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

config: the dead record parameters go (#36974)

Browse Source 
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Cheng Wan
2026-08-29 04:20:55 -07:00
committed by GitHub
co-authored by Claude Opus 5
parent 1a3e152f03
commit f0d621cfa6
11 changed files with 14 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/arg_groups/cuda_graph_hook.py
+2 -2
View File
@@ -376,7 +376,7 @@ def apply_deepep_adjustments(server_args: Any):
if bs is None:
# 2048 = documented prefill default; max_bs unresolved here.
max_bs = cfg.cuda_graph_config.prefill.max_bs or 2048
bs = generate_prefill_cuda_graph_batch_sizes(server_args, max_bs)
bs = generate_prefill_cuda_graph_batch_sizes(max_bs)
aligned = sorted({((b + 7) // 8) * 8 for b in bs})
if aligned != sorted(bs):
logger.info(
@@ -473,7 +473,7 @@ def validate_cuda_graph_config(server_args: Any):
)
def generate_prefill_cuda_graph_batch_sizes(server_args: Any, max_bs: int):
def generate_prefill_cuda_graph_batch_sizes(max_bs: int):
"""
Generate the list of batch sizes for prefill CUDA graph capture
based on max_bs. For tc_piecewise prefill, bs carries the
python/sglang/srt/arg_groups/memory_hook.py
+1 -1
View File
@@ -206,7 +206,7 @@ def handle_gpu_memory_settings(server_args: Any, gpu_mem):
if prefill_cuda_graph_config.bs is None:
prefill_cuda_graph_config.bs = generate_prefill_cuda_graph_batch_sizes(
server_args, prefill_cuda_graph_config.max_bs
prefill_cuda_graph_config.max_bs
)
if cuda_graph_config != cfg.cuda_graph_config:
python/sglang/srt/arg_groups/model_hook.py
+1 -1
View File
@@ -786,7 +786,7 @@ def handle_model_capability_adjustments(server_args: Any):
}
if (Phase.PREFILL, "bs") not in cuda_graph_config_locked:
sizing["bs"] = generate_prefill_cuda_graph_batch_sizes(
server_args, sizing["max_bs"]
sizing["max_bs"]
)
declare_resolution(
server_args,
python/sglang/srt/arg_groups/parallel_hook.py
+2 -4
View File
@@ -221,7 +221,7 @@ def handle_data_parallelism(server_args: Any):
clamped = {"max_bs": cfg.chunked_prefill_size}
if (Phase.PREFILL, "bs") not in server_args._cuda_graph_config_locked:
clamped["bs"] = generate_prefill_cuda_graph_batch_sizes(
server_args, clamped["max_bs"]
clamped["max_bs"]
)
declare_resolution(
server_args,
@@ -370,9 +370,7 @@ def handle_elastic_ep(server_args: Any):
declare_resolution(
server_args,
"_handle_elastic_ep",
mooncake_ib_device=validate_ib_devices(
server_args, cfg.mooncake_ib_device
),
mooncake_ib_device=validate_ib_devices(cfg.mooncake_ib_device),
)
if cfg.ep_join_mode is not None:
assert (
python/sglang/srt/arg_groups/pd_disaggregation_hook.py
+1 -3
View File
@@ -218,9 +218,7 @@ def handle_encoder_disaggregation(server_args: Any):
declare_resolution(
server_args,
"_handle_encoder_disaggregation",
disaggregation_ib_device=validate_ib_devices(
server_args, cfg.disaggregation_ib_device
),
disaggregation_ib_device=validate_ib_devices(cfg.disaggregation_ib_device),
)
# Validate model type for encoder disaggregation
python/sglang/srt/arg_groups/pipeline.py
+1 -1
View File
@@ -104,7 +104,7 @@ def run_resolution_pipeline(server_args: Any) -> None:
handle_hardware_runtime_validation,
)
handle_hardware_runtime_validation(server_args)
handle_hardware_runtime_validation()
if cfg.model_path.lower() in ["none", "dummy"]:
return
python/sglang/srt/arg_groups/platform_hook.py
+1 -1
View File
