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

Disable all two stream overlap on amd (#6475)

Browse Source
This commit is contained in:
Lianmin Zheng
2025-05-20 19:06:59 -07:00
committed by GitHub
parent 66324895c6
commit 03886917bd
3 changed files with 21 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/models/deepseek_v2.py
View File

@@ -76,13 +76,12 @@ from sglang.srt.layers.vocab_parallel_embedding import (
VocabParallelEmbedding,
)
from sglang.srt.managers.expert_distribution import (
ExpertDistributionRecorder,
get_global_expert_distribution_recorder,
)
from sglang.srt.managers.expert_location import ModelConfigForExpertLocation
from sglang.srt.managers.expert_location_dispatch import ExpertLocationDispatchInfo
from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.operations import execute_operations
from sglang.srt.operations_strategy import compute_layer_operations
@@ -1321,8 +1320,7 @@ class DeepseekV2Model(nn.Module):
config.hidden_size,
enable_tp=not global_server_args_dict["enable_dp_attention"],
)
# TODO(haishaw): multi-stream performance on ROCm
self.alt_stream = None if _is_hip else torch.cuda.Stream()
self.alt_stream = torch.cuda.Stream() if _is_cuda else None
self.layers = nn.ModuleList(
[
DeepseekV2DecoderLayer(