phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-08-02 16:54:42 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add eval_id for jobs

Browse source
This commit is contained in:
Xi Yan 2024-11-07 14:30:46 -08:00
parent ea80f623fb
commit f05db9a25c
4 changed files with 30 additions and 41 deletions
Download patch file Download diff file Expand all files Collapse all files

12
llama_stack/apis/eval/eval.py
View file

@ -85,21 +85,17 @@ class Eval(Protocol):
self, self,
input_rows: List[Dict[str, Any]], input_rows: List[Dict[str, Any]],
scoring_functions: List[str], scoring_functions: List[str],
eval_task_config: EvalTaskConfig, task_config: EvalTaskConfig,
eval_task_id: Optional[str] = None, eval_task_id: Optional[str] = None,
) -> EvaluateResponse: ... ) -> EvaluateResponse: ...
@webmethod(route="/eval/job/status", method="GET") @webmethod(route="/eval/job/status", method="GET")
async def job_status( async def job_status(
self, job_id: str, eval_task_id: Optional[str] = None self, job_id: str, eval_task_id: str
) -> Optional[JobStatus]: ... ) -> Optional[JobStatus]: ...
@webmethod(route="/eval/job/cancel", method="POST") @webmethod(route="/eval/job/cancel", method="POST")
async def job_cancel( async def job_cancel(self, job_id: str, eval_task_id: str) -> None: ...
self, job_id: str, eval_task_id: Optional[str] = None
) -> None: ...
@webmethod(route="/eval/job/result", method="GET") @webmethod(route="/eval/job/result", method="GET")
async def job_result( async def job_result(self, job_id: str, eval_task_id: str) -> EvaluateResponse: ...
self, job_id: str, eval_task_id: Optional[str] = None
) -> EvaluateResponse: ...

24
llama_stack/distribution/routers/routers.py
View file

@ -290,7 +290,7 @@ class EvalRouter(Eval):
self, self,
input_rows: List[Dict[str, Any]], input_rows: List[Dict[str, Any]],
scoring_functions: List[str], scoring_functions: List[str],
eval_task_config: EvalTaskConfig, task_config: EvalTaskConfig,
eval_task_id: Optional[str] = None, eval_task_id: Optional[str] = None,
) -> EvaluateResponse: ) -> EvaluateResponse:
# NOTE: This is to deal with the case where we do not pre-register an eval benchmark_task # NOTE: This is to deal with the case where we do not pre-register an eval benchmark_task
@ -300,36 +300,32 @@ class EvalRouter(Eval):
return await self.routing_table.get_provider_impl(eval_task_id).evaluate_rows( return await self.routing_table.get_provider_impl(eval_task_id).evaluate_rows(
input_rows=input_rows, input_rows=input_rows,
scoring_functions=scoring_functions, scoring_functions=scoring_functions,
eval_task_config=eval_task_config, task_config=task_config,
) )
async def job_status( async def job_status(
self, self,
job_id: str, job_id: str,
eval_task_id: Optional[str] = None, eval_task_id: str,
) -> Optional[JobStatus]: ) -> Optional[JobStatus]:
if eval_task_id is None:
eval_task_id = DEFAULT_EVAL_TASK_IDENTIFIER
return await self.routing_table.get_provider_impl(eval_task_id).job_status( return await self.routing_table.get_provider_impl(eval_task_id).job_status(
job_id job_id, eval_task_id
) )
async def job_cancel( async def job_cancel(
self, self,
job_id: str, job_id: str,
eval_task_id: Optional[str] = None, eval_task_id: str,
) -> None: ) -> None:
if eval_task_id is None: await self.routing_table.get_provider_impl(eval_task_id).job_cancel(
eval_task_id = DEFAULT_EVAL_TASK_IDENTIFIER job_id, eval_task_id
await self.routing_table.get_provider_impl(eval_task_id).job_cancel(job_id) )
async def job_result( async def job_result(
self, self,
job_id: str, job_id: str,
eval_task_id: Optional[str] = None, eval_task_id: str,
) -> EvaluateResponse: ) -> EvaluateResponse:
if eval_task_id is None:
eval_task_id = DEFAULT_EVAL_TASK_IDENTIFIER
return await self.routing_table.get_provider_impl(eval_task_id).job_result( return await self.routing_table.get_provider_impl(eval_task_id).job_result(
job_id job_id, eval_task_id
) )

