[Bugfix] fix decode oom with EAGLE

python/sglang/srt/managers/schedule_batch.py

@@ -1558,6 +1558,38 @@ def new_page_count_next_decode(self, selected_indices: Optional[List[int]] = Non
             else sum(1 for req in requests if (req.seqlen - 1) % page_size == 0)
         )
 
+    def _calculate_eagle_extra_tokens(
+        self, selected_indices: Optional[List[int]] = None
+    ) -> int:
+        """Calculate extra tokens required for EAGLE speculative decoding.
+
+        For speculative decoding (EAGLE), we need additional memory for:
+        1. Draft stage: num_seqs * speculative_num_steps * topk tokens
+        2. Verify stage: num_seqs * speculative_num_draft_tokens tokens
+
+        Args:
+            selected_indices: Optional list of selected request indices.
+
+        Returns:
+            Extra tokens required for EAGLE, or 0 if EAGLE is not enabled.
+        """
+        if not self.spec_algorithm.is_eagle():
+            return 0
+
+        server_args = get_global_server_args()
+        speculative_num_steps = server_args.speculative_num_steps
+        topk = server_args.speculative_eagle_topk
+        speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+
+        num_seqs = (
+            len(selected_indices) if selected_indices is not None else self.batch_size()
+        )
+        # Draft/Verify stage memory requirement
+        eagle_draft_tokens = num_seqs * speculative_num_steps * topk
+        eagle_verify_tokens = num_seqs * speculative_num_draft_tokens
+
+        return eagle_draft_tokens + eagle_verify_tokens
+
     def check_decode_mem(
         self, buf_multiplier=1, selected_indices: Optional[List[int]] = None
     ):
@@ -1567,6 +1599,9 @@ def check_decode_mem(
             * self.token_to_kv_pool_allocator.page_size
         )
 
+        if self.spec_algorithm.is_eagle():
+            num_tokens += self._calculate_eagle_extra_tokens(selected_indices)
+
         evict_from_tree_cache(self.tree_cache, num_tokens)
         return self._is_available_size_sufficient(num_tokens)