[Spec] NGRAMWorker on BaseSpecWorker; algo-owned verify-tree shape params (#27799)

python/sglang/srt/mem_cache/common.py

@@ -169,11 +169,10 @@ def get_alloc_len_per_decode(server_args: Optional[ServerArgs] = None) -> int:
     spec_tokens = server_args.max_speculative_num_draft_tokens
     page_size = server_args.page_size
 
-    # NGRAM drafts are a flat candidate list written to contiguous slots after
+    from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
-    # seq_lens (no per-topk page duplication), so the flat formula applies at
+
-    # any page_size.
+    spec_algo = SpeculativeAlgorithm.from_string(server_args.speculative_algorithm)
-    is_ngram = (server_args.speculative_algorithm or "").upper() == "NGRAM"
+    if page_size == 1 or spec_topk == 1 or not spec_algo.has_draft_kv():
-    if page_size == 1 or spec_topk == 1 or is_ngram:
         return max(spec_steps * spec_topk, spec_tokens)
     else:
         # page_size > 1 + topk > 1 (spec v2 tree): worst-case page-aligned tree

python/sglang/srt/speculative/eagle_info_v2.py

@@ -368,6 +368,19 @@ class EagleDraftInputV2Mixin:
 
 @dataclass
 class EagleVerifyInputV2Mixin:
+    @property
+    def max_tree_depth(self: EagleVerifyInput) -> int:
+        """Longest root-to-leaf chain of the verify tree, incl. the root;
+        bounds the accept_index row width. EAGLE trees are depth-bounded by
+        the draft loop. Algorithms with other tree shapes override this."""
+        return self.spec_steps + 1
+
+    @property
+    def tree_topk(self: EagleVerifyInput) -> int:
+        """Branching factor passed to the tree-verify kernels; -1 means an
+        irregular tree (no fixed per-level branching)."""
+        return self.topk
+
     def prepare_for_v2_verify(
         self: EagleVerifyInput,
         req_to_token_pool: ReqToTokenPool,
@@ -493,17 +506,8 @@ class EagleVerifyInputV2Mixin:
         candidates = self.draft_token.reshape(bs, self.draft_token_num)
         predict_shape = list(next_token_logits.shape)[:-1]
         predict = torch.zeros(predict_shape, dtype=torch.int32, device=device).flatten()
-        # Longest root-to-leaf chain of the verify tree, incl. the root; bounds
-        # the accept_index row width. EAGLE trees are depth-bounded by the draft
-        # loop (spec_steps + 1); NGRAM trees are node-budgeted with no depth cap
-        # (a single corpus match can chain all draft_token_num nodes).
-        max_tree_depth = (
-            self.draft_token_num
-            if batch.spec_algorithm.is_ngram()
-            else self.spec_steps + 1
-        )
         accept_index = torch.full(
-            (bs, max_tree_depth), -1, dtype=torch.int32, device=device
+            (bs, self.max_tree_depth), -1, dtype=torch.int32, device=device
         )
         num_correct_drafts = torch.empty((bs,), dtype=torch.int32, device=device)
 
@@ -520,7 +524,7 @@ class EagleVerifyInputV2Mixin:
                 retrieve_next_token=self.retrieve_next_token,
                 retrieve_next_sibling=self.retrieve_next_sibling,
                 target_predict=target_predict,
-                topk=-1 if batch.spec_algorithm.is_ngram() else self.topk,
+                topk=self.tree_topk,
             )
         else:
             # Apply temperature and get target probs
@@ -598,7 +602,7 @@ class EagleVerifyInputV2Mixin:
                 num_correct_drafts=num_correct_drafts,  # mutable
                 simulate_acc_len=SIMULATE_ACC_LEN,
                 bs=bs,
-                spec_steps=max_tree_depth - 1,
+                spec_steps=self.max_tree_depth - 1,
             )
 
         # `num_correct_drafts` stays drafts-only inside this function; the returned

python/sglang/srt/speculative/ngram_info.py

@@ -49,6 +49,18 @@ class NgramVerifyInput(SpecInput, EagleDraftInputV2Mixin, EagleVerifyInputV2Mixi
             custom_mask.device if custom_mask is not None else new_seq_lens.device
         )
 
+    @property
+    def max_tree_depth(self) -> int:
+        # NGRAM trees are node-budgeted with no depth cap: the corpus BFS only
+        # stops on the node budget, so a single long match can chain all
+        # draft_token_num nodes (spec_steps is meaningless for this tree).
+        return self.draft_token_num
+
+    @property
+    def tree_topk(self) -> int:
+        # Irregular tree: per-level branching follows the corpus matches.
+        return -1
+
     def get_spec_adjust_token_coefficient(self) -> Tuple[int, int]:
         return self.draft_token_num, self.draft_token_num
 

python/sglang/srt/speculative/ngram_worker.py

@@ -12,6 +12,7 @@ from sglang.srt.managers.tp_worker import TpModelWorker
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
 from sglang.srt.observability.req_time_stats import set_time_batch
 from sglang.srt.server_args import ServerArgs
+from sglang.srt.speculative.base_spec_worker import BaseDraftWorker, BaseSpecWorker
 from sglang.srt.speculative.cpp_ngram.ngram_corpus import NgramCorpus
 from sglang.srt.speculative.ngram_info import NgramVerifyInput
 from sglang.srt.speculative.spec_utils import (
@@ -30,7 +31,7 @@ logger = logging.getLogger(__name__)
 USE_FULL_MASK = True
 
 
-class NGRAMWorker:
+class NGRAMWorker(BaseSpecWorker):
     def __init__(
         self,
         server_args: ServerArgs,
@@ -45,7 +46,7 @@ class NGRAMWorker:
     ):
         self.server_args = server_args
         self.enable_overlap = not server_args.disable_overlap_schedule
-        self.target_worker = target_worker
+        self._target_worker = target_worker
         self.model_runner = target_worker.model_runner
         self.tp_rank = tp_rank
         self.page_size = server_args.page_size
@@ -99,6 +100,15 @@ class NGRAMWorker:
                 loaded,
             )
 
+    @property
+    def target_worker(self) -> TpModelWorker:
+        return self._target_worker
+
+    @property
+    def draft_worker(self) -> Optional[BaseDraftWorker]:
+        # NGRAM has no draft model; drafts come from the CPU-side corpus.
+        return None
+
     def clear_cache_pool(self):
         self.ngram_corpus.reset()
         self._prev_decode_rids = set()

python/sglang/srt/speculative/spec_info.py

@@ -118,6 +118,12 @@ class SpeculativeAlgorithm(Enum):
     def supports_target_verify_for_draft(self) -> bool:
         return self.is_dflash()
 
+    def has_draft_kv(self) -> bool:
+        """Whether the draft phase writes KV chains. NGRAM does not (its tree
+        lives only in the verify mask), so per-decode KV sizing needs no
+        per-topk page rounding; see get_alloc_len_per_decode."""
+        return not self.is_ngram()
+
     def create_future_map(
         self,
         device: torch.device,

python/sglang/srt/speculative/spec_registry.py

@@ -78,6 +78,10 @@ class CustomSpecAlgo:
     def supports_target_verify_for_draft(self) -> bool:
         return False
 
+    def has_draft_kv(self) -> bool:
+        # Conservative default: the larger KV reserve.
+        return True
+
     def supports_spec_v2(self) -> bool:
         return self.supports_overlap