Simplify prepare_extend_after_decode (#6987)

python/sglang/srt/managers/schedule_batch.py

@@ -1636,7 +1636,9 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         if self.spec_info:
             self.spec_info.merge_batch(other.spec_info)
 
-    def get_model_worker_batch(self) -> ModelWorkerBatch:
+    def get_model_worker_batch(
+        self, seq_lens_cpu_cache: Optional[torch.Tensor] = None
+    ) -> ModelWorkerBatch:
         if self.forward_mode.is_decode_or_idle():
             extend_seq_lens = extend_prefix_lens = extend_logprob_start_lens = None
         else:
@@ -1646,16 +1648,20 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
 
         # Create seq_lens_cpu when needed
         if (
-            (
+            global_server_args_dict["attention_backend"] == "fa3"
+            or (
                 global_server_args_dict["use_mla_backend"]
                 and global_server_args_dict["attention_backend"] == "flashinfer"
             )
             or global_server_args_dict["attention_backend"] == "flashmla"
-            or global_server_args_dict["attention_backend"] == "fa3"
             or global_server_args_dict["attention_backend"] == "cutlass_mla"
             or global_server_args_dict["enable_two_batch_overlap"]
         ):
-            seq_lens_cpu = self.seq_lens.cpu()
+            seq_lens_cpu = (
+                seq_lens_cpu_cache
+                if seq_lens_cpu_cache is not None
+                else self.seq_lens.cpu()
+            )
         else:
             seq_lens_cpu = None
 

python/sglang/srt/managers/scheduler.py

@@ -1575,10 +1575,9 @@ class Scheduler(
                     num_accepted_tokens,
                     can_run_cuda_graph,
                 ) = self.draft_worker.forward_batch_speculative_generation(batch)
-                self.spec_num_total_accepted_tokens += (
-                    num_accepted_tokens + batch.batch_size()
-                )
-                self.spec_num_total_forward_ct += batch.batch_size()
+                bs = batch.batch_size()
+                self.spec_num_total_accepted_tokens += num_accepted_tokens + bs
+                self.spec_num_total_forward_ct += bs
                 self.num_generated_tokens += num_accepted_tokens
 
             if self.pp_group.is_last_rank:

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -56,6 +56,16 @@ def get_is_capture_mode():
     return is_capture_mode
 
 
+@contextmanager
+def model_capture_mode():
+    global is_capture_mode
+    is_capture_mode = True
+
+    yield
+
+    is_capture_mode = False
+
+
 def _to_torch(model: torch.nn.Module, reverse: bool, num_tokens: int):
     for sub in model._modules.values():
         if isinstance(sub, CustomOp):
@@ -291,22 +301,13 @@ class CudaGraphRunner:
 
         # Capture
         try:
-            with self.model_capture_mode():
+            with model_capture_mode():
                 self.capture()
         except RuntimeError as e:
             raise Exception(
                 f"Capture cuda graph failed: {e}\n{CUDA_GRAPH_CAPTURE_FAILED_MSG}"
             )
 
-    @contextmanager
-    def model_capture_mode(self):
-        global is_capture_mode
-        is_capture_mode = True
-
-        yield
-
-        is_capture_mode = False
-
     def can_run(self, forward_batch: ForwardBatch):
         if self.enable_dp_attention or self.enable_sp_layernorm:
             total_global_tokens = sum(forward_batch.global_num_tokens_cpu)
@@ -650,6 +651,8 @@ class CudaGraphRunner:
                     topk=self.model_runner.server_args.speculative_eagle_topk,
                     draft_token_num=self.model_runner.server_args.speculative_num_draft_tokens,
                     capture_hidden_mode=CaptureHiddenMode.FULL,
+                    seq_lens_sum=None,
+                    seq_lens_cpu=None,
                 )
 
