total_params):.2f}% of all params are trainable.\n \"\"\"\n )\n\n 6742609920 total params,\n 4194304 trainable params,\n 0.06% of all params are trainable.\n\n.. note::\n If you are directly using the LoRA recipe (as detailed :ref:`here<lora_recipe_label>`), you need only pass the\n relevant checkpoint path. Loading model weights and setting trainable parameters will be taken care\n of in the recipe.\n\n\n.. _lora_recipe_label:\n\nLoRA finetuning recipe in torchtune\n-----------------------------------\n\nFinally, we can put it all together and finetune a model using torchtune's `LoRA recipe <https://github.com/pytorch/torchtune/blob/48626d19d2108f92\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 5:\nDocument_id:cc646\nContent: ,\n and (b) the memory constraints of your hardware.\n\nThe preceding command will run a LoRA finetune with torchtune's factory settings, but we may want to experiment a bit.\nLet's take a closer look at some of the :code:`lora_finetune_distributed` config.\n\n.. code-block:: yaml\n\n # Model Arguments\n model:\n _component_: lora_llama2_7b\n lora_attn_modules: ['q_proj', 'v_proj']\n lora_rank: 8\n lora_alpha: 16\n ...\n\nWe see that the default is to apply LoRA to Q and V projections with a rank of 8.\nSome experiments with LoRA have found that it can be beneficial to apply LoRA to all linear layers in\nthe self-attention, and to increase the rank to 16 or 32. Note that this is likely to increase our max memory,\nbut as long as we keep :code:`rank<<embed_dim`, the impact should be relatively minor.\n\nLet's run this experiment. We can also increase alpha (in general it is good practice to scale alpha and rank together).\n\n.. code-block:: bash\n\n tune run --nnodes 1 --nproc_per_node 2 lora_finetune_distributed --config llama2/7B_lora \\\n lora_attn_modules=['q_proj','k_proj','v_proj','output_proj'] \\\n lora_rank=32 lora_alpha=64 output_dir=./lora_experiment_1\n\nA comparison of the (smoothed) loss curves between this run and our baseline over the first 500 steps can be seen below.\n\n.. image:: /_static/img/lora_experiment_loss_curves.png\n\n.. note::\n The above figure was generated with W&B. You can use torchtune's :class:`~torchtune.training.metric_logging.WandBLogger`\n to generate similar loss curves, but you will need to install W&B and setup an account separately. For more details on\n using W&B in torchtune, see our \":ref:`wandb_logging`\" recipe.\n\n.. _lora_tutorial_memory_tradeoff_label:\n\nTrading off memory and model performance with LoRA\n--------------------------------------------------\n\nIn the preceding example, we ran LoRA on two devices. But given LoRA's low memory footprint, we can run fine-tuning\non a single device using most commodity GPUs which support `bfloat16 <https://\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "END of knowledge_search tool results.\n",
|
||||
"type": "text"
|
||||
}
|
||||
],
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": {
|
||||
"document_ids": [
|
||||
"cc6460bf-74ab-4d11-8d32-bc02144a4e79",
|
||||
"cc6460bf-74ab-4d11-8d32-bc02144a4e79",
|
||||
"cc6460bf-74ab-4d11-8d32-bc02144a4e79",
|
||||
"cc6460bf-74ab-4d11-8d32-bc02144a4e79",
|
||||
"cc6460bf-74ab-4d11-8d32-bc02144a4e79"
|
||||
]
|
||||
}
|
||||
}
|
||||
},
|
||||
"()_[('kwargs', {'session_id': '<UUID>', 'query': 'How to use LoRA', 'vector_db_ids': ['vector_db_<UUID>']}), ('tool_name', 'knowledge_search')]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
|
|
@ -307,23 +371,23 @@
|
