Fix Req fill_len (fill_ids) having dual semantics by restricting to truncated/committed semantics (#26659)

python/sglang/srt/dllm/mixin/req.py

@@ -42,11 +42,11 @@ class ReqDllmMixin:
         prefix_length = len(self.prefix_indices)
         min_required_length = prefix_length + self.dllm_config.block_size
 
-        if self.fill_len < min_required_length:
+        if len(self.full_untruncated_fill_ids) < min_required_length:
             # still incoming stage
             return
 
-        input_block = self.get_fill_ids()[prefix_length:min_required_length]
+        input_block = self.full_untruncated_fill_ids[prefix_length:min_required_length]
         is_prefill_phase = self.dllm_config.mask_id not in input_block
 
         if is_prefill_phase:

python/sglang/srt/managers/schedule_batch.py

@@ -1078,9 +1078,8 @@ class Req(ReqDllmMixin):
             self.determine_dllm_phase()
         else:
             self.full_untruncated_fill_ids = self.origin_input_ids + self.output_ids
-            self.fill_len = len(self.full_untruncated_fill_ids)
 
-        input_len = self.fill_len
+        input_len = len(self.full_untruncated_fill_ids)
 
         # Streaming sessions reuse committed KV from the session slot, so
         # custom logprob_start_len is not supported — override to -1.
@@ -1098,7 +1097,7 @@ class Req(ReqDllmMixin):
             )
             self.logprob_start_len = -1
 
-        token_ids_to_match = self.get_fill_ids()[
+        token_ids_to_match = self.full_untruncated_fill_ids[
             : self._compute_max_prefix_len(input_len)
         ]
 
@@ -1163,7 +1162,7 @@ class Req(ReqDllmMixin):
                 )
             )
 
-        self.set_extend_input_len(self.fill_len - len(self.prefix_indices))
+        self.set_extend_input_len(input_len - len(self.prefix_indices))
 
     def _compute_max_prefix_len(self, input_len: int) -> int:
         # NOTE: the matched length is at most 1 less than the input length to enable logprob computation
@@ -1371,6 +1370,7 @@ class Req(ReqDllmMixin):
         self.swa_evicted_seqlen = 0
         self.extend_batch_idx = 0
         self.decode_batch_idx = 0
+        self.fill_len = 0
 
         # When using input_embeds, we cannot easily mix the original input embeddings
         # with the newly generated output token IDs during re-prefill of retracted request.
@@ -1429,7 +1429,7 @@ class Req(ReqDllmMixin):
         # - extend_input_len: Number of tokens that need to be processed in this extend batch
         self.extend_input_len = extend_input_len
         if self.logprob_start_len == -1:
-            logprob_start_len = self.fill_len
+            logprob_start_len = len(self.full_untruncated_fill_ids)
         else:
             # logprob_start_len should be at least the length of the prefix indices
             logprob_start_len = max(self.logprob_start_len, len(self.prefix_indices))

python/sglang/srt/managers/schedule_policy.py

@@ -827,7 +827,11 @@ class PrefillAdder:
             self.rem_chunk_tokens is None  # chunked prefill is disabled
             or req.extend_input_len <= self.rem_chunk_tokens  # it is the last chunk
         ):
-            # Non-chunked prefill
+            # Non-chunked prefill — the whole sequence is committed this iter.
+            req.fill_len = len(req.full_untruncated_fill_ids)
+            assert (
+                req.fill_len == len(req.prefix_indices) + req.extend_input_len
+            ), f"{req.fill_len=} {len(req.prefix_indices)=} {req.extend_input_len=}"
             self.can_run_list.append(req)
             self._update_prefill_budget(
                 0,
@@ -932,7 +936,9 @@ class PrefillAdder:
                     )
                 )
                 req.prefix_indices = torch.cat([req.prefix_indices, new_indices])
-                req.set_extend_input_len(req.fill_len - len(req.prefix_indices))
+                req.set_extend_input_len(
+                    len(req.full_untruncated_fill_ids) - len(req.prefix_indices)
+                )
                 prefix_len = len(req.prefix_indices)
                 req.cache_protected_len = prefix_len
 
@@ -959,7 +965,11 @@ class PrefillAdder:
                 self._add_dllm_req(req, prefix_len)
                 self._req_inc_lock_ref(req)
             elif self.rem_chunk_tokens is None or input_tokens <= self.rem_chunk_tokens:
-                # Non-chunked prefill
+                # Non-chunked prefill — the whole sequence is committed this iter.
+                req.fill_len = len(req.full_untruncated_fill_ids)
+                assert (
+                    req.fill_len == len(req.prefix_indices) + req.extend_input_len
+                ), f"{req.fill_len=} {len(req.prefix_indices)=} {req.extend_input_len=}"
                 self.can_run_list.append(req)
 
                 self._req_inc_lock_ref(req)

python/sglang/srt/managers/scheduler.py

@@ -2159,7 +2159,7 @@ class Scheduler(
             if last_host_node.backuped or last_host_node is self.tree_cache.root_node:
                 last_hash = last_host_node.get_last_hash_value()
                 matched_len = len(req.prefix_indices) + req.host_hit_length
-                new_input_tokens = req.get_fill_ids()[matched_len:]
+                new_input_tokens = req.full_untruncated_fill_ids[matched_len:]
 
                 prefix_keys = (
                     last_host_node.get_prefix_hash_values(last_host_node.parent)