diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index 358f30a5d..8cccf094c 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -850,9 +850,6 @@ class Scheduler(
     def init_model_worker(self):
         # Load model weights.
         self.init_tp_model_worker()
-        if self.spec_algorithm.is_frozen_kv_mtp():
-            # Frozen-KV MTP draft construction needs the target KV pool.
-            self.init_target_memory_pool()
         self.maybe_init_draft_worker()
 
         # Allocate KV cache pools for all workers.
diff --git a/python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py b/python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py
index 3f1b88a87..30fa31cb1 100644
--- a/python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py
+++ b/python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py
@@ -114,16 +114,11 @@ class FrozenKVMTPDraftWorker(EagleDraftWorkerBase, TpModelWorker):
             f"{self.speculative_algorithm.name}."
         )
 
-        # Draft attention uses target req_to_token + KV allocator (read-only).
-        self.req_to_token_pool, self.token_to_kv_pool_allocator = (
-            target_worker.get_memory_pool()
-        )
-
-        target_cfg = target_worker.model_runner.memory_pool_config
-        self.draft_pool_config = MemoryPoolConfig(
-            max_total_num_tokens=64,  # Dummy value
-            max_running_requests=target_cfg.max_running_requests,
-        )
+        # Target pools (read-only) are bound in alloc_memory_pool(), not here, so
+        # the worker can be built before the target pool exists (see #29021).
+        self.req_to_token_pool = None
+        self.token_to_kv_pool_allocator = None
+        self.draft_pool_config: Optional[MemoryPoolConfig] = None
 
         self.hot_token_id = None
 
@@ -144,9 +139,6 @@ class FrozenKVMTPDraftWorker(EagleDraftWorkerBase, TpModelWorker):
                 moe_dp_rank=moe_dp_rank,
                 nccl_port=nccl_port,
                 is_draft_worker=True,
-                req_to_token_pool=self.req_to_token_pool,
-                token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
-                memory_pool_config=self.draft_pool_config,
             )
 
         embed, head = self.target_worker.model_runner.model.get_embed_and_head()
@@ -160,8 +152,6 @@ class FrozenKVMTPDraftWorker(EagleDraftWorkerBase, TpModelWorker):
             )
 
         self.kv_context: Optional[FrozenKVMTPContext] = None
-        if hasattr(self.draft_model_runner.model, "bind_frozen_kv_context"):
-            self._bind_kv_context()
 
         self.draft_tp_context = (
             draft_tp_context if server_args.enable_dp_attention else empty_context
@@ -180,23 +170,26 @@ class FrozenKVMTPDraftWorker(EagleDraftWorkerBase, TpModelWorker):
         req_to_token_pool=None,
         token_to_kv_pool_allocator=None,
     ):
+        self.req_to_token_pool = req_to_token_pool
+        self.token_to_kv_pool_allocator = token_to_kv_pool_allocator
+
+        self.draft_pool_config = MemoryPoolConfig(
+            max_total_num_tokens=64,  # Dummy value
+            max_running_requests=memory_pool_config.max_running_requests,
+        )
+
         # NOTE: call TpModelWorker explicitly -- EagleDraftWorkerBase precedes it in
         # the MRO and its alloc_memory_pool is a no-op stub.
         TpModelWorker.alloc_memory_pool(
             self,
             memory_pool_config=self.draft_pool_config,
-            req_to_token_pool=(
-                req_to_token_pool
-                if req_to_token_pool is not None
-                else self.req_to_token_pool
-            ),
-            token_to_kv_pool_allocator=(
-                token_to_kv_pool_allocator
-                if token_to_kv_pool_allocator is not None
-                else self.token_to_kv_pool_allocator
-            ),
+            req_to_token_pool=req_to_token_pool,
+            token_to_kv_pool_allocator=token_to_kv_pool_allocator,
         )
 
+        if hasattr(self.draft_model_runner.model, "bind_frozen_kv_context"):
+            self._bind_kv_context()
+
     def init_attention_backends(self):
         with (
             self.draft_tp_context(self.draft_model_runner.tp_group),
@@ -442,6 +435,9 @@ class FrozenKVMTPDraftWorker(EagleDraftWorkerBase, TpModelWorker):
         self._set_positions(forward_batch)
         self._expand_for_topk_draft(forward_batch)
 
+        # Frozen draft never writes KV; None signals fill_from to skip the slot.
+        forward_batch.out_cache_loc = None
+
         can_run_cuda_graph = (
             self.cuda_graph_runner
             and self.cuda_graph_runner.can_run_graph(forward_batch)