feat(SpecEagleV2): add standalone_worker_v2 (#12625)

Co-authored-by: liugaoji.lgj <liugaoji.lgj@alibaba-inc.com>

python/sglang/srt/managers/scheduler.py

@@ -499,7 +499,7 @@ class Scheduler(
 
         # Draft workers are looked up via `SpeculativeAlgorithm` registry; new
         # algorithms should register their factory instead of patching this code.
-        if self.spec_algorithm.is_eagle():
+        if self.spec_algorithm.supports_spec_v2():
             draft_worker_kwargs["enable_overlap"] = self.enable_overlap
 
         # FIXME: refactor the draft worker registration logic
@@ -852,7 +852,7 @@ class Scheduler(
 
         if self.draft_worker is None or self.spec_algorithm.is_ngram():
             draft_token_to_kv_pool = None
-        elif self.spec_algorithm.is_eagle() and self.enable_overlap:
+        elif self.spec_algorithm.supports_spec_v2() and self.enable_overlap:
             if self.server_args.enable_multi_layer_eagle:
                 draft_runner = self.draft_worker.draft_worker.draft_runner_list[0]
             else:
@@ -930,11 +930,13 @@ class Scheduler(
                 hidden_size=(
                     model_config.hidden_size
                     if self.spec_algorithm.is_eagle()
+                    or self.spec_algorithm.is_standalone()
                     else 16  # minimal padding size for RDMA
                 ),
                 hidden_states_dtype=(
                     model_config.dtype
                     if self.spec_algorithm.is_eagle()
+                    or self.spec_algorithm.is_standalone()
                     else torch.float32
                 ),
                 custom_mem_pool=self.token_to_kv_pool_allocator.get_kvcache().maybe_get_custom_mem_pool(),