phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-10-04 04:04:14 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

chore(apis): unpublish deprecated /v1/inference apis

Browse source
This commit is contained in:
Matthew Farrellee 2025-01-09 02:03:04 -05:00
parent 478b4ff1e6
commit 26f4f3fe14
6 changed files with 1286 additions and 3770 deletions
Download patch file Download diff file Expand all files Collapse all files

2319
docs/_static/llama-stack-spec.html vendored
View file

File diff suppressed because it is too large Load diff

1810
docs/_static/llama-stack-spec.yaml vendored
View file

File diff suppressed because it is too large Load diff

5
llama_stack/apis/inference/inference.py
View file

@ -1026,7 +1026,6 @@ class InferenceProvider(Protocol):
model_store: ModelStore | None = None
@webmethod(route="/inference/completion", method="POST")
async def completion(
self,
model_id: str,
@ -1049,7 +1048,6 @@ class InferenceProvider(Protocol):
"""
...
@webmethod(route="/inference/batch-completion", method="POST", experimental=True)
async def batch_completion(
self,
model_id: str,
@ -1070,7 +1068,6 @@ class InferenceProvider(Protocol):
raise NotImplementedError("Batch completion is not implemented")
return # this is so mypy's safe-super rule will consider the method concrete
@webmethod(route="/inference/chat-completion", method="POST")
async def chat_completion(
self,
model_id: str,
@ -1110,7 +1107,6 @@ class InferenceProvider(Protocol):
"""
...
@webmethod(route="/inference/batch-chat-completion", method="POST", experimental=True)
async def batch_chat_completion(
self,
model_id: str,
@ -1135,7 +1131,6 @@ class InferenceProvider(Protocol):
raise NotImplementedError("Batch chat completion is not implemented")
return # this is so mypy's safe-super rule will consider the method concrete
@webmethod(route="/inference/embeddings", method="POST")
async def embeddings(
self,
model_id: str,

76
tests/integration/inference/test_batch_inference.py
View file

@ -1,76 +0,0 @@
# 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 ..test_cases.test_case import TestCase
def skip_if_provider_doesnt_support_batch_inference(client_with_models, model_id):
models = {m.identifier: m for m in client_with_models.models.list()}
models.update({m.provider_resource_id: m for m in client_with_models.models.list()})
provider_id = models[model_id].provider_id
providers = {p.provider_id: p for p in client_with_models.providers.list()}
provider = providers[provider_id]
if provider.provider_type not in ("inline::meta-reference",):
pytest.skip(f"Model {model_id} hosted by {provider.provider_type} doesn't support batch inference")
@pytest.mark.parametrize(
"test_case",
[
"inference:completion:batch_completion",
],
)
def test_batch_completion_non_streaming(client_with_models, text_model_id, test_case):
skip_if_provider_doesnt_support_batch_inference(client_with_models, text_model_id)
tc = TestCase(test_case)
content_batch = tc["contents"]
response = client_with_models.inference.batch_completion(
content_batch=content_batch,
model_id=text_model_id,
sampling_params={
"max_tokens": 50,
},
)
assert len(response.batch) == len(content_batch)
for i, r in enumerate(response.batch):
print(f"response {i}: {r.content}")
assert len(r.content) > 10
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:batch_completion",
],
)
def test_batch_chat_completion_non_streaming(client_with_models, text_model_id, test_case):
skip_if_provider_doesnt_support_batch_inference(client_with_models, text_model_id)
tc = TestCase(test_case)
qa_pairs = tc["qa_pairs"]
message_batch = [
[
{
"role": "user",
"content": qa["question"],
}
]
for qa in qa_pairs
]
response = client_with_models.inference.batch_chat_completion(
messages_batch=message_batch,
model_id=text_model_id,
)
assert len(response.batch) == len(qa_pairs)
for i, r in enumerate(response.batch):
print(f"response {i}: {r.completion_message.content}")
assert len(r.completion_message.content) > 0
assert qa_pairs[i]["answer"].lower() in r.completion_message.content.lower()

