[feature] extend DBO to XBO (#30120)

Signed-off-by: jiangkuaixue123 <[email protected]>
Co-authored-by: root <[email protected]>

Author: jiangkuaixue123

tests/v1/attention/test_attention_splitting.py

@@ -323,6 +323,7 @@ def test_prefill_split_across_ubatches(
         num_tokens,
         batch_spec.batch_size,
         split_point=split_point,
+        num_ubatches=2,
     )
     assert ubatch_slices is not None and len(ubatch_slices) == 2
 

vllm/config/parallel.py

@@ -156,6 +156,8 @@ class ParallelConfig:
 
     enable_dbo: bool = False
     """Enable dual batch overlap for the model executor."""
+    ubatch_size: int = 0
+    """Number of ubatch size."""
 
     dbo_decode_token_threshold: int = 32
     """The threshold for dual batch overlap for batches only containing decodes.
@@ -325,6 +327,14 @@ def world_size_across_dp(self) -> int:
         including data parallelism."""
         return self.world_size * self.data_parallel_size
 
+    @property
+    def use_ubatching(self) -> bool:
+        return self.enable_dbo or self.ubatch_size > 1
+
+    @property
+    def num_ubatches(self) -> int:
+        return 2 if self.enable_dbo else self.ubatch_size
+
     def get_next_dp_init_port(self) -> int:
         """
         We might need to initialize process groups in multiple

vllm/config/vllm.py

@@ -870,9 +870,12 @@ def has_blocked_weights():
                     f"cudagraph_mode={self.compilation_config.cudagraph_mode}"
                 )
 
-        if self.parallel_config.enable_dbo:
+        if self.parallel_config.use_ubatching:
             a2a_backend = self.parallel_config.all2all_backend
-            assert a2a_backend in ["deepep_low_latency", "deepep_high_throughput"], (
+            assert a2a_backend in [
+                "deepep_low_latency",
+                "deepep_high_throughput",
+            ], (
                 "Microbatching currently only supports the deepep_low_latency and "
                 f"deepep_high_throughput all2all backend. {a2a_backend} is not "
                 "supported. To fix use --all2all-backend=deepep_low_latency or "

vllm/engine/arg_utils.py

@@ -408,6 +408,7 @@ class EngineArgs:
     enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
     all2all_backend: str | None = ParallelConfig.all2all_backend
     enable_dbo: bool = ParallelConfig.enable_dbo
+    ubatch_size: int = ParallelConfig.ubatch_size
     dbo_decode_token_threshold: int = ParallelConfig.dbo_decode_token_threshold
     dbo_prefill_token_threshold: int = ParallelConfig.dbo_prefill_token_threshold
     disable_nccl_for_dp_synchronization: bool = (
@@ -841,6 +842,10 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             "--all2all-backend", **parallel_kwargs["all2all_backend"]
         )
         parallel_group.add_argument("--enable-dbo", **parallel_kwargs["enable_dbo"])
+        parallel_group.add_argument(
+            "--ubatch-size",
+            **parallel_kwargs["