phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-06-27 18:50:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

test: introduce recordable mocks for Agent tests (#1268)

Browse source 
Summary:

Agent tests shouldn't need to run inference and tools calls repeatedly.
This PR introduces a way to record inference/tool calls and reuse them
in subsequent test runs, which makes the tests more reliable and saves
costs.

Test Plan:
Run when there's no recorded calls created (fails):
```
LLAMA_STACK_CONFIG=fireworks pytest -s -v tests/client-sdk/agents/test_agents.py --safety-shield meta-llama/Llama-Guard-3-8B
```

Run when `--record-responses` to record calls:
```
LLAMA_STACK_CONFIG=fireworks pytest -s -v tests/client-sdk/agents/test_agents.py --safety-shield meta-llama/Llama-Guard-3-8B --record-responses
```

Run without `--record-responses` again (succeeds):
```
LLAMA_STACK_CONFIG=fireworks pytest -s -v tests/client-sdk/agents/test_agents.py --safety-shield meta-llama/Llama-Guard-3-8B
```
This commit is contained in:
ehhuang 2025-03-03 14:48:32 -08:00 • committed by GitHub
parent 816fdf289a
commit 386c806c70
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 6893 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

58
tests/client-sdk/agents/test_agents.py
View file

@ -41,8 +41,8 @@ def get_boiling_point(liquid_name: str, celcius: bool = True) -> int:
@pytest.fixture(scope="session")
def agent_config(llama_stack_client, text_model_id):
available_shields = [shield.identifier for shield in llama_stack_client.shields.list()]
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()]
available_shields = available_shields[:1]
agent_config = AgentConfig(
model=text_model_id,
@ -62,8 +62,8 @@ def agent_config(llama_stack_client, text_model_id):
return agent_config
def test_agent_simple(llama_stack_client, agent_config):
agent = Agent(llama_stack_client, agent_config)
def test_agent_simple(llama_stack_client_with_mocked_inference, agent_config):
agent = Agent(llama_stack_client_with_mocked_inference, agent_config)
session_id = agent.create_session(f"test-session-{uuid4()}")
simple_hello = agent.create_turn(
@ -100,7 +100,7 @@ def test_agent_simple(llama_stack_client, agent_config):
assert "I can't" in logs_str
def test_tool_config(llama_stack_client, agent_config):
def test_tool_config(llama_stack_client_with_mocked_inference, agent_config):
common_params = dict(
model="meta-llama/Llama-3.2-3B-Instruct",
instructions="You are a helpful assistant",
@ -156,14 +156,14 @@ def test_tool_config(llama_stack_client, agent_config):
Server__AgentConfig(**agent_config)
def test_builtin_tool_web_search(llama_stack_client, agent_config):
def test_builtin_tool_web_search(llama_stack_client_with_mocked_inference, agent_config):
agent_config = {
**agent_config,
"toolgroups": [
"builtin::websearch",
],
}
agent = Agent(llama_stack_client, agent_config)
agent = Agent(llama_stack_client_with_mocked_inference, agent_config)
session_id = agent.create_session(f"test-session-{uuid4()}")
response = agent.create_turn(
@ -186,14 +186,14 @@ def test_builtin_tool_web_search(llama_stack_client, agent_config):
assert "No Violation" in logs_str
def test_builtin_tool_code_execution(llama_stack_client, agent_config):
def test_builtin_tool_code_execution(llama_stack_client_with_mocked_inference, agent_config):
agent_config = {
**agent_config,
"toolgroups": [
"builtin::code_interpreter",
],
}
agent = Agent(llama_stack_client, agent_config)
agent = Agent(llama_stack_client_with_mocked_inference, agent_config)
session_id = agent.create_session(f"test-session-{uuid4()}")
response = agent.create_turn(