         return spec_info

python/sglang/srt/server_args.py

@@ -1013,13 +1013,13 @@ class ServerArgs:
             type=str,
             choices=[
                 "aiter",
-                "flashinfer",
-                "triton",
-                "torch_native",
+                "cutlass_mla",
                 "fa3",
+                "flashinfer",
                 "flashmla",
-                "cutlass_mla",
                 "intel_amx",
+                "torch_native",
+                "triton",
             ],
             default=ServerArgs.attention_backend,
             help="Choose the kernels for attention layers.",

python/sglang/srt/speculative/eagle_draft_cuda_graph_runner.py

@@ -10,6 +10,7 @@ from sglang.srt.model_executor.cuda_graph_runner import (
     CudaGraphRunner,
     get_batch_sizes_to_capture,
     get_global_graph_memory_pool,
+    model_capture_mode,
     set_global_graph_memory_pool,
     set_torch_compile_config,
 )
@@ -80,7 +81,8 @@ class EAGLEDraftCudaGraphRunner:
 
         # Capture
         try:
-            self.capture()
+            with model_capture_mode():
+                self.capture()
         except RuntimeError as e:
             raise Exception(
                 f"Capture cuda graph failed: {e}\n{CUDA_GRAPH_CAPTURE_FAILED_MSG}"

python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py

@@ -11,6 +11,7 @@ from sglang.srt.model_executor.cuda_graph_runner import (
     LogitsProcessorOutput,
     get_batch_sizes_to_capture,
     get_global_graph_memory_pool,
+    model_capture_mode,
     set_global_graph_memory_pool,
     set_torch_compile_config,
 )
@@ -19,7 +20,7 @@ from sglang.srt.model_executor.forward_batch_info import (
     ForwardBatch,
     ForwardMode,
 )
-from sglang.srt.speculative.eagle_utils import EagleDraftInput
+from sglang.srt.speculative.eagle_utils import EagleDraftInput, fast_topk
 
 if TYPE_CHECKING:
     from sglang.srt.speculative.eagle_worker import EAGLEWorker
@@ -37,6 +38,7 @@ class EAGLEDraftExtendCudaGraphRunner:
         self.tp_size = self.model_runner.tp_size
         self.dp_size = model_runner.server_args.dp_size
         self.speculative_num_steps = model_runner.server_args.speculative_num_steps
+        self.topk = model_runner.server_args.speculative_eagle_topk
         self.capture_bs, self.compile_bs = get_batch_sizes_to_capture(model_runner)
         self.padded_static_len = -1
 
@@ -87,7 +89,8 @@ class EAGLEDraftExtendCudaGraphRunner:
 
         # Capture
         try:
-            self.capture()
+            with model_capture_mode():
+                self.capture()
         except RuntimeError as e:
             raise Exception(
                 f"Capture cuda graph failed: {e}\n{CUDA_GRAPH_CAPTURE_FAILED_MSG}"
@@ -170,6 +173,8 @@ class EAGLEDraftExtendCudaGraphRunner:
                 forward_batch.positions,
                 forward_batch,
             )
+            probs = torch.softmax(ret.next_token_logits, dim=-1)
+            ret.topk_p, ret.topk_index = fast_topk(probs, self.topk, dim=-1)
 
             forward_batch.out_cache_loc = output_cache_loc_backup
             forward_batch.spec_info.hidden_states = hidden_states_backup
@@ -198,7 +203,7 @@ class EAGLEDraftExtendCudaGraphRunner:
 
         index = bisect.bisect_left(self.capture_bs, raw_bs)
         bs = self.capture_bs[index]
-        if bs != raw_bs:
+        if bs * self.num_tokens_per_bs != num_tokens:
             self.seq_lens.fill_(1)
             self.accept_length.fill_(1)
             self.out_cache_loc.zero_()
@@ -238,8 +243,11 @@ class EAGLEDraftExtendCudaGraphRunner:
         out = self.output_buffers[bs]
         if bs != raw_bs:
             forward_batch.spec_info.accept_length = self.accept_length[:raw_bs]
+            out_copy = out
             out = LogitsProcessorOutput(
                 next_token_logits=out.next_token_logits[:raw_bs],
                 hidden_states=out.hidden_states[:raw_bs],
             )
+            out.topk_p = out_copy.topk_p[:raw_bs]
+            out.topk_index = out_copy.topk_index[:raw_bs]
         return out

