[Bugfix] Fix computed-token metadata for requests split across DBO microbatches

tests/v1/attention/test_attention_splitting.py

@@ -390,6 +390,20 @@ def test_prefill_split_across_ubatches(
     # Second ubatch: first request (continuation) seq_len should be full
     #  original
     assert int(second_meta.seq_lens[0]) == seq_lens[split_req_idx]
+
+    # Tokens processed for the first chunk become computed context for the
+    # continuation. The original metadata must remain unchanged.
+    assert first_meta._num_computed_tokens_cpu is not None
+    assert second_meta._num_computed_tokens_cpu is not None
+    assert common._num_computed_tokens_cpu is not None
+    assert int(first_meta._num_computed_tokens_cpu[-1]) == context_lens[split_req_idx]
+    assert int(second_meta._num_computed_tokens_cpu[0]) == (
+        context_lens[split_req_idx] + tokens_in_first_chunk
+    )
+    assert int(common._num_computed_tokens_cpu[split_req_idx]) == context_lens[
+        split_req_idx
+    ]
+
     # Any following full requests in second ubatch should match originals
     for j in range(1, second_meta.num_reqs):
         # Map to original request index

vllm/v1/worker/ubatch_utils.py

@@ -296,6 +296,11 @@ def _make_metadata_with_slice(
         if attn_metadata._num_computed_tokens_cpu is not None
         else None
     )
+    if splits_first_request and num_computed_tokens_cpu is not None:
+        # Tokens from the preceding microbatch are now part of this request's
+        # computed context. Clone to avoid mutating the parent metadata.
+        num_computed_tokens_cpu = num_computed_tokens_cpu.clone()
+        num_computed_tokens_cpu[0] += first_tok - start_locs[first_req]
 
     if splits_last_request:
         # NOTE: We use start_locs (the original query_start_loc_cpu) to calculate