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

config: stop handing the record to code that does not read it (#36252)

Browse Source
This commit is contained in:
Cheng Wan
2026-08-26 05:02:17 -07:00
committed by GitHub
parent d7b144f64e
commit ae5feb4b9c
60 changed files with 168 additions and 280 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/unit/entrypoints/test_http_server_warmup.py
-1
View File
@@ -57,7 +57,6 @@ class TestDisaggregationServerWarmup(unittest.IsolatedAsyncioTestCase):
with patch("sglang.srt.entrypoints.http_server.aiohttp.ClientSession", Session):
status_codes = await _send_disaggregation_warmup_requests(
server_args=server_args,
url="http://localhost:30000",
headers={"Authorization": "Bearer token"},
ssl_verify=False,
test/registered/unit/hardware_backend/mlx/test_attention_patching.py
-4
View File
@@ -1200,10 +1200,6 @@ class TestMlxOverlapScheduler(unittest.TestCase):
disaggregation_mode=None,
enable_overlap=False,
enable_overlap_mlx=False,
server_args=SimpleNamespace(
disaggregation_decode_enable_offload_kvcache=False,
enable_hisparse=False,
),
model_config=None,
token_to_kv_pool_allocator=None,
tree_cache=tree_cache,
test/registered/unit/managers/test_batch_result_processor_hidden_states.py
-4
View File
@@ -31,10 +31,6 @@ def _make_processor(case, server_mode: str = "full") -> SchedulerBatchResultProc
disaggregation_mode=None,
enable_overlap=False,
enable_overlap_mlx=False,
server_args=SimpleNamespace(
enable_metrics=False,
enable_hisparse=False,
),
model_config=SimpleNamespace(think_end_ids=None),
token_to_kv_pool_allocator=Mock(),
tree_cache=None,
test/registered/unit/managers/test_batch_result_processor_mamba_boundary.py
-1
View File
@@ -60,7 +60,6 @@ def _make_processor() -> SchedulerBatchResultProcessor:
disaggregation_mode=None,
enable_overlap=True,
enable_overlap_mlx=False,
server_args=SimpleNamespace(),
model_config=SimpleNamespace(think_end_ids=None),
token_to_kv_pool_allocator=MagicMock(),
tree_cache=SimpleNamespace(page_size=TRACK_INTERVAL),
test/registered/unit/managers/test_batch_result_processor_spec_grammar.py
-1
View File
@@ -65,7 +65,6 @@ def _make_processor() -> SchedulerBatchResultProcessor:
disaggregation_mode=None,
enable_overlap=False,
enable_overlap_mlx=False,
server_args=SimpleNamespace(enable_metrics=False),
model_config=SimpleNamespace(think_end_ids=None),
token_to_kv_pool_allocator=None,
tree_cache=None,
test/registered/unit/managers/test_mm_process_config.py
+6 -7
View File
@@ -81,21 +81,20 @@ class TestBaseProcessorConfigExtraction(CustomTestCase):
BaseMultimodalProcessor,
)
# The multimodal config comes from the bags.
override = get_context().override_server_args(
mm_process_config=mm_process_config,
allowed_media_domains=[],
mm_processor_worker_num=mm_processor_worker_num,
mm_io_worker_num=mm_io_worker_num,
mm_preprocess_cache_size_mb=None,
tokenizer_worker_num=1,
trust_mm_content_hashes=False,
media_url_max_file_size_mb=64,
)
override.install()
self.addCleanup(override.restore)
server_args = MagicMock()
server_args.mm_processor_worker_num = mm_processor_worker_num
server_args.mm_io_worker_num = mm_io_worker_num
server_args.mm_preprocess_cache_size_mb = None
server_args.tokenizer_worker_num = 1
server_args.trust_mm_content_hashes = False
server_args.media_url_max_file_size_mb = 64
hf_config = MagicMock()
mock_hf_processor = MagicMock()
test/registered/unit/model_executor/model_runner_components/test_startup_weight_load.py
+1 -2
View File
@@ -209,8 +209,7 @@ class TestStartupWeightLoadSelector(CustomTestCase):
