mirror of
https://github.com/meta-llama/llama-stack.git
synced 2025-12-28 06:00:23 +00:00
Merge branch 'meta-llama:main' into feat/litellm_sambanova_usage
This commit is contained in:
Jorge Piedrahita Ortiz
GitHub
commit
716cb09056
145 changed files with 21384 additions and 1283 deletions
|
|
@ -10,8 +10,7 @@ from uuid import uuid4
|
|||
import pytest
|
||||
from llama_stack_client.lib.agents.agent import Agent
|
||||
from llama_stack_client.lib.agents.event_logger import EventLogger
|
||||
from llama_stack_client.types.agents.turn_create_params import Document as AgentDocument
|
||||
from llama_stack_client.types.memory_insert_params import Document
|
||||
from llama_stack_client.types.agents.turn_create_params import Document
|
||||
from llama_stack_client.types.shared_params.agent_config import AgentConfig, ToolConfig
|
||||
|
||||
from llama_stack.apis.agents.agents import (
|
||||
|
|
@ -187,7 +186,7 @@ def test_builtin_tool_web_search(llama_stack_client_with_mocked_inference, agent
|
|||
messages=[
|
||||
{
|
||||
"role": "user",
|
||||
"content": "Search the web and tell me who the current CEO of Meta is.",
|
||||
"content": "Search the web and tell me who the founder of Meta is.",
|
||||
}
|
||||
],
|
||||
session_id=session_id,
|
||||
|
|
@ -242,7 +241,7 @@ def test_code_interpreter_for_attachments(llama_stack_client_with_mocked_inferen
|
|||
|
||||
codex_agent = Agent(llama_stack_client_with_mocked_inference, **agent_config)
|
||||
session_id = codex_agent.create_session(f"test-session-{uuid4()}")
|
||||
inflation_doc = AgentDocument(
|
||||
inflation_doc = Document(
|
||||
content="https://raw.githubusercontent.com/meta-llama/llama-stack-apps/main/examples/resources/inflation.csv",
|
||||
mime_type="text/csv",
|
||||
)
|
||||
|
|
|
|||
|
|
@ -6,12 +6,24 @@
|
|||
import inspect
|
||||
import itertools
|
||||
import os
|
||||
import platform
|
||||
import textwrap
|
||||
import time
|
||||
|
||||
from dotenv import load_dotenv
|
||||
|
||||
from llama_stack.log import get_logger
|
||||
|
||||
from .report import Report
|
||||
|
||||
logger = get_logger(__name__, category="tests")
|
||||
|
||||
|
||||
def pytest_runtest_teardown(item):
|
||||
interval_seconds = os.getenv("LLAMA_STACK_TEST_INTERVAL_SECONDS")
|
||||
if interval_seconds:
|
||||
time.sleep(float(interval_seconds))
|
||||
|
||||
|
||||
def pytest_configure(config):
|
||||
config.option.tbstyle = "short"
|
||||
|
|
@ -24,6 +36,10 @@ def pytest_configure(config):
|
|||
key, value = env_var.split("=", 1)
|
||||
os.environ[key] = value
|
||||
|
||||
if platform.system() == "Darwin": # Darwin is the system name for macOS
|
||||
os.environ["DISABLE_CODE_SANDBOX"] = "1"
|
||||
logger.info("Setting DISABLE_CODE_SANDBOX=1 for macOS")
|
||||
|
||||
if config.getoption("--report"):
|
||||
config.pluginmanager.register(Report(config))
|
||||
|
||||
|
|
|
|||
|
|
@ -9,11 +9,25 @@ import mimetypes
|
|||
import os
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
|
||||
# How to run this test:
|
||||
#
|
||||
# LLAMA_STACK_CONFIG="template-name" pytest -v tests/integration/datasetio
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def dataset_for_test(llama_stack_client):
|
||||
dataset_id = "test_dataset"
|
||||
register_dataset(llama_stack_client, dataset_id=dataset_id)
|
||||
yield
|
||||
# Teardown - this always runs, even if the test fails
|
||||
try:
|
||||
llama_stack_client.datasets.unregister(dataset_id)
|
||||
except Exception as e:
|
||||
print(f"Warning: Failed to unregister test_dataset: {e}")
|
||||
|
||||
|
||||
def data_url_from_file(file_path: str) -> str:
|
||||
if not os.path.exists(file_path):
|
||||
raise FileNotFoundError(f"File not found: {file_path}")
|
||||
|
|
@ -80,8 +94,7 @@ def test_register_unregister_dataset(llama_stack_client):
|
|||
assert len(response) == 0
|
||||
|
||||
|
||||
def test_get_rows_paginated(llama_stack_client):
|
||||
register_dataset(llama_stack_client)
|
||||
def test_get_rows_paginated(llama_stack_client, dataset_for_test):
|
||||
response = llama_stack_client.datasetio.get_rows_paginated(
|
||||
dataset_id="test_dataset",
|
||||
rows_in_page=3,
|
||||
|
|
|
|||
|
|
@ -52,6 +52,8 @@ def llama_stack_client_with_mocked_inference(llama_stack_client, request):
|
|||
|
||||
If --record-responses is passed, it will call the real APIs and record the responses.
