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 16, 22:00 UTC
[CI/Build] Cleanup VLM tests (#6107)
Cyrus Leung committed
Jul 4, 2024 at 09:58 UTC
3dd507083f4d8416d5fed9827e91d22f29b0b723
4 files changed
+5
-8
tests/models/test_llava_next.py
+2
-5
@@ -1,4 +1,3 @@
1
-import re
1
from typing import List, Optional, Tuple
2
3
import pytest
@@ -36,7 +35,6 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
35
output_ids, output_str, out_logprobs = vllm_output
36
37
tokenizer = AutoTokenizer.from_pretrained(model)
39
- image_token_str = tokenizer.decode(IMAGE_TOKEN_ID)
38
eos_token_id = tokenizer.eos_token_id
39
40
hf_output_ids = [
@@ -44,9 +42,8 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
42
if token_id != IMAGE_TOKEN_ID or output_ids[idx - 1] != IMAGE_TOKEN_ID
43
]
44
47
- hf_output_str = re.sub(fr"({image_token_str})+", "", output_str)
48
- assert hf_output_str[0] == " "
49
- hf_output_str = hf_output_str[1:]
45
+ assert output_str[0] == " "
46
+ hf_output_str = output_str[1:]
47
if hf_output_ids[-1] == eos_token_id:
48
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
49
tests/models/test_phi3v.py
+1
-2
@@ -35,8 +35,7 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
35
assert output_str_without_image[0] == " "
36
output_str_without_image = output_str_without_image[1:]
37
38
- hf_output_str = output_str_without_image.replace("<|user|>", "") \
39
- .replace("<|end|>\n<|assistant|>", " ")
38
+ hf_output_str = output_str_without_image + "<|end|><|endoftext|>"
39
40
tokenizer = AutoTokenizer.from_pretrained(model)
41
hf_output_ids = tokenizer.encode(output_str_without_image)
tests/models/utils.py
+1
@@ -77,6 +77,7 @@ def check_logprobs_close(
77
# Each predicted token must be in top N logprobs of the other
78
fail_msg = (
79
f"Test{prompt_idx}:"
80
+ f"\nMatched tokens:\t{output_ids_0[:idx]}"
81
f"\n{name_0}:\t{output_str_0!r}\t{logprobs_elem_0}"
82
f"\n{name_1}:\t{output_str_1!r}\t{logprobs_elem_1}")
83
vllm/multimodal/image.py
+1
-1
@@ -115,7 +115,7 @@ class ImagePlugin(MultiModalPlugin):
115
if isinstance(data, Image.Image):
116
image_processor = self._get_hf_image_processor(model_config)
117
if image_processor is None:
118
- raise RuntimeError("No HuggingFace processor is available"
118
+ raise RuntimeError("No HuggingFace processor is available "
119
"to process the image object")
120
try:
121
batch_data = image_processor \