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, 08:30 UTC
[Bugfix][FailingTest]Fix test_model_load_with_params.py (#18758)
Signed-off-by: rabi <ramishra@redhat.com>
Rabi Mishra committed
May 28, 2025 at 11:12 UTC
b78f844a6743732b58022f2f84858d61b40b5913
4 files changed
+19
-22
.buildkite/test-pipeline.yaml
+11
-11
@@ -274,17 +274,6 @@ steps:
274
- pytest -v -s samplers
275
- VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
276
277
-- label: LogitsProcessor Test # 5min
278
- mirror_hardwares: [amdexperimental, amdproduction]
279
- source_file_dependencies:
280
- - vllm/model_executor/layers
281
- - vllm/model_executor/guided_decoding
282
- - tests/test_logits_processor
283
- - tests/model_executor/test_guided_processors
284
- commands:
285
- - pytest -v -s test_logits_processor.py
286
- - pytest -v -s model_executor/test_guided_processors.py
287
-
277
- label: Speculative decoding tests # 40min
278
mirror_hardwares: [amdexperimental]
279
source_file_dependencies:
@@ -397,6 +386,17 @@ steps:
386
- pytest -v -s tensorizer_loader
387
- pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
388
389
+- label: Model Executor Test
390
+ mirror_hardwares: [amdexperimental, amdproduction]
391
+ soft_fail: true
392
+ source_file_dependencies:
393
+ - vllm/model_executor
394
+ - tests/model_executor
395
+ commands:
396
+ - apt-get update && apt-get install -y curl libsodium23
397
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
398
+ - pytest -v -s model_executor
399
+
400
- label: Benchmarks # 9min
401
mirror_hardwares: [amdexperimental, amdproduction]
402
working_dir: "/vllm-workspace/.buildkite"
tests/model_executor/test_logits_processor.py
renamed
tests/model_executor/test_model_load_with_params.py
+8
-11
@@ -4,7 +4,7 @@ import os
4
5
import pytest
6
7
-from vllm.model_executor.layers.pooler import CLSPool, PoolingType
7
+from vllm.model_executor.layers.pooler import CLSPool, MeanPool, PoolingType
8
from vllm.model_executor.models.bert import BertEmbeddingModel
9
from vllm.model_executor.models.roberta import RobertaEmbeddingModel
10
from vllm.platforms import current_platform
@@ -14,7 +14,7 @@ MODEL_NAME = os.environ.get("MODEL_NAME", "BAAI/bge-base-en-v1.5")
14
REVISION = os.environ.get("REVISION", "main")
15
16
MODEL_NAME_ROBERTA = os.environ.get("MODEL_NAME",
17
- "intfloat/multilingual-e5-small")
17
+ "intfloat/multilingual-e5-base")
18
REVISION_ROBERTA = os.environ.get("REVISION", "main")
19
20
@@ -40,17 +40,15 @@ def test_model_loading_with_params(vllm_runner):
40
41
# asserts on the pooling config files
42
assert model_config.pooler_config.pooling_type == PoolingType.CLS.name
43
- assert model_config.pooler_config.pooling_norm
43
+ assert model_config.pooler_config.normalize
44
45
# asserts on the tokenizer loaded
46
assert model_tokenizer.tokenizer_id == "BAAI/bge-base-en-v1.5"
47
- assert model_tokenizer.tokenizer_config["do_lower_case"]
47
assert model_tokenizer.tokenizer.model_max_length == 512
48
49
def check_model(model):
50
assert isinstance(model, BertEmbeddingModel)
52
- assert model._pooler.pooling_type == PoolingType.CLS
53
- assert model._pooler.normalize
51
+ assert isinstance(model._pooler, CLSPool)
52
53
vllm_model.apply_model(check_model)
54
@@ -80,16 +78,15 @@ def test_roberta_model_loading_with_params(vllm_runner):
78
79
# asserts on the pooling config files
80
assert model_config.pooler_config.pooling_type == PoolingType.MEAN.name
83
- assert model_config.pooler_config.pooling_norm
81
+ assert model_config.pooler_config.normalize
82
83
# asserts on the tokenizer loaded
86
- assert model_tokenizer.tokenizer_id == "intfloat/multilingual-e5-small"
87
- assert not model_tokenizer.tokenizer_config["do_lower_case"]
84
+ assert model_tokenizer.tokenizer_id == "intfloat/multilingual-e5-base"
85
+ assert model_tokenizer.tokenizer.model_max_length == 512
86
87
def check_model(model):
88
assert isinstance(model, RobertaEmbeddingModel)
91
- assert model._pooler.pooling_type == PoolingType.MEAN
92
- assert model._pooler.normalize
89
+ assert isinstance(model._pooler, MeanPool)
90
91
vllm_model.apply_model(check_model)
92