chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Move torch.compile configs into cuda_graph_runner.py (#993)

Browse Source
This commit is contained in:
Ying Sheng
2024-08-08 13:20:30 -07:00
committed by GitHub
parent ab7875941b
commit 9f662501a3
4 changed files with 15 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
python/sglang/srt/utils.py
View File

@@ -622,19 +622,6 @@ def receive_addrs(model_port_args, server_args):
dist.destroy_process_group()
def set_torch_compile_config():
# The following configurations are for torch compile optimizations
import torch._dynamo.config
import torch._inductor.config
torch._inductor.config.coordinate_descent_tuning = True
torch._inductor.config.triton.unique_kernel_names = True
torch._inductor.config.fx_graph_cache = True # Experimental feature to reduce compilation times, will be on by default in future
# FIXME: tmp workaround
torch._dynamo.config.accumulated_cache_size_limit = 256
def set_ulimit(target_soft_limit=65535):
resource_type = resource.RLIMIT_NOFILE
current_soft, current_hard = resource.getrlimit(resource_type)