# The parallel sizes come from the bags, so the config has to be published.
publish(server_args, role="test")
self.addCleanup(reset_context)
options = StartupWeightLoadOptions.from_server_args(
server_args=server_args,
options = StartupWeightLoadOptions.from_published_config(
is_draft_worker=False,
)
test/registered/unit/models/test_kimi_k25.py
+8 -6
View File
@@ -648,6 +648,14 @@ def _k3_preprocess_config(
)
def test_kimi_processor_workers_clone_the_gpu_wrapper(processor_cls, wrapper_cls):
server_args = SimpleNamespace(
base_gpu_id=0,
rl_on_policy_target=None,
tp_size=1,
)
with get_context().override_server_args(
mm_feature_transport="cpu",
mm_process_config={},
allowed_media_domains=[],
image_processor_backend="auto",
disable_fast_image_processor=False,
skip_tokenizer_init=False,
@@ -656,13 +664,7 @@ def test_kimi_processor_workers_clone_the_gpu_wrapper(processor_cls, wrapper_cls
tokenizer_worker_num=1,
mm_preprocess_cache_size_mb=0,
trust_mm_content_hashes=False,
base_gpu_id=0,
rl_on_policy_target=None,
media_url_max_file_size_mb=64,
)
# The multimodal config comes from the bags.
with get_context().override_server_args(
mm_feature_transport="cpu", mm_process_config={}, allowed_media_domains=[]
):
processor = processor_cls(
hf_config=SimpleNamespace(media_placeholder_token_id=42),
test/registered/unit/multimodal/rust/qwen/_fixtures.py
+3 -2
View File
@@ -89,14 +89,15 @@ def make_processor(case, config, image_processor_cls=None):
allowed_media_domains=[],
media_url_max_file_size_mb=64,
)
# The processor reads its media policy, transport and per-modality limits
# from the mm bag, so the fixture publishes before building it.
# Left at the default backend, the fast image processor sends the tensor to
# `cuda:<base_gpu_id>`, which a CPU-only host cannot do.
publish(
ServerArgs(
model_path="dummy",
mm_feature_transport=server_args.mm_feature_transport,
mm_process_config=server_args.mm_process_config,
allowed_media_domains=server_args.allowed_media_domains,
disable_fast_image_processor=server_args.disable_fast_image_processor,
),
role="tokenizer",
)
test/registered/unit/spec/test_draft_per_runner_config.py
+1 -2
View File
@@ -134,14 +134,13 @@ class TestDraftPerRunnerConfig(CustomTestCase):
def test_an_unresolved_draft_falls_back_to_the_config_field(self):
"""The v2 workers pass no backend: --speculative-draft-attention-backend."""
server_args = self._seed(
self._seed(
attention_backend="fa3", speculative_draft_attention_backend="triton"
)
def effective(*, is_draft_worker, passed=None):
return resolve_draft_attention_backend(
draft_attention_backend=passed,
server_args=server_args,
is_draft_worker=is_draft_worker,
)
test/registered/unit/test_supplied_instance_exposure_ratchet.py
-5
View File
@@ -137,7 +137,6 @@ _PASSED = frozenset({"model_path", "device", "random_seed"})
_EXPOSED = {
("dllm/config.py", "max_running_requests"),
("dllm/config.py", "model_path"),
("multimodal/processors/base_processor.py", "image_processor_backend"),
("speculative/spec_registry.py", "disable_overlap_schedule"),
("disaggregation/encoder/server.py", "model_loader_extra_config"),
("layers/moe/utils.py", "deepep_mode"),
@@ -165,8 +164,6 @@ _EXPOSED = {
("entrypoints/engine.py", "enable_symm_mem"),
("entrypoints/engine.py", "reasoning_parser"),
("entrypoints/engine.py", "tool_call_parser"),
("eplb/eplb_manager.py", "ep_dispatch_algorithm"),
("eplb/eplb_manager.py", "expert_distribution_recorder_buffer_size"),
("layers/cp/base.py", "attn_cp_size"),
("layers/cp/base.py", "cp_strategy"),
("layers/cp/base.py", "enable_prefill_cp"),
@@ -178,11 +175,9 @@ _EXPOSED = {
("layers/moe/utils.py", "moe_runner_backend"),
("layers/moe/utils.py", "quantization"),
("layers/moe/utils.py", "speculative_moe_runner_backend"),
("lora/lora_manager.py", "enable_lora_overlap_loading"),
("lora/marlin_lora_temp/policy.py", "lora_paths"),
("model_loader/expert_pack_runtime.py", "model_path"),
("model_loader/expert_pack_runtime.py", "tokenizer_path"),
("multimodal/processors/base_processor.py", "image_processor_backend"),
("parser/template_detection.py", "model_path"),
("speculative/adaptive_spec_params.py", "speculative_algorithm"),
("speculative/adaptive_spec_params.py", "speculative_eagle_topk"),