llama-stack/llama_stack/distribution/resolver.py
Botao Chen aeb76390fc
[1/n] torchtune <> llama-stack integration skeleton (#540)
### Context 
This is the 1st of series PRs that integrate torchtune with llama-stack
as meta reference post-training implementation. For MVP, we will focus
on single device LoRA SFT.

Though this PR is still WIP, we want to get early feedback on the high
level design of this skeleton while still working on several details

### Scope
To limit the scope of this PR, we focus on the skeleton of the
implementation.

**What are included?**
- refine the post-training SFT apis
- skeleton of supervised_fine_tune implementation. We verified that we
can call the supervised_fine_tune API successfully from llama stack
client SDK (client side PR:
https://github.com/meta-llama/llama-stack-client-python/pull/51)
- a very basic single device LoRA training recipe based on torchtune
core components
- parity check with torchtune library and post training api unit test

**What are not includes?**
- implementation of other job management, get training artifacts apis
(separate PR)
- refactor the meta reference inference logic to support eval on
finetuned model (separate PR)
- several necessary functionality in the training recipe such as
logging, validation etc (separate PR)
- interop with telemetry for tracing and metrics logging, currently
temporarily log to local disk (separate PR)

### Testing
**e2e test**
Although we haven't added detailed testing and numerical parity check
with torchtune yet, we did a simple E2E test from client to server
1. setup server with` llama stack build --template
experimental-post-training --image-type conda` and `llama stack run
experimental-post-training `
2. On client, run `llama-stack-client --endpoint
http://devgpu018.nha2.facebook.com:5000 post_training
supervised_fine_tune`
3. Training finishes successfully. On server side, get the finetune
checkpoints under output dir. On client side, get the job uuid

server 
<img width="1110" alt="Screenshot 2024-12-02 at 5 52 32 PM"
src="https://github.com/user-attachments/assets/b548eb90-7a9b-4edc-a858-ee237cc4361d">

client 
<img width="807" alt="Screenshot 2024-12-02 at 5 52 37 PM"
src="https://github.com/user-attachments/assets/1138ffa8-4698-40fa-b190-3d7b99646838">

**parity check**
torchtune dataloader output and llama-stack post training dataloader
output are same
<img width="1116" alt="Screenshot 2024-12-04 at 8 18 46 PM"
src="https://github.com/user-attachments/assets/5e295cdc-4c24-4ea6-82c0-ca96ef1bd6ee">

torchtune LoRA SFT and llama-stack post training LoRA SFT on alpaca
dataset with llama3.2 3B instruct model are numerical match

<img width="860" alt="Screenshot 2024-12-04 at 8 17 01 PM"
src="https://github.com/user-attachments/assets/c05cf0a8-c674-4d2e-9f0a-c5d01b2dca99">

<img width="1049" alt="Screenshot 2024-12-04 at 8 17 06 PM"
src="https://github.com/user-attachments/assets/b911d4e2-e7b1-41a9-b62c-d75529b6d443">