303
tests/integration/inference/test_embedding.py
View file

@ -1,303 +0,0 @@
# 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.
#
# Test plan:
#
# Types of input:
# - array of a string
# - array of a image (ImageContentItem, either URL or base64 string)
# - array of a text (TextContentItem)
# Types of output:
# - list of list of floats
# Params:
# - text_truncation
# - absent w/ long text -> error
# - none w/ long text -> error
# - absent w/ short text -> ok
# - none w/ short text -> ok
# - end w/ long text -> ok
# - end w/ short text -> ok
# - start w/ long text -> ok
# - start w/ short text -> ok
# - output_dimension
# - response dimension matches
# - task_type, only for asymmetric models
# - query embedding != passage embedding
# Negative:
# - long string
# - long text
#
# Todo:
# - negative tests
# - empty
# - empty list
# - empty string
# - empty text
# - empty image
# - long
# - large image
# - appropriate combinations
# - batch size
# - many inputs
# - invalid
# - invalid URL
# - invalid base64
#
# Notes:
# - use llama_stack_client fixture
# - use pytest.mark.parametrize when possible
# - no accuracy tests: only check the type of output, not the content
#
import pytest
from llama_stack_client import BadRequestError as LlamaStackBadRequestError
from llama_stack_client.types import EmbeddingsResponse
from llama_stack_client.types.shared.interleaved_content import (
ImageContentItem,
ImageContentItemImage,
ImageContentItemImageURL,
TextContentItem,
)
from openai import BadRequestError as OpenAIBadRequestError
from llama_stack.core.library_client import LlamaStackAsLibraryClient
DUMMY_STRING = "hello"
DUMMY_STRING2 = "world"
DUMMY_LONG_STRING = "NVDA " * 10240
DUMMY_TEXT = TextContentItem(text=DUMMY_STRING, type="text")
DUMMY_TEXT2 = TextContentItem(text=DUMMY_STRING2, type="text")
DUMMY_LONG_TEXT = TextContentItem(text=DUMMY_LONG_STRING, type="text")
# TODO(mf): add a real image URL and base64 string
DUMMY_IMAGE_URL = ImageContentItem(
image=ImageContentItemImage(url=ImageContentItemImageURL(uri="https://example.com/image.jpg")), type="image"
)
DUMMY_IMAGE_BASE64 = ImageContentItem(image=ImageContentItemImage(data="base64string"), type="image")
SUPPORTED_PROVIDERS = {"remote::nvidia"}
MODELS_SUPPORTING_MEDIA = {}
MODELS_SUPPORTING_OUTPUT_DIMENSION = {"nvidia/llama-3.2-nv-embedqa-1b-v2"}
MODELS_REQUIRING_TASK_TYPE = {
"nvidia/llama-3.2-nv-embedqa-1b-v2",
"nvidia/nv-embedqa-e5-v5",
"nvidia/nv-embedqa-mistral-7b-v2",
"snowflake/arctic-embed-l",
}
MODELS_SUPPORTING_TASK_TYPE = MODELS_REQUIRING_TASK_TYPE
def default_task_type(model_id):
"""
Some models require a task type parameter. This provides a default value for
testing those models.
"""
if model_id in MODELS_REQUIRING_TASK_TYPE:
return {"task_type": "query"}
return {}
@pytest.mark.parametrize(
"contents",
[
[DUMMY_STRING, DUMMY_STRING2],
[DUMMY_TEXT, DUMMY_TEXT2],
],
ids=[
"list[string]",
"list[text]",
],
)
def test_embedding_text(llama_stack_client, embedding_model_id, contents, inference_provider_type):
if inference_provider_type not in SUPPORTED_PROVIDERS:
pytest.xfail(f"{inference_provider_type} doesn't support embedding model yet")
response = llama_stack_client.inference.embeddings(
model_id=embedding_model_id, contents=contents, **default_task_type(embedding_model_id)
)
assert isinstance(response, EmbeddingsResponse)
assert len(response.embeddings) == sum(len(content) if isinstance(content, list) else 1 for content in contents)
assert isinstance(response.embeddings[0], list)
assert isinstance(response.embeddings[0][0], float)
@pytest.mark.parametrize(
"contents",
[
[DUMMY_IMAGE_URL, DUMMY_IMAGE_BASE64],
[DUMMY_IMAGE_URL, DUMMY_STRING, DUMMY_IMAGE_BASE64, DUMMY_TEXT],
],
ids=[
"list[url,base64]",
"list[url,string,base64,text]",
],
)
def test_embedding_image(llama_stack_client, embedding_model_id, contents, inference_provider_type):
if inference_provider_type not in SUPPORTED_PROVIDERS:
pytest.xfail(f"{inference_provider_type} doesn't support embedding model yet")
if embedding_model_id not in MODELS_SUPPORTING_MEDIA:
pytest.xfail(f"{embedding_model_id} doesn't support media")
response = llama_stack_client.inference.embeddings(
model_id=embedding_model_id, contents=contents, **default_task_type(embedding_model_id)
)
assert isinstance(response, EmbeddingsResponse)
assert len(response.embeddings) == sum(len(content) if isinstance(content, list) else 1 for content in contents)
assert isinstance(response.embeddings[0], list)
assert isinstance(response.embeddings[0][0], float)
@pytest.mark.parametrize(
"text_truncation",
[
"end",
"start",
],
)
@pytest.mark.parametrize(
