phoenix-oss/llama-stack
1
0
Fork 0
forked from phoenix-oss/llama-stack-mirror
Code Pull requests Releases Packages 2 Activity Actions

Remove "routing_table" and "routing_key" concepts for the user (#201)

Browse source 
This PR makes several core changes to the developer experience surrounding Llama Stack.

Background: PR #92 introduced the notion of "routing" to the Llama Stack. It introduces three object types: (1) models, (2) shields and (3) memory banks. Each of these objects can be associated with a distinct provider. So you can get model A to be inferenced locally while model B, C can be inference remotely (e.g.)

However, this had a few drawbacks:

you could not address the provider instances -- i.e., if you configured "meta-reference" with a given model, you could not assign an identifier to this instance which you could re-use later.
the above meant that you could not register a "routing_key" (e.g. model) dynamically and say "please use this existing provider I have already configured" for a new model.
the terms "routing_table" and "routing_key" were exposed directly to the user. in my view, this is way too much overhead for a new user (which almost everyone is.) people come to the stack wanting to do ML and encounter a completely unexpected term.
What this PR does: This PR structures the run config with only a single prominent key:

- providers
Providers are instances of configured provider types. Here's an example which shows two instances of the remote::tgi provider which are serving two different models.

providers:
  inference:
  - provider_id: foo
    provider_type: remote::tgi
    config: { ... }
  - provider_id: bar
    provider_type: remote::tgi
    config: { ... }
Secondly, the PR adds dynamic registration of { models | shields | memory_banks } to the API surface. The distribution still acts like a "routing table" (as previously) except that it asks the backing providers for a listing of these objects. For example it asks a TGI or Ollama inference adapter what models it is serving. Only the models that are being actually served can be requested by the user for inference. Otherwise, the Stack server will throw an error.

When dynamically registering these objects, you can use the provider IDs shown above. Info about providers can be obtained using the Api.inspect set of endpoints (/providers, /routes, etc.)

The above examples shows the correspondence between inference providers and models registry items. Things work similarly for the safety <=> shields and memory <=> memory_banks pairs.

Registry: This PR also makes it so that Providers need to implement additional methods for registering and listing objects. For example, each Inference provider is now expected to implement the ModelsProtocolPrivate protocol (naming is not great!) which consists of two methods

register_model
list_models
The goal is to inform the provider that a certain model needs to be supported so the provider can make any relevant backend changes if needed (or throw an error if the model cannot be supported.)

There are many other cleanups included some of which are detailed in a follow-up comment.
This commit is contained in:
Ashwin Bharambe 2024-10-10 10:24:13 -07:00 • committed by GitHub
parent 8c3010553f
commit 6bb57e72a7
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
93 changed files with 4697 additions and 4457 deletions
Download patch file Download diff file Expand all files Collapse all files