@@ -17,7 +17,7 @@ from sglang.srt.utils.common import is_cuda, is_hip, is_host_cpu_arm64, is_npu
logger = logging.getLogger(__name__)
def handle_hardware_runtime_validation(server_args: Any):
def handle_hardware_runtime_validation():
# This is intentionally independent of `server_args.device`: setting
# SGLANG_USE_MLX opts into the MLX backend and must fail immediately if
# the environment cannot honor that request. With the flag unset,
python/sglang/srt/arg_groups/validation_hook.py
+4 -8
View File
@@ -122,12 +122,8 @@ def check_server_args(server_args: Any):
assert cfg.detokenizer_worker_num > 0, "Detokenizer worker num must >= 1"
assert cfg.mm_processor_worker_num >= 0, "Multimodal processor worker num must >= 0"
assert cfg.mm_io_worker_num >= 0, "Multimodal I/O worker num must >= 0"
validate_buckets_rule(
server_args, "--prompt-tokens-buckets", cfg.prompt_tokens_buckets
)
validate_buckets_rule(
server_args, "--generation-tokens-buckets", cfg.generation_tokens_buckets
)
validate_buckets_rule("--prompt-tokens-buckets", cfg.prompt_tokens_buckets)
validate_buckets_rule("--generation-tokens-buckets", cfg.generation_tokens_buckets)
# Check scheduling policy
if cfg.enable_priority_scheduling:
@@ -222,7 +218,7 @@ def check_server_args(server_args: Any):
check_load_publish_args(server_args)
def validate_buckets_rule(server_args: Any, arg_name: str, buckets_rule: List[str]):
def validate_buckets_rule(arg_name: str, buckets_rule: List[str]):
if not buckets_rule:
return
@@ -313,7 +309,7 @@ def check_load_publish_args(server_args: Any):
raise ValueError(reason)
def validate_ib_devices(server_args: Any, device_str: Optional[str]) -> Optional[str]:
def validate_ib_devices(device_str: Optional[str]) -> Optional[str]:
"""
Validate IB devices before passing to mooncake.
python/sglang/srt/mem_cache/hiradix_cache.py
-2
View File
@@ -148,7 +148,6 @@ class HiRadixCache(RadixCache):
attach_hybrid_dsa_pool_to_hiradix_cache(
self,
params,
server_args,
extra_config=extra_config,
prefetch_threshold=prefetch_threshold,
enable_storage_metrics=self.enable_storage_metrics,
@@ -162,7 +161,6 @@ class HiRadixCache(RadixCache):
attach_hybrid_minimax_sparse_pool_to_hiradix_cache(
self,
params,
server_args,
extra_config=extra_config,
prefetch_threshold=prefetch_threshold,
enable_storage_metrics=self.enable_storage_metrics,
python/sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py
-16
View File
@@ -289,7 +289,6 @@ def build_hybrid_swa_group(
def build_kv_only_stack(
*,
params: CacheInitParams,
server_args: ServerArgs,
kv_pool: Any,
full_layer_mapping: dict[int, int],
load_cache_event,
@@ -335,7 +334,6 @@ def build_kv_only_stack(
def build_hybrid_swa_stack(
*,
params: CacheInitParams,
server_args: ServerArgs,
full_kv_pool: Any,
swa_kv_pool: Any,
full_layer_mapping: dict[int, int],
@@ -437,7 +435,6 @@ def _dsv4_compressed_region_buffers(kvcache: Any, ratio: int) -> tuple[list, int
def build_deepseek_v4_hicache_stack(
*,
params: CacheInitParams,
server_args: ServerArgs,
kvcache: Any,
load_cache_event,
storage_backend: Optional[str],
@@ -687,7 +684,6 @@ def build_deepseek_v4_hicache_stack(
def build_hybrid_mamba_stack(
*,
params: CacheInitParams,
server_args: ServerArgs,
kv_pool: Any,
mamba_pool: Any,
full_layer_mapping: dict[int, int],
@@ -781,7 +777,6 @@ def build_hybrid_mamba_stack(
def build_hybrid_mamba_swa_stack(
*,
params: CacheInitParams,