"contents",
[
[DUMMY_LONG_TEXT],
[DUMMY_STRING],
],
ids=[
"long",
"short",
],
)
def test_embedding_truncation(
llama_stack_client, embedding_model_id, text_truncation, contents, inference_provider_type
):
if inference_provider_type not in SUPPORTED_PROVIDERS:
pytest.xfail(f"{inference_provider_type} doesn't support embedding model yet")
response = llama_stack_client.inference.embeddings(
model_id=embedding_model_id,
contents=contents,
text_truncation=text_truncation,
**default_task_type(embedding_model_id),
)
assert isinstance(response, EmbeddingsResponse)
assert len(response.embeddings) == 1
assert isinstance(response.embeddings[0], list)
assert isinstance(response.embeddings[0][0], float)
@pytest.mark.parametrize(
"text_truncation",
[
None,
"none",
],
)
@pytest.mark.parametrize(
"contents",
[
[DUMMY_LONG_TEXT],
[DUMMY_LONG_STRING],
],
ids=[
"long-text",
"long-str",
],
)
def test_embedding_truncation_error(
llama_stack_client, embedding_model_id, text_truncation, contents, inference_provider_type
):
if inference_provider_type not in SUPPORTED_PROVIDERS:
pytest.xfail(f"{inference_provider_type} doesn't support embedding model yet")
# Using LlamaStackClient from llama_stack_client will raise llama_stack_client.BadRequestError
# While using LlamaStackAsLibraryClient from llama_stack.distribution.library_client will raise the error that the backend raises
error_type = (
OpenAIBadRequestError
if isinstance(llama_stack_client, LlamaStackAsLibraryClient)
else LlamaStackBadRequestError
)
with pytest.raises(error_type):
llama_stack_client.inference.embeddings(
model_id=embedding_model_id,
contents=[DUMMY_LONG_TEXT],
text_truncation=text_truncation,
**default_task_type(embedding_model_id),
)
def test_embedding_output_dimension(llama_stack_client, embedding_model_id, inference_provider_type):
if inference_provider_type not in SUPPORTED_PROVIDERS:
pytest.xfail(f"{inference_provider_type} doesn't support embedding model yet")
if embedding_model_id not in MODELS_SUPPORTING_OUTPUT_DIMENSION:
pytest.xfail(f"{embedding_model_id} doesn't support output_dimension")
base_response = llama_stack_client.inference.embeddings(
model_id=embedding_model_id, contents=[DUMMY_STRING], **default_task_type(embedding_model_id)
)
test_response = llama_stack_client.inference.embeddings(
model_id=embedding_model_id,
contents=[DUMMY_STRING],
**default_task_type(embedding_model_id),
output_dimension=32,
)
assert len(base_response.embeddings[0]) != len(test_response.embeddings[0])
assert len(test_response.embeddings[0]) == 32
def test_embedding_task_type(llama_stack_client, embedding_model_id, inference_provider_type):
if inference_provider_type not in SUPPORTED_PROVIDERS:
pytest.xfail(f"{inference_provider_type} doesn't support embedding model yet")
if embedding_model_id not in MODELS_SUPPORTING_TASK_TYPE:
pytest.xfail(f"{embedding_model_id} doesn't support task_type")
query_embedding = llama_stack_client.inference.embeddings(
model_id=embedding_model_id, contents=[DUMMY_STRING], task_type="query"
)
document_embedding = llama_stack_client.inference.embeddings(
model_id=embedding_model_id, contents=[DUMMY_STRING], task_type="document"
)
assert query_embedding.embeddings != document_embedding.embeddings
@pytest.mark.parametrize(
"text_truncation",
[
None,
"none",
"end",
"start",
],
)
def test_embedding_text_truncation(llama_stack_client, embedding_model_id, text_truncation, inference_provider_type):
if inference_provider_type not in SUPPORTED_PROVIDERS:
pytest.xfail(f"{inference_provider_type} doesn't support embedding model yet")
response = llama_stack_client.inference.embeddings(
model_id=embedding_model_id,
contents=[DUMMY_STRING],
text_truncation=text_truncation,
**default_task_type(embedding_model_id),
)
assert isinstance(response, EmbeddingsResponse)
assert len(response.embeddings) == 1
assert isinstance(response.embeddings[0], list)
assert isinstance(response.embeddings[0][0], float)
@pytest.mark.parametrize(
"text_truncation",
[
"NONE",
"END",
"START",
"left",
"right",
],
)
def test_embedding_text_truncation_error(
llama_stack_client, embedding_model_id, text_truncation, inference_provider_type
):
if inference_provider_type not in SUPPORTED_PROVIDERS:
pytest.xfail(f"{inference_provider_type} doesn't support embedding model yet")
error_type = ValueError if isinstance(llama_stack_client, LlamaStackAsLibraryClient) else LlamaStackBadRequestError
with pytest.raises(error_type):
llama_stack_client.inference.embeddings(
model_id=embedding_model_id,
contents=[DUMMY_STRING],
text_truncation=text_truncation,
**default_task_type(embedding_model_id),
)

