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, 08:15 UTC
[Bugfix] Fix edge-case crash when using chat with the Mistral Tekken Tokenizer (#10051)
Signed-off-by: Travis Johnson <tsjohnso@us.ibm.com>
Travis Johnson committed
Nov 5, 2024 at 21:28 UTC
2bcbae704c0d52913c6a2887260fc6bde6c20361
2 files changed
+12
-5
tests/models/decoder_only/language/test_mistral.py
+6
-3
@@ -10,19 +10,22 @@ from ...utils import check_logprobs_close
10
11
MODELS = [
12
"mistralai/Mistral-7B-Instruct-v0.1",
13
- "mistralai/Mistral-7B-Instruct-v0.3",
14
- # Mistral-Nemo is to big for CI, but passes locally
15
- # "mistralai/Mistral-Nemo-Instruct-2407"
13
]
14
15
MISTRAL_FORMAT_MODELS = [
16
"mistralai/Mistral-7B-Instruct-v0.3",
17
+ # uses the v3-Tekken tokenizer
18
+ "mistralai/Ministral-8B-Instruct-2410",
19
+ # Mistral-Nemo is to big for CI, but passes locally
20
+ # "mistralai/Mistral-Nemo-Instruct-2407"
21
]
22
23
SAMPLING_PARAMS = SamplingParams(max_tokens=512, temperature=0.0, logprobs=5)
24
SYMBOLIC_LANG_PROMPTS = [
25
"勇敢な船乗りについての詩を書く", # japanese
26
"寫一首關於勇敢的水手的詩", # chinese
27
+ "ပုံပြင်လေးပြောပြပါ်:\n", # burmese
28
+ "Repeat the phrase 'URGENCY🌶️':\nURGENCY🌶️\nURGENCY🌶️\n", # see https://github.com/vllm-project/vllm/pull/9625
29
]
30
31
# for function calling
vllm/transformers_utils/tokenizers/mistral.py
+6
-2
@@ -254,7 +254,7 @@ class MistralTokenizer:
254
skip_special_tokens: bool = True) -> str:
255
assert (
256
skip_special_tokens
257
- ), "Skipping special tokens is not supported for Mistral tokenizers."
257
+ ), "skip_special_tokens=False is not supported for Mistral tokenizers."
258
259
if isinstance(ids, int):
260
ids = [ids]
@@ -268,12 +268,16 @@ class MistralTokenizer:
268
# TODO(Patrick) - potentially allow special tokens to not be skipped
269
assert (
270
skip_special_tokens
271
- ), "Skipping special tokens is not supported for Mistral tokenizers."
271
+ ), "skip_special_tokens=False is not supported for Mistral tokenizers."
272
273
assert isinstance(self.tokenizer,
274
(Tekkenizer, SentencePieceTokenizer)), type(
275
self.tokenizer)
276
277
+ if isinstance(self.tokenizer, Tekkenizer):
278
+ # skip special tokens
279
+ ids = [i for i in ids if i > self.tokenizer.num_special_tokens]
280
+
281
tokens = [self.tokenizer.id_to_piece(id) for id in ids]
282
283
if any("�" in t for t in tokens):