A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 14:45 UTC
[CI][Bugfix] Fix bad tolerance for test_batch_base64_embedding (#16221)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Apr 8, 2025 at 22:14 UTC
87b4ac56c2f9c61bd1d2e0cdcfd2db43c8c52d91
1 file changed
+17
-11
tests/entrypoints/openai/test_embedding.py
+17
-11
@@ -11,6 +11,7 @@ import requests
11
from vllm.entrypoints.openai.protocol import EmbeddingResponse
12
from vllm.transformers_utils.tokenizer import get_tokenizer
13
14
+from ...models.embedding.utils import check_embeddings_close
15
from ...utils import RemoteOpenAIServer
16
17
MODEL_NAME = "intfloat/multilingual-e5-small"
@@ -190,30 +191,35 @@ async def test_batch_base64_embedding(client: openai.AsyncOpenAI,
191
responses_float = await client.embeddings.create(input=input_texts,
192
model=model_name,
193
encoding_format="float")
194
+ float_data = [d.embedding for d in responses_float.data]
195
196
responses_base64 = await client.embeddings.create(input=input_texts,
197
model=model_name,
198
encoding_format="base64")
197
-
198
- decoded_responses_base64_data = []
199
+ base64_data = []
200
for data in responses_base64.data:
200
- decoded_responses_base64_data.append(
201
+ base64_data.append(
202
np.frombuffer(base64.b64decode(data.embedding),
203
dtype="float32").tolist())
204
204
- assert responses_float.data[0].embedding == decoded_responses_base64_data[
205
- 0]
206
- assert responses_float.data[1].embedding == decoded_responses_base64_data[
207
- 1]
205
+ check_embeddings_close(
206
+ embeddings_0_lst=float_data,
207
+ embeddings_1_lst=base64_data,
208
+ name_0="float",
209
+ name_1="base64",
210
+ )
211
212
# Default response is float32 decoded from base64 by OpenAI Client
213
responses_default = await client.embeddings.create(input=input_texts,
214
model=model_name)
215
+ default_data = [d.embedding for d in responses_default.data]
216
213
- assert responses_float.data[0].embedding == responses_default.data[
214
- 0].embedding
215
- assert responses_float.data[1].embedding == responses_default.data[
216
- 1].embedding
217
+ check_embeddings_close(
218
+ embeddings_0_lst=float_data,
219
+ embeddings_1_lst=default_data,
220
+ name_0="float",
221
+ name_1="default",
222
+ )
223
224
225
@pytest.mark.asyncio