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

Register allgather/reducescatter buffers with symm memory (#12572)

Browse Source
This commit is contained in:
Nicolas Castet
2025-11-04 19:11:36 -06:00
committed by GitHub
parent 1357ab025a
commit 2340798353
19 changed files with 250 additions and 114 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -660,7 +660,11 @@ class CudaGraphRunner:
def run_once():
# Clean intermediate result cache for DP attention
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(),
)
set_is_extend_in_batch(False)
kwargs = {}

4
python/sglang/srt/model_executor/forward_batch_info.py
View File

@@ -719,7 +719,9 @@ class ForwardBatch:
num_tokens = global_num_tokens[0]
self.global_dp_buffer_len = buffer_len
set_dp_buffer_len(buffer_len, num_tokens, global_num_tokens)
set_dp_buffer_len(
buffer_len, num_tokens, dp_padding_mode.is_max_len(), global_num_tokens
)
set_is_extend_in_batch(self.is_extend_in_batch)
bs = self.batch_size