543
tests/integration/inference/test_text_inference.py
View file

@ -1,543 +0,0 @@
# 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.
from time import sleep
import pytest
from pydantic import BaseModel
from llama_stack.models.llama.sku_list import resolve_model
from ..test_cases.test_case import TestCase
PROVIDER_LOGPROBS_TOP_K = {"remote::together", "remote::fireworks", "remote::vllm"}
def skip_if_model_doesnt_support_completion(client_with_models, model_id):
models = {m.identifier: m for m in client_with_models.models.list()}
models.update({m.provider_resource_id: m for m in client_with_models.models.list()})
provider_id = models[model_id].provider_id
providers = {p.provider_id: p for p in client_with_models.providers.list()}
provider = providers[provider_id]
if (
provider.provider_type
in (
"remote::openai",
"remote::anthropic",
"remote::gemini",
"remote::vertexai",
"remote::groq",
"remote::sambanova",
)
or "openai-compat" in provider.provider_type
):
pytest.skip(f"Model {model_id} hosted by {provider.provider_type} doesn't support completion")
def skip_if_model_doesnt_support_json_schema_structured_output(client_with_models, model_id):
models = {m.identifier: m for m in client_with_models.models.list()}
models.update({m.provider_resource_id: m for m in client_with_models.models.list()})
provider_id = models[model_id].provider_id
providers = {p.provider_id: p for p in client_with_models.providers.list()}
provider = providers[provider_id]
if provider.provider_type in ("remote::sambanova",):
pytest.skip(
f"Model {model_id} hosted by {provider.provider_type} doesn't support json_schema structured output"
)
def get_llama_model(client_with_models, model_id):
models = {}
for m in client_with_models.models.list():
models[m.identifier] = m
models[m.provider_resource_id] = m
assert model_id in models, f"Model {model_id} not found"
model = models[model_id]
ids = (model.identifier, model.provider_resource_id)
for mid in ids:
if resolve_model(mid):
return mid
return model.metadata.get("llama_model", None)
@pytest.mark.parametrize(
"test_case",
[
"inference:completion:sanity",
],
)
def test_text_completion_non_streaming(client_with_models, text_model_id, test_case):
skip_if_model_doesnt_support_completion(client_with_models, text_model_id)
tc = TestCase(test_case)
response = client_with_models.inference.completion(
content=tc["content"],
stream=False,
model_id=text_model_id,
sampling_params={
"max_tokens": 50,
},
)
assert len(response.content) > 10
# assert "blue" in response.content.lower().strip()
@pytest.mark.parametrize(
"test_case",
[
"inference:completion:sanity",
],
)
def test_text_completion_streaming(client_with_models, text_model_id, test_case):
skip_if_model_doesnt_support_completion(client_with_models, text_model_id)
tc = TestCase(test_case)
response = client_with_models.inference.completion(
content=tc["content"],
stream=True,
model_id=text_model_id,
sampling_params={
"max_tokens": 50,
},
)
streamed_content = [chunk.delta for chunk in response]
content_str = "".join(streamed_content).lower().strip()
# assert "blue" in content_str
assert len(content_str) > 10
@pytest.mark.parametrize(
"test_case",
[
"inference:completion:stop_sequence",
],
)
def test_text_completion_stop_sequence(client_with_models, text_model_id, inference_provider_type, test_case):
skip_if_model_doesnt_support_completion(client_with_models, text_model_id)
# This is only supported/tested for remote vLLM: https://github.com/meta-llama/llama-stack/issues/1771
if inference_provider_type != "remote::vllm":
pytest.xfail(f"{inference_provider_type} doesn't support 'stop' parameter yet")
tc = TestCase(test_case)
response = client_with_models.inference.completion(
content=tc["content"],
stream=True,
model_id=text_model_id,
sampling_params={
"max_tokens": 50,
"stop": ["1963"],