|
||||
"""
|
||||
# TODO: will rework this to be more stable
|
||||
return llama_stack_client
|
||||
if not isinstance(llama_stack_client, LlamaStackAsLibraryClient):
|
||||
logging.warning(
|
||||
"llama_stack_client_with_mocked_inference is not supported for this client, returning original client without mocking"
|
||||
|
|
|
|||
File diff suppressed because one or more lines are too long
|
|
@ -5,7 +5,7 @@
|
|||
"__module__": "llama_stack.apis.tools.tools",
|
||||
"__pydantic__": "ToolInvocationResult",
|
||||
"data": {
|
||||
"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]",
|
||||
"content": "completed\n[stdout]\n541\n[/stdout]",
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": null
|
||||
|
|
@ -31,7 +31,7 @@
|
|||
"__module__": "llama_stack.apis.tools.tools",
|
||||
"__pydantic__": "ToolInvocationResult",
|
||||
"data": {
|
||||
"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]",
|
||||
"content": "completed\n[stdout]\nNumber of rows and columns in the data: (10, 13)\nColumns of the data are: 13\nColumns of the data are: Index(['Year', 'Jan', 'Feb', 'Mar', 'Apr', 'May', 'Jun', 'Jul', 'Aug', 'Sep',\n 'Oct', 'Nov', 'Dec'],\n dtype='object')\nDatatype of the columns are: Year int64\nJan float64\nFeb float64\nMar float64\nApr float64\nMay float64\nJun float64\nJul float64\nAug float64\nSep float64\nOct float64\nNov float64\nDec float64\ndtype: object\n[/stdout]",
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": null
|
||||
|
|
@ -70,7 +70,7 @@
|
|||
"__module__": "llama_stack.apis.tools.tools",
|
||||
"__pydantic__": "ToolInvocationResult",
|
||||
"data": {
|
||||
"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]",
|
||||
"content": "completed\n[stdout]\nYear Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec\n0 2014 1.6 1.6 1.7 1.8 2.0 1.9 1.9 1.7 1.7 1.8 1.7 1.6\n1 2015 1.6 1.7 1.8 1.8 1.7 1.8 1.8 1.8 1.9 1.9 2.0 2.1\n2 2016 2.2 2.3 2.2 2.1 2.2 2.2 2.2 2.3 2.2 2.1 2.1 2.2\n3 2017 2.3 2.2 2.0 1.9 1.7 1.7 1.7 1.7 1.7 1.8 1.7 1.8\n4 2018 1.8 1.8 2.1 2.1 2.2 2.3 2.4 2.2 2.2 2.1 2.2 2.2\n[/stdout]",
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": null
|
||||
|
|
@ -83,7 +83,7 @@
|
|||
"__module__": "llama_stack.apis.tools.tools",
|
||||
"__pydantic__": "ToolInvocationResult",
|
||||
"data": {
|
||||
"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]",
|
||||
"content": "completed\n[stderr]\nTraceback (most recent call last):\n line 142, in <module>\n line 23, in <module>\n from .code_execution import CodeExecutionContext, CodeExecutionRequest, CodeExecutor\nImportError: attempted relative import with no known parent package\n[/stderr]",
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": null
|
||||
|
|
@ -116,6 +116,19 @@
|
|||
}
|
||||
}
|
||||
},
|
||||
"[[], {\"kwargs\": {\"code\": \"import pandas as pd\\nimport matplotlib.pyplot as plt\\n\\n# Load data\\ndf = pd.read_csv(\\\"<TEMP_FILE>\")\\n\\n# Calculate average yearly inflation\\ndf['Average'] = df[['Jan', 'Feb', 'Mar', 'Apr', 'May', 'Jun', 'Jul', 'Aug', 'Sep', 'Oct', 'Nov', 'Dec']].mean(axis=1)\\n\\n# Plot time series\\nplt.figure(figsize=(10,6))\\nplt.plot(df['Year'], df['Average'])\\nplt.xlabel('Year')\\nplt.ylabel('Average Yearly Inflation')\\nplt.title('Average Yearly Inflation Over Time')\\nplt.grid(True)\\nplt.show()\", \"session_id\": \"<UUID>\"}, \"tool_name\": \"code_interpreter\"}]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
"__module__": "llama_stack.apis.tools.tools",
|
||||
"__pydantic__": "ToolInvocationResult",
|
||||
"data": {
|
||||
"content": "completed",
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": null
|
||||
}
|
||||
}
|
||||
},
|
||||