python/sglang/srt/speculative/eagle_utils.py

@@ -22,8 +22,7 @@ from sglang.srt.managers.schedule_batch import (
     global_server_args_dict,
 )
 from sglang.srt.mem_cache.memory_pool import TokenToKVPoolAllocator
-from sglang.srt.model_executor.forward_batch_info import CaptureHiddenMode
-from sglang.srt.speculative.build_eagle_tree import build_tree_kernel_efficient
+from sglang.srt.model_executor.forward_batch_info import CaptureHiddenMode, ForwardMode
 from sglang.srt.utils import fast_topk, is_cuda, is_hip, next_power_of_2
 
 if is_cuda():
@@ -86,78 +85,29 @@ class EagleDraftInput:
         self,
         batch: ScheduleBatch,
         speculative_num_steps: int,
-        context_length: int,
-        pad_input: bool = False,
     ):
-        accept_length_cpu = batch.spec_info.accept_length_cpu
-        batch.extend_lens = [x + 1 for x in accept_length_cpu]
+        batch.forward_mode = ForwardMode.DRAFT_EXTEND
+        batch.input_ids = self.verified_id
+        batch.extend_lens = [x + 1 for x in batch.spec_info.accept_length_cpu]
         batch.extend_num_tokens = sum(batch.extend_lens)
         batch.seq_lens = batch.spec_info.seq_lens_for_draft_extend
         batch.req_pool_indices = batch.spec_info.req_pool_indices_for_draft_extend
-        seq_lens_cpu = batch.seq_lens.tolist()
+        batch.return_logprob = False
 
-        self.positions = torch.empty_like(self.verified_id, dtype=torch.long)
-        new_verified_id = torch.empty_like(self.accept_length, dtype=torch.int32)
+        self.capture_hidden_mode = CaptureHiddenMode.LAST
         self.accept_length.add_(1)
+        self.positions = torch.empty_like(batch.input_ids, dtype=torch.long)
+        self.verified_id = torch.empty_like(self.accept_length, dtype=torch.int32)
 
-        create_extend_spec_info[(self.accept_length.numel(),)](
-            self.verified_id,
+        create_extend_after_decode_spec_info[(len(batch.seq_lens),)](
+            batch.input_ids,
             batch.seq_lens,
             self.accept_length,
-            torch.cumsum(self.accept_length, axis=0, dtype=torch.int),
             self.positions,
-            new_verified_id,
-            next_power_of_2(speculative_num_steps + 1),
+            self.verified_id,
+            next_power_of_2(max(speculative_num_steps + 1, len(batch.seq_lens))),
         )
 
-        batch.seq_lens_sum = sum(seq_lens_cpu)
-        batch.input_ids = self.verified_id
-        self.verified_id = new_verified_id
-
-        if not pad_input:
-            return
-
-        batch_size = sum(not req.finished() for req in batch.reqs)
-        # Total constant input length after padding
-        static_len = speculative_num_steps + 1
-        # Total size after padding
-        padded_input_size = batch_size * static_len
-
-        padded_len = padded_input_size - batch.input_ids.shape[0]
-        if padded_len > 0:
-            new_input_ids = torch.nn.functional.pad(
-                batch.input_ids, (0, padded_len), value=0
-            )
-            position_padding = torch.arange(padded_len, device=self.positions.device)
-            new_positions = torch.cat([self.positions, position_padding])
-
-            # need dummy hidden states for the padded positions
-            hidden_states_dim = self.hidden_states.shape[-1]
-            new_hidden_states = torch.cat(
-                [
-                    self.hidden_states,
-                    torch.zeros(
-                        (padded_len, hidden_states_dim),
-                        dtype=self.hidden_states.dtype,
-                        device=self.hidden_states.device,
-                    ),
-                ],
-                dim=0,
-            )
-
-            # allocate KV cache location for the padded tokens
-            padded_cache_loc = torch.zeros(
-                padded_len,
-                dtype=batch.out_cache_loc.dtype,
-                device=batch.out_cache_loc.device,
-            )
-            new_out_cache_loc = torch.cat([batch.out_cache_loc, padded_cache_loc])
-
-            batch.input_ids = new_input_ids
-            self.hidden_states = new_hidden_states
-            self.positions = new_positions
-            batch.out_cache_loc = new_out_cache_loc
-
     def generate_attn_arg_prefill(
         self,
         req_pool_indices: torch.Tensor,
@@ -173,8 +123,9 @@ class EagleDraftInput:
         cum_kv_seq_len = torch.zeros((bs + 1,), dtype=torch.int32, device="cuda")
         cum_kv_seq_len[1:] = torch.cumsum(paged_kernel_lens, dim=0)
 