@ -215,7 +215,7 @@ def test_builtin_tool_code_execution(llama_stack_client, agent_config):
# This test must be run in an environment where `bwrap` is available. If you are running against a
# server, this means the _server_ must have `bwrap` available. If you are using library client, then
# you must have `bwrap` available in test's environment.
def test_code_interpreter_for_attachments(llama_stack_client, agent_config):
def test_code_interpreter_for_attachments(llama_stack_client_with_mocked_inference, agent_config):
agent_config = {
**agent_config,
"toolgroups": [
@ -223,7 +223,7 @@ def test_code_interpreter_for_attachments(llama_stack_client, agent_config):
],
}
codex_agent = Agent(llama_stack_client, agent_config)
codex_agent = Agent(llama_stack_client_with_mocked_inference, agent_config)
session_id = codex_agent.create_session(f"test-session-{uuid4()}")
inflation_doc = AgentDocument(
content="https://raw.githubusercontent.com/meta-llama/llama-stack-apps/main/examples/resources/inflation.csv",
@ -251,7 +251,7 @@ def test_code_interpreter_for_attachments(llama_stack_client, agent_config):
assert "Tool:code_interpreter" in logs_str
def test_custom_tool(llama_stack_client, agent_config):
def test_custom_tool(llama_stack_client_with_mocked_inference, agent_config):
client_tool = get_boiling_point
agent_config = {
**agent_config,
@ -259,7 +259,7 @@ def test_custom_tool(llama_stack_client, agent_config):
"client_tools": [client_tool.get_tool_definition()],
}
agent = Agent(llama_stack_client, agent_config, client_tools=(client_tool,))
agent = Agent(llama_stack_client_with_mocked_inference, agent_config, client_tools=(client_tool,))
session_id = agent.create_session(f"test-session-{uuid4()}")
response = agent.create_turn(
@ -278,7 +278,7 @@ def test_custom_tool(llama_stack_client, agent_config):
assert "get_boiling_point" in logs_str
def test_custom_tool_infinite_loop(llama_stack_client, agent_config):
def test_custom_tool_infinite_loop(llama_stack_client_with_mocked_inference, agent_config):
client_tool = get_boiling_point
agent_config = {
**agent_config,
@ -287,7 +287,7 @@ def test_custom_tool_infinite_loop(llama_stack_client, agent_config):
"max_infer_iters": 5,
}
agent = Agent(llama_stack_client, agent_config, client_tools=(client_tool,))
agent = Agent(llama_stack_client_with_mocked_inference, agent_config, client_tools=(client_tool,))
session_id = agent.create_session(f"test-session-{uuid4()}")
response = agent.create_turn(
@ -305,7 +305,7 @@ def test_custom_tool_infinite_loop(llama_stack_client, agent_config):
assert num_tool_calls <= 5
def test_tool_choice(llama_stack_client, agent_config):
def test_tool_choice(llama_stack_client_with_mocked_inference, agent_config):
def run_agent(tool_choice):
client_tool = get_boiling_point
@ -315,7 +315,7 @@ def test_tool_choice(llama_stack_client, agent_config):
"client_tools": [client_tool.get_tool_definition()],
}
agent = Agent(llama_stack_client, test_agent_config, client_tools=(client_tool,))
agent = Agent(llama_stack_client_with_mocked_inference, test_agent_config, client_tools=(client_tool,))
session_id = agent.create_session(f"test-session-{uuid4()}")
response = agent.create_turn(
@ -342,7 +342,7 @@ def test_tool_choice(llama_stack_client, agent_config):
@pytest.mark.parametrize("rag_tool_name", ["builtin::rag/knowledge_search", "builtin::rag"])
def test_rag_agent(llama_stack_client, agent_config, rag_tool_name):
def test_rag_agent(llama_stack_client_with_mocked_inference, agent_config, rag_tool_name):
urls = ["chat.rst", "llama3.rst", "memory_optimizations.rst", "lora_finetune.rst"]
documents = [
Document(