"[[], {\"kwargs\": {\"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 average yearly inflation as a 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()\", \"session_id\": \"<UUID>\"}, \"tool_name\": \"code_interpreter\"}]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
|
|
@ -154,23 +167,23 @@
|
|||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 1:\nDocument_id:961ff\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 1:\nDocument_id:c4e00\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:961ff\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",
|
||||
"text": "Result 2:\nDocument_id:c4e00\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:961ff\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 3:\nDocument_id:c4e00\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:961ff\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",
|
||||
"text": "Result 4:\nDocument_id:c4e00\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:961ff\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",
|
||||
"text": "Result 5:\nDocument_id:c4e00\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"
|
||||
},
|
||||
{
|
||||
|
|
@ -182,11 +195,11 @@
|
|||
"error_message": null,
|
||||
"metadata": {
|
||||
"document_ids": [
|
||||
"961ff2d1-8887-41ef-a4fe-fa4cbab7b932",
|
||||
"961ff2d1-8887-41ef-a4fe-fa4cbab7b932",
|
||||
"961ff2d1-8887-41ef-a4fe-fa4cbab7b932",
|
||||
"961ff2d1-8887-41ef-a4fe-fa4cbab7b932",
|
||||
"961ff2d1-8887-41ef-a4fe-fa4cbab7b932"
|
||||
"c4e00391-aeb8-4d32-ac41-ae3242f38a19",
|
||||
"c4e00391-aeb8-4d32-ac41-ae3242f38a19",
|
||||
"c4e00391-aeb8-4d32-ac41-ae3242f38a19",
|
||||
"c4e00391-aeb8-4d32-ac41-ae3242f38a19",
|
||||
"c4e00391-aeb8-4d32-ac41-ae3242f38a19"
|
||||
]
|
||||
}
|
||||
}
|
||||
|
|
@ -242,6 +255,19 @@
|
|||
}
|
||||
}
|
||||
},
|
||||
"[[], {\"kwargs\": {\"query\": \"Meta founder\", \"session_id\": \"<UUID>\"}, \"tool_name\": \"web_search\"}]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
"__module__": "llama_stack.apis.tools.tools",
|
||||
"__pydantic__": "ToolInvocationResult",
|
||||
"data": {
|
||||
"content": "{\"query\": \"Meta founder\", \"top_k\": [{\"title\": \"Mark Zuckerberg, Founder, Chairman and Chief Executive Officer - Meta\", \"url\": \"https://about.meta.com/media-gallery/executives/mark-zuckerberg/\", \"content\": \"Mark Zuckerberg, Founder, Chairman and Chief Executive Officer | Meta Meta Quest Ray-Ban Meta Meta Horizon Meta AI Meta Verified Meta Pay Meta Horizon Workrooms Meta and you Learn about our community Shop Meta Meta Quest Meta Portal Meta Horizon Mark Zuckerberg is the founder, chairman and CEO of Meta, which he originally founded as Facebook in 2004. In October 2021, Facebook rebranded to Meta to reflect all of its products and services across its family of apps and a focus on developing social experiences for the metaverse \\u2014 moving beyond 2D screens toward immersive experiences like augmented and virtual reality to help build the next evolution in social technology. Shop Ray-Ban Meta glassesRay-Ban StoriesPrivacy informationSupported countries \\u00a9 2025 Meta\", \"score\": 0.81595254, \"raw_content\": null}, {\"title\": \"Meta - Leadership & Governance\", \"url\": \"https://investor.atmeta.com/leadership-and-governance/\", \"content\": \"Mr. Andreessen was a co-founder of Netscape Communications Corporation, a software company, serving in various positions, including Chief Technology Officer and Executive Vice President of Products. Ms. Killefer also served as Assistant Secretary for Management, Chief Financial Officer, and Chief Operating Officer of the U.S. Department of the Treasury from 1997 to 2000 and as a member of the IRS Oversight Board from 2000 to 2005, including as Chair of the IRS Oversight Board from 2002 to 2004. Ms. Travis has served as Executive Vice President and Chief Financial Officer of The Estee Lauder Companies Inc., a global manufacturer and marketer of skin care, makeup, fragrance and hair care products, since August 2012.