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 17, 06:45 UTC
add support for tokenizer revision (#1163)
Co-authored-by: Zhuohan Li <zhuohan123@gmail.com>
Federico Cassano committed
Oct 2, 2023 at 22:19 UTC
66d18a7fb07d3a9d6e6e46352aea05355a51cfed
5 files changed
+23
-1
vllm/config.py
+5
@@ -41,6 +41,9 @@ class ModelConfig:
41
revision: The specific model version to use. It can be a branch name,
42
a tag name, or a commit id. If unspecified, will use the default
43
version.
44
+ tokenizer_revision: The specific tokenizer version to use. It can be a
45
+ branch name, a tag name, or a commit id. If unspecified, will use
46
+ the default version.
47
max_model_len: Maximum length of a sequence (including prompt and
48
output). If None, will be derived from the model.
49
quantization: Quantization method that was used to quantize the model
@@ -58,6 +61,7 @@ class ModelConfig:
61
dtype: str,
62
seed: int,
63
revision: Optional[str] = None,
64
+ tokenizer_revision: Optional[str] = None,
65
max_model_len: Optional[int] = None,
66
quantization: Optional[str] = None,
67
) -> None:
@@ -69,6 +73,7 @@ class ModelConfig:
73
self.load_format = load_format
74
self.seed = seed
75
self.revision = revision
76
+ self.tokenizer_revision = tokenizer_revision
77
self.quantization = quantization
78
79
self.hf_config = get_config(model, trust_remote_code, revision)
vllm/engine/arg_utils.py
+10
-1
@@ -29,6 +29,7 @@ class EngineArgs:
29
max_num_seqs: int = 256
30
disable_log_stats: bool = False
31
revision: Optional[str] = None
32
+ tokenizer_revision: Optional[str] = None
33
quantization: Optional[str] = None
34
35
def __post_init__(self):
@@ -57,6 +58,13 @@ class EngineArgs:
58
help='the specific model version to use. It can be a branch '
59
'name, a tag name, or a commit id. If unspecified, will use '
60
'the default version.')
61
+ parser.add_argument(
62
+ '--tokenizer-revision',
63
+ type=str,
64
+ default=None,
65
+ help='the specific tokenizer version to use. It can be a branch '
66
+ 'name, a tag name, or a commit id. If unspecified, will use '
67
+ 'the default version.')
68
parser.add_argument('--tokenizer-mode',
69
type=str,
70
default=EngineArgs.tokenizer_mode,
@@ -175,7 +183,8 @@ class EngineArgs:
183
self.tokenizer_mode, self.trust_remote_code,
184
self.download_dir, self.load_format,
185
self.dtype, self.seed, self.revision,
178
- self.max_model_len, self.quantization)
186
+ self.tokenizer_revision, self.max_model_len,
187
+ self.quantization)
188
cache_config = CacheConfig(
189
self.block_size, self.gpu_memory_utilization, self.swap_space,
190
getattr(model_config.hf_config, 'sliding_window', None))
vllm/engine/llm_engine.py
+2
@@ -75,6 +75,7 @@ class LLMEngine:
75
f"tokenizer={model_config.tokenizer!r}, "
76
f"tokenizer_mode={model_config.tokenizer_mode}, "
77
f"revision={model_config.revision}, "
78
+ f"tokenizer_revision={model_config.tokenizer_revision}, "
79
f"trust_remote_code={model_config.trust_remote_code}, "
80
f"dtype={model_config.dtype}, "
81
f"max_seq_len={model_config.max_model_len}, "
@@ -98,6 +99,7 @@ class LLMEngine:
99
model_config.tokenizer,
100
tokenizer_mode=model_config.tokenizer_mode,
101
trust_remote_code=model_config.trust_remote_code,
102
+ tokenizer_revision=model_config.tokenizer_revision,
103
revision=model_config.revision)
104
self.seq_counter = Counter()
105
vllm/entrypoints/llm.py
+4
@@ -42,6 +42,8 @@ class LLM:
42
quantized and use `dtype` to determine the data type of the weights.
43
revision: The specific model version to use. It can be a branch name,
44
a tag name, or a commit id.
45
+ tokenizer_revision: The specific tokenizer version to use. It can be a
46
+ branch name, a tag name, or a commit id.
47
seed: The seed to initialize the random number generator for sampling.
48
gpu_memory_utilization: The ratio (between 0 and 1) of GPU memory to
49
reserve for the model weights, activations, and KV cache. Higher
@@ -65,6 +67,7 @@ class LLM:
67
dtype: str = "auto",
68
quantization: Optional[str] = None,
69
revision: Optional[str] = None,
70
+ tokenizer_revision: Optional[str] = None,
71
seed: int = 0,
72
gpu_memory_utilization: float = 0.9,
73
swap_space: int = 4,
@@ -81,6 +84,7 @@ class LLM:
84
dtype=dtype,
85
quantization=quantization,
86
revision=revision,
87
+ tokenizer_revision=tokenizer_revision,
88
seed=seed,
89
gpu_memory_utilization=gpu_memory_utilization,
90
swap_space=swap_space,
vllm/transformers_utils/tokenizer.py
+2
@@ -16,6 +16,7 @@ def get_tokenizer(
16
*args,
17
tokenizer_mode: str = "auto",
18
trust_remote_code: bool = False,
19
+ tokenizer_revision: Optional[str] = None,
20
**kwargs,
21
) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
22
"""Gets a tokenizer for the given model name via Huggingface."""
@@ -37,6 +38,7 @@ def get_tokenizer(
38
tokenizer_name,
39
*args,
40
trust_remote_code=trust_remote_code,
41
+ tokenizer_revision=tokenizer_revision,
42
**kwargs)
43
except TypeError as e:
44
# The LLaMA tokenizer causes a protobuf error in some environments.