},
)
streamed_content = [chunk.delta for chunk in response]
content_str = "".join(streamed_content).lower().strip()
assert "1963" not in content_str
@pytest.mark.parametrize(
"test_case",
[
"inference:completion:log_probs",
],
)
def test_text_completion_log_probs_non_streaming(client_with_models, text_model_id, inference_provider_type, test_case):
skip_if_model_doesnt_support_completion(client_with_models, text_model_id)
if inference_provider_type not in PROVIDER_LOGPROBS_TOP_K:
pytest.xfail(f"{inference_provider_type} doesn't support log probs yet")
tc = TestCase(test_case)
response = client_with_models.inference.completion(
content=tc["content"],
stream=False,
model_id=text_model_id,
sampling_params={
"max_tokens": 5,
},
logprobs={
"top_k": 1,
},
)
assert response.logprobs, "Logprobs should not be empty"
assert 1 <= len(response.logprobs) <= 5 # each token has 1 logprob and here max_tokens=5
assert all(len(logprob.logprobs_by_token) == 1 for logprob in response.logprobs)
@pytest.mark.parametrize(
"test_case",
[
"inference:completion:log_probs",
],
)
def test_text_completion_log_probs_streaming(client_with_models, text_model_id, inference_provider_type, test_case):
skip_if_model_doesnt_support_completion(client_with_models, text_model_id)
if inference_provider_type not in PROVIDER_LOGPROBS_TOP_K:
pytest.xfail(f"{inference_provider_type} doesn't support log probs yet")
tc = TestCase(test_case)
response = client_with_models.inference.completion(
content=tc["content"],
stream=True,
model_id=text_model_id,
sampling_params={
"max_tokens": 5,
},
logprobs={
"top_k": 1,
},
)
streamed_content = list(response)
for chunk in streamed_content:
if chunk.delta: # if there's a token, we expect logprobs
assert chunk.logprobs, "Logprobs should not be empty"
assert all(len(logprob.logprobs_by_token) == 1 for logprob in chunk.logprobs)
else: # no token, no logprobs
assert not chunk.logprobs, "Logprobs should be empty"
@pytest.mark.parametrize(
"test_case",
[
"inference:completion:structured_output",
],
)
def test_text_completion_structured_output(client_with_models, text_model_id, test_case):
skip_if_model_doesnt_support_completion(client_with_models, text_model_id)
class AnswerFormat(BaseModel):
name: str
year_born: str
year_retired: str
tc = TestCase(test_case)
user_input = tc["user_input"]
response = client_with_models.inference.completion(
model_id=text_model_id,
content=user_input,
stream=False,
sampling_params={
"max_tokens": 50,
},
response_format={
"type": "json_schema",
"json_schema": AnswerFormat.model_json_schema(),
},
)
answer = AnswerFormat.model_validate_json(response.content)
expected = tc["expected"]
assert answer.name == expected["name"]
assert answer.year_born == expected["year_born"]
assert answer.year_retired == expected["year_retired"]
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:non_streaming_01",
"inference:chat_completion:non_streaming_02",
],
)
def test_text_chat_completion_non_streaming(client_with_models, text_model_id, test_case):
tc = TestCase(test_case)
question = tc["question"]
expected = tc["expected"]
response = client_with_models.inference.chat_completion(
model_id=text_model_id,
messages=[
{
"role": "user",
"content": question,
}
],
stream=False,
)
message_content = response.completion_message.content.lower().strip()
assert len(message_content) > 0
assert expected.lower() in message_content
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:streaming_01",
"inference:chat_completion:streaming_02",
],
)
def test_text_chat_completion_streaming(client_with_models, text_model_id, test_case):
tc = TestCase(test_case)
question = tc["question"]
expected = tc["expected"]
response = client_with_models.inference.chat_completion(
model_id=text_model_id,
messages=[{"role": "user", "content": question}],
stream=True,