**unit test ** 
![Uploading Screenshot 2024-12-09 at 1.35.10 PM.png…]()
2024-12-13 11:05:35 -08:00

389 lines
13 KiB
Python

# 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 importlib
import inspect
from typing import Any, Dict, List, Set
from llama_stack.providers.datatypes import * # noqa: F403
from llama_stack.distribution.datatypes import * # noqa: F403
import logging
from llama_stack.apis.agents import Agents
from llama_stack.apis.datasetio import DatasetIO
from llama_stack.apis.datasets import Datasets
from llama_stack.apis.eval import Eval
from llama_stack.apis.eval_tasks import EvalTasks
from llama_stack.apis.inference import Inference
from llama_stack.apis.inspect import Inspect
from llama_stack.apis.memory import Memory
from llama_stack.apis.memory_banks import MemoryBanks
from llama_stack.apis.models import Models
from llama_stack.apis.post_training import PostTraining
from llama_stack.apis.safety import Safety
from llama_stack.apis.scoring import Scoring
from llama_stack.apis.scoring_functions import ScoringFunctions
from llama_stack.apis.shields import Shields
from llama_stack.apis.telemetry import Telemetry
from llama_stack.distribution.client import get_client_impl
from llama_stack.distribution.distribution import builtin_automatically_routed_apis
from llama_stack.distribution.store import DistributionRegistry
from llama_stack.distribution.utils.dynamic import instantiate_class_type
log = logging.getLogger(__name__)
class InvalidProviderError(Exception):
pass
def api_protocol_map() -> Dict[Api, Any]:
return {
Api.agents: Agents,
Api.inference: Inference,
Api.inspect: Inspect,
Api.memory: Memory,
Api.memory_banks: MemoryBanks,
Api.models: Models,
Api.safety: Safety,
Api.shields: Shields,
Api.telemetry: Telemetry,
Api.datasetio: DatasetIO,
Api.datasets: Datasets,
Api.scoring: Scoring,
Api.scoring_functions: ScoringFunctions,
Api.eval: Eval,
Api.eval_tasks: EvalTasks,
Api.post_training: PostTraining,
}
def additional_protocols_map() -> Dict[Api, Any]:
return {
Api.inference: (ModelsProtocolPrivate, Models, Api.models),
Api.memory: (MemoryBanksProtocolPrivate, MemoryBanks, Api.memory_banks),
Api.safety: (ShieldsProtocolPrivate, Shields, Api.shields),
Api.datasetio: (DatasetsProtocolPrivate, Datasets, Api.datasets),
Api.scoring: (
ScoringFunctionsProtocolPrivate,
ScoringFunctions,
Api.scoring_functions,
),
Api.eval: (EvalTasksProtocolPrivate, EvalTasks, Api.eval_tasks),
}
# TODO: make all this naming far less atrocious. Provider. ProviderSpec. ProviderWithSpec. WTF!
class ProviderWithSpec(Provider):
spec: ProviderSpec
ProviderRegistry = Dict[Api, Dict[str, ProviderSpec]]
# TODO: this code is not very straightforward to follow and needs one more round of refactoring
async def resolve_impls(
run_config: StackRunConfig,
provider_registry: ProviderRegistry,
dist_registry: DistributionRegistry,
) -> Dict[Api, Any]:
"""
Does two things:
- flatmaps, sorts and resolves the providers in dependency order
- for each API, produces either a (local, passthrough or router) implementation
"""
routing_table_apis = set(
x.routing_table_api for x in builtin_automatically_routed_apis()
)
router_apis = set(x.router_api for x in builtin_automatically_routed_apis())
providers_with_specs = {}
for api_str, providers in run_config.providers.items():
api = Api(api_str)
if api in routing_table_apis:
raise ValueError(
f"Provider for `{api_str}` is automatically provided and cannot be overridden"
)
specs = {}
for provider in providers:
if provider.provider_type not in provider_registry[api]:
raise ValueError(
f"Provider `{provider.provider_type}` is not available for API `{api}`"
)
p = provider_registry[api][provider.provider_type]
if p.deprecation_error:
log.error(p.deprecation_error, "red", attrs=["bold"])
raise InvalidProviderError(p.deprecation_error)
elif p.deprecation_warning:
log.warning(
f"Provider `{provider.provider_type}` for API `{api}` is deprecated and will be removed in a future release: {p.deprecation_warning}",
)
p.deps__ = [a.value for a in p.api_dependencies]
spec = ProviderWithSpec(
spec=p,
**(provider.model_dump()),
)
specs[provider.provider_id] = spec
key = api_str if api not in router_apis else f"inner-{api_str}"
providers_with_specs[key] = specs
apis_to_serve = run_config.apis or set(
list(providers_with_specs.keys())
+ [x.value for x in routing_table_apis]
+ [x.value for x in router_apis]
)
for info in builtin_automatically_routed_apis():
if info.router_api.value not in apis_to_serve:
continue
providers_with_specs[info.routing_table_api.value] = {
"__builtin__": ProviderWithSpec(
provider_id="__routing_table__",
provider_type="__routing_table__",
config={},
spec=RoutingTableProviderSpec(
api=info.routing_table_api,
router_api=info.router_api,
module="llama_stack.distribution.routers",
api_dependencies=[],
deps__=([f"inner-{info.router_api.value}"]),
),
)
}
providers_with_specs[info.router_api.value] = {
"__builtin__": ProviderWithSpec(
provider_id="__autorouted__",
provider_type="__autorouted__",
config={},
spec=AutoRoutedProviderSpec(
api=info.router_api,
module="llama_stack.distribution.routers",
routing_table_api=info.routing_table_api,
api_dependencies=[info.routing_table_api],
deps__=([info.routing_table_api.value]),
),
)
}
sorted_providers = topological_sort(
{k: v.values() for k, v in providers_with_specs.items()}