@ -354,12 +354,12 @@ def test_rag_agent(llama_stack_client, agent_config, rag_tool_name):
for i, url in enumerate(urls)
]
vector_db_id = f"test-vector-db-{uuid4()}"
llama_stack_client.vector_dbs.register(
llama_stack_client_with_mocked_inference.vector_dbs.register(
vector_db_id=vector_db_id,
embedding_model="all-MiniLM-L6-v2",
embedding_dimension=384,
)
llama_stack_client.tool_runtime.rag_tool.insert(
llama_stack_client_with_mocked_inference.tool_runtime.rag_tool.insert(
documents=documents,
vector_db_id=vector_db_id,
# small chunks help to get specific info out of the docs
@ -376,7 +376,7 @@ def test_rag_agent(llama_stack_client, agent_config, rag_tool_name):
)
],
}
rag_agent = Agent(llama_stack_client, agent_config)
rag_agent = Agent(llama_stack_client_with_mocked_inference, agent_config)
session_id = rag_agent.create_session(f"test-session-{uuid4()}")
user_prompts = [
(
@ -401,7 +401,7 @@ def test_rag_agent(llama_stack_client, agent_config, rag_tool_name):
assert expected_kw in response.output_message.content.lower()
def test_rag_agent_with_attachments(llama_stack_client, agent_config):
def test_rag_agent_with_attachments(llama_stack_client_with_mocked_inference, agent_config):
urls = ["chat.rst", "llama3.rst", "memory_optimizations.rst", "lora_finetune.rst"]
documents = [
Document(
@ -423,7 +423,7 @@ def test_rag_agent_with_attachments(llama_stack_client, agent_config):
)
],
}
rag_agent = Agent(llama_stack_client, agent_config)
rag_agent = Agent(llama_stack_client_with_mocked_inference, agent_config)
session_id = rag_agent.create_session(f"test-session-{uuid4()}")
user_prompts = [
(
@ -462,7 +462,7 @@ def test_rag_agent_with_attachments(llama_stack_client, agent_config):
assert "lora" in response.output_message.content.lower()
def test_rag_and_code_agent(llama_stack_client, agent_config):
def test_rag_and_code_agent(llama_stack_client_with_mocked_inference, agent_config):
documents = []
documents.append(
Document(
@ -484,12 +484,12 @@ def test_rag_and_code_agent(llama_stack_client, agent_config):
)
)
vector_db_id = f"test-vector-db-{uuid4()}"
llama_stack_client.vector_dbs.register(
llama_stack_client_with_mocked_inference.vector_dbs.register(
vector_db_id=vector_db_id,
embedding_model="all-MiniLM-L6-v2",
embedding_dimension=384,
)
llama_stack_client.tool_runtime.rag_tool.insert(
llama_stack_client_with_mocked_inference.tool_runtime.rag_tool.insert(
documents=documents,
vector_db_id=vector_db_id,
chunk_size_in_tokens=128,
@ -504,7 +504,7 @@ def test_rag_and_code_agent(llama_stack_client, agent_config):
"builtin::code_interpreter",
],
}
agent = Agent(llama_stack_client, agent_config)
agent = Agent(llama_stack_client_with_mocked_inference, agent_config)
inflation_doc = Document(
document_id="test_csv",
content="https://raw.githubusercontent.com/meta-llama/llama-stack-apps/main/examples/resources/inflation.csv",
@ -546,7 +546,7 @@ def test_rag_and_code_agent(llama_stack_client, agent_config):
assert expected_kw in response.output_message.content.lower()
def test_create_turn_response(llama_stack_client, agent_config):
def test_create_turn_response(llama_stack_client_with_mocked_inference, agent_config):
client_tool = get_boiling_point
agent_config = {
**agent_config,
@ -555,7 +555,7 @@ def test_create_turn_response(llama_stack_client, agent_config):
"client_tools": [client_tool.get_tool_definition()],
}
agent = Agent(llama_stack_client, agent_config, client_tools=(client_tool,))
agent = Agent(llama_stack_client_with_mocked_inference, agent_config, client_tools=(client_tool,))
session_id = agent.create_session(f"test-session-{uuid4()}")
response = agent.create_turn(