|||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 1:\nDocument_id:f76dc\nContent: conversational data, :func:`~torchtune.datasets.chat_dataset` seems to be a good fit. For any\ncustom local dataset we always need to specify ``source``, ``data_files``, and ``split`` for any dataset\nbuilder in torchtune. For :func:`~torchtune.datasets.chat_dataset`, we additionally need to specify\n``conversation_column`` and ``conversation_style``. Our data follows the ``\"sharegpt\"`` format, so\nwe can specify that here. Altogether, our :func:`~torchtune.datasets.chat_dataset` call should\nlook like so:\n\n.. code-block:: python\n\n from torchtune.datasets import chat_dataset\n from torchtune.models.llama3 import llama3_tokenizer\n\n tokenizer = llama3_tokenizer(\"/tmp/Meta-Llama-3-8B-Instruct/original/tokenizer.model\")\n ds = chat_dataset(\n tokenizer=tokenizer,\n source=\"json\",\n data_files=\"data/my_data.json\",\n split=\"train\",\n conversation_column=\"dialogue\",\n conversation_style=\"sharegpt\",\n )\n\n.. code-block:: yaml\n\n # In config\n tokenizer:\n _component_: torchtune.models.llama3.llama3_tokenizer\n path: /tmp/Meta-Llama-3-8B-Instruct/original/tokenizer.model\n\n dataset:\n _component_: torchtune.datasets.chat_dataset\n source: json\n data_files: data/my_data.json\n split: train\n conversation_column: dialogue\n conversation_style: sharegpt\n\n.. note::\n You can pass in any keyword argument for `load_dataset <https://huggingface.co/docs/datasets/v2.20.0/en/package_reference/loading_methods#datasets.load_dataset>`_ into all our\n Dataset classes and they will honor them. This is useful for common parameters\n such as specifying the data split with :code:`split` or configuration with\n :code:`name`\n\nIf you needed to add a prompt template, you would simply pass it into the tokenizer.\nSince we're fine-tuning Llama3, the tokenizer will handle all formatting for\nus and prompt templates are optional. Other models such as Mistral's :class:`~torchtune.models.mistral._tokenizer.MistralTokenizer`,\nuse a chat template by default (:class:`~torchtune.models.mistral.MistralChatTemplate`) to format\nall messages according to their `recommendations <https://\n",
|
||||
"text": "Result 1:\nDocument_id:ab1b9\nContent: conversational data, :func:`~torchtune.datasets.chat_dataset` seems to be a good fit. For any\ncustom local dataset we always need to specify ``source``, ``data_files``, and ``split`` for any dataset\nbuilder in torchtune. For :func:`~torchtune.datasets.chat_dataset`, we additionally need to specify\n``conversation_column`` and ``conversation_style``. Our data follows the ``\"sharegpt\"`` format, so\nwe can specify that here. Altogether, our :func:`~torchtune.datasets.chat_dataset` call should\nlook like so:\n\n.. code-block:: python\n\n from torchtune.datasets import chat_dataset\n from torchtune.models.llama3 import llama3_tokenizer\n\n tokenizer = llama3_tokenizer(\"/tmp/Meta-Llama-3-8B-Instruct/original/tokenizer.model\")\n ds = chat_dataset(\n tokenizer=tokenizer,\n source=\"json\",\n data_files=\"data/my_data.json\",\n split=\"train\",\n conversation_column=\"dialogue\",\n conversation_style=\"sharegpt\",\n )\n\n.. code-block:: yaml\n\n # In config\n tokenizer:\n _component_: torchtune.models.llama3.llama3_tokenizer\n path: /tmp/Meta-Llama-3-8B-Instruct/original/tokenizer.model\n\n dataset:\n _component_: torchtune.datasets.chat_dataset\n source: json\n data_files: data/my_data.json\n split: train\n conversation_column: dialogue\n conversation_style: sharegpt\n\n.. note::\n You can pass in any keyword argument for `load_dataset <https://huggingface.co/docs/datasets/v2.20.0/en/package_reference/loading_methods#datasets.load_dataset>`_ into all our\n Dataset classes and they will honor them. This is useful for common parameters\n such as specifying the data split with :code:`split` or configuration with\n :code:`name`\n\nIf you needed to add a prompt template, you would simply pass it into the tokenizer.