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 6, 18:15 UTC
[Bugfix] Fix DeepSeek-V3.2 tokenizer stripping spaces (#37004)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Mar 13, 2026 at 18:55 UTC
9efc4db9658a987390b809dbcc13a9a771701b7f
2 files changed
+4
-2
vllm/config/model.py
+2
@@ -540,6 +540,8 @@ class ModelConfig:
540
self.tokenizer_mode = "kimi_audio"
541
elif arch == "QwenVLForConditionalGeneration":
542
self.tokenizer_mode = "qwen_vl"
543
+ elif arch == "DeepseekV32ForCausalLM":
544
+ self.tokenizer_mode = "deepseek_v32"
545
546
if self.tokenizer_mode != "auto":
547
logger.info(
vllm/tokenizers/deepseek_v32.py
+2
-2
@@ -3,7 +3,7 @@
3
import copy
4
from typing import Any
5
6
-from transformers import AutoTokenizer
6
+from transformers import PreTrainedTokenizerFast
7
8
from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
9
@@ -85,5 +85,5 @@ def get_deepseek_v32_tokenizer(tokenizer: HfTokenizer) -> HfTokenizer:
85
class DeepseekV32Tokenizer(TokenizerLike):
86
@classmethod
87
def from_pretrained(cls, *args, **kwargs) -> HfTokenizer:
88
- tokenizer = AutoTokenizer.from_pretrained(*args, **kwargs)
88
+ tokenizer = PreTrainedTokenizerFast.from_pretrained(*args, **kwargs)
89
return get_cached_tokenizer(get_deepseek_v32_tokenizer(tokenizer))