66
tests/client-sdk/conftest.py
View file

@ -3,13 +3,18 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import copy
import logging
import os
from pathlib import Path
import pytest
from fixtures.recordable_mock import RecordableMock
from llama_stack_client import LlamaStackClient
from report import Report
from llama_stack import LlamaStackAsLibraryClient
from llama_stack.apis.datatypes import Api
from llama_stack.providers.tests.env import get_env_or_fail
@ -66,6 +71,12 @@ def pytest_addoption(parser):
default=384,
help="Output dimensionality of the embedding model to use for testing",
)
parser.addoption(
"--record-responses",
action="store_true",
default=False,
help="Record new API responses instead of using cached ones.",
)
@pytest.fixture(scope="session")
@ -101,6 +112,61 @@ def llama_stack_client(provider_data, text_model_id):
return client
@pytest.fixture(scope="session")
def llama_stack_client_with_mocked_inference(llama_stack_client, request):
"""
Returns a client with mocked inference APIs and tool runtime APIs that use recorded responses by default.
If --record-responses is passed, it will call the real APIs and record the responses.
"""
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"
)
return llama_stack_client
record_responses = request.config.getoption("--record-responses")
cache_dir = Path(__file__).parent / "fixtures" / "recorded_responses"
# Create a shallow copy of the client to avoid modifying the original
client = copy.copy(llama_stack_client)
# Get the inference API used by the agents implementation
agents_impl = client.async_client.impls[Api.agents]
original_inference = agents_impl.inference_api
# Create a new inference object with the same attributes
inference_mock = copy.copy(original_inference)
# Replace the methods with recordable mocks
inference_mock.chat_completion = RecordableMock(
original_inference.chat_completion, cache_dir, "chat_completion", record=record_responses
)
inference_mock.completion = RecordableMock(
original_inference.completion, cache_dir, "text_completion", record=record_responses
)
inference_mock.embeddings = RecordableMock(
original_inference.embeddings, cache_dir, "embeddings", record=record_responses
)
# Replace the inference API in the agents implementation
agents_impl.inference_api = inference_mock
original_tool_runtime_api = agents_impl.tool_runtime_api
tool_runtime_mock = copy.copy(original_tool_runtime_api)
# Replace the methods with recordable mocks
tool_runtime_mock.invoke_tool = RecordableMock(
original_tool_runtime_api.invoke_tool, cache_dir, "invoke_tool", record=record_responses
)
agents_impl.tool_runtime_api = tool_runtime_mock
# Also update the client.inference for consistency
client.inference = inference_mock
return client
@pytest.fixture(scope="session")
def inference_provider_type(llama_stack_client):
providers = llama_stack_client.providers.list()