\nSince we're fine-tuning Llama3, the tokenizer will handle all formatting for\nus and prompt templates are optional. Other models such as Mistral's :class:`~torchtune.models.mistral._tokenizer.MistralTokenizer`,\nuse a chat template by default (:class:`~torchtune.models.mistral.MistralChatTemplate`) to format\nall messages according to their `recommendations <https://\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 2:\nDocument_id:c4fc3\nContent: .. _lora_finetune_label:\n\n============================\nFine-Tuning Llama2 with LoRA\n============================\n\nThis guide will teach you about `LoRA <https://arxiv.org/abs/2106.09685>`_, a parameter-efficient finetuning technique,\nand show you how you can use torchtune to finetune a Llama2 model with LoRA.\nIf you already know what LoRA is and want to get straight to running\nyour own LoRA finetune in torchtune, you can jump to :ref:`LoRA finetuning recipe in torchtune<lora_recipe_label>`.\n\n.. grid:: 2\n\n .. grid-item-card:: :octicon:`mortar-board;1em;` What you will learn\n\n * What LoRA is and how it saves memory during finetuning\n * An overview of LoRA components in torchtune\n * How to run a LoRA finetune using torchtune\n * How to experiment with different LoRA configurations\n\n .. grid-item-card:: :octicon:`list-unordered;1em;` Prerequisites\n\n * Be familiar with :ref:`torchtune<overview_label>`\n * Make sure to :ref:`install torchtune<install_label>`\n * Make sure you have downloaded the :ref:`Llama2-7B model weights<download_llama_label>`\n\nWhat is LoRA?\n-------------\n\n`LoRA <https://arxiv.org/abs/2106.09685>`_ is an adapter-based method for\nparameter-efficient finetuning that adds trainable low-rank decomposition matrices to different layers of a neural network,\nthen freezes the network's remaining parameters. LoRA is most commonly applied to\ntransformer models, in which case it is common to add the low-rank matrices\nto some of the linear projections in each transformer layer's self-attention.\n\n.. note::\n\n If you're unfamiliar, check out these references for the `definition of rank <https://en.wikipedia.org/wiki/Rank_(linear_algebra)>`_\n and discussion of `low-rank approximations <https://en.wikipedia.org/wiki/Low-rank_approximation>`_.\n\nBy finetuning with LoRA (as opposed to finetuning all model parameters),\nyou can expect to see memory savings due to a substantial reduction in the\nnumber of parameters with gradients. When using an optimizer with momentum,\nlike `AdamW <https://py\n",
|
||||
"text": "Result 2:\nDocument_id:cc646\nContent: .. _lora_finetune_label:\n\n============================\nFine-Tuning Llama2 with LoRA\n============================\n\nThis guide will teach you about `LoRA <https://arxiv.org/abs/2106.09685>`_, a parameter-efficient finetuning technique,\nand show you how you can use torchtune to finetune a Llama2 model with LoRA.\nIf you already know what LoRA is and want to get straight to running\nyour own LoRA finetune in torchtune, you can jump to :ref:`LoRA finetuning recipe in torchtune<lora_recipe_label>`.\n\n.. grid:: 2\n\n .. grid-item-card:: :octicon:`mortar-board;1em;` What you will learn\n\n * What LoRA is and how it saves memory during finetuning\n * An overview of LoRA components in torchtune\n * How to run a LoRA finetune using torchtune\n * How to experiment with different LoRA configurations\n\n .. grid-item-card:: :octicon:`list-unordered;1em;` Prerequisites\n\n * Be familiar with :ref:`torchtune<overview_label>`\n * Make sure to :ref:`install torchtune<install_label>`\n * Make sure you have downloaded the :ref:`Llama2-7B model weights<download_llama_label>`\n\nWhat is LoRA?\n-------------\n\n`LoRA <https://arxiv.org/abs/2106.09685>`_ is an adapter-based method for\nparameter-efficient finetuning that adds trainable low-rank decomposition matrices to different layers of a neural network,\nthen freezes the network's remaining parameters. LoRA is most commonly applied to\ntransformer models, in which case it is common to add the low-rank matrices\nto some of the linear projections in each transformer layer's self-attention.