phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-10-04 12:07:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

feat(batches, completions): add /v1/completions support to /v1/batches (#3309)

Browse source 
# What does this PR do?

add support for /v1/completions to the /v1/batches api


## Test Plan

ci
This commit is contained in:
Matthew Farrellee 2025-09-05 14:59:57 -04:00 • committed by GitHub
parent e2fe39aee1
commit df1526991f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 205 additions and 26 deletions
Download patch file Download diff file Expand all files Collapse all files

69
llama_stack/providers/inline/batches/reference/batches.py
View file

@ -178,9 +178,9 @@ class ReferenceBatchesImpl(Batches):
# TODO: set expiration time for garbage collection
if endpoint not in ["/v1/chat/completions"]:
if endpoint not in ["/v1/chat/completions", "/v1/completions"]:
raise ValueError(
f"Invalid endpoint: {endpoint}. Supported values: /v1/chat/completions. Code: invalid_value. Param: endpoint",
f"Invalid endpoint: {endpoint}. Supported values: /v1/chat/completions, /v1/completions. Code: invalid_value. Param: endpoint",
)
if completion_window != "24h":
@ -424,13 +424,21 @@ class ReferenceBatchesImpl(Batches):
)
valid = False
for param, expected_type, type_string in [
("model", str, "a string"),
# messages is specific to /v1/chat/completions
# we could skip validating messages here and let inference fail. however,
# that would be a very expensive way to find out messages is wrong.
("messages", list, "an array"), # TODO: allow messages to be a string?
]:
if batch.endpoint == "/v1/chat/completions":
required_params = [
("model", str, "a string"),
# messages is specific to /v1/chat/completions
# we could skip validating messages here and let inference fail. however,
# that would be a very expensive way to find out messages is wrong.
("messages", list, "an array"), # TODO: allow messages to be a string?
]
else: # /v1/completions
required_params = [
("model", str, "a string"),
("prompt", str, "a string"), # TODO: allow prompt to be a list of strings??
]
for param, expected_type, type_string in required_params:
if param not in body:
errors.append(
BatchError(
@ -591,20 +599,37 @@ class ReferenceBatchesImpl(Batches):
try:
# TODO(SECURITY): review body for security issues
request.body["messages"] = [convert_to_openai_message_param(msg) for msg in request.body["messages"]]
chat_response = await self.inference_api.openai_chat_completion(**request.body)
if request.url == "/v1/chat/completions":
request.body["messages"] = [convert_to_openai_message_param(msg) for msg in request.body["messages"]]
chat_response = await self.inference_api.openai_chat_completion(**request.body)
# this is for mypy, we don't allow streaming so we'll get the right type
assert hasattr(chat_response, "model_dump_json"), "Chat response must have model_dump_json method"
return {
"id": request_id,
"custom_id": request.custom_id,
"response": {
"status_code": 200,
"request_id": request_id, # TODO: should this be different?
"body": chat_response.model_dump_json(),
},
}
# this is for mypy, we don't allow streaming so we'll get the right type
assert hasattr(chat_response, "model_dump_json"), "Chat response must have model_dump_json method"
return {
"id": request_id,
"custom_id": request.custom_id,
"response": {
"status_code": 200,
"request_id": request_id, # TODO: should this be different?
"body": chat_response.model_dump_json(),
},
}
else: # /v1/completions
completion_response = await self.inference_api.openai_completion(**request.body)
# this is for mypy, we don't allow streaming so we'll get the right type
assert hasattr(completion_response, "model_dump_json"), (
"Completion response must have model_dump_json method"
)
return {
"id": request_id,
"custom_id": request.custom_id,
"response": {
"status_code": 200,
"request_id": request_id,
"body": completion_response.model_dump_json(),
},
}
except Exception as e:
logger.info(f"Error processing request {request.custom_id} in batch {batch_id}: {e}")
return {