biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Revert "[Renderer] Separate out RendererConfig from ModelConfig (#30145)" (#30199)

Browse Source
This commit is contained in:
Cyrus Leung
2025-12-07 16:00:22 +08:00
committed by GitHub
parent 27f4c2fd46
commit e83b7e379c
105 changed files with 797 additions and 969 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/model_executor/models/nano_nemotron_vl.py
View File

@@ -1169,17 +1169,16 @@ class NemotronH_Nano_VL_V2(
self.mlp1 = self.mlp1.to(self.language_model.config.dtype)
self.config = config
self.model_config = vllm_config.model_config
# Pre-tokenize special tokens for video processing
# to avoid repeated tokenization
self._tokenizer = cached_tokenizer_from_config(vllm_config.renderer_config)
self._img_start_token_ids = self._tokenizer.encode(
tokenizer = cached_tokenizer_from_config(vllm_config.model_config)
self._img_start_token_ids = tokenizer.encode(
IMG_START, add_special_tokens=False
)
self._img_end_token_ids = self._tokenizer.encode(
IMG_END, add_special_tokens=False
)
self._img_context_token_ids = self._tokenizer.encode(
self._img_end_token_ids = tokenizer.encode(IMG_END, add_special_tokens=False)
self._img_context_token_ids = tokenizer.encode(
IMG_CONTEXT, add_special_tokens=False
)
@@ -1365,7 +1364,7 @@ class NemotronH_Nano_VL_V2(
input_embeds for the LLM.
"""
device = video_embeddings.device
tokenizer = self._tokenizer
tokenizer = cached_tokenizer_from_config(self.model_config)
# Generate video replacement token IDs using get_video_repl
# This tokenizes each frame separator independently, then uses pre-tokenized