[Spec] Support DSPARK in disaggregation decode mode

python/sglang/srt/speculative/dflash_disaggregation.py

@@ -0,0 +1,37 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+
+from sglang.srt.managers.overlap_utils import RelayPayload
+from sglang.srt.speculative.draft_worker_common import make_draft_input_v2
+
+if TYPE_CHECKING:
+    from sglang.srt.managers.overlap_utils import FutureMap
+    from sglang.srt.managers.schedule_batch import ScheduleBatch
+    from sglang.srt.server_args import ServerArgs
+    from sglang.srt.speculative.dflash_info_v2 import DFlashDraftInputV2
+
+
+def build_dflash_disagg_draft_input(
+    batch: ScheduleBatch,
+    server_args: ServerArgs,
+    last_tokens_tensor: torch.Tensor,
+    future_map: FutureMap,
+) -> "DFlashDraftInputV2":
+    spec_info = make_draft_input_v2(
+        bonus_tokens=last_tokens_tensor,
+        new_seq_lens=batch.seq_lens,
+    )
+
+    if batch.enable_overlap:
+        spec_info.future_indices = batch.req_pool_indices
+        # Seed the relay buf with the known seq_lens; publish's chained record
+        # keeps the in-flight forward's fence intact (see FutureMap.publish).
+        future_map.publish(spec_info.future_indices, batch.seq_lens)
+        future_map.stash(
+            spec_info.future_indices, RelayPayload.from_draft_input(spec_info)
+        )
+
+    return spec_info

python/sglang/srt/speculative/dspark_components/dspark_draft.py

@@ -415,6 +415,7 @@ def _fill_dp_moe_sync_metadata(
             batch.global_num_tokens_for_logprob,
         )
         device = self.draft_model_runner.device
+        forward_batch.original_global_num_tokens_cpu = batch.global_num_tokens
         forward_batch.global_num_tokens_cpu = gnt
         forward_batch.global_num_tokens_for_logprob_cpu = gnt_logprob
         forward_batch.global_num_tokens_gpu = torch.tensor(gnt, dtype=torch.int64).to(

python/sglang/srt/speculative/spec_info.py

@@ -176,6 +176,14 @@ def build_disagg_draft_input(
             return build_eagle_disagg_draft_input(
                 batch, server_args, last_tokens_tensor, future_map
             )
+        if self.is_dflash_family():
+            from sglang.srt.speculative.dflash_disaggregation import (
+                build_dflash_disagg_draft_input,
+            )
+
+            return build_dflash_disagg_draft_input(
+                batch, server_args, last_tokens_tensor, future_map
+            )
         return None
 
     def need_topk(self) -> bool: