EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Initialize image processor for skip-tokenizer-init codepath (#4479)

Browse Source 
Co-authored-by: Alex Kirillov <alex@iterationlab.org>
This commit is contained in:
Yinghai Lu
2025-03-16 18:10:09 -07:00
committed by GitHub
parent 9d02bb3e2a
commit 2e4a1e2d05
1 changed files with 22 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
python/sglang/srt/managers/tokenizer_manager.py
View File

@@ -168,27 +168,32 @@ class TokenizerManager:
self.context_len = self.model_config.context_len self.context_len = self.model_config.context_len
self.image_token_id = self.model_config.image_token_id self.image_token_id = self.model_config.image_token_id
# Create image processor placeholder if self.model_config.is_multimodal:
self.image_processor = get_dummy_image_processor() _processor = get_processor(
server_args.tokenizer_path,
tokenizer_mode=server_args.tokenizer_mode,
trust_remote_code=server_args.trust_remote_code,
revision=server_args.revision,
)
# Create tokenizer # We want to parallelize the image pre-processing so we create an executor for it
if server_args.skip_tokenizer_init: # We creat image_processor for any skip_tokenizer_init to make sure we still encode
self.tokenizer = self.processor = None # images even with skip_tokenizer_init=False.
else: self.image_processor = get_image_processor(
if self.model_config.is_multimodal: self.model_config.hf_config, server_args, _processor
self.processor = get_processor( )
server_args.tokenizer_path,
tokenizer_mode=server_args.tokenizer_mode, if server_args.skip_tokenizer_init:
trust_remote_code=server_args.trust_remote_code, self.tokenizer = self.processor = None
revision=server_args.revision, else:
) self.processor = _processor
self.tokenizer = self.processor.tokenizer self.tokenizer = self.processor.tokenizer
os.environ["TOKENIZERS_PARALLELISM"] = "false" os.environ["TOKENIZERS_PARALLELISM"] = "false"
else:
self.image_processor = get_dummy_image_processor()
# We want to parallelize the image pre-processing so we create an executor for it if server_args.skip_tokenizer_init:
self.image_processor = get_image_processor( self.tokenizer = self.processor = None
self.model_config.hf_config, server_args, self.processor
)
else: else:
self.tokenizer = get_tokenizer( self.tokenizer = get_tokenizer(
server_args.tokenizer_path, server_args.tokenizer_path,