\", \"score\": 0.46759978, \"raw_content\": null}, {\"title\": \"Executives - Meta\", \"url\": \"https://about.meta.com/media-gallery/executives/\", \"content\": \"Meta leadership: images of senior executives for download to use in articles about the company. ... Mark Zuckerberg, Founder, Chairman and Chief Executive Officer. Nick Clegg, President, Global Affairs. Joel Kaplan, Chief Global Affairs Officer. Susan Li, Chief Financial Officer.\", \"score\": 0.46482924, \"raw_content\": null}, {\"title\": \"Meta Platforms - Wikipedia\", \"url\": \"https://en.wikipedia.org/wiki/Meta_Platforms\", \"content\": \"Following a period of intense scrutiny and damaging whistleblower leaks, news started to emerge on October 21, 2021, about Facebook's plan to rebrand the company and change its name.[15][54] In the Q3 2021 Earnings Call on October 25, Mark Zuckerberg discussed the ongoing criticism of the company's social services and the way it operates, and pointed to the pivoting efforts to building the metaverse \\u2013 without mentioning the rebranding and the name change.[55] The metaverse vision and the name change from Facebook, Inc. to Meta Platforms was introduced at Facebook Connect on October 28, 2021.[16] Based on Facebook's PR campaign, the name change reflects the company's shifting long term focus of building the metaverse, a digital extension of the physical world by social media, virtual reality and augmented reality features.[16][56]\", \"score\": 0.14999175, \"raw_content\": null}, {\"title\": \"Mark Zuckerberg - Wikipedia\", \"url\": \"https://en.wikipedia.org/wiki/Mark_Zuckerberg\", \"content\": \"They began dating in 2003.[175] In September 2010, Chan, who was a medical student at the University of California, San Francisco at the time,[176] moved into his rented house in Palo Alto, California.[177][178] They married on May 19, 2012, in the grounds of his mansion in an event that also celebrated her graduation from medical school.[179][180] Zuckerberg revealed in July 2015 that they were expecting a baby girl and that Chan had previously experienced three miscarriages.[181] Their first daughter was born in December 2015.[182] They announced in a Chinese New Year video that their daughter's Chinese name is Chen Mingyu (Chinese: \\u9648\\u660e\\u5b87).[183] Their second daughter was born in August 2017.[184] Zuckerberg and his wife welcomed their third daughter in March 2023 and announced the news across his social media pages.[185] The couple also have a Puli dog named Beast,[186] who has over two million followers on Facebook.[187] Zuckerberg commissioned the visual artist Daniel Arsham to build a 7-foot-tall sculpture of his wife, which was unveiled in 2024.[188]\", \"score\": 0.036911618, \"raw_content\": null}]}",
|
||||
"error_code": null,
|
||||
"error_message": null,
|
||||
"metadata": null
|
||||
}
|
||||
}
|
||||
},
|
||||
"[[], {\"kwargs\": {\"query\": \"NBA creation date\", \"session_id\": \"<UUID>\", \"vector_db_ids\": [\"test-vector-db-<UUID>\"]}, \"tool_name\": \"knowledge_search\"}]": {
|
||||
"type": "value",
|
||||
"value": {
|
||||
|
|
@ -374,23 +400,23 @@
|
|||
"type": "text"
|
||||
},
|
||||
{
|
||||
"text": "Result 1:\nDocument_id:24443\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:9050a\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:961ff\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:c4e00\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:b49f7\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:15efa\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:961ff\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:c4e00\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:b49f7\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:15efa\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"
