fix(spec_v2): support EAGLE with spec_v2

1. add pass for spec_v2 in base_attn

2. fix: EAGLE with spec_v2 overlap Grammar accept_token failed

Signed-off-by: wxiwnd <wxiwnd@outlook.com>

python/sglang/srt/disaggregation/decode.py

@@ -137,9 +137,9 @@ class DecodeReqToTokenPool:
         # Indices of reqs that already have a req_pool_idx and will reuse
         # their existing slot (e.g. chunked prefill continuing across chunks).
         reusing = [i for i, r in enumerate(reqs) if r.req_pool_idx is not None]
-        assert (
-            len(reusing) <= 1
-        ), "only one chunked request may reuse req_pool_idx in a batch"
+        assert len(reusing) <= 1, (
+            "only one chunked request may reuse req_pool_idx in a batch"
+        )
         assert all(
             reqs[i].is_chunked > 0 or reqs[i].kv_committed_len > 0 for i in reusing
         ), "reusing request must be chunked or have committed KV"
@@ -166,7 +166,6 @@ class DecodeReqToTokenPool:
 
 
 class HybridMambaDecodeReqToTokenPool(HybridReqToTokenPool):
-
     def __init__(
         self,
         size: int,
@@ -793,9 +792,9 @@ class DecodePreallocQueue:
         """Pre-allocate the memory for req_to_token and token_kv_pool"""
         req_pool_indices = self.req_to_token_pool.alloc([req])
 
-        assert (
-            req_pool_indices is not None
-        ), "req_pool_indices is full! There is a bug in memory estimation."
+        assert req_pool_indices is not None, (
+            "req_pool_indices is full! There is a bug in memory estimation."
+        )
 
         # Alloc all tokens for the prebuilt req (except for the reserved input token for decoding)
         fill_len = len(req.origin_input_ids) + max(len(req.output_ids) - 1, 0)
@@ -814,9 +813,9 @@ class DecodePreallocQueue:
                 extend_num_tokens=fill_len,
             )
 
-        assert (
-            kv_loc is not None
-        ), "KV cache is full! There is a bug in memory estimation."
+        assert kv_loc is not None, (
+            "KV cache is full! There is a bug in memory estimation."
+        )
 
         self.req_to_token_pool.write((req.req_pool_idx, slice(0, len(kv_loc))), kv_loc)
 
@@ -1008,7 +1007,6 @@ class DecodeTransferQueue:
 
 
 class SchedulerDisaggregationDecodeMixin:
-
     @torch.no_grad()
     def event_loop_normal_disagg_decode(self: Scheduler):
         """A normal scheduler loop for decode worker in disaggregation mode."""
@@ -1023,6 +1021,18 @@ class SchedulerDisaggregationDecodeMixin:
             # Get the next batch to run
             batch = self.get_next_disagg_decode_batch_to_run()
             self.cur_batch = batch
+            disable_overlap_for_batch = self.is_disable_overlap_for_batch(batch)
+
+            def pop_and_process():
+                # Process the results of the last batch
+                tmp_batch, tmp_result = self.result_queue.popleft()
+                self.process_batch_result(tmp_batch, tmp_result)
+
+            # If we need grammar sync (spec + grammar), process the last batch
+            # results first so that the grammar state is up-to-date before
+            # generating the bitmask for the current batch.
+            if disable_overlap_for_batch:
+                pop_and_process()
 
             # Launch the current batch
             if batch:
@@ -1040,6 +1050,11 @@ class SchedulerDisaggregationDecodeMixin:
         self.result_queue = deque()
         self.last_batch: Optional[ScheduleBatch] = None
 
+        def pop_and_process():
+            # Process the results of the last batch
+            tmp_batch, tmp_result = self.result_queue.popleft()
+            self.process_batch_result(tmp_batch, tmp_result)
+
         while True:
             # Receive requests
             recv_reqs = self.recv_requests()
@@ -1050,6 +1065,13 @@ class SchedulerDisaggregationDecodeMixin:
             # Get the next batch to run
             batch = self.get_next_disagg_decode_batch_to_run()
             self.cur_batch = batch
+            disable_overlap_for_batch = self.is_disable_overlap_for_batch(batch)
+
+            # If we need grammar sync (spec + grammar), process the last batch
+            # results first so that the grammar state is up-to-date before
+            # generating the bitmask for the current batch.
+            if self.last_batch and disable_overlap_for_batch:
+                pop_and_process()
 
             # Launch the current batch
             if batch:
@@ -1060,8 +1082,8 @@ class SchedulerDisaggregationDecodeMixin:
 
             # Process the last batch
             if self.last_batch:
-                tmp_batch, tmp_result = self.result_queue.popleft()
-                self.process_batch_result(tmp_batch, tmp_result)
+                if not disable_overlap_for_batch:
+                    pop_and_process()
             elif batch is None:
                 self.self_check_during_idle()
 

python/sglang/srt/layers/attention/base_attn_backend.py

@@ -75,7 +75,10 @@ class AttentionBackend(ABC):
         Here, we need to redo the computation of all metadata of the attention backend
         that depends on tree mask and position buffers.
         """
-        raise NotImplementedError()
+        # if self.topk <= 1:
+        #     return
+        # raise NotImplementedError()
+        pass
 
     @debug_kernel_api
     def forward(