[Tokenizer] Add an option to specify tokenizer #284
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This was
linked to
issues
Jun 28, 2023
|
This pr is very useful, my local test is always hard code tokenizer path |
zhuohan123
approved these changes
Jun 28, 2023
Comment on lines
+20
to
+25
| if "llama" in tokenizer_name.lower() and kwargs.get("use_fast", True): | ||
| logger.info( | ||
| "OpenLLaMA models do not support the fast tokenizer. " | ||
| "Using the slow tokenizer instead.") | ||
| elif config.model_type == "llama" and kwargs.get("use_fast", True): | ||
| # LLaMA fast tokenizer causes protobuf errors in some environments. | ||
| # However, we found that the below LLaMA fast tokenizer works well in | ||
| # most environments. | ||
| model_name = "hf-internal-testing/llama-tokenizer" | ||
| logger.info( | ||
| f"Using the LLaMA fast tokenizer in '{model_name}' to avoid " | ||
| "potential protobuf errors.") | ||
| elif config.model_type in _MODEL_TYPES_WITH_SLOW_TOKENIZER: | ||
| if kwargs.get("use_fast", False) == True: | ||
| raise ValueError( | ||
| f"Cannot use the fast tokenizer for {config.model_type} due to " | ||
| "bugs in the fast tokenizer.") | ||
| logger.info( | ||
| f"Using the slow tokenizer for {config.model_type} due to bugs in " | ||
| "the fast tokenizer. This could potentially lead to performance " | ||
| "degradation.") | ||
| kwargs["use_fast"] = False | ||
| return AutoTokenizer.from_pretrained(model_name, *args, **kwargs) | ||
| "For some LLaMA-based models, initializing the fast tokenizer may " | ||
| "take a long time. To eliminate the initialization time, consider " | ||
| f"using '{_FAST_LLAMA_TOKENIZER}' instead of the original " | ||
| "tokenizer.") |
There was a problem hiding this comment.
After this PR, do we need to manually specify llama to use the fast tokenizer for benchmarking?
There was a problem hiding this comment.
It depends. Actually, LLaMA fast tokenizers in lmsys/vicuna-7b-v1.3 or huggyllama/llama-7b work in my docker environment. So hf-internal-testing/llama-tokenizer is not needed when I use vLLM in my docker environment.
|
wow is cool，bro you are so cool |
|
THANKS VERY MUCH! |
michaelfeil
pushed a commit
to michaelfeil/vllm
that referenced
this pull request
Jul 1, 2023
hongxiayang
pushed a commit
to hongxiayang/vllm
that referenced
this pull request
Feb 13, 2024
sjchoi1
pushed a commit
to casys-kaist-internal/vllm
that referenced
this pull request
May 7, 2024
yukavio
pushed a commit
to yukavio/vllm
that referenced
this pull request
Jul 3, 2024
SUMMARY: * only run 4 x a10 tests for python 3.10.12 NOTE: AWS looks to be having availability issues with these instances. i'm day to day with this repo being migrated to GCP, so in the meantime let's reduce demand. TEST PLAN: n/a Co-authored-by: andy-neuma <andy@neuralmagic.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fixes #111 #246 #259 #270 #281
This PR adds
tokenizerto the input/cli arguments. If it isNone, vLLM uses the model name/path as the tokenizer name/path. In addition, from this PR, vLLM does not usehf-internal-testing/llama-tokenizeras the default tokenizer for llama models.