-        # TODO: replace cum_kv_seq_len[-1] with paged_kernel_lens_sum to avoid the device sync.
-        kv_indices = torch.empty(cum_kv_seq_len[-1], dtype=torch.int32, device="cuda")
+        kv_indices = torch.empty(
+            paged_kernel_lens_sum, dtype=torch.int32, device="cuda"
+        )
 
         create_flashinfer_kv_indices_triton[(bs,)](
             req_to_token,
@@ -238,54 +189,10 @@ class EagleVerifyInput:
     topk: int
     draft_token_num: int
     capture_hidden_mode: CaptureHiddenMode
+    seq_lens_sum: int
+    seq_lens_cpu: torch.Tensor
     grammar: BaseGrammarObject = None
 
-    @classmethod
-    def create(
-        cls,
-        verified_id: torch.Tensor,
-        score_list: List[torch.Tensor],
-        token_list: List[torch.Tensor],
-        parents_list: List[torch.Tensor],
-        seq_lens: torch.Tensor,
-        seq_lens_sum: int,
-        topk: int,
-        spec_steps: int,
-        num_verify_tokens: int,
-    ):
-        (
-            tree_mask,
-            position,
-            retrive_index,
-            retrive_next_token,
-            retrive_next_sibling,
-            draft_tokens,
-        ) = build_tree_kernel_efficient(
-            verified_id,
-            score_list,
-            token_list,
-            parents_list,
-            seq_lens,
-            seq_lens_sum,
-            topk,
-            spec_steps,
-            num_verify_tokens,
-        )
-
-        return cls(
-            draft_token=draft_tokens,
-            custom_mask=tree_mask,
-            positions=position,
-            retrive_index=retrive_index,
-            retrive_next_token=retrive_next_token,
-            retrive_next_sibling=retrive_next_sibling,
-            retrive_cum_len=None,
-            spec_steps=spec_steps,
-            topk=topk,
-            draft_token_num=num_verify_tokens,
-            capture_hidden_mode=CaptureHiddenMode.FULL,
-        )
-
     def prepare_for_verify(self, batch: ScheduleBatch, page_size: int):
         batch.input_ids = self.draft_token
 
@@ -614,26 +521,28 @@ class EagleVerifyInput:
 
 
 @triton.jit
-def create_extend_spec_info(
+def create_extend_after_decode_spec_info(
     verified_id,
-    seq_len,
-    accept_len,
-    accept_len_cum,
+    seq_lens,
+    accept_lens,
     positions,
     new_verified_id,
-    accept_len_upper: tl.constexpr,
+    bs_upper: tl.constexpr,
 ):
     pid = tl.program_id(axis=0)
-    offset = 0 if pid == 0 else tl.load(accept_len_cum + pid - 1)
-    seq_length = tl.load(seq_len + pid)
-    accept_length = tl.load(accept_len + pid)
-    positions_ptr = positions + offset
-    data = tl.arange(0, accept_len_upper)
-    mask = data < accept_length
-    tl.store(positions_ptr + data, seq_length - accept_length + data, mask)
-
-    offset = tl.load(accept_len_cum + pid) - 1
-    verified_id_data = tl.load(verified_id + offset)
+    offsets = tl.arange(0, bs_upper)
+    seq_length = tl.load(seq_lens + pid)
+    accept_length = tl.load(accept_lens + pid)
+
+    accept_len_cumsum = tl.sum(
+        tl.load(accept_lens + offsets, mask=offsets < pid, other=0)
+    )
+    positions_ptr = positions + accept_len_cumsum
+    mask = offsets < accept_length
+    tl.store(positions_ptr + offsets, seq_length - accept_length + offsets, mask)
+
+    accept_len_cumsum += accept_length - 1
+    verified_id_data = tl.load(verified_id + accept_len_cumsum)
     tl.store(new_verified_id + pid, verified_id_data)
 