208
tests/client-sdk/fixtures/recordable_mock.py Normal file
View file

@ -0,0 +1,208 @@
# 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 json
import os
import pickle
import re
from pathlib import Path
class RecordableMock:
"""A mock that can record and replay API responses."""
def __init__(self, real_func, cache_dir, func_name, record=False):
self.real_func = real_func
self.pickle_path = Path(cache_dir) / f"{func_name}.pickle"
self.json_path = Path(cache_dir) / f"{func_name}.json"
self.record = record
self.cache = {}
# Load existing cache if available and not recording
if self.pickle_path.exists():
try:
with open(self.pickle_path, "rb") as f:
self.cache = pickle.load(f)
except Exception as e:
print(f"Error loading cache from {self.pickle_path}: {e}")
async def __call__(self, *args, **kwargs):
"""
Returns a coroutine that when awaited returns the result or an async generator,
matching the behavior of the original function.
"""
# Create a cache key from the arguments
key = self._create_cache_key(args, kwargs)
if self.record:
# In record mode, always call the real function
real_result = self.real_func(*args, **kwargs)
# If it's a coroutine, we need to create a wrapper coroutine
if hasattr(real_result, "__await__"):
# Define a coroutine function that will record the result
async def record_coroutine():
try:
# Await the real coroutine
result = await real_result
# Check if the result is an async generator
if hasattr(result, "__aiter__"):
# It's an async generator, so we need to record its chunks
chunks = []
# Create and return a new async generator that records chunks
async def recording_generator():
nonlocal chunks
async for chunk in result:
chunks.append(chunk)
yield chunk
# After all chunks are yielded, save to cache
self.cache[key] = {"type": "generator", "chunks": chunks}
self._save_cache()
return recording_generator()
else:
# It's a regular result, save it to cache
self.cache[key] = {"type": "value", "value": result}
self._save_cache()
return result
except Exception as e:
print(f"Error in recording mode: {e}")
raise
return await record_coroutine()
else:
# It's already an async generator, so we need to record its chunks
async def record_generator():
chunks = []
async for chunk in real_result:
chunks.append(chunk)
yield chunk
# After all chunks are yielded, save to cache
self.cache[key] = {"type": "generator", "chunks": chunks}
self._save_cache()
return record_generator()
elif key not in self.cache:
# In replay mode, if the key is not in the cache, throw an error
raise KeyError(
f"No cached response found for key: {key}\nRun with --record-responses to record this response."
)
else:
# In replay mode with a cached response
cached_data = self.cache[key]
# Check if it's a value or chunks
if cached_data.get("type") == "value":
# It's a regular value
return cached_data["value"]
else:
# It's chunks from an async generator
async def replay_generator():
for chunk in cached_data["chunks"]:
yield chunk
return replay_generator()
def _create_cache_key(self, args, kwargs):
"""Create a hashable key from the function arguments, ignoring auto-generated IDs."""
# Convert args and kwargs to a string representation directly
args_str = str(args)
kwargs_str = str(sorted([(k, kwargs[k]) for k in kwargs]))
# Combine into a single key
key = f"{args_str}_{kwargs_str}"
# Post-process the key with regex to replace IDs with placeholders
# Replace UUIDs and similar patterns
key = re.sub(r"[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}", "<UUID>", key)
# Replace temporary file paths created by tempfile.mkdtemp()
key = re.sub(r"/var/folders/[^,'\"\s]+", "<TEMP_FILE>", key)
return key
def _save_cache(self):
"""Save the cache to disk in both pickle and JSON formats."""
os.makedirs(self.pickle_path.parent, exist_ok=True)
# Save as pickle for exact object preservation
with open(self.pickle_path, "wb") as f:
pickle.dump(self.cache, f)
# Also save as JSON for human readability and diffing
try:
# Create a simplified version of the cache for JSON
json_cache = {}
for key, value in self.cache.items():
if value.get("type") == "generator":
# For generators, create a simplified representation of each chunk
chunks = []
for chunk in value["chunks"]:
chunk_dict = self._object_to_json_safe_dict(chunk)
chunks.append(chunk_dict)
json_cache[key] = {"type": "generator", "chunks": chunks}
else:
# For values, create a simplified representation
val = value["value"]
val_dict = self._object_to_json_safe_dict(val)
json_cache[key] = {"type": "value", "value": val_dict}
# Write the JSON file with pretty formatting
with open(self.json_path, "w") as f:
json.dump(json_cache, f, indent=2, sort_keys=True)
except Exception as e:
print(f"Error saving JSON cache: {e}")
def _object_to_json_safe_dict(self, obj):
"""Convert an object to a JSON-safe dictionary."""
# Handle enum types
if hasattr(obj, "value") and hasattr(obj.__class__, "__members__"):
return {"__enum__": obj.__class__.__name__, "value": obj.value}
# Handle Pydantic models
if hasattr(obj, "model_dump"):
return self._process_dict(obj.model_dump())
elif hasattr(obj, "dict"):
return self._process_dict(obj.dict())
# Handle regular objects with __dict__
try:
return self._process_dict(vars(obj))
except Exception as e:
print(f"Error converting object to JSON-safe dict: {e}")
# If we can't get a dict, convert to string
return str(obj)
def _process_dict(self, d):
"""Process a dictionary to make all values JSON-safe."""
if not isinstance(d, dict):
return d
result = {}
for k, v in d.items():
if isinstance(v, dict):
result[k] = self._process_dict(v)
elif isinstance(v, list):
result[k] = [
self._process_dict(item)
if isinstance(item, dict)
else self._object_to_json_safe_dict(item)
if hasattr(item, "__dict__")
else item
for item in v
]
elif hasattr(v, "value") and hasattr(v.__class__, "__members__"):
# Handle enum
result[k] = {"__enum__": v.__class__.__name__, "value": v.value}
elif hasattr(v, "__dict__"):
# Handle nested objects
result[k] = self._object_to_json_safe_dict(v)
else:
# Basic types
result[k] = v
return result

