[DCP] Bound a request by the aggregate KV pool, not one rank's share (#33448)

python/sglang/srt/managers/scheduler.py

@@ -2080,7 +2080,10 @@ class Scheduler(
             min(
                 max_new_tokens,
                 self.max_req_len - input_len - 1,
-                self.max_total_num_tokens - paged_input_len - self.page_size - 1,
+                self.max_total_num_tokens * self.server_args.dcp_size
+                - paged_input_len
+                - self.page_size
+                - 1,
             ),
         )
         # Clipping above can push max_new_tokens below min_new_tokens, which

python/sglang/srt/managers/tp_worker.py

@@ -385,7 +385,9 @@ class TpModelWorker(BaseTpWorker):
         assert self.model_runner.max_running_requests > 0, "max_running_request is zero"
         max_req_len = min(
             self.model_config.context_len - 1,
-            self.model_runner.effective_max_total_num_tokens - 1,
+            self.model_runner.effective_max_total_num_tokens
+            * self.model_runner.dcp_size
+            - 1,
         )
         assert max_req_len > 0, "Memory pool size is too small"
 
@@ -485,7 +487,9 @@ class TpModelWorker(BaseTpWorker):
     def get_worker_info(self):
         max_req_len = min(
             self.model_config.context_len - 1,
-            self.model_runner.effective_max_total_num_tokens - 1,
+            self.model_runner.effective_max_total_num_tokens
+            * self.model_runner.dcp_size
+            - 1,
         )
         return (
             self.model_runner.max_total_num_tokens,