streaming session: spec v2 bonus accounting + comprehensive test matrix (#22651)

python/sglang/srt/managers/scheduler_output_processor_mixin.py

@@ -358,7 +358,8 @@ class SchedulerOutputProcessorMixin:
         stride = self.draft_worker.speculative_num_draft_tokens
 
         for i, req in enumerate(batch.reqs):
-            req.kv_committed_len += accept_lens[i]
+            # -1 because prepare_for_decode pre-claimed the bonus slot.
+            req.kv_committed_len += accept_lens[i] - 1
             predict_tokens.append(
                 next_token_ids[i * stride : i * stride + accept_lens[i]]
             )

python/sglang/srt/managers/tokenizer_communicator_mixin.py

@@ -1075,15 +1075,6 @@ class TokenizerCommunicatorMixin:
                     "Streaming sessions are disabled. "
                     "Please relaunch with --enable-streaming-session."
                 )
-            if (
-                self.server_args.speculative_algorithm is not None
-                and not self.server_args.disable_overlap_schedule
-            ):
-                raise ValueError(
-                    "Streaming sessions are incompatible with speculative decoding v2 "
-                    "(overlap + speculative). Use --disable-overlap-schedule or "
-                    "disable speculative decoding."
-                )
 
         if obj.session_id is None:
             obj.session_id = uuid.uuid4().hex

python/sglang/srt/speculative/eagle_info_v2.py

@@ -127,6 +127,8 @@ class EagleDraftInputV2Mixin:
             num_needed_tokens += x
             r.kv_allocated_len += x
             r.decode_batch_idx += 1
+            # Pre-claim bonus slot here (like normal decode); resolve subtracts 1.
+            r.kv_committed_len += 1
 
         cur_kv_lens_cpu = torch.tensor(cur_kv_lens_cpu, dtype=torch.int32, device="cpu")
         nxt_kv_lens_cpu = torch.tensor(nxt_kv_lens_cpu, dtype=torch.int32, device="cpu")