timeout=120, # Increase timeout to 2 minutes for large conversation history
)
streamed_content = [str(chunk.event.delta.text.lower().strip()) for chunk in response]
assert len(streamed_content) > 0
assert expected.lower() in "".join(streamed_content)
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:tool_calling",
],
)
def test_text_chat_completion_with_tool_calling_and_non_streaming(client_with_models, text_model_id, test_case):
tc = TestCase(test_case)
response = client_with_models.inference.chat_completion(
model_id=text_model_id,
messages=tc["messages"],
tools=tc["tools"],
tool_choice="auto",
stream=False,
)
# some models can return content for the response in addition to the tool call
assert response.completion_message.role == "assistant"
assert len(response.completion_message.tool_calls) == 1
assert response.completion_message.tool_calls[0].tool_name == tc["tools"][0]["tool_name"]
assert response.completion_message.tool_calls[0].arguments == tc["expected"]
# Will extract streamed text and separate it from tool invocation content
# The returned tool inovcation content will be a string so it's easy to comapare with expected value
# e.g. "[get_weather, {'location': 'San Francisco, CA'}]"
def extract_tool_invocation_content(response):
tool_invocation_content: str = ""
for chunk in response:
delta = chunk.event.delta
if delta.type == "tool_call" and delta.parse_status == "succeeded":
call = delta.tool_call
tool_invocation_content += f"[{call.tool_name}, {call.arguments}]"
return tool_invocation_content
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:tool_calling",
],
)
def test_text_chat_completion_with_tool_calling_and_streaming(client_with_models, text_model_id, test_case):
tc = TestCase(test_case)
response = client_with_models.inference.chat_completion(
model_id=text_model_id,
messages=tc["messages"],
tools=tc["tools"],
tool_choice="auto",
stream=True,
)
tool_invocation_content = extract_tool_invocation_content(response)
expected_tool_name = tc["tools"][0]["tool_name"]
expected_argument = tc["expected"]
assert tool_invocation_content == f"[{expected_tool_name}, {expected_argument}]"
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:tool_calling",
],
)
def test_text_chat_completion_with_tool_choice_required(client_with_models, text_model_id, test_case):
tc = TestCase(test_case)
response = client_with_models.inference.chat_completion(
model_id=text_model_id,
messages=tc["messages"],
tools=tc["tools"],
tool_config={
"tool_choice": "required",
},
stream=True,
)
tool_invocation_content = extract_tool_invocation_content(response)
expected_tool_name = tc["tools"][0]["tool_name"]
expected_argument = tc["expected"]
assert tool_invocation_content == f"[{expected_tool_name}, {expected_argument}]"
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:tool_calling",
],
)
def test_text_chat_completion_with_tool_choice_none(client_with_models, text_model_id, test_case):
tc = TestCase(test_case)
response = client_with_models.inference.chat_completion(
model_id=text_model_id,
messages=tc["messages"],
tools=tc["tools"],
tool_config={"tool_choice": "none"},
stream=True,
)
tool_invocation_content = extract_tool_invocation_content(response)
assert tool_invocation_content == ""
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:structured_output",
],
)
def test_text_chat_completion_structured_output(client_with_models, text_model_id, test_case):
skip_if_model_doesnt_support_json_schema_structured_output(client_with_models, text_model_id)
class NBAStats(BaseModel):
year_for_draft: int
num_seasons_in_nba: int
class AnswerFormat(BaseModel):
first_name: str
last_name: str
year_of_birth: int
nba_stats: NBAStats
tc = TestCase(test_case)
response = client_with_models.inference.chat_completion(
model_id=text_model_id,
messages=tc["messages"],
response_format={