server_args: ServerArgs,
full_kv_pool: Any,
swa_kv_pool: Any,
mamba_pool: Any,
@@ -893,7 +888,6 @@ def build_hybrid_mamba_swa_stack(
def build_anchor_sidecar_stack(
*,
params: CacheInitParams,
server_args: ServerArgs,
kv_pool: Any,
sidecar_pool_name: PoolName,
full_layer_mapping: dict[int, int],
@@ -1214,7 +1208,6 @@ class _DeepSeekV4Strategy(StackStrategy):
):
host_pool_group, cache_controller = build_deepseek_v4_hicache_stack(
params=params,
server_args=server_args,
kvcache=kvcache,
load_cache_event=load_cache_event,
storage_backend=storage_backend,
@@ -1290,7 +1283,6 @@ class _MambaStrategy(StackStrategy):
mamba_layer_mapping = dict(params.req_to_token_pool.mamba_map)
host_pool_group, cache_controller = build_hybrid_mamba_stack(
params=params,
server_args=server_args,
kv_pool=kvcache.full_kv_pool,
mamba_pool=params.req_to_token_pool.mamba_pool,
full_layer_mapping=full_layer_mapping,
@@ -1356,7 +1348,6 @@ class _SwaStrategy(StackStrategy):
full_layer_mapping, swa_layer_mapping = _swa_layer_mappings(kvcache)
host_pool_group, cache_controller = build_hybrid_swa_stack(
params=params,
server_args=server_args,
full_kv_pool=kvcache.full_kv_pool,
swa_kv_pool=kvcache.swa_kv_pool,
full_layer_mapping=full_layer_mapping,
@@ -1417,7 +1408,6 @@ class _MambaSwaStrategy(StackStrategy):
mamba_layer_mapping = dict(params.req_to_token_pool.mamba_map)
host_pool_group, cache_controller = build_hybrid_mamba_swa_stack(
params=params,
server_args=server_args,
full_kv_pool=kvcache.full_kv_pool,
swa_kv_pool=kvcache.swa_kv_pool,
mamba_pool=params.req_to_token_pool.mamba_pool,
@@ -1485,7 +1475,6 @@ class _DsaStrategy(StackStrategy):
full_layer_mapping = {i: i for i in range(full_kv_pool.layer_num)}
host_pool_group, cache_controller = build_anchor_sidecar_stack(
params=params,
server_args=server_args,
kv_pool=full_kv_pool,
sidecar_pool_name=PoolName.INDEXER,
full_layer_mapping=full_layer_mapping,
@@ -1621,7 +1610,6 @@ class _PlainKvStrategy(StackStrategy):
full_layer_mapping = {i: i for i in range(full_kv_pool.layer_num)}
host_pool_group, cache_controller = build_kv_only_stack(
params=params,
server_args=server_args,
kv_pool=full_kv_pool,
full_layer_mapping=full_layer_mapping,
load_cache_event=load_cache_event,
@@ -1828,7 +1816,6 @@ def build_minimax_sparse_hicache_stack(
def attach_hybrid_minimax_sparse_pool_to_hiradix_cache(
radix_cache: HiRadixCache,
params: CacheInitParams,
server_args: ServerArgs,
*,
extra_config: dict,
prefetch_threshold: int,
@@ -1856,7 +1843,6 @@ def attach_hybrid_minimax_sparse_pool_to_hiradix_cache(
if sparse_pool.index_k_pool is None:
host_pool_group, cache_controller = build_kv_only_stack(
params=params,
server_args=server_args,
kv_pool=main_pool,
full_layer_mapping={
layer_id: layer_id for layer_id in range(main_pool.layer_num)
@@ -1902,7 +1888,6 @@ def attach_hybrid_minimax_sparse_pool_to_hiradix_cache(
def attach_hybrid_dsa_pool_to_hiradix_cache(
radix_cache: HiRadixCache,
params: CacheInitParams,
server_args: ServerArgs,
*,
extra_config: dict,
prefetch_threshold: int,
@@ -1918,7 +1903,6 @@ def attach_hybrid_dsa_pool_to_hiradix_cache(
layer_mapping = {layer_id: layer_id for layer_id in range(kv.layer_num)}
host_pool_group, cache_controller = build_anchor_sidecar_stack(
params=params,
server_args=server_args,
kv_pool=kv,
sidecar_pool_name=PoolName.INDEXER,
full_layer_mapping=layer_mapping,