25
llama_stack/providers/inline/meta_reference/eval/eval.py
View file

@ -122,7 +122,7 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
res = await self.evaluate_rows( res = await self.evaluate_rows(
input_rows=all_rows.rows, input_rows=all_rows.rows,
scoring_functions=scoring_functions, scoring_functions=scoring_functions,
eval_task_config=task_config, task_config=task_config,
) )
# TODO: currently needs to wait for generation before returning # TODO: currently needs to wait for generation before returning
@ -135,10 +135,10 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
self, self,
input_rows: List[Dict[str, Any]], input_rows: List[Dict[str, Any]],
scoring_functions: List[str], scoring_functions: List[str],
eval_task_config: EvalTaskConfig, task_config: EvalTaskConfig,
eval_task_id: Optional[str] = None, eval_task_id: Optional[str] = None,
) -> EvaluateResponse: ) -> EvaluateResponse:
candidate = eval_task_config.eval_candidate candidate = task_config.eval_candidate
if candidate.type == "agent": if candidate.type == "agent":
raise NotImplementedError( raise NotImplementedError(
"Evaluation with generation has not been implemented for agents" "Evaluation with generation has not been implemented for agents"
@ -190,12 +190,9 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
for input_r, generated_r in zip(input_rows, generations) for input_r, generated_r in zip(input_rows, generations)
] ]
if ( if task_config.type == "app" and task_config.scoring_params is not None:
eval_task_config.type == "app"
and eval_task_config.scoring_params is not None
):
scoring_functions_dict = { scoring_functions_dict = {
scoring_fn_id: eval_task_config.scoring_params.get(scoring_fn_id, None) scoring_fn_id: task_config.scoring_params.get(scoring_fn_id, None)
for scoring_fn_id in scoring_functions for scoring_fn_id in scoring_functions
} }
else: else:
@ -209,21 +206,17 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
return EvaluateResponse(generations=generations, scores=score_response.results) return EvaluateResponse(generations=generations, scores=score_response.results)
async def job_status( async def job_status(self, job_id: str, eval_task_id: str) -> Optional[JobStatus]:
self, job_id: str, eval_task_id: Optional[str] = None
) -> Optional[JobStatus]:
if job_id in self.jobs: if job_id in self.jobs:
return JobStatus.completed return JobStatus.completed
return None return None
async def job_cancel(self, job_id: str, eval_task_id: Optional[str] = None) -> None: async def job_cancel(self, job_id: str, eval_task_id: str) -> None:
raise NotImplementedError("Job cancel is not implemented yet") raise NotImplementedError("Job cancel is not implemented yet")
async def job_result( async def job_result(self, job_id: str, eval_task_id: str) -> EvaluateResponse:
self, job_id: str, eval_task_id: Optional[str] = None status = await self.job_status(job_id, eval_task_id)
) -> EvaluateResponse:
status = await self.job_status(job_id)
if not status or status != JobStatus.completed: if not status or status != JobStatus.completed:
raise ValueError(f"Job is not completed, Status: {status.value}") raise ValueError(f"Job is not completed, Status: {status.value}")

10
llama_stack/providers/tests/eval/test_eval.py
View file

@ -52,7 +52,7 @@ class Testeval:
response = await eval_impl.evaluate_rows( response = await eval_impl.evaluate_rows(
input_rows=rows.rows, input_rows=rows.rows,
scoring_functions=scoring_functions, scoring_functions=scoring_functions,
eval_task_config=AppEvalTaskConfig( task_config=AppEvalTaskConfig(
eval_candidate=ModelCandidate( eval_candidate=ModelCandidate(
model="Llama3.2-3B-Instruct", model="Llama3.2-3B-Instruct",
sampling_params=SamplingParams(), sampling_params=SamplingParams(),
@ -90,9 +90,13 @@ class Testeval:
), ),
) )
assert response.job_id == "0" assert response.job_id == "0"
job_status = await eval_impl.job_status(response.job_id) job_status = await eval_impl.job_status(
response.job_id, "meta-reference::app_eval"
)
assert job_status and job_status.value == "completed" assert job_status and job_status.value == "completed"
eval_response = await eval_impl.job_result(response.job_id) eval_response = await eval_impl.job_result(
response.job_id, "meta-reference::app_eval"
)
assert eval_response is not None assert eval_response is not None
assert len(eval_response.generations) == 5 assert len(eval_response.generations) == 5