)
apis = [x[1].spec.api for x in sorted_providers]
sorted_providers.append(
(
"inspect",
ProviderWithSpec(
provider_id="__builtin__",
provider_type="__builtin__",
config={
"run_config": run_config.dict(),
},
spec=InlineProviderSpec(
api=Api.inspect,
provider_type="__builtin__",
config_class="llama_stack.distribution.inspect.DistributionInspectConfig",
module="llama_stack.distribution.inspect",
api_dependencies=apis,
deps__=([x.value for x in apis]),
),
),
)
)
log.info(f"Resolved {len(sorted_providers)} providers")
for api_str, provider in sorted_providers:
log.info(f" {api_str} => {provider.provider_id}")
log.info("")
impls = {}
inner_impls_by_provider_id = {f"inner-{x.value}": {} for x in router_apis}
for api_str, provider in sorted_providers:
deps = {a: impls[a] for a in provider.spec.api_dependencies}
inner_impls = {}
if isinstance(provider.spec, RoutingTableProviderSpec):
inner_impls = inner_impls_by_provider_id[
f"inner-{provider.spec.router_api.value}"
]
impl = await instantiate_provider(
provider,
deps,
inner_impls,
dist_registry,
)
# TODO: ugh slightly redesign this shady looking code
if "inner-" in api_str:
inner_impls_by_provider_id[api_str][provider.provider_id] = impl
else:
api = Api(api_str)
impls[api] = impl
return impls
def topological_sort(
providers_with_specs: Dict[str, List[ProviderWithSpec]],
) -> List[ProviderWithSpec]:
def dfs(kv, visited: Set[str], stack: List[str]):
api_str, providers = kv
visited.add(api_str)
deps = []
for provider in providers:
for dep in provider.spec.deps__:
deps.append(dep)
for dep in deps:
if dep not in visited:
dfs((dep, providers_with_specs[dep]), visited, stack)
stack.append(api_str)
visited = set()
stack = []
for api_str, providers in providers_with_specs.items():
if api_str not in visited:
dfs((api_str, providers), visited, stack)
flattened = []
for api_str in stack:
for provider in providers_with_specs[api_str]:
flattened.append((api_str, provider))
return flattened
# returns a class implementing the protocol corresponding to the Api
async def instantiate_provider(
provider: ProviderWithSpec,
deps: Dict[str, Any],
inner_impls: Dict[str, Any],
dist_registry: DistributionRegistry,
):
protocols = api_protocol_map()
additional_protocols = additional_protocols_map()
provider_spec = provider.spec
module = importlib.import_module(provider_spec.module)
args = []
if isinstance(provider_spec, RemoteProviderSpec):
config_type = instantiate_class_type(provider_spec.config_class)
config = config_type(**provider.config)
method = "get_adapter_impl"
args = [config, deps]
elif isinstance(provider_spec, AutoRoutedProviderSpec):
method = "get_auto_router_impl"
config = None
args = [provider_spec.api, deps[provider_spec.routing_table_api], deps]
elif isinstance(provider_spec, RoutingTableProviderSpec):
method = "get_routing_table_impl"
config = None
args = [provider_spec.api, inner_impls, deps, dist_registry]
else:
method = "get_provider_impl"
config_type = instantiate_class_type(provider_spec.config_class)
config = config_type(**provider.config)
args = [config, deps]
fn = getattr(module, method)
impl = await fn(*args)
impl.__provider_id__ = provider.provider_id
impl.__provider_spec__ = provider_spec
impl.__provider_config__ = config
check_protocol_compliance(impl, protocols[provider_spec.api])
if (
not isinstance(provider_spec, AutoRoutedProviderSpec)
and provider_spec.api in additional_protocols
):
additional_api, _, _ = additional_protocols[provider_spec.api]
check_protocol_compliance(impl, additional_api)
return impl
def check_protocol_compliance(obj: Any, protocol: Any) -> None:
missing_methods = []
mro = type(obj).__mro__
for name, value in inspect.getmembers(protocol):
if inspect.isfunction(value) and hasattr(value, "__webmethod__"):
if not hasattr(obj, name):
missing_methods.append((name, "missing"))
elif not callable(getattr(obj, name)):
missing_methods.append((name, "not_callable"))
else:
# Check if the method signatures are compatible
obj_method = getattr(obj, name)
proto_sig = inspect.signature(value)
obj_sig = inspect.signature(obj_method)
proto_params = set(proto_sig.parameters)
proto_params.discard("self")
obj_params = set(obj_sig.parameters)
obj_params.discard("self")
if not (proto_params <= obj_params):
log.error(
f"Method {name} incompatible proto: {proto_params} vs. obj: {obj_params}"
)
missing_methods.append((name, "signature_mismatch"))
else:
# Check if the method is actually implemented in the class
method_owner = next(
(cls for cls in mro if name in cls.__dict__), None
)
if (
method_owner is None
or method_owner.__name__ == protocol.__name__
):
missing_methods.append((name, "not_actually_implemented"))
if missing_methods:
raise ValueError(
f"Provider `{obj.__provider_id__} ({obj.__provider_spec__.api})` does not implement the following methods:\n{missing_methods}"
)
async def resolve_remote_stack_impls(
config: RemoteProviderConfig,
apis: List[str],
) -> Dict[Api, Any]:
protocols = api_protocol_map()
additional_protocols = additional_protocols_map()
impls = {}
for api_str in apis:
api = Api(api_str)
impls[api] = await get_client_impl(
protocols[api],
config,
{},
)
if api in additional_protocols:
_, additional_protocol, additional_api = additional_protocols[api]
impls[additional_api] = await get_client_impl(
additional_protocol,
config,
{},
)
return impls