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

fix: second_per_grid_ts should be used to get mrope position (#3682)

Browse Source
This commit is contained in:
Mick
2025-03-18 09:12:38 +08:00
committed by GitHub
parent 98be3bd306
commit d373a48c98
8 changed files with 93 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/model_executor/forward_batch_info.py
View File

@@ -402,9 +402,16 @@ class ForwardBatch:
extend_start_loc : extend_start_loc + extend_seq_len
],
image_grid_thw=image_inputs.image_grid_thws,
video_grid_thw=image_inputs.video_grid_thws,
image_token_id=image_inputs.im_token_id,
video_token_id=image_inputs.video_token_id,
vision_start_token_id=hf_config.vision_start_token_id,
vision_end_token_id=hf_config.vision_end_token_id,
spatial_merge_size=hf_config.vision_config.spatial_merge_size,
context_len=0,
seq_len=len(self.input_ids),
second_per_grid_ts=image_inputs.second_per_grid_ts,
tokens_per_second=hf_config.vision_config.tokens_per_second,
)
)
batch.image_inputs[i].mrope_position_delta = mrope_position_delta

6
python/sglang/srt/model_executor/model_runner.py
View File

@@ -258,10 +258,12 @@ class ModelRunner:
if self.model_config.hf_config.architectures == [
"Qwen2VLForConditionalGeneration"
] or self.model_config.hf_config.architectures == [
"Qwen2_5_VLForConditionalGeneration"
]:
# TODO: qwen2-vl does not support radix cache now, set disable_radix_cache=True automatically
# TODO: qwen2-vl series does not support radix cache now, set disable_radix_cache=True automatically
logger.info(
"Automatically turn off --chunked-prefill-size and disable radix cache for qwen2-vl."
"Automatically turn off --chunked-prefill-size and disable radix cache for qwen-vl series."
)
server_args.chunked_prefill_size = -1
server_args.disable_radix_cache = True