\n\n.. note::\n\n If you're unfamiliar, check out these references for the `definition of rank <https://en.wikipedia.org/wiki/Rank_(linear_algebra)>`_\n and discussion of `low-rank approximations <https://en.wikipedia.org/wiki/Low-rank_approximation>`_.\n\nBy finetuning with LoRA (as opposed to finetuning all model parameters),\nyou can expect to see memory savings due to a substantial reduction in the\nnumber of parameters with gradients. When using an optimizer with momentum,\nlike `AdamW <https://py\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 3:\nDocument_id:de2d4\nContent: ` module, which we swap\n out for :class:`~torchtune.modules.peft.LoRALinear` when ``use_dora=True``.\n\n.. _glossary_distrib:\n\n\n.. TODO\n\n.. Distributed\n.. -----------\n\n.. .. _glossary_fsdp:\n\n.. Fully Sharded Data Parallel (FSDP)\n.. ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n\n.. All our ``_distributed`` recipes use `FSDP <https://pytorch.org/docs/stable/fsdp.html>`.\n.. .. _glossary_fsdp2:\n\n",
|
||||
"text": "Result 3:\nDocument_id:8bcf6\nContent: ` module, which we swap\n out for :class:`~torchtune.modules.peft.LoRALinear` when ``use_dora=True``.\n\n.. _glossary_distrib:\n\n\n.. TODO\n\n.. Distributed\n.. -----------\n\n.. .. _glossary_fsdp:\n\n.. Fully Sharded Data Parallel (FSDP)\n.. ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n\n.. All our ``_distributed`` recipes use `FSDP <https://pytorch.org/docs/stable/fsdp.html>`.\n.. .. _glossary_fsdp2:\n\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 4:\nDocument_id:c4fc3\nContent: 06% of all params are trainable.\n\n.. note::\n If you are directly using the LoRA recipe (as detailed :ref:`here<lora_recipe_label>`), you need only pass the\n relevant checkpoint path. Loading model weights and setting trainable parameters will be taken care\n of in the recipe.\n\n\n.. _lora_recipe_label:\n\nLoRA finetuning recipe in torchtune\n-----------------------------------\n\nFinally, we can put it all together and finetune a model using torchtune's `LoRA recipe <https://github.com/pytorch/torchtune/blob/48626d19d2108f92c749411fbd5f0ff140023a25/recipes/lora_finetune.py>`_.\nMake sure that you have first downloaded the Llama2 weights and tokenizer by following :ref:`these instructions<download_llama_label>`.\nYou can then run the following command to perform a LoRA finetune of Llama2-7B with two GPUs (each having VRAM of at least 16GB):\n\n.. code-block:: bash\n\n tune run --nnodes 1 --nproc_per_node 2 lora_finetune_distributed --config llama2/7B_lora\n\n.. note::\n Make sure to point to the location of your Llama2 weights and tokenizer. This can be done\n either by adding :code:`checkpointer.checkpoint_files=[my_model_checkpoint_path] tokenizer_checkpoint=my_tokenizer_checkpoint_path`\n or by directly modifying the :code:`7B_lora.yaml` file. See our \"\":ref:`config_tutorial_label`\" recipe\n for more details on how you can easily clone and modify torchtune configs.\n\n.. note::\n You can modify the value of :code:`nproc_per_node` depending on (a) the number of GPUs you have available,\n and (b) the memory constraints of your hardware.\n\nThe preceding command will run a LoRA finetune with torchtune's factory settings, but we may want to experiment a bit.\nLet's take a closer look at some of the :code:`lora_finetune_distributed` config.\n\n.. code-block:: yaml\n\n # Model Arguments\n model:\n _component_: lora_llama2_7b\n lora_attn_modules: ['q_proj', 'v_proj']\n lora_rank: 8\n lora_alpha: 16\n ...\n\nWe see that the\n",
|
||||