 
@@ -654,8 +563,8 @@ def assign_req_to_token_pool(
     token_pool = req_to_token + tl.load(req_pool_indices + pid) * pool_len
 
     length_offset = tl.arange(0, bs_upper)
-    start = tl.load(start_offset + length_offset, mask=length_offset < pid)
-    end = tl.load(end_offset + length_offset, mask=length_offset < pid)
+    start = tl.load(start_offset + length_offset, mask=length_offset < pid, other=0)
+    end = tl.load(end_offset + length_offset, mask=length_offset < pid, other=0)
     out_offset = tl.sum(end - start, axis=0)
 
     out_cache_ptr = out_cache_loc + out_offset
@@ -736,7 +645,7 @@ def generate_draft_decode_kv_indices(
     iters += 1
 
     load_offset = tl.arange(0, bs_upper)
-    seq_lens = tl.load(paged_kernel_lens + load_offset, mask=load_offset < bid)
+    seq_lens = tl.load(paged_kernel_lens + load_offset, mask=load_offset < bid, other=0)
     seq_len = tl.load(paged_kernel_lens + bid)
     cum_seq_len = tl.sum(seq_lens)
 
@@ -765,7 +674,7 @@ def generate_draft_decode_kv_indices(
     zid = bid * topk + topk_id
     if zid == 0:
         zid = num_seqs * topk
-    positions = tl.load(positions + bs_offset, mask=bs_offset < zid)
+    positions = tl.load(positions + bs_offset, mask=bs_offset < zid, other=0)
     base = tl.sum(positions)
     tl.store(kv_indptr + zid, base + zid * iters)
 
@@ -783,7 +692,9 @@ def align_evict_mask_to_page_size(
     bid = tl.program_id(axis=0)
     seq_len = tl.load(seq_lens + bid)
     io_mask = t_range < num_draft_tokens
-    mask_row = tl.load(evict_mask + bid * num_draft_tokens + t_range, mask=io_mask)
+    mask_row = tl.load(
+        evict_mask + bid * num_draft_tokens + t_range, mask=io_mask, other=0
+    )
 
     num_trues = tl.sum(mask_row)
     num_false = num_draft_tokens - num_trues

python/sglang/srt/speculative/eagle_worker.py

@@ -23,6 +23,7 @@ from sglang.srt.model_executor.forward_batch_info import (
     ForwardMode,
 )
 from sglang.srt.server_args import ServerArgs
+from sglang.srt.speculative.build_eagle_tree import build_tree_kernel_efficient
 from sglang.srt.speculative.eagle_draft_cuda_graph_runner import (
     EAGLEDraftCudaGraphRunner,
 )
@@ -69,7 +70,6 @@ class EAGLEWorker(TpModelWorker):
         self.server_args = server_args
         self.topk = server_args.speculative_eagle_topk
         self.speculative_num_steps = server_args.speculative_num_steps
-        self.padded_static_len = self.speculative_num_steps + 1
         self.enable_nan_detection = server_args.enable_nan_detection
         self.gpu_id = gpu_id
         self.device = server_args.device
@@ -78,6 +78,7 @@ class EAGLEWorker(TpModelWorker):
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
             server_args.speculative_algorithm
         )
+        self.padded_static_len = -1
 
