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

Refactor global_server_args_dict (#6866)

Browse Source
This commit is contained in:
fzyzcjy
2025-06-07 18:10:35 +08:00
committed by GitHub
parent c499591ac8
commit f5599ef124
2 changed files with 38 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
python/sglang/srt/managers/schedule_batch.py
View File

@@ -70,33 +70,38 @@ if TYPE_CHECKING:
INIT_INCREMENTAL_DETOKENIZATION_OFFSET = 5
GLOBAL_SERVER_ARGS_KEYS = [
"attention_backend",
"debug_tensor_dump_inject",
"debug_tensor_dump_output_folder",
"chunked_prefill_size",
"deepep_mode",
"device",
"disable_chunked_prefix_cache",
"disable_radix_cache",
"enable_deepep_moe",
"enable_dp_attention",
"enable_two_batch_overlap",
"enable_dp_lm_head",
"enable_ep_moe",
"deepep_config",
"enable_nan_detection",
"flashinfer_mla_disable_ragged",
"max_micro_batch_size",
"moe_dense_tp_size",
"ep_dispatch_algorithm",
"disable_shared_experts_fusion",
"sampling_backend",
"speculative_accept_threshold_acc",
"speculative_accept_threshold_single",
"torchao_config",
"triton_attention_reduce_in_fp32",
"ep_num_redundant_experts",
"mm_attention_backend",
]
# Put some global args for easy access
global_server_args_dict = {
"attention_backend": ServerArgs.attention_backend,
"chunked_prefill_size": ServerArgs.chunked_prefill_size,
"deepep_mode": ServerArgs.deepep_mode,
"device": ServerArgs.device,
"disable_chunked_prefix_cache": ServerArgs.disable_chunked_prefix_cache,
"disable_radix_cache": ServerArgs.disable_radix_cache,
"enable_deepep_moe": ServerArgs.enable_deepep_moe,
"enable_dp_attention": ServerArgs.enable_dp_attention,
"enable_two_batch_overlap": ServerArgs.enable_two_batch_overlap,
"enable_dp_lm_head": ServerArgs.enable_dp_lm_head,
"enable_ep_moe": ServerArgs.enable_ep_moe,
"deepep_config": ServerArgs.deepep_config,
"enable_nan_detection": ServerArgs.enable_nan_detection,
"flashinfer_mla_disable_ragged": ServerArgs.flashinfer_mla_disable_ragged,
"max_micro_batch_size": ServerArgs.max_micro_batch_size,
"moe_dense_tp_size": ServerArgs.moe_dense_tp_size,
"ep_dispatch_algorithm": ServerArgs.ep_dispatch_algorithm,
"disable_shared_experts_fusion": ServerArgs.disable_shared_experts_fusion,
"sampling_backend": ServerArgs.sampling_backend,
"speculative_accept_threshold_acc": ServerArgs.speculative_accept_threshold_acc,
"speculative_accept_threshold_single": ServerArgs.speculative_accept_threshold_single,
"torchao_config": ServerArgs.torchao_config,
"triton_attention_reduce_in_fp32": ServerArgs.triton_attention_reduce_in_fp32,
"ep_num_redundant_experts": ServerArgs.ep_num_redundant_experts,
}
global_server_args_dict = {k: getattr(ServerArgs, k) for k in GLOBAL_SERVER_ARGS_KEYS}
logger = logging.getLogger(__name__)