219
llama_stack/providers/impls/meta_reference/inference/inference.py
View file

@ -6,15 +6,15 @@
import asyncio
from typing import AsyncIterator, List, Union
from typing import AsyncGenerator, List
from llama_models.sku_list import resolve_model
from llama_models.llama3.api.datatypes import * # noqa: F403
from llama_stack.apis.inference import * # noqa: F403
from llama_stack.distribution.datatypes import RoutableProvider
from llama_stack.providers.utils.inference.augment_messages import (
augment_messages_for_tools,
from llama_stack.providers.datatypes import ModelDef, ModelsProtocolPrivate
from llama_stack.providers.utils.inference.prompt_adapter import (
chat_completion_request_to_messages,
)
from .config import MetaReferenceImplConfig
@ -25,7 +25,7 @@ from .model_parallel import LlamaModelParallelGenerator
SEMAPHORE = asyncio.Semaphore(1)
class MetaReferenceInferenceImpl(Inference, RoutableProvider):
class MetaReferenceInferenceImpl(Inference, ModelsProtocolPrivate):
def __init__(self, config: MetaReferenceImplConfig) -> None:
self.config = config
model = resolve_model(config.model)
@ -35,21 +35,35 @@ class MetaReferenceInferenceImpl(Inference, RoutableProvider):
# verify that the checkpoint actually is for this model lol
async def initialize(self) -> None:
print(f"Loading model `{self.model.descriptor()}`")
self.generator = LlamaModelParallelGenerator(self.config)
self.generator.start()
async def validate_routing_keys(self, routing_keys: List[str]) -> None:
assert (
len(routing_keys) == 1
), f"Only one routing key is supported {routing_keys}"
assert routing_keys[0] == self.config.model
async def register_model(self, model: ModelDef) -> None:
raise ValueError("Dynamic model registration is not supported")
async def list_models(self) -> List[ModelDef]:
return [
ModelDef(
identifier=self.model.descriptor(),
llama_model=self.model.descriptor(),
)
]
async def shutdown(self) -> None:
self.generator.stop()
# hm, when stream=False, we should not be doing SSE :/ which is what the
# top-level server is going to do. make the typing more specific here
async def chat_completion(
def completion(
self,
model: str,
content: InterleavedTextMedia,
sampling_params: Optional[SamplingParams] = SamplingParams(),
stream: Optional[bool] = False,
logprobs: Optional[LogProbConfig] = None,
) -> Union[CompletionResponse, CompletionResponseStreamChunk]:
raise NotImplementedError()
def chat_completion(
self,
model: str,
messages: List[Message],
@ -59,9 +73,10 @@ class MetaReferenceInferenceImpl(Inference, RoutableProvider):
tool_prompt_format: Optional[ToolPromptFormat] = ToolPromptFormat.json,
stream: Optional[bool] = False,
logprobs: Optional[LogProbConfig] = None,
) -> AsyncIterator[
Union[ChatCompletionResponseStreamChunk, ChatCompletionResponse]
]:
) -> AsyncGenerator:
if logprobs:
assert logprobs.top_k == 1, f"Unexpected top_k={logprobs.top_k}"
# wrapper request to make it easier to pass around (internal only, not exposed to API)
request = ChatCompletionRequest(
model=model,
@ -74,7 +89,6 @@ class MetaReferenceInferenceImpl(Inference, RoutableProvider):
logprobs=logprobs,
)
messages = augment_messages_for_tools(request)
model = resolve_model(request.model)
if model is None:
raise RuntimeError(
@ -88,21 +102,74 @@ class MetaReferenceInferenceImpl(Inference, RoutableProvider):
if SEMAPHORE.locked():
raise RuntimeError("Only one concurrent request is supported")
if request.stream:
return self._stream_chat_completion(request)
else:
return self._nonstream_chat_completion(request)
async def _nonstream_chat_completion(
self, request: ChatCompletionRequest
) -> ChatCompletionResponse:
async with SEMAPHORE:
if request.stream:
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.start,
delta="",
)
)
messages = chat_completion_request_to_messages(request)
tokens = []
logprobs = []
stop_reason = None
buffer = ""
for token_result in self.generator.chat_completion(
messages=messages,
temperature=request.sampling_params.temperature,
top_p=request.sampling_params.top_p,
max_gen_len=request.sampling_params.max_tokens,
logprobs=request.logprobs,
tool_prompt_format=request.tool_prompt_format,
):
tokens.append(token_result.token)
if token_result.text == "<|eot_id|>":
stop_reason = StopReason.end_of_turn
elif token_result.text == "<|eom_id|>":
stop_reason = StopReason.end_of_message
if request.logprobs:
assert len(token_result.logprobs) == 1
logprobs.append(
TokenLogProbs(
logprobs_by_token={
token_result.text: token_result.logprobs[0]
}
)
)
if stop_reason is None:
stop_reason = StopReason.out_of_tokens
message = self.generator.formatter.decode_assistant_message(
tokens, stop_reason
)
return ChatCompletionResponse(
completion_message=message,
logprobs=logprobs if request.logprobs else None,
)
async def _stream_chat_completion(
self, request: ChatCompletionRequest
) -> AsyncGenerator:
async with SEMAPHORE:
messages = chat_completion_request_to_messages(request)
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.start,
delta="",
)
)
tokens = []
logprobs = []
stop_reason = None
ipython = False
for token_result in self.generator.chat_completion(
@ -113,10 +180,9 @@ class MetaReferenceInferenceImpl(Inference, RoutableProvider):
logprobs=request.logprobs,
tool_prompt_format=request.tool_prompt_format,
):
buffer += token_result.text
tokens.append(token_result.token)
if not ipython and buffer.startswith("<|python_tag|>"):
if not ipython and token_result.text.startswith("<|python_tag|>"):
ipython = True
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
@ -127,26 +193,6 @@ class MetaReferenceInferenceImpl(Inference, RoutableProvider):
),
)
)
buffer = buffer[len("<|python_tag|>") :]
continue
if not request.stream:
if request.logprobs:
assert (
len(token_result.logprobs) == 1
), "Expected logprob to contain 1 result for the current token"
assert (
request.logprobs.top_k == 1
), "Only top_k=1 is supported for LogProbConfig"
logprobs.append(
TokenLogProbs(
logprobs_by_token={
token_result.text: token_result.logprobs[0]
}
)
)
continue
if token_result.text == "<|eot_id|>":
@ -167,59 +213,68 @@ class MetaReferenceInferenceImpl(Inference, RoutableProvider):
delta = text
if stop_reason is None:
if request.logprobs:
assert len(token_result.logprobs) == 1
logprobs.append(
TokenLogProbs(
logprobs_by_token={
token_result.text: token_result.logprobs[0]
}
)
)
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.progress,
delta=delta,
stop_reason=stop_reason,
logprobs=logprobs if request.logprobs else None,
)
)
if stop_reason is None:
stop_reason = StopReason.out_of_tokens
# TODO(ashwin): parse tool calls separately here and report errors?
# if someone breaks the iteration before coming here we are toast
message = self.generator.formatter.decode_assistant_message(
tokens, stop_reason
)
if request.stream:
parsed_tool_calls = len(message.tool_calls) > 0
if ipython and not parsed_tool_calls:
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.progress,
delta=ToolCallDelta(
content="",
parse_status=ToolCallParseStatus.failure,
),
stop_reason=stop_reason,
)
)
for tool_call in message.tool_calls:
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.progress,
delta=ToolCallDelta(
content=tool_call,
parse_status=ToolCallParseStatus.success,
),
stop_reason=stop_reason,
)
)
parsed_tool_calls = len(message.tool_calls) > 0
if ipython and not parsed_tool_calls:
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.complete,
delta="",
event_type=ChatCompletionResponseEventType.progress,
delta=ToolCallDelta(
content="",
parse_status=ToolCallParseStatus.failure,
),
stop_reason=stop_reason,
)
)
# TODO(ashwin): what else do we need to send out here when everything finishes?
else:
yield ChatCompletionResponse(
completion_message=message,
logprobs=logprobs if request.logprobs else None,
for tool_call in message.tool_calls:
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.progress,
delta=ToolCallDelta(
content=tool_call,
parse_status=ToolCallParseStatus.success,
),
stop_reason=stop_reason,
)
)
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(
event_type=ChatCompletionResponseEventType.complete,
delta="",
stop_reason=stop_reason,
)
)
async def embeddings(
self,
model: str,
contents: List[InterleavedTextMedia],
) -> EmbeddingsResponse:
raise NotImplementedError()