         # Override context length with target model's context length
         server_args.context_length = target_worker.model_runner.model_config.context_len
@@ -184,7 +185,6 @@ class EAGLEWorker(TpModelWorker):
                     self.draft_model_runner,
                     skip_prefill=False,
                 )
-            self.padded_static_len = self.speculative_num_steps + 1
             self.has_prefill_wrapper_verify = True
         elif self.server_args.attention_backend == "triton":
             from sglang.srt.layers.attention.triton_backend import (
@@ -201,7 +201,6 @@ class EAGLEWorker(TpModelWorker):
                 self.draft_model_runner,
                 skip_prefill=False,
             )
-            self.padded_static_len = self.speculative_num_steps + 1
             self.has_prefill_wrapper_verify = False
         elif self.server_args.attention_backend == "fa3":
             from sglang.srt.layers.attention.flashattention_backend import (
@@ -218,7 +217,6 @@ class EAGLEWorker(TpModelWorker):
                 self.draft_model_runner,
                 skip_prefill=False,
             )
-            self.padded_static_len = self.speculative_num_steps + 1
             self.has_prefill_wrapper_verify = False
         elif self.server_args.attention_backend == "flashmla":
             from sglang.srt.layers.attention.flashmla_backend import (
@@ -231,7 +229,6 @@ class EAGLEWorker(TpModelWorker):
                 self.speculative_num_steps,
             )
             self.draft_extend_attn_backend = None
-            self.padded_static_len = self.speculative_num_steps + 1
             self.has_prefill_wrapper_verify = False
         else:
             raise ValueError(
@@ -319,10 +316,12 @@ class EAGLEWorker(TpModelWorker):
 
             return logits_output, next_token_ids, model_worker_batch.bid, 0, False
         else:
-            logits_output, next_token_ids, bid = self.forward_target_extend(batch)
+            logits_output, next_token_ids, bid, seq_lens_cpu = (
+                self.forward_target_extend(batch)
+            )
             with self.draft_tp_context(self.draft_model_runner.tp_group):
                 self.forward_draft_extend(
-                    batch, logits_output.hidden_states, next_token_ids
+                    batch, logits_output.hidden_states, next_token_ids, seq_lens_cpu
                 )
             return logits_output, next_token_ids, bid, 0, False
 
@@ -346,7 +345,12 @@ class EAGLEWorker(TpModelWorker):
         logits_output, next_token_ids, _ = self.target_worker.forward_batch_generation(
             model_worker_batch
         )
-        return logits_output, next_token_ids, model_worker_batch.bid
+        return (
+            logits_output,
+            next_token_ids,
+            model_worker_batch.bid,
+            model_worker_batch.seq_lens_cpu,
+        )
 
     def draft(self, batch: ScheduleBatch):
         # Parse args
@@ -452,7 +456,14 @@ class EAGLEWorker(TpModelWorker):
 
         self.token_to_kv_pool_allocator.restore_state(token_to_kv_pool_state_backup)
 
-        ret = EagleVerifyInput.create(
+        (
+            tree_mask,
+            position,
+            retrive_index,
+            retrive_next_token,
+            retrive_next_sibling,
+            draft_tokens,
+        ) = build_tree_kernel_efficient(
             spec_info.verified_id,
             score_list,
             token_list,
@@ -463,7 +474,22 @@ class EAGLEWorker(TpModelWorker):
             self.speculative_num_steps,
             self.server_args.speculative_num_draft_tokens,
         )
-        return ret
+
+        return EagleVerifyInput(
+            draft_token=draft_tokens,
+            custom_mask=tree_mask,
+            positions=position,
+            retrive_index=retrive_index,
+            retrive_next_token=retrive_next_token,
+            retrive_next_sibling=retrive_next_sibling,
+            retrive_cum_len=None,
+            spec_steps=self.speculative_num_steps,
+            topk=self.topk,
+            draft_token_num=self.server_args.speculative_num_draft_tokens,
+            capture_hidden_mode=CaptureHiddenMode.FULL,
+            seq_lens_sum=forward_batch.seq_lens_sum,
+            seq_lens_cpu=forward_batch.seq_lens_cpu,
+        )
 
