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, 23:30 UTC
[Misc] Remove unnecessary detokenization in multimodal processing (#12868)
Cyrus Leung committed
Feb 7, 2025 at 22:21 UTC
ce26b16268ef8d7db5c1346c482b899f49dcd3cd
4 files changed
+7
-10
tests/entrypoints/openai/test_audio.py
+3
-3
@@ -83,7 +83,7 @@ async def test_single_chat_session_audio(client: openai.AsyncOpenAI,
83
choice = chat_completion.choices[0]
84
assert choice.finish_reason == "length"
85
assert chat_completion.usage == openai.types.CompletionUsage(
86
- completion_tokens=10, prompt_tokens=202, total_tokens=212)
86
+ completion_tokens=10, prompt_tokens=201, total_tokens=211)
87
88
message = choice.message
89
message = chat_completion.choices[0].message
@@ -140,7 +140,7 @@ async def test_single_chat_session_audio_base64encoded(
140
choice = chat_completion.choices[0]
141
assert choice.finish_reason == "length"
142
assert chat_completion.usage == openai.types.CompletionUsage(
143
- completion_tokens=10, prompt_tokens=202, total_tokens=212)
143
+ completion_tokens=10, prompt_tokens=201, total_tokens=211)
144
145
message = choice.message
146
message = chat_completion.choices[0].message
@@ -196,7 +196,7 @@ async def test_single_chat_session_input_audio(
196
choice = chat_completion.choices[0]
197
assert choice.finish_reason == "length"
198
assert chat_completion.usage == openai.types.CompletionUsage(
199
- completion_tokens=10, prompt_tokens=202, total_tokens=212)
199
+ completion_tokens=10, prompt_tokens=201, total_tokens=211)
200
201
message = choice.message
202
message = chat_completion.choices[0].message
tests/entrypoints/openai/test_vision.py
+2
-2
@@ -92,7 +92,7 @@ async def test_single_chat_session_image(client: openai.AsyncOpenAI,
92
choice = chat_completion.choices[0]
93
assert choice.finish_reason == "length"
94
assert chat_completion.usage == openai.types.CompletionUsage(
95
- completion_tokens=10, prompt_tokens=775, total_tokens=785)
95
+ completion_tokens=10, prompt_tokens=774, total_tokens=784)
96
97
message = choice.message
98
message = chat_completion.choices[0].message
@@ -185,7 +185,7 @@ async def test_single_chat_session_image_base64encoded(
185
choice = chat_completion.choices[0]
186
assert choice.finish_reason == "length"
187
assert chat_completion.usage == openai.types.CompletionUsage(
188
- completion_tokens=10, prompt_tokens=775, total_tokens=785)
188
+ completion_tokens=10, prompt_tokens=774, total_tokens=784)
189
190
message = choice.message
191
message = chat_completion.choices[0].message
tests/entrypoints/openai/test_vision_embedding.py
+2
-2
@@ -93,5 +93,5 @@ async def test_image_embedding(server: RemoteOpenAIServer, model_name: str,
93
assert len(embeddings.data) == 1
94
assert len(embeddings.data[0].embedding) == 3072
95
assert embeddings.usage.completion_tokens == 0
96
- assert embeddings.usage.prompt_tokens == 764
97
- assert embeddings.usage.total_tokens == 764
96
+ assert embeddings.usage.prompt_tokens == 763
97
+ assert embeddings.usage.total_tokens == 763
vllm/inputs/preprocess.py
-3
@@ -260,9 +260,6 @@ class InputPreprocessor:
260
mm_processor = self.mm_registry.create_processor(
261
self.model_config, tokenizer)
262
263
- if isinstance(prompt, list):
264
- prompt = tokenizer.decode(prompt)
265
-
263
if mm_processor_kwargs is None:
264
mm_processor_kwargs = {}
265