6324
tests/client-sdk/fixtures/recorded_responses/chat_completion.json Normal file
View file

File diff suppressed because one or more lines are too long

BIN
tests/client-sdk/fixtures/recorded_responses/chat_completion.pickle Normal file
View file

Binary file not shown.

266
tests/client-sdk/fixtures/recorded_responses/invoke_tool.json Normal file
View file

@ -0,0 +1,266 @@
{
"()_[('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": {
"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': 'def is_prime(n):\\n if n <= 1:\\n return False\\n if n <= 3:\\n return True\\n if n % 2 == 0 or n % 3 == 0:\\n return False\\n i = 5\\n while i * i <= n:\\n if n % i == 0 or n % (i + 2) == 0:\\n return False\\n i += 6\\n return True\\n\\ndef get_nth_prime(n):\\n count = 0\\n num = 2\\n while True:\\n if is_prime(num):\\n count += 1\\n if count == n:\\n return num\\n num += 1\\n\\nprint(get_nth_prime(100))'}), ('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\\n# Load data\\ndf = pd.read_csv(\"<TEMP_FILE>\")\\n# Rows\\nprint(\"Number of rows and columns in the data:\", df.shape)\\n# Columns\\nprint(\"Columns of the data are:\", len(df.columns))\\n# Column names\\nprint(\"Columns of the data are:\", df.columns)\\n# Column dtypes\\nprint(\"Datatype of the columns are:\", df.dtypes)'}), ('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\\ndf = pd.read_csv(\"<TEMP_FILE>\")\\nprint(df.head())'}), ('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\\ndf = pd.read_csv(\"<TEMP_FILE>\")\\nprint(df.head())\\nprint(df.info())\\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>', 'query': 'How to use LoRA', '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:cbc88\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:cbc88\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 3:\nDocument_id:8892b\nContent: with training with LoRA quickly,\njust specify any config with ``_lora`` in its name, e.g:\n\n.. code-block:: bash\n\n tune run lora_finetune_single_device --config llama3/8B_lora_single_device\n\n\nThere are two sets of parameters to customize LoRA to suit your needs. Firstly, the parameters which control\nwhich linear layers LoRA should be applied to in the model:\n\n* ``lora_attn_modules: List[str]`` accepts a list of strings specifying which layers of the model to apply\n LoRA to:\n\n * ``q_proj`` applies LoRA to the query projection layer.\n * ``k_proj`` applies LoRA to the key projection layer.\n * ``v_proj`` applies LoRA to the value projection layer.\n * ``output_proj`` applies LoRA to the attention output projection layer.\n\n Whilst adding more layers to be fine-tuned may improve model accuracy,\n this will come at the cost of increased memory usage and reduced training speed.\n\n* ``apply_lora_to_mlp: Bool`` applies LoRA to the MLP in each transformer layer.\n* ``apply_lora_to_output: Bool`` applies LoRA to the model's final output projection.\n This is usually a projection to vocabulary space (e.g. in language models), but\n other modelling tasks may have different projections - classifier models will project\n to the number of classes, for example\n\n.. note::\n\n Models which use tied embeddings (such as Gemma and Qwen2 1.5B and 0.5B) for the\n final output projection do not support ``apply_lora_to_output``.\n\nThese are all specified under the ``model`` flag or config entry, i.e:\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\",\"output_proj\"]\n\n.. code-block:: yaml\n\n model:\n _component_: torchtune.models.llama3.lora_llama3_8b\n apply_lora_to_mlp: True\n model.lora_attn_modules: [\"q_proj\", \"k_proj\", \"v_proj\",\"output_proj\"]\n\nSecondly, parameters which control the scale of the impact of LoRA on the model:\n\n* ``lora_rank: int`` affects the scale of\n",
"type": "text"
},
{
"text": "Result 4:\nDocument_id:cbc88\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 5:\nDocument_id:9dcb7\nContent: ora_finetune_label>`.\nFor more on QLoRA in torchtune, see our :ref:`QLoRA Tutorial <qlora_finetune_label>`.\n\nLet's take a look at how we can fine-tune Llama3-8B-Instruct with LoRA on a single device using torchtune. In this example, we will fine-tune\nfor one epoch on a common instruct dataset for illustrative purposes. The basic command for a single-device LoRA fine-tune is\n\n.. code-block:: bash\n\n tune run lora_finetune_single_device --config llama3/8B_lora_single_device\n\n.. note::\n To see a full list of recipes and their corresponding configs, simply run ``tune ls`` from the command line.\n\nWe can also add :ref:`command-line overrides <cli_override>` as needed, e.g.\n\n.. code-block:: bash\n\n tune run lora_finetune_single_device --config llama3/8B_lora_single_device \\\n checkpointer.checkpoint_dir=<checkpoint_dir> \\\n tokenizer.path=<checkpoint_dir>/tokenizer.model \\\n checkpointer.output_dir=<checkpoint_dir>\n\nThis will load the Llama3-8B-Instruct checkpoint and tokenizer from ``<checkpoint_dir>`` used in the :ref:`tune download <tune_download_label>` command above,\nthen save a final checkpoint in the same directory following the original format. For more details on the\ncheckpoint formats supported in torchtune, see our :ref:`checkpointing deep-dive <understand_checkpointer>`.\n\n.. note::\n To see the full set of configurable parameters for this (and other) configs we can use :ref:`tune cp <tune_cp_cli_label>` to copy (and modify)\n the default config. :ref:`tune cp <tune_cp_cli_label>` can be used with recipe scripts too, in case you want to make more custom changes\n that cannot be achieved by directly modifying existing configurable parameters. For more on :ref:`tune cp <tune_cp_cli_label>` see the section on\n :ref:`modifying configs <tune_cp_label>` in our \":ref:`finetune_llama_label`\" tutorial.\n\nOnce training is complete, the model checkpoints will be saved and their locations will be logged. For\nLoRA fine-tuning, the final checkpoint will contain the merged weights, and a copy of just the (much smaller) LoRA weights\nwill\n",