     def draft_forward(self, forward_batch: ForwardBatch):
         # Parse args
@@ -523,7 +549,9 @@ class EAGLEWorker(TpModelWorker):
         spec_info.prepare_for_verify(batch, self.page_size)
         batch.forward_mode = ForwardMode.TARGET_VERIFY
         batch.spec_info = spec_info
-        model_worker_batch = batch.get_model_worker_batch()
+        model_worker_batch = batch.get_model_worker_batch(
+            seq_lens_cpu_cache=spec_info.seq_lens_cpu
+        )
 
         if batch.has_grammar:
             retrieve_next_token_cpu = spec_info.retrive_next_token.cpu()
@@ -650,6 +678,7 @@ class EAGLEWorker(TpModelWorker):
         batch: ScheduleBatch,
         hidden_states: torch.Tensor,
         next_token_ids: List[int],
+        seq_lens_cpu: torch.Tensor,
     ):
         """Run draft model extend. This API modifies the states of the batch.
 
@@ -664,7 +693,9 @@ class EAGLEWorker(TpModelWorker):
         )
         batch.spec_info.prepare_for_extend(batch)
         batch.spec_info.capture_hidden_mode = CaptureHiddenMode.LAST
-        model_worker_batch = batch.get_model_worker_batch()
+        model_worker_batch = batch.get_model_worker_batch(
+            seq_lens_cpu_cache=seq_lens_cpu
+        )
         forward_batch = ForwardBatch.init_new(
             model_worker_batch, self.draft_model_runner
         )
@@ -683,19 +714,18 @@ class EAGLEWorker(TpModelWorker):
         return_logprob_backup = batch.return_logprob
 
         # Prepare metadata
-        batch.forward_mode = ForwardMode.DRAFT_EXTEND
         batch.spec_info.prepare_extend_after_decode(
             batch,
             self.speculative_num_steps,
-            self.server_args.context_length,
-            pad_input=self.cuda_graph_runner_for_draft_extend is not None,
         )
-        batch.spec_info.capture_hidden_mode = CaptureHiddenMode.LAST
-        batch.return_logprob = False
         model_worker_batch = batch.get_model_worker_batch()
         forward_batch = ForwardBatch.init_new(
             model_worker_batch, self.draft_model_runner
         )
+        if forward_batch.seq_lens_cpu is not None:
+            forward_batch.seq_lens_sum = forward_batch.seq_lens_cpu.sum().item()
+        else:
+            forward_batch.seq_lens_sum = batch.seq_lens.sum().item()
 
         # Run
         can_cuda_graph = (
@@ -706,14 +736,19 @@ class EAGLEWorker(TpModelWorker):
             logits_output = self.cuda_graph_runner_for_draft_extend.replay(
                 forward_batch
             )
+            forward_batch.spec_info.topk_p, forward_batch.spec_info.topk_index = (
+                logits_output.topk_p,
+                logits_output.topk_index,
+            )
+            forward_batch.spec_info.hidden_states = logits_output.hidden_states
         else:
             self.draft_model_runner.attn_backend.init_forward_metadata(forward_batch)
             logits_output = self.draft_model_runner.model.forward(
                 forward_batch.input_ids, forward_batch.positions, forward_batch
             )
+            self.capture_for_decode(logits_output, forward_batch.spec_info)
 
         self._detect_nan_if_needed(logits_output)
-        self.capture_for_decode(logits_output, forward_batch.spec_info)
 
         # Restore backup.
         # This is because `seq_lens` can be modified in `prepare_extend_after_decode`

test/srt/test_full_deepseek_v3.py

@@ -87,7 +87,7 @@ class TestDeepseekV3MTP(CustomTestCase):
             "--speculative-num-steps",
             "3",
             "--speculative-eagle-topk",
-            "2",
+            "1",
             "--speculative-num-draft-tokens",
             "4",
         ]
@@ -155,7 +155,7 @@ class TestDeepseekV3MTP(CustomTestCase):
             if is_in_amd_ci():
                 self.assertGreater(speed, 15)
             else:
-                self.assertGreater(speed, 105)
+                self.assertGreater(speed, 130)
 
 
 if __name__ == "__main__":