Clean up startup log noise (#28807)

python/sglang/srt/arg_groups/speculative_hook.py

@@ -278,10 +278,6 @@ def _handle_eagle_family(server_args: ServerArgs) -> None:
             "Non-overlap (synchronous) spec v2 is used for eagle/eagle3/standalone "
             "speculative decoding."
         )
-    else:
-        logger.warning(
-            "Overlap spec v2 is enabled by default for eagle/eagle3/standalone speculative decoding."
-        )
 
     if server_args.enable_mixed_chunk:
         server_args.enable_mixed_chunk = False

python/sglang/srt/distributed/device_communicators/custom_all_reduce_v2.py

@@ -19,7 +19,7 @@ from sglang.srt.distributed.device_communicators.custom_all_reduce_vmm_utils imp
 from sglang.srt.model_executor.runner_backend_utils.tc_piecewise_cuda_graph import (
     is_in_tc_piecewise_cuda_graph,
 )
-from sglang.srt.utils import is_sm100_supported, log_info_on_rank0
+from sglang.srt.utils import is_sm100_supported
 
 logger = logging.getLogger(__name__)
 
@@ -79,7 +79,6 @@ class CustomAllReduceV2:
         )
         self._post_init_obj()
         self.disabled = False
-        log_info_on_rank0(logger, "Custom allreduce v2 initialized successfully")
 
     def override_shot(self, shot: int | None):
         if shot is None:
@@ -129,9 +128,6 @@ class CustomAllReduceV2:
         offsets_all = self._share_list(offsets)
         result = [list(zip(o, h)) for o, h in zip(offsets_all, handles_all)]
         self.obj.register_inputs(result)
-        log_info_on_rank0(
-            logger, f"Registered {len(pairs)} cuda graph addresses via IPC"
-        )
 
     def should_custom_ar(self, inp: torch.Tensor) -> bool:
         """Check if the input tensor is suitable for custom all-reduce."""

python/sglang/srt/layers/attention/flashinfer_backend.py

@@ -314,14 +314,8 @@ class FlashInferAttnBackend(AttentionBackend):
         fmha_backend = "auto"
         if is_sm100_supported():
             # Disable CUTLASS backend when piecewise cuda graph is enabled
-            # due to TMA descriptor initialization issues on B200
+            # due to TMA descriptor initialization issues on SM100 GPUs.
-            if check_cuda_graph_backend(Phase.PREFILL, Backend.TC_PIECEWISE):
+            if not check_cuda_graph_backend(Phase.PREFILL, Backend.TC_PIECEWISE):
-                logger.info(
-                    "CUTLASS backend is disabled when piecewise cuda graph is enabled "
-                    "due to TMA descriptor initialization issues on SM100 GPUs. "
-                    "Using auto backend instead for stability."
-                )
-            else:
                 fmha_backend = "cutlass"
         self.prefill_wrapper_ragged = BatchPrefillWithRaggedKVCacheWrapper(
             self.workspace_buffer, "NHD", backend=fmha_backend

python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py

@@ -1003,10 +1003,6 @@ class ModelRunnerKVCacheMixin:
                 requested_per_worker,
                 max_num_reqs,
             )
-        logger.info(
-            f"Max concurrent requests (per dp worker) from the finalized token capacity: "
-            f"max_num_reqs={max_num_reqs}."
-        )
         return max_num_reqs
 
     def _apply_memory_pool_config(self: ModelRunner, config: MemoryPoolConfig):