"type": "text"
},
{
"text": "END of knowledge_search tool results.\n",
"type": "text"
}
],
"error_code": null,
"error_message": null,
"metadata": {
"document_ids": [
"cbc884b1-9d88-4d5c-aff4-7a4b3a56618c",
"cbc884b1-9d88-4d5c-aff4-7a4b3a56618c",
"8892b092-6394-471e-b143-a23c6cc374f8",
"cbc884b1-9d88-4d5c-aff4-7a4b3a56618c",
"9dcb747d-0627-40cc-a23c-0bee2b6b05af"
]
}
}
},
"()_[('kwargs', {'session_id': '<UUID>', 'query': 'Llama3-8B attention type', 'vector_db_ids': ['test-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:num-1\nContent: 3 <https://llama.meta.com/llama3>`_ is a new family of models released by Meta AI that improves upon the performance of the Llama2 family\nof models across a `range of different benchmarks <https://huggingface.co/meta-llama/Meta-Llama-3-8B#base-pretrained-models>`_.\nCurrently there are two different sizes of Meta Llama 3: 8B and 70B. In this tutorial we will focus on the 8B size model.\nThere are a few main changes between Llama2-7B and Llama3-8B models:\n\n- Llama3-8B uses `grouped-query attention <https://arxiv.org/abs/2305.13245>`_ instead of the standard multi-head attention from Llama2-7B\n- Llama3-8B has a larger vocab size (128,256 instead of 32,000 from Llama2 models)\n- Llama3-8B uses a different tokenizer than Llama2 models (`tiktoken <https://github.com/openai/tiktoken>`_ instead of `sentencepiece <https://github.com/google/sentencepiece>`_)\n- Llama3-\n",
"type": "text"
},
{
"text": "Result 2:\nDocument_id:num-1\nContent: instead of 32,000 from Llama2 models)\n- Llama3-8B uses a different tokenizer than Llama2 models (`tiktoken <https://github.com/openai/tiktoken>`_ instead of `sentencepiece <https://github.com/google/sentencepiece>`_)\n- Llama3-8B uses a larger intermediate dimension in its MLP layers than Llama2-7B\n- Llama3-8B uses a higher base value to calculate theta in its `rotary positional embeddings <https://arxiv.org/abs/2104.09864>`_\n\n|\n\nGetting access to Llama3-8B-Instruct\n------------------------------------\n\nFor this tutorial, we will be using the instruction-tuned version of Llama3-8B. First, let's download the model from Hugging Face. You will need to follow the instructions\non the `official Meta page <https://github.com/meta-llama/llama3/blob/main/README.md>`_ to gain access to the model.\nNext, make sure you grab your Hugging Face token from `here <https://huggingface.co/settings/tokens>`_.\n\n\n.. code-block:: bash\n\n tune download meta-llama/Meta-Llama-3\n",
"type": "text"
},
{
"text": "Result 3:\nDocument_id:num-0\nContent: :`download Llama3 Instruct weights <llama3_label>`\n\n\nTemplate changes from Llama2 to Llama3\n--------------------------------------\n\nThe Llama2 chat model requires a specific template when prompting the pre-trained\nmodel. Since the chat model was pretrained with this prompt template, if you want to run\ninference on the model, you'll need to use the same template for optimal performance\non chat data. Otherwise, the model will just perform standard text completion, which\nmay or may not align with your intended use case.\n\nFrom the `official Llama2 prompt\ntemplate guide <https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-2>`_\nfor the Llama2 chat model, we can see that special tags are added:\n\n.. code-block:: text\n\n <s>[INST] <<SYS>>\n You are a helpful, respectful, and honest assistant.\n <</SYS>>\n\n Hi! I am a human. [/INST] Hello there! Nice to meet you! I'm Meta AI, your friendly AI assistant </s>\n\nLlama3 Instruct `overhauled <https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3>`\n",
"type": "text"
},
{
"text": "Result 4:\nDocument_id:num-0\nContent: 'm Meta AI, your friendly AI assistant<|eot_id|>\n\nThe tags are entirely different, and they are actually encoded differently than in\nLlama2. Let's walk through tokenizing an example with the Llama2 template and the\nLlama3 template to understand how.\n\n.. note::\n The Llama3 Base model uses a `different prompt template\n <https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3>`_ than Llama3 Instruct\n because it has not yet been instruct tuned and the extra special tokens are untrained. If you\n are running inference on the Llama3 Base model without fine-tuning we recommend the base\n template for optimal performance. Generally, for instruct and chat data, we recommend using\n Llama3 Instruct with its prompt template. The rest of this tutorial assumes you are using\n Llama3 Instruct.\n\n.. _prompt_template_vs_special_tokens:\n\nTokenizing prompt templates & special tokens\n--------------------------------------------\n\nLet's say I have a sample of a single user-assistant turn accompanied with a system\nprompt:\n\n.. code-block:: python\n\n sample = [\n {\n \"role\": \"system\",\n \"\n",