|
||||
},
|
||||
{
|
||||
|
|
@ -402,11 +428,11 @@
|
|||
"error_message": null,
|
||||
"metadata": {
|
||||
"document_ids": [
|
||||
"24443dfb-a0b3-4ce8-820e-3fb1f12364bb",
|
||||
"961ff2d1-8887-41ef-a4fe-fa4cbab7b932",
|
||||
"b49f7985-6615-4dcf-99be-d1765b6a6fc6",
|
||||
"961ff2d1-8887-41ef-a4fe-fa4cbab7b932",
|
||||
"b49f7985-6615-4dcf-99be-d1765b6a6fc6"
|
||||
"9050ae1c-eba1-4846-b550-2db1957fee7d",
|
||||
"c4e00391-aeb8-4d32-ac41-ae3242f38a19",
|
||||
"15efa3d7-f804-4d31-ab05-a5524d82b96a",
|
||||
"c4e00391-aeb8-4d32-ac41-ae3242f38a19",
|
||||
"15efa3d7-f804-4d31-ab05-a5524d82b96a"
|
||||
]
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -36,7 +36,7 @@ def test_image_chat_completion_non_streaming(client_with_models, vision_model_id
|
|||
"type": "image",
|
||||
"image": {
|
||||
"url": {
|
||||
"uri": "https://raw.githubusercontent.com/meta-llama/llama-stack/main/tests/api/inference/dog.png"
|
||||
"uri": "https://raw.githubusercontent.com/meta-llama/llama-stack/main/tests/integration/inference/dog.png"
|
||||
},
|
||||
},
|
||||
},
|
||||
|
|
@ -65,7 +65,7 @@ def test_image_chat_completion_streaming(client_with_models, vision_model_id):
|
|||
"type": "image",
|
||||
"image": {
|
||||
"url": {
|
||||
"uri": "https://raw.githubusercontent.com/meta-llama/llama-stack/main/tests/api/inference/dog.png"
|
||||
"uri": "https://raw.githubusercontent.com/meta-llama/llama-stack/main/tests/integration/inference/dog.png"
|
||||
},
|
||||
},
|
||||
},
|
||||
|
|
|
|||
5
tests/integration/providers/__init__.py
Normal file
5
tests/integration/providers/__init__.py
Normal file
|
|
@ -0,0 +1,5 @@
|
|||
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
||||
# All rights reserved.
|
||||
#
|
||||
# This source code is licensed under the terms described in the LICENSE file in
|
||||
# the root directory of this source tree.
|
||||
17
tests/integration/providers/test_providers.py
Normal file
17
tests/integration/providers/test_providers.py
Normal file
|
|
@ -0,0 +1,17 @@
|
|||
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
||||
# All rights reserved.
|
||||
#
|
||||
# This source code is licensed under the terms described in the LICENSE file in
|
||||
# the root directory of this source tree.
|
||||
|
||||
import pytest
|
||||
from llama_stack_client import LlamaStackClient
|
||||
|
||||
from llama_stack import LlamaStackAsLibraryClient
|
||||
|
||||
|
||||
class TestProviders:
|
||||
@pytest.mark.asyncio
|
||||
def test_list(self, llama_stack_client: LlamaStackAsLibraryClient | LlamaStackClient):
|
||||
provider_list = llama_stack_client.providers.list()
|
||||
assert provider_list is not None
|
||||
|
|
@ -10,6 +10,19 @@ import pytest
|
|||
from ..datasetio.test_datasetio import register_dataset
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def rag_dataset_for_test(llama_stack_client):
|
||||
dataset_id = "test_dataset"
|
||||
register_dataset(llama_stack_client, for_rag=True, dataset_id=dataset_id)
|
||||
yield # This is where the test function will run
|
||||
|
||||
# Teardown - this always runs, even if the test fails
|
||||
try:
|
||||
llama_stack_client.datasets.unregister(dataset_id)
|
||||
except Exception as e:
|
||||
print(f"Warning: Failed to unregister test_dataset: {e}")
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def sample_judge_prompt_template():
|
||||
return "Output a number response in the following format: Score: <number>, where <number> is the number between 0 and 9."
