Skip to content

Refactor global_server_args_dict #6866

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
zhyncs merged 9 commits into from
Jun 7, 2025
Merged

Refactor global_server_args_dict #6866

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
5cc3937
more
fzyzcjy Jun 4, 2025
a6489ab
more
fzyzcjy Jun 4, 2025
0e86679
more
fzyzcjy Jun 4, 2025
fd3c474
more
fzyzcjy Jun 4, 2025
597414b
fmt
fzyzcjy Jun 4, 2025
8bf9367
more
fzyzcjy Jun 4, 2025
b4e6bfc
Merge branch 'main-upstream' into feat/refactor_args_dict
fzyzcjy Jun 5, 2025
038645a
more
fzyzcjy Jun 5, 2025
bc657be
Update schedule_batch.py
fzyzcjy Jun 5, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
57 changes: 31 additions & 26 deletions python/sglang/srt/managers/schedule_batch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,33 +70,38 @@

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",
"num_fused_shared_experts",
"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,
"num_fused_shared_experts": ServerArgs.num_fused_shared_experts,
"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}
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Jun 4, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

This line initializes a module-level global dictionary global_server_args_dict. This dictionary is later updated in model_runner.py. Using mutable global state initialized at module load time and updated elsewhere can lead to issues with initialization order and make the system harder to reason about. If other parts of the code access this dictionary before the update in model_runner.py, they might get default or incomplete configuration values. Consider managing configuration via explicit object passing or a properly initialized singleton.


logger = logging.getLogger(__name__)

Expand Down
33 changes: 6 additions & 27 deletions python/sglang/srt/model_executor/model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,7 +64,10 @@
get_global_expert_location_metadata,
set_global_expert_location_metadata,
)
from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.managers.schedule_batch import (
GLOBAL_SERVER_ARGS_KEYS,
global_server_args_dict,
)
from sglang.srt.mem_cache.memory_pool import (
DoubleSparseTokenToKVPool,
MHATokenToKVPool,
Expand Down Expand Up @@ -186,33 +189,9 @@ def __init__(

# Global vars
global_server_args_dict.update(
{
"attention_backend": server_args.attention_backend,
"debug_tensor_dump_inject": server_args.debug_tensor_dump_inject,
"debug_tensor_dump_output_folder": server_args.debug_tensor_dump_output_folder,
"deepep_mode": server_args.deepep_mode,
"device": server_args.device,
"disable_chunked_prefix_cache": server_args.disable_chunked_prefix_cache,
"disable_radix_cache": server_args.disable_radix_cache,
"enable_nan_detection": server_args.enable_nan_detection,
"enable_dp_attention": server_args.enable_dp_attention,
"enable_two_batch_overlap": server_args.enable_two_batch_overlap,
"enable_dp_lm_head": server_args.enable_dp_lm_head,
"enable_ep_moe": server_args.enable_ep_moe,
"enable_deepep_moe": server_args.enable_deepep_moe,
"deepep_config": server_args.deepep_config,
"flashinfer_mla_disable_ragged": server_args.flashinfer_mla_disable_ragged,
"moe_dense_tp_size": server_args.moe_dense_tp_size,
"ep_dispatch_algorithm": server_args.ep_dispatch_algorithm,
"num_fused_shared_experts": server_args.num_fused_shared_experts,
"triton_attention_reduce_in_fp32": server_args.triton_attention_reduce_in_fp32,
"torchao_config": server_args.torchao_config,
"sampling_backend": server_args.sampling_backend,
"speculative_accept_threshold_single": server_args.speculative_accept_threshold_single,
"speculative_accept_threshold_acc": server_args.speculative_accept_threshold_acc,
{k: getattr(server_args, k) for k in GLOBAL_SERVER_ARGS_KEYS}
| {
"use_mla_backend": self.use_mla_backend,
"mm_attention_backend": server_args.mm_attention_backend,
"ep_num_redundant_experts": server_args.ep_num_redundant_experts,
}
)
Comment on lines 191 to 198
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Jun 4, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

This code updates the global global_server_args_dict initialized in schedule_batch.py. This pattern of updating a global mutable state initialized elsewhere can lead to dependencies on import order and make the system harder to understand and test. It's generally preferable to manage configuration via explicit object passing or a properly initialized singleton.


Expand Down
Loading