[PD][LoRA] Gate decode admission on adapter slots (#39332)

python/sglang/srt/disaggregation/decode.py

@@ -1188,6 +1188,19 @@ class DecodePreallocQueue(DecodeHiCachePreallocMixin):
                 - len(self.transfer_queue.queue),
             )
 
+        if self.scheduler.enable_lora:
+            running_batches = (
+                self.scheduler.running_mbs
+                if is_pp_mode
+                else (self.scheduler.running_batch,)
+            )
+            # Include finished requests; GPU work may still use their adapters.
+            running_loras = {
+                req.lora_id for batch in running_batches for req in batch.reqs
+            }
+            running_loras.update(r.req.lora_id for r in self.transfer_queue.queue)
+            running_loras.update(req.lora_id for req in self.scheduler.waiting_queue)
+
         # Then, preallocate the remaining requests if possible
         for i, decode_req in enumerate(self.queue):
             if rids_to_check is not None and decode_req.req.rid not in rids_to_check:
@@ -1208,6 +1221,11 @@ class DecodePreallocQueue(DecodeHiCachePreallocMixin):
             if hisparse_req_budget <= 0:
                 break
 
+            if self.scheduler.enable_lora and not self.scheduler.can_schedule_lora_req(
+                decode_req.req, running_loras
+            ):
+                continue
+
             # Memory estimation: don't add if the projected memory cannot be met
             # TODO: add new_token ratio
             origin_input_len = self._rebootstrap_prefill_len(decode_req.req)
@@ -1559,6 +1577,8 @@ class DecodePreallocQueue(DecodeHiCachePreallocMixin):
             self._num_published_destinations += 1
             preallocated_reqs.append(decode_req)
             indices_to_remove.add(i)
+            if self.scheduler.enable_lora:
+                running_loras.add(decode_req.req.lora_id)
             decode_req.req.time_stats.set_decode_transfer_queue_entry_time()
 
         if failed_reqs:

python/sglang/srt/managers/scheduler.py

@@ -3825,7 +3825,7 @@ class Scheduler(
             mamba_allocator.alloc_group_begin(len(self.waiting_queue))
         # Get requests from the waiting queue to a new prefill batch
         for req in self.waiting_queue:
-            if self.enable_lora and not self._can_schedule_lora_req(req, running_loras):
+            if self.enable_lora and not self.can_schedule_lora_req(req, running_loras):
                 continue
 
             running_bs = len(running_batch.reqs)
@@ -4035,7 +4035,7 @@ class Scheduler(
 
         return new_batch, running_batch
 
-    def _can_schedule_lora_req(
+    def can_schedule_lora_req(
         self, req: Req, running_loras: set[Optional[str]]
     ) -> bool:
         """