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 15, 17:15 UTC
fix: Comments to English for better dev experience (#15768)
Signed-off-by: yihong0618 <zouzou0208@gmail.com>
yihong committed
Mar 31, 2025 at 01:47 UTC
70fedd0f7954079ebee36a7ca834cdf2f3e5d568
2 files changed
+2
-2
vllm/model_executor/models/deepseek_vl2.py
+1
-1
@@ -509,7 +509,7 @@ class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
509
_, hw, n_dim = images_embeds.shape
510
h = w = int(hw**0.5)
511
512
- # 根据self.tile_tag & self.global_view_pos填充image token sequence
512
+ # fill image token based on self.tile_tag & self.global_view_pos
513
tile_index = 0
514
vision_embeddings = []
515
for jdx in range(images_spatial_crop.size(0)):
vllm/transformers_utils/processors/deepseek_vl2.py
+1
-1
@@ -226,7 +226,7 @@ class DeepseekVLV2Processor(ProcessorMixin):
226
input_ids[input_ids < 0] = self.pad_id
227
228
if inference_mode:
229
- # 去掉结尾的eos token
229
+ # Remove the ending eos token
230
assert input_ids[-1] == self.eos_id
231
input_ids = input_ids[:-1]
232
target_ids = target_ids[:-1]