Fix mrope_positions size when req is retracted (#13700)
Signed-off-by: Kun(llfl) <i@imux.top> Co-authored-by: Xuchun Shang <xuchun.shang@gmail.com>
This commit is contained in:
@@ -810,6 +810,22 @@ class Req:
|
||||
match_result.host_hit_length,
|
||||
)
|
||||
self.cache_protected_len = len(self.prefix_indices)
|
||||
|
||||
if (
|
||||
self.is_retracted
|
||||
and self.multimodal_inputs is not None
|
||||
and self.multimodal_inputs.mrope_positions is not None
|
||||
):
|
||||
from sglang.srt.managers.mm_utils import (
|
||||
extend_mrope_positions_for_retracted_request,
|
||||
)
|
||||
|
||||
self.multimodal_inputs.mrope_positions = (
|
||||
extend_mrope_positions_for_retracted_request(
|
||||
self.multimodal_inputs.mrope_positions, len(self.output_ids)
|
||||
)
|
||||
)
|
||||
|
||||
self.extend_input_len = len(self.fill_ids) - len(self.prefix_indices)
|
||||
|
||||
# Based on https://github.com/vllm-project/vllm/blob/7a64d24aad69e4d2548aa0bf528d9fe63428ab01/vllm/transformers_utils/detokenizer.py#L194-L313
|
||||
|
||||
Reference in New Issue
Block a user