"type": "text"
},
{
"text": "Result 5:\nDocument_id:num-3\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\n",
"type": "text"
},
{
"text": "END of knowledge_search tool results.\n",
"type": "text"
}
],
"error_code": null,
"error_message": null,
"metadata": {
"document_ids": [
"num-1",
"num-1",
"num-0",
"num-0",
"num-3"
]
}
}
},
"()_[('kwargs', {'session_id': '<UUID>', 'query': 'NBA creation date', '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"
]
}
}
},
"()_[('kwargs', {'session_id': '<UUID>', 'query': 'Perplexity company founding date', '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: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 2:\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": "Result 3:\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": "END of knowledge_search tool results.\n",
"type": "text"
}
],
"error_code": null,
"error_message": null,
"metadata": {
"document_ids": [
"perplexity_wiki",
"perplexity_wiki",
"nba_wiki"
]
}
}
},
"()_[('kwargs', {'session_id': '<UUID>', 'query': 'Torchtune documentation', '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:f4fd3\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:cbc88\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:8892b\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:cbc88\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:8892b\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"
},
{
"text": "END of knowledge_search tool results.\n",
"type": "text"
}
],
"error_code": null,
"error_message": null,
"metadata": {
"document_ids": [
"f4fd30bb-23d3-4ff8-bb8a-846041ae22cf",
"cbc884b1-9d88-4d5c-aff4-7a4b3a56618c",
"8892b092-6394-471e-b143-a23c6cc374f8",
"cbc884b1-9d88-4d5c-aff4-7a4b3a56618c",
"8892b092-6394-471e-b143-a23c6cc374f8"
]
}
}
},
"()_[('kwargs', {'session_id': '<UUID>', 'query': 'current CEO of Meta'}), ('tool_name', 'web_search')]": {
"type": "value",
"value": {
"content": "{\"query\": \"current CEO of Meta\", \"top_k\": [{\"title\": \"Executives - Meta\", \"url\": \"https://about.meta.com/media-gallery/executives/\", \"content\": \"Mark Zuckerberg, Founder, Chairman and Chief Executive Officer Joel Kaplan, Chief Global Affairs Officer Susan Li, Chief Financial Officer Javier Olivan, Chief Operating Officer Chris Cox, Chief Product Officer Andrew \\u2018Boz\\u2019 Bosworth, Chief Technology Officer Jennifer Newstead, Chief Legal Officer Dave Wehner, Chief Strategy Officer Will Cathcart, Head of WhatsApp Naomi Gleit, Head of Product John Hegeman, Chief Revenue Officer Adam Mosseri, Head of Instagram Erin Egan, Chief Privacy Officer, Policy Michel Protti, Chief Privacy Officer, Product Alex Schultz, Chief Marketing Officer and VP of Analytics Tom Alison, Head of Facebook Nicola Mendelsohn, Head of Global Business Group Ahmad Al-Dahle, VP and Head of GenAI at Meta Joelle Pineau, Vice President of AI Research and Head of FAIR at Meta\", \"score\": 0.8190992, \"raw_content\": null}, {\"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\": 0.79099923, \"raw_content\": null}, {\"title\": \"Zuckerberg's political pivot targets Apple, puts Meta staffers on edge\", \"url\": \"https://www.cnbc.com/2025/02/14/zuckerbergs-rightward-policy-shift-hits-meta-staffers-targets-apple.html\", \"content\": \"Meta CEO Mark Zuckerberg's actions to curry favor with the president have rattled employees, but people familiar with his efforts say there's a clear strategy.\", \"score\": 0.77179235, \"raw_content\": null}, {\"title\": \"Meet the Executive CSuite Team of Meta (Facebook) [2025]\", \"url\": \"https://digitaldefynd.com/IQ/meet-the-executive-csuite-team-of-meta-facebook/\", \"content\": \"Harvard University Executive Programs Free Harvard University Courses As a chief financial officer of Meta, Susan Li oversees the firm\\u2019s finance and facilities team to keep track of the company\\u2019s overall financial health. The chief operating officer of Meta, Javier Olivan, oversees the firm\\u2019s business team, infrastructure, and other products. Andrew Bosworth, called Boz, serves as chief technology officer at Meta and is responsible for leading the firm\\u2019s AR/VR organization, Reality Labs. Andrew has also served as engineering director to oversee events, mobile monetization, and feed ads and as VP of ads and business platforms to lead engineering, design, analytics, and product teams. Meta\\u2019s c-suite team comprises experienced and diverse executives, having extensive experience in technology, finance, legal, and all major industries.\", \"score\": 0.7602419, \"raw_content\": null}, {\"title\": \"Meta to spend up to $65 billion this year to power AI goals, Zuckerberg ...\", \"url\": \"https://www.reuters.com/technology/meta-invest-up-65-bln-capital-expenditure-this-year-2025-01-24/\", \"content\": \"Meta Platforms plans to spend as much as $65 billion this year to expand its AI infrastructure, CEO Mark Zuckerberg said on Friday, aiming to bolster the company's position against rivals OpenAI\", \"score\": 0.73914057, \"raw_content\": null}]}",
"error_code": null,
"error_message": null,
"metadata": null
}
}
}

BIN
tests/client-sdk/fixtures/recorded_responses/invoke_tool.pickle Normal file
View file

Binary file not shown.