"text": "Result 4:\nDocument_id:cc646\nContent: 06% of all params are trainable.\n\n.. note::\n If you are directly using the LoRA recipe (as detailed :ref:`here<lora_recipe_label>`), you need only pass the\n relevant checkpoint path. Loading model weights and setting trainable parameters will be taken care\n of in the recipe.\n\n\n.. _lora_recipe_label:\n\nLoRA finetuning recipe in torchtune\n-----------------------------------\n\nFinally, we can put it all together and finetune a model using torchtune's `LoRA recipe <https://github.com/pytorch/torchtune/blob/48626d19d2108f92c749411fbd5f0ff140023a25/recipes/lora_finetune.py>`_.\nMake sure that you have first downloaded the Llama2 weights and tokenizer by following :ref:`these instructions<download_llama_label>`.\nYou can then run the following command to perform a LoRA finetune of Llama2-7B with two GPUs (each having VRAM of at least 16GB):\n\n.. code-block:: bash\n\n tune run --nnodes 1 --nproc_per_node 2 lora_finetune_distributed --config llama2/7B_lora\n\n.. note::\n Make sure to point to the location of your Llama2 weights and tokenizer. This can be done\n either by adding :code:`checkpointer.checkpoint_files=[my_model_checkpoint_path] tokenizer_checkpoint=my_tokenizer_checkpoint_path`\n or by directly modifying the :code:`7B_lora.yaml` file. See our \"\":ref:`config_tutorial_label`\" recipe\n for more details on how you can easily clone and modify torchtune configs.\n\n.. note::\n You can modify the value of :code:`nproc_per_node` depending on (a) the number of GPUs you have available,\n and (b) the memory constraints of your hardware.\n\nThe preceding command will run a LoRA finetune with torchtune's factory settings, but we may want to experiment a bit.\nLet's take a closer look at some of the :code:`lora_finetune_distributed` config.\n\n.. code-block:: yaml\n\n # Model Arguments\n model:\n _component_: lora_llama2_7b\n lora_attn_modules: ['q_proj', 'v_proj']\n lora_rank: 8\n lora_alpha: 16\n ...\n\nWe see that the\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 5:\nDocument_id:de2d4\nContent: etune\n:func:`torchtune.models.llama3.llama3_8b` with DoRA, you would use :func:`torchtune.models.llama3.lora_llama3_8b` with ``use_dora=True``:\n\n.. code-block:: bash\n\n tune run lora_finetune_single_device --config llama3/8B_lora_single_device \\\n model.use_dora=True\n\n.. code-block:: yaml\n\n model:\n _component_: torchtune.models.lora_llama3_8b\n use_dora: True\n\nSince DoRA extends LoRA, the parameters for :ref:`customizing LoRA <glossary_lora>` are identical. You can also quantize the base model weights like in :ref:`glossary_qlora` by using ``quantize=True`` to reap\neven more memory savings!\n\n.. code-block:: bash\n\n tune run lora_finetune_single_device --config llama3/8B_lora_single_device \\\n model.apply_lora_to_mlp=True \\\n model.lora_attn_modules=[\"q_proj\",\"k_proj\",\"v_proj\"] \\\n model.lora_rank=16 \\\n model.lora_alpha=32 \\\n model.use_dora=True \\\n model.quantize_base=True\n\n.. code-block:: yaml\n\n model:\n _component_: torchtune.models.lora_llama3_8b\n apply_lora_to_mlp: True\n lora_attn_modules: [\"q_proj\", \"k_proj\", \"v_proj\"]\n lora_rank: 16\n lora_alpha: 32\n use_dora: True\n quantize_base: True\n\n\n.. note::\n\n Under the hood, we've enabled DoRA by adding the :class:`~torchtune.modules.peft.DoRALinear` module, which we swap\n out for :class:`~torchtune.modules.peft.LoRALinear` when ``use_dora=True``.\n\n.. _glossary_distrib:\n\n\n.. TODO\n\n.. Distributed\n.. -----------\n\n.. .. _glossary_fsdp:\n\n.. Fully Sharded Data Parallel (FSDP)\n.. ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n\n.. All our ``_distributed`` recipes use `FSDP <https://pytorch.org/docs/stable/fsdp.html>`.\n.. .. _glossary_fsdp2:\n\n",
|
||||
