[Frontend] Separate pooling APIs in offline inference (#11129)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2024-12-13 10:40:07 +00:00
committed by GitHub
parent f93bf2b189
commit eeec9e3390
21 changed files with 669 additions and 304 deletions
+7 -11
View File
@@ -719,14 +719,6 @@ class VllmRunner:
return inputs
def classify(self, prompts: List[str]) -> List[str]:
req_outputs = self.model.encode(prompts)
outputs = []
for req_output in req_outputs:
embedding = req_output.outputs.embedding
outputs.append(embedding)
return outputs
def generate(
self,
prompts: List[str],
@@ -897,6 +889,10 @@ class VllmRunner:
returned_outputs.append((token_ids, texts))
return returned_outputs
def classify(self, prompts: List[str]) -> List[List[float]]:
req_outputs = self.model.classify(prompts)
return [req_output.outputs.probs for req_output in req_outputs]
def encode(
self,
prompts: List[str],
@@ -909,16 +905,16 @@ class VllmRunner:
videos=videos,
audios=audios)
req_outputs = self.model.encode(inputs)
req_outputs = self.model.embed(inputs)
return [req_output.outputs.embedding for req_output in req_outputs]
def score(
self,
text_1: Union[str, List[str]],
text_2: Union[str, List[str]],
) -> List[List[float]]:
) -> List[float]:
req_outputs = self.model.score(text_1, text_2)
return [req_output.outputs.embedding for req_output in req_outputs]
return [req_output.outputs.score for req_output in req_outputs]
def __enter__(self):
return self
+5 -5
View File
@@ -39,8 +39,8 @@ async def test_text_1_str_text_2_list(server: RemoteOpenAIServer,
assert score.id is not None
assert score.data is not None
assert len(score.data) == 2
assert score.data[0].score[0] <= 0.01
assert score.data[1].score[0] >= 0.9
assert score.data[0].score <= 0.01
assert score.data[1].score >= 0.9
@pytest.mark.asyncio
@@ -67,8 +67,8 @@ async def test_text_1_list_text_2_list(server: RemoteOpenAIServer,
assert score.id is not None
assert score.data is not None
assert len(score.data) == 2
assert score.data[0].score[0] <= 0.01
assert score.data[1].score[0] >= 0.9
assert score.data[0].score <= 0.01
assert score.data[1].score >= 0.9
@pytest.mark.asyncio
@@ -90,4 +90,4 @@ async def test_text_1_str_text_2_str(server: RemoteOpenAIServer,
assert score.id is not None
assert score.data is not None
assert len(score.data) == 1
assert score.data[0].score[0] >= 0.9
assert score.data[0].score >= 0.9
@@ -42,7 +42,7 @@ def test_llm_1_to_1(vllm_runner, hf_runner, model_name, dtype: str):
assert len(vllm_outputs) == 1
assert len(hf_outputs) == 1
assert math.isclose(hf_outputs[0], vllm_outputs[0][0], rel_tol=0.01)
assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)
@pytest.mark.parametrize("dtype", ["half"])
@@ -63,8 +63,8 @@ def test_llm_1_to_N(vllm_runner, hf_runner, model_name, dtype: str):
assert len(vllm_outputs) == 2
assert len(hf_outputs) == 2
assert math.isclose(hf_outputs[0], vllm_outputs[0][0], rel_tol=0.01)
assert math.isclose(hf_outputs[1], vllm_outputs[1][0], rel_tol=0.01)
assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)
assert math.isclose(hf_outputs[1], vllm_outputs[1], rel_tol=0.01)
@pytest.mark.parametrize("dtype", ["half"])
@@ -85,5 +85,5 @@ def test_llm_N_to_N(vllm_runner, hf_runner, model_name, dtype: str):
assert len(vllm_outputs) == 2
assert len(hf_outputs) == 2
assert math.isclose(hf_outputs[0], vllm_outputs[0][0], rel_tol=0.01)
assert math.isclose(hf_outputs[1], vllm_outputs[1][0], rel_tol=0.01)
assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)
assert math.isclose(hf_outputs[1], vllm_outputs[1], rel_tol=0.01)
+2 -3
View File
@@ -2,7 +2,7 @@ import os
import pytest
from vllm import LLM, PoolingParams, SamplingParams
from vllm import LLM, SamplingParams
from vllm.assets.image import ImageAsset
from ..utils import fork_new_process_for_each_test
@@ -36,9 +36,8 @@ def test_oot_registration_text_generation(dummy_opt_path):
def test_oot_registration_embedding(dummy_gemma2_embedding_path):
os.environ["VLLM_PLUGINS"] = "register_dummy_model"
prompts = ["Hello, my name is", "The text does not matter"]
sampling_params = PoolingParams()
llm = LLM(model=dummy_gemma2_embedding_path, load_format="dummy")
outputs = llm.encode(prompts, sampling_params)
outputs = llm.embed(prompts)
for output in outputs:
assert all(v == 0 for v in output.outputs.embedding)