Fix/hisparse host backed max request length (#28753)

Co-authored-by: huangtingwei <141888744+huangtingwei9988@users.noreply.github.com>
Co-authored-by: Zhangheng <hzh0425@apache.org>

python/sglang/srt/disaggregation/decode.py

@@ -656,9 +656,14 @@ class DecodePreallocQueue(DecodeHiCachePreallocMixin):
         return len(req.origin_input_ids) + max(len(req.output_ids) - 1, 0)
 
     def _check_if_req_exceed_kv_capacity(self, req: Req) -> bool:
+        # HiSparse admits up to the host-backed logical capacity.
+        if self.scheduler.enable_hisparse:
+            capacity = self.scheduler.tp_worker.model_runner.max_token_pool_size
+        else:
+            capacity = self.max_total_num_tokens
         input_len = self._rebootstrap_prefill_len(req)
-        if input_len > self.max_total_num_tokens:
-            message = f"Request {req.rid} exceeds the maximum number of tokens: {input_len} > {self.max_total_num_tokens}"
+        if input_len > capacity:
+            message = f"Request {req.rid} exceeds the maximum number of tokens: {input_len} > {capacity}"
             logger.error(message)
             prepare_abort(req, message, status_code=HTTPStatus.BAD_REQUEST)
             self.scheduler.output_streamer.stream_output([req], req.return_logprob)

python/sglang/srt/model_executor/model_runner.py

@@ -997,7 +997,7 @@ class ModelRunner:
             RoutedExpertsCapturer.create(
                 model=self.model,
                 model_config=self.model_config,
-                num_tokens=self.max_total_num_tokens + self.page_size,
+                num_tokens=self.max_token_pool_size + self.page_size,
                 max_running_requests=self.max_running_requests,
                 device=self.device,
             )
@@ -1007,7 +1007,7 @@ class ModelRunner:
         set_global_indexer_capturer(
             create_indexer_capturer(
                 model_config=self.model_config,
-                num_tokens=self.max_total_num_tokens + self.page_size,
+                num_tokens=self.max_token_pool_size + self.page_size,
                 max_running_requests=self.max_running_requests,
                 device=self.device,
             )
@@ -1244,6 +1244,16 @@ class ModelRunner:
         else:
             return self.max_total_num_tokens
 
+    @property
+    def max_token_pool_size(self):
+        """Return the max token pool size considering hybrid swa and hisparse settings."""
+        if self.enable_hisparse:
+            # HiSparse uses the host-backed full pool capacity.
+            size_full = getattr(self.token_to_kv_pool_allocator, "size_full", None)
+            if size_full is not None:
+                return size_full
+        return self.effective_max_total_num_tokens
+
     def _load_format_scope(self, load_format: Optional[str]):
         """Make this runner's load format the published one while it loads.