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

Cleanup server_args.py and minor code tidying (#22820)

Browse Source
This commit is contained in:
Lianmin Zheng
2026-04-14 18:52:41 -07:00
committed by GitHub
parent ea05ea5abe
commit adb310b976
4 changed files with 62 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+37 -40
View File
@@ -77,7 +77,9 @@ logger = logging.getLogger(__name__)
# Define constants
DEFAULT_UVICORN_ACCESS_LOG_EXCLUDE_PREFIXES = ()
SAMPLING_BACKEND_CHOICES = {"flashinfer", "pytorch", "ascend"}
LOAD_FORMAT_CHOICES = [
"auto",
"pt",
@@ -151,37 +153,13 @@ ATTENTION_BACKEND_CHOICES = [
"intel_xpu",
]
LORA_BACKEND_CHOICES = ["triton", "csgmv", "ascend", "torch_native"]
DISAGG_TRANSFER_BACKEND_CHOICES = ["mooncake", "nixl", "ascend", "fake", "mori"]
ENCODER_TRANSFER_BACKEND_CHOICES = ["zmq_to_scheduler", "zmq_to_tokenizer", "mooncake"]
GRAMMAR_BACKEND_CHOICES = ["xgrammar", "outlines", "llguidance", "none"]
DETERMINISTIC_ATTENTION_BACKEND_CHOICES = ["flashinfer", "fa3", "triton"]
RADIX_SUPPORTED_DETERMINISTIC_ATTENTION_BACKEND = ["fa3", "triton"]
NSA_PREFILL_CP_SPLIT_CHOICES = ["in-seq-split", "round-robin-split"]
DISAGG_TRANSFER_BACKEND_CHOICES = ["mooncake", "nixl", "ascend", "fake", "mori"]
PREFILL_CP_SPLIT_CHOICES = ["in-seq-split"]
DEFAULT_LORA_EVICTION_POLICY = "lru"
NSA_CHOICES = [
"flashmla_sparse",
"flashmla_kv",
"flashmla_auto",
"fa3",
"tilelang",
"aiter",
"trtllm",
]
RADIX_EVICTION_POLICY_CHOICES = ["lru", "lfu", "slru"]
RL_ON_POLICY_TARGET_CHOICES = ["fsdp"]
GRAMMAR_BACKEND_CHOICES = ["xgrammar", "outlines", "llguidance", "none"]
MOE_RUNNER_BACKEND_CHOICES = [
"auto",
@@ -225,11 +203,34 @@ FP4_GEMM_RUNNER_BACKEND_CHOICES = [
"flashinfer_trtllm",
]
MAMBA_SSM_DTYPE_CHOICES = ["float32", "bfloat16", "float16"]
RADIX_EVICTION_POLICY_CHOICES = ["lru", "lfu", "slru"]
RL_ON_POLICY_TARGET_CHOICES = ["fsdp"]
LORA_BACKEND_CHOICES = ["triton", "csgmv", "ascend", "torch_native"]
ENCODER_TRANSFER_BACKEND_CHOICES = ["zmq_to_scheduler", "zmq_to_tokenizer", "mooncake"]
NSA_PREFILL_CP_SPLIT_CHOICES = ["in-seq-split", "round-robin-split"]
PREFILL_CP_SPLIT_CHOICES = ["in-seq-split"]
DEFAULT_LORA_EVICTION_POLICY = "lru"
NSA_CHOICES = [
"flashmla_sparse",
"flashmla_kv",
"flashmla_auto",
"fa3",
"tilelang",
"aiter",
"trtllm",
]
MAMBA_SCHEDULER_STRATEGY_CHOICES = ["auto", "no_buffer", "extra_buffer"]
MAMBA_BACKEND_CHOICES = ["triton", "flashinfer"]
LINEAR_ATTN_KERNEL_BACKEND_CHOICES = ["triton", "cutedsl", "flashinfer"]
@@ -246,6 +247,14 @@ def add_attention_backend_choices(choices):
ATTENTION_BACKEND_CHOICES.extend(choices)
def add_deterministic_attention_backend_choices(choices):
DETERMINISTIC_ATTENTION_BACKEND_CHOICES.extend(choices)
def add_radix_supported_deterministic_attention_backend_choices(choices):
RADIX_SUPPORTED_DETERMINISTIC_ATTENTION_BACKEND.extend(choices)
def add_disagg_transfer_backend_choices(choices):
DISAGG_TRANSFER_BACKEND_CHOICES.extend(choices)
@@ -266,14 +275,6 @@ def add_fp4_gemm_runner_backend_choices(choices):
FP4_GEMM_RUNNER_BACKEND_CHOICES.extend(choices)
def add_deterministic_attention_backend_choices(choices):
DETERMINISTIC_ATTENTION_BACKEND_CHOICES.extend(choices)
def add_radix_supported_deterministic_attention_backend_choices(choices):
RADIX_SUPPORTED_DETERMINISTIC_ATTENTION_BACKEND.extend(choices)
def add_radix_eviction_policy_choices(choices):
RADIX_EVICTION_POLICY_CHOICES.extend(choices)
@@ -282,10 +283,6 @@ def add_rl_on_policy_target_choices(choices):
RL_ON_POLICY_TARGET_CHOICES.extend(choices)
def add_mamba_ssm_dtype_choices(choices):
MAMBA_SSM_DTYPE_CHOICES.extend(choices)
@dataclasses.dataclass
class ServerArgs:
"""
@@ -5416,7 +5413,7 @@ class ServerArgs:
"--mamba-ssm-dtype",
type=str,
default=None,
choices=MAMBA_SSM_DTYPE_CHOICES,
choices=["float32", "bfloat16", "float16"],
help="The data type of the SSM states in mamba cache. "
"If not set, will be read from model config (mamba_ssm_dtype).",
)