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

Revert "Deprecate global_server_args_dict" (#11520)

Browse Source
This commit is contained in:
Cheng Wan
2025-10-12 17:40:40 -07:00
committed by GitHub
parent 6cd296940a
commit 1bdd010291
54 changed files with 321 additions and 240 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
python/sglang/srt/mem_cache/common.py
View File

@@ -11,7 +11,7 @@ from sglang.srt.mem_cache.allocator import SWATokenToKVPoolAllocator
from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
from sglang.srt.mem_cache.chunk_cache import ChunkCache, SWAChunkCache
from sglang.srt.mem_cache.memory_pool import HybridReqToTokenPool, ReqToTokenPool
from sglang.srt.server_args import get_global_server_args
from sglang.srt.server_args import ServerArgs
from sglang.srt.utils import support_triton
if TYPE_CHECKING:
@@ -19,6 +19,10 @@ if TYPE_CHECKING:
logger = logging.getLogger(__name__)
GLOBAL_SERVER_ARGS_KEYS = ["attention_backend"]
global_server_args_dict = {k: getattr(ServerArgs, k) for k in GLOBAL_SERVER_ARGS_KEYS}
@triton.jit
def write_req_to_token_pool_triton(
@@ -84,7 +88,7 @@ def write_cache_indices(
prefix_tensors: list[torch.Tensor],
req_to_token_pool: ReqToTokenPool,
):
if support_triton(get_global_server_args().attention_backend):
if support_triton(global_server_args_dict.get("attention_backend")):
prefix_pointers = torch.tensor(
[t.data_ptr() for t in prefix_tensors],
device=req_to_token_pool.device,
@@ -125,8 +129,8 @@ def get_last_loc(
prefix_lens_tensor: torch.Tensor,
) -> torch.Tensor:
if (
get_global_server_args().attention_backend != "ascend"
and get_global_server_args().attention_backend != "torch_native"
global_server_args_dict["attention_backend"] != "ascend"
and global_server_args_dict["attention_backend"] != "torch_native"
):
impl = get_last_loc_triton
else: