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

Save peak memory in logits processor (#8343)

Browse Source
This commit is contained in:
Cheng Wan
2025-07-26 01:46:42 -07:00
committed by GitHub
parent 4fa44d63c6
commit e236d8fee8
1 changed files with 14 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
python/sglang/srt/layers/logits_processor.py
View File

@@ -170,8 +170,6 @@ class LogitsMetadata:
)
def compute_dp_attention_metadata(self):
# TODO(ch-wan): gathered_buffer here is larger than the actual required size in draft extend,
# we may use a smaller buffer in draft extend.
cumtokens = torch.cumsum(self.global_num_tokens_for_logprob_gpu, dim=0)
dp_rank = get_attention_dp_rank()
@@ -186,6 +184,19 @@ class LogitsMetadata:
self.dp_local_start_pos = dp_local_start_pos
self.dp_local_num_tokens = dp_local_num_tokens
if self.global_num_tokens_for_logprob_cpu is not None:
# create a smaller buffer to reduce peak memory usage
self.gathered_buffer = torch.empty(
(
sum(self.global_num_tokens_for_logprob_cpu),
self.gathered_buffer.shape[1],
),
dtype=self.gathered_buffer.dtype,
device=self.gathered_buffer.device,
)
else:
self.gathered_buffer = torch.empty_like(self.gathered_buffer)
class LogitsProcessor(nn.Module):
def __init__(
@@ -430,7 +441,7 @@ class LogitsProcessor(nn.Module):
if self.do_tensor_parallel_all_gather_dp_attn:
logits_metadata.compute_dp_attention_metadata()
hidden_states, local_hidden_states = (
torch.empty_like(logits_metadata.gathered_buffer),
logits_metadata.gathered_buffer,
hidden_states,
)
dp_gather_replicate(hidden_states, local_hidden_states, logits_metadata)