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, 18:30 UTC
[CI/Build] Delete LoRA bias test (#14849)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Mar 15, 2025 at 13:09 UTC
e0fdfa1608b988f3b6767117a9adf2f9a3831cb9
2 files changed
-68
tests/lora/conftest.py
-5
@@ -173,11 +173,6 @@ def sql_lora_files(sql_lora_huggingface_id):
173
return snapshot_download(repo_id=sql_lora_huggingface_id)
174
175
176
-@pytest.fixture(scope="session")
177
-def lora_bias_files():
178
- return snapshot_download(repo_id="followumesh/granite-3b-lora8-bias")
179
-
180
-
176
@pytest.fixture(scope="session")
177
def mixtral_lora_files():
178
# Note: this module has incorrect adapter_config.json to test
tests/lora/test_lora_bias_e2e.py
deleted
-63
@@ -1,63 +0,0 @@
1
-# SPDX-License-Identifier: Apache-2.0
2
-
3
-import pytest
4
-
5
-import vllm
6
-from vllm.lora.request import LoRARequest
7
-
8
-MODEL_PATH = "ibm-granite/granite-3b-code-base"
9
-
10
-
11
-def do_sample(llm: vllm.LLM, lora_path: str, lora_id: int) -> list[str]:
12
- prompts = [
13
- "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE candidate (people_id VARCHAR, unsure_rate INTEGER); CREATE TABLE people (sex VARCHAR, people_id VARCHAR)\n\n question: which gender got the highest average uncertain ratio. [/user] [assistant]", # noqa: E501
14
- "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the women's doubles for werner schlager [/user] [assistant]" # noqa: E501
15
- ]
16
- sampling_params = vllm.SamplingParams(temperature=0,
17
- max_tokens=256,
18
- stop=["[/assistant]"])
19
- outputs = llm.generate(
20
- prompts,
21
- sampling_params,
22
- lora_request=LoRARequest(str(lora_id), lora_id, lora_path)
23
- if lora_id else None)
24
- generated_texts: list[str] = []
25
- for output in outputs:
26
- generated_text = output.outputs[0].text
27
- generated_texts.append(generated_text)
28
- return generated_texts
29
-
30
-
31
-@pytest.fixture(autouse=True)
32
-def v1(run_with_both_engines_lora):
33
- # Simple autouse wrapper to run both engines for each test
34
- # This can be promoted up to conftest.py to run for every
35
- # test in a package
36
- pass
37
-
38
-
39
-# Skipping for V1 for now as we are hitting,
40
-# "Head size 80 is not supported by FlashAttention." error.
41
-@pytest.mark.skip_v1
42
-@pytest.mark.parametrize("lora_bias", [True])
43
-@pytest.mark.parametrize("fully_sharded", [True, False])
44
-def test_lora_bias(lora_bias_files: str, lora_bias: bool, fully_sharded: bool):
45
- llm = vllm.LLM(MODEL_PATH,
46
- enable_lora=True,
47
- max_num_seqs=16,
48
- max_lora_rank=8,
49
- max_loras=1,
50
- enable_lora_bias=lora_bias,
51
- tensor_parallel_size=1,
52
- fully_sharded_loras=fully_sharded)
53
-
54
- print("lora adapter created")
55
- output1 = do_sample(llm, lora_bias_files, lora_id=0)
56
-
57
- print("lora")
58
- output2 = do_sample(llm, lora_bias_files, lora_id=1)
59
-
60
- if lora_bias:
61
- assert output1 != output2
62
- else:
63
- assert output1 == output2