[lora] Remove synchronous .any().item() guard in LoRA MoE prefill path (#25531)

Co-authored-by: Cursor <cursoragent@cursor.com>

python/sglang/srt/lora/layers.py

@@ -980,6 +980,10 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
         )
         req_to_lora = wi
 
+        # Single source of truth: lora_manager precomputes this per-batch from
+        # the Python weight_indices list, no GPU sync needed.
+        has_active_lora = bool(getattr(batch_info, "has_active_lora", False))
+
         return LoRAInfo(
             gate_up_lora_a_weights=self.gate_up_lora_a_weights,
             gate_up_lora_b_weights=self.gate_up_lora_b_weights,
@@ -989,6 +993,7 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
             req_to_lora=req_to_lora,
             lora_ranks=lora_ranks,
             adapter_enabled=adapter_enabled,
+            has_active_lora=has_active_lora,
             max_lora_rank=max_lora_rank,
             num_experts=self.base_layer.num_experts,
             experts_shared_outer_loras=self.experts_shared_outer_loras,

python/sglang/srt/lora/lora_moe_runners.py

@@ -179,6 +179,7 @@ class LoRAInfo:
     max_lora_rank: int  # Maximum LoRA rank across all adapters
 
     num_experts: int
+    has_active_lora: bool = True
     experts_shared_outer_loras: bool = False
     cg_buffers: dict | None = None
 
@@ -335,24 +336,9 @@ def _add_lora_gate_up_delta(
         merged_experts_fused_moe_lora_add,
     )
 
-    if get_is_capture_mode():
-        # During CUDA graph capture, always enter the LoRA path so that
-        # the LoRA kernels are recorded in the graph.  adapter_enabled is
-        # all-zeros during capture, so the Triton kernel early-exits per
-        # program (zero overhead).  During replay the tensor is updated
-        # in-place with the real adapter mask before graph.replay().
-        has_active_lora = True
-    else:
-        num_loras = len(lora_info.lora_ranks)
-        has_active_lora = (
-            (
-                lora_info.adapter_enabled[:num_loras]
-                * (lora_info.lora_ranks > 0).to(lora_info.adapter_enabled.dtype)
-            )
-            .any()
-            .item()
-        )
-    if not has_active_lora or lora_info is None or lora_info.max_lora_rank == 0:
+    if lora_info is None or lora_info.max_lora_rank == 0:
+        return
+    if not get_is_capture_mode() and not lora_info.has_active_lora:
         return
 
     M, top_k, gate_up_dim = intermediate_cache.shape
@@ -520,6 +506,11 @@ def build_lora_hooks(
     """
     if lora_info is None or lora_info.max_lora_rank == 0:
         return LoRAHooks()
+    # Skip alignment/mapping work entirely when the batch has no active adapter.
+    # During CUDA graph capture we still need to record the kernels into the
+    # graph (adapter_enabled is all-zero, kernels early-exit on GPU).
+    if not get_is_capture_mode() and not lora_info.has_active_lora:
+        return LoRAHooks()
 
     # Compute alignment / mapping (once, shared by both hooks)
     token_lora_mapping: torch.Tensor | None = None