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 14, 05:30 UTC
[Misc] Fix warnings for mistral model (#23552)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> Signed-off-by: Jiangyun Zhu <riverclouds.zhu@qq.com> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Jiangyun Zhu committed
Aug 29, 2025 at 14:58 UTC
885ca6d31db8816ee08e3fa634fbb58add289898
3 files changed
+31
-23
vllm/model_executor/models/pixtral.py
+7
-5
@@ -15,7 +15,7 @@ from mistral_common.protocol.instruct.messages import (ImageChunk, TextChunk,
15
from mistral_common.protocol.instruct.request import ChatCompletionRequest
16
from mistral_common.tokens.tokenizers.multimodal import ImageEncoder
17
from PIL import Image
18
-from transformers import PixtralVisionConfig, TensorType
18
+from transformers import BatchFeature, PixtralVisionConfig, TensorType
19
from transformers.image_utils import ImageInput
20
from transformers.models.pixtral.image_processing_pixtral import (
21
_num_image_tokens as _get_pixtral_hf_num_image_tokens)
@@ -163,10 +163,12 @@ class PixtralProcessorAdapter:
163
images_processed.append(image_processed)
164
images_tokens.append(image_tokens)
165
166
- return {
167
- "input_ids": torch.cat(images_tokens)[None].expand(len(text), -1),
168
- "images": images_processed,
169
- }
166
+ return BatchFeature({
167
+ "input_ids":
168
+ torch.cat(images_tokens)[None].expand(len(text), -1),
169
+ "images":
170
+ images_processed,
171
+ })
172
173
174
class PixtralProcessingInfo(BaseProcessingInfo):
vllm/model_executor/models/voxtral.py
+7
-5
@@ -17,7 +17,7 @@ from mistral_common.protocol.instruct.messages import (AudioChunk, RawAudio,
17
from mistral_common.protocol.instruct.request import ChatCompletionRequest
18
from mistral_common.protocol.transcription.request import TranscriptionRequest
19
from mistral_common.tokens.tokenizers.audio import Audio, AudioEncoder
20
-from transformers import TensorType, WhisperConfig
20
+from transformers import BatchFeature, TensorType, WhisperConfig
21
from transformers.tokenization_utils_base import TextInput
22
23
from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig
@@ -156,10 +156,12 @@ class VoxtralProcessorAdapter:
156
audios_tokens.append(torch.tensor(audio_tokens))
157
audios_processed.append(torch.tensor(audio))
158
159
- return {
160
- "input_ids": torch.cat(audios_tokens)[None].expand(len(text), -1),
161
- "audio_arrays": audios_processed,
162
- }
159
+ return BatchFeature({
160
+ "input_ids":
161
+ torch.cat(audios_tokens)[None].expand(len(text), -1),
162
+ "audio_arrays":
163
+ audios_processed,
164
+ })
165
166
167
class VoxtralProcessingInfo(BaseProcessingInfo):
vllm/transformers_utils/tokenizers/mistral.py
+17
-13
@@ -204,18 +204,16 @@ class MistralTokenizer(TokenizerBase):
204
self.version: int = int(_mistral_version_str.split("v")[-1])
205
206
tokenizer_ = tokenizer.instruct_tokenizer.tokenizer
207
- from mistral_common.tokens.tokenizers.tekken import (
208
- SpecialTokenPolicy, Tekkenizer)
207
+ from mistral_common.tokens.tokenizers.base import SpecialTokenPolicy
208
+ from mistral_common.tokens.tokenizers.tekken import Tekkenizer
209
+
210
self.is_tekken = isinstance(tokenizer_, Tekkenizer)
211
from mistral_common.tokens.tokenizers.sentencepiece import (
212
SentencePieceTokenizer)
213
self.is_spm = isinstance(tokenizer_, SentencePieceTokenizer)
213
- if self.is_tekken:
214
- # Make sure special tokens will not raise
215
- tokenizer_.special_token_policy = SpecialTokenPolicy.IGNORE
216
- elif self.is_spm:
217
- pass
218
- else:
214
+ self._special_token_policy = (SpecialTokenPolicy.IGNORE
215
+ if self.is_tekken else None)
216
+ if not (self.is_tekken or self.is_spm):
217
raise TypeError(f"Unsupported tokenizer: {type(tokenizer_)}")
218
219
self._vocab = tokenizer_.vocab()
@@ -430,7 +428,8 @@ class MistralTokenizer(TokenizerBase):
428
return self.tokenizer.unk_id
429
430
ids = [_token_to_id(t) for t in tokens]
433
- decoded = self.tokenizer.decode(ids)
431
+ decoded = self.tokenizer.decode(ids,
432
+ self._special_token_policy)
433
else:
434
decoded = "".join(tokens)
435
else:
@@ -444,7 +443,8 @@ class MistralTokenizer(TokenizerBase):
443
if token in special_tokens:
444
if regular_tokens:
445
decoded_list.append(
447
- self.tokenizer.decode(regular_tokens))
446
+ self.tokenizer.decode(regular_tokens,
447
+ self._special_token_policy))
448
regular_tokens = []
449
decoded_list.append(token)
450
else:
@@ -452,7 +452,8 @@ class MistralTokenizer(TokenizerBase):
452
453
if regular_tokens:
454
decoded_list.append(
455
- self.tokenizer.decode(regular_tokens)) # type: ignore
455
+ self.tokenizer.decode(regular_tokens,
456
+ self._special_token_policy))
457
458
decoded = ''.join(decoded_list)
459
@@ -470,7 +471,7 @@ class MistralTokenizer(TokenizerBase):
471
472
if isinstance(ids, int):
473
ids = [ids]
473
- return self.tokenizer.decode(ids)
474
+ return self.tokenizer.decode(ids, self._special_token_policy)
475
476
def convert_ids_to_tokens(
477
self,
@@ -511,6 +512,9 @@ class MistralTokenizer(TokenizerBase):
512
# See: https://github.com/vllm-project/vllm/pull/8640
513
# https://github.com/vllm-project/vllm/pull/9625
514
# if underlying tokenizeir is sentencepiece, we just add "�"
514
- tokens = [self.tokenizer.id_to_byte_piece(id) for id in ids]
515
+ tokens = [
516
+ self.tokenizer.id_to_byte_piece(id, self._special_token_policy)
517
+ for id in ids
518
+ ]
519
520
return tokens