"type": "json_schema",
"json_schema": AnswerFormat.model_json_schema(),
},
stream=False,
)
answer = AnswerFormat.model_validate_json(response.completion_message.content)
expected = tc["expected"]
assert answer.first_name == expected["first_name"]
assert answer.last_name == expected["last_name"]
assert answer.year_of_birth == expected["year_of_birth"]
assert answer.nba_stats.num_seasons_in_nba == expected["num_seasons_in_nba"]
assert answer.nba_stats.year_for_draft == expected["year_for_draft"]
@pytest.mark.parametrize("streaming", [True, False])
@pytest.mark.parametrize(
"test_case",
[
"inference:chat_completion:tool_calling_tools_absent",
],
)
def test_text_chat_completion_tool_calling_tools_not_in_request(
client_with_models, text_model_id, test_case, streaming
):
tc = TestCase(test_case)
# TODO: more dynamic lookup on tool_prompt_format for model family
tool_prompt_format = "json" if "3.1" in text_model_id else "python_list"
request = {
"model_id": text_model_id,
"messages": tc["messages"],
"tools": tc["tools"],
"tool_choice": "auto",
"tool_prompt_format": tool_prompt_format,
"stream": streaming,
}
response = client_with_models.inference.chat_completion(**request)
if streaming:
for chunk in response:
delta = chunk.event.delta
if delta.type == "tool_call" and delta.parse_status == "succeeded":
assert delta.tool_call.tool_name == "get_object_namespace_list"
if delta.type == "tool_call" and delta.parse_status == "failed":
# expect raw message that failed to parse in tool_call
assert isinstance(delta.tool_call, str)
assert len(delta.tool_call) > 0
else:
for tc in response.completion_message.tool_calls:
assert tc.tool_name == "get_object_namespace_list"
@pytest.mark.parametrize(
"test_case",
[
# Tests if the model can handle simple messages like "Hi" or
# a message unrelated to one of the tool calls
"inference:chat_completion:text_then_tool",
# Tests if the model can do full tool call with responses correctly
"inference:chat_completion:tool_then_answer",
# Tests if model can generate multiple params and
# read outputs correctly
"inference:chat_completion:array_parameter",
],
)
def test_text_chat_completion_with_multi_turn_tool_calling(client_with_models, text_model_id, test_case):
"""This test tests the model's tool calling loop in various scenarios"""
if "llama-4" not in text_model_id.lower() and "llama4" not in text_model_id.lower():
pytest.xfail("Not tested for non-llama4 models yet")
tc = TestCase(test_case)
messages = []
# keep going until either
# 1. we have messages to test in multi-turn
# 2. no messages bust last message is tool response
while len(tc["messages"]) > 0 or (len(messages) > 0 and messages[-1]["role"] == "tool"):
# do not take new messages if last message is tool response
if len(messages) == 0 or messages[-1]["role"] != "tool":
new_messages = tc["messages"].pop(0)
messages += new_messages
# pprint(messages)
response = client_with_models.inference.chat_completion(
model_id=text_model_id,
messages=messages,
tools=tc["tools"],
stream=False,
sampling_params={
"strategy": {
"type": "top_p",
"top_p": 0.9,
"temperature": 0.6,
}
},
)
op_msg = response.completion_message
messages.append(op_msg.model_dump())
# print(op_msg)
assert op_msg.role == "assistant"
expected = tc["expected"].pop(0)
assert len(op_msg.tool_calls) == expected["num_tool_calls"]
if expected["num_tool_calls"] > 0:
assert op_msg.tool_calls[0].tool_name == expected["tool_name"]
assert op_msg.tool_calls[0].arguments == expected["tool_arguments"]
tool_response = tc["tool_responses"].pop(0)
messages.append(
# Tool Response Message
{
"role": "tool",
"call_id": op_msg.tool_calls[0].call_id,
"content": tool_response["response"],
}
)
else:
actual_answer = op_msg.content.lower()
# pprint(actual_answer)
assert expected["answer"] in actual_answer
# sleep to avoid rate limit
sleep(1)