minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix missing dp_max_padding argument in set_dp_buffer_len (#12812)

Browse Source
This commit is contained in:
Chen1022
2025-11-07 21:27:43 +08:00
committed by GitHub
parent bef37d6de8
commit c67fce160e
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
+5 -1
View File
@@ -382,7 +382,11 @@ class PiecewiseCudaGraphRunner:
def run_once(): def run_once():
# Clean intermediate result cache for DP attention # Clean intermediate result cache for DP attention
forward_batch.dp_local_start_pos = forward_batch.dp_local_num_tokens = None forward_batch.dp_local_start_pos = forward_batch.dp_local_num_tokens = None
set_dp_buffer_len(global_dp_buffer_len, num_tokens) set_dp_buffer_len(
global_dp_buffer_len,
num_tokens,
forward_batch.dp_padding_mode.is_max_len(),
)
# FIXME: the implementation is hacky. `is_extend_in_batch`` is for determining the deepep mode. # FIXME: the implementation is hacky. `is_extend_in_batch`` is for determining the deepep mode.
# It is True in this context but we need to set it to use low latency deepep mode. # It is True in this context but we need to set it to use low latency deepep mode.
set_is_extend_in_batch(False) set_is_extend_in_batch(False)