|
||||
|
|
@ -79,9 +92,7 @@ def test_scoring_functions_register(
|
|||
# TODO: add unregister api for scoring functions
|
||||
|
||||
|
||||
def test_scoring_score(llama_stack_client):
|
||||
register_dataset(llama_stack_client, for_rag=True)
|
||||
|
||||
def test_scoring_score(llama_stack_client, rag_dataset_for_test):
|
||||
# scoring individual rows
|
||||
rows = llama_stack_client.datasetio.get_rows_paginated(
|
||||
dataset_id="test_dataset",
|
||||
|
|
@ -115,9 +126,9 @@ def test_scoring_score(llama_stack_client):
|
|||
assert len(response.results[x].score_rows) == 5
|
||||
|
||||
|
||||
def test_scoring_score_with_params_llm_as_judge(llama_stack_client, sample_judge_prompt_template, judge_model_id):
|
||||
register_dataset(llama_stack_client, for_rag=True)
|
||||
|
||||
def test_scoring_score_with_params_llm_as_judge(
|
||||
llama_stack_client, sample_judge_prompt_template, judge_model_id, rag_dataset_for_test
|
||||
):
|
||||
# scoring individual rows
|
||||
rows = llama_stack_client.datasetio.get_rows_paginated(
|
||||
dataset_id="test_dataset",
|
||||
|
|
@ -167,9 +178,8 @@ def test_scoring_score_with_params_llm_as_judge(llama_stack_client, sample_judge
|
|||
],
|
||||
)
|
||||
def test_scoring_score_with_aggregation_functions(
|
||||
llama_stack_client, sample_judge_prompt_template, judge_model_id, provider_id
|
||||
llama_stack_client, sample_judge_prompt_template, judge_model_id, provider_id, rag_dataset_for_test
|
||||
):
|
||||
register_dataset(llama_stack_client, for_rag=True)
|
||||
rows = llama_stack_client.datasetio.get_rows_paginated(
|
||||
dataset_id="test_dataset",
|
||||
rows_in_page=3,
|
||||
|
|
|
|||
50
tests/unit/providers/test_configs.py
Normal file
50
tests/unit/providers/test_configs.py
Normal file
|
|
@ -0,0 +1,50 @@
|
|||
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
||||
# All rights reserved.
|
||||
#
|
||||
# This source code is licensed under the terms described in the LICENSE file in
|
||||
# the root directory of this source tree.
|
||||
|
||||
import pytest
|
||||
from pydantic import BaseModel
|
||||
|
||||
from llama_stack.distribution.distribution import get_provider_registry, providable_apis
|
||||
from llama_stack.distribution.utils.dynamic import instantiate_class_type
|
||||
|
||||
|
||||
class TestProviderConfigurations:
|
||||
"""Test suite for testing provider configurations across all API types."""
|
||||
|
||||
def test_all_api_providers_exist(self):
|
||||
provider_registry = get_provider_registry()
|
||||
for api in providable_apis():
|
||||
providers = provider_registry.get(api, {})
|
||||
assert providers, f"No providers found for API type: {api}"
|
||||
|
||||
@pytest.mark.parametrize("api", providable_apis())
|
||||
def test_api_providers(self, api):
|
||||
provider_registry = get_provider_registry()
|
||||
providers = provider_registry.get(api, {})
|
||||
assert providers, f"No providers found for API type: {api}"
|
||||
|
||||
failures = []
|
||||
for provider_type, provider_spec in providers.items():
|
||||
try:
|
||||
self._verify_provider_config(provider_type, provider_spec)
|
||||
except Exception as e:
|
||||
failures.append(f"Failed to verify {provider_type} config: {str(e)}")
|
||||
|
||||
if failures:
|
||||
pytest.fail("\n".join(failures))
|
||||
|
||||
def _verify_provider_config(self, provider_type, provider_spec):
|
||||
"""Helper method to verify a single provider configuration."""
|
||||
# Get the config class
|
||||
config_class_name = provider_spec.config_class
|
||||
config_type = instantiate_class_type(config_class_name)
|
||||
|
||||
assert issubclass(config_type, BaseModel), f"{config_class_name} is not a subclass of BaseModel"
|
||||
|
||||
assert hasattr(config_type, "sample_run_config"), f"{config_class_name} does not have sample_run_config method"
|
||||
|
||||
sample_config = config_type.sample_run_config(__distro_dir__="foobarbaz")
|
||||
assert isinstance(sample_config, dict), f"{config_class_name}.sample_run_config() did not return a dict"
|
||||
Loading…
Add table
Add a link
Reference in a new issue