"text": "Result 5:\nDocument_id:8bcf6\nContent: etune\n:func:`torchtune.models.llama3.llama3_8b` with DoRA, you would use :func:`torchtune.models.llama3.lora_llama3_8b` with ``use_dora=True``:\n\n.. code-block:: bash\n\n tune run lora_finetune_single_device --config llama3/8B_lora_single_device \\\n model.use_dora=True\n\n.. code-block:: yaml\n\n model:\n _component_: torchtune.models.lora_llama3_8b\n use_dora: True\n\nSince DoRA extends LoRA, the parameters for :ref:`customizing LoRA <glossary_lora>` are identical. You can also quantize the base model weights like in :ref:`glossary_qlora` by using ``quantize=True`` to reap\neven more memory savings!\n\n.. code-block:: bash\n\n tune run lora_finetune_single_device --config llama3/8B_lora_single_device \\\n model.apply_lora_to_mlp=True \\\n model.lora_attn_modules=[\"q_proj\",\"k_proj\",\"v_proj\"] \\\n model.lora_rank=16 \\\n model.lora_alpha=32 \\\n model.use_dora=True \\\n model.quantize_base=True\n\n.. code-block:: yaml\n\n model:\n _component_: torchtune.models.lora_llama3_8b\n apply_lora_to_mlp: True\n lora_attn_modules: [\"q_proj\", \"k_proj\", \"v_proj\"]\n lora_rank: 16\n lora_alpha: 32\n use_dora: True\n quantize_base: True\n\n\n.. note::\n\n Under the hood, we've enabled DoRA by adding the :class:`~torchtune.modules.peft.DoRALinear` module, which we swap\n out for :class:`~torchtune.modules.peft.LoRALinear` when ``use_dora=True``.\n\n.. _glossary_distrib:\n\n\n.. TODO\n\n.. Distributed\n.. -----------\n\n.. .. _glossary_fsdp:\n\n.. Fully Sharded Data Parallel (FSDP)\n.. ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n\n.. All our ``_distributed`` recipes use `FSDP <https://pytorch.org/docs/stable/fsdp.html>`.\n.. .. _glossary_fsdp2:\n\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
|
|
@ -335,11 +399,11 @@
|
|||
"error_message": null,
|
||||
"metadata": {
|
||||
"document_ids": [
|
||||
"f76dc7f5-9648-4272-a579-c8387fb1408a",
|
||||
"c4fc3cb6-6172-489e-90a7-b39d343e14c0",
|
||||
"de2d49de-55de-44dd-9bca-6f4f6d633b0a",
|
||||
"c4fc3cb6-6172-489e-90a7-b39d343e14c0",
|
||||
"de2d49de-55de-44dd-9bca-6f4f6d633b0a"
|
||||
"ab1b9c78-180f-48cb-bbef-c70a4a59e42d",
|
||||
"cc6460bf-74ab-4d11-8d32-bc02144a4e79",
|
||||
"8bcf61e4-98c4-41a7-87f9-833c1a4d2b28",
|
||||
"cc6460bf-74ab-4d11-8d32-bc02144a4e79",
|
||||
"8bcf61e4-98c4-41a7-87f9-833c1a4d2b28"
|
||||
]
|
||||
}
|
||||
}
|
||||
|
|
@ -398,5 +462,41 @@
|
|||
]
|
||||
}
|
||||
}
|
||||
},
|
||||
"()_[('kwargs', {'session_id': '<UUID>', 'query': 'when was the nba created', 'vector_db_ids': ['test-vector-db-<UUID>']}), ('tool_name', 'knowledge_search')]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
"content": [
|
||||
{
|
||||
"text": "knowledge_search tool found 3 chunks:\nBEGIN of knowledge_search tool results.\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 1:\nDocument_id:nba_w\nContent: The NBA was created on August 3, 1949, with the merger of the Basketball Association of America (BAA) and the National Basketball League (NBL).\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 2:\nDocument_id:perpl\nContent: Perplexity the company was founded in 2022 by Aravind Srinivas, Andy Konwinski, Denis Yarats and Johnny Ho, engineers with backgrounds in back-end systems, artificial intelligence (AI) and machine learning:\n\n Srinivas, the CEO, worked at OpenAI as an AI researcher.\n Konwinski was among the founding team at Databricks.\n Yarats, the CTO, was an AI research scientist at Meta.\n Ho, the CSO, worked as an engineer at Quora, then as a quantitative trader on Wall Street.[5]\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 3:\nDocument_id:perpl\nContent: Ho, the CSO, worked as an engineer at Quora, then as a quantitative trader on Wall Street.[5]\n",
|
||||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "END of knowledge_search tool results.\n",
|
||||
"type": "text"
|
||||
}
|
||||
],
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": {
|
||||
"document_ids": [
|
||||
"nba_wiki",
|
||||
"perplexity_wiki",
|
||||
"perplexity_wiki"
|
||||
]
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
|
|||
Binary file not shown.
Loading…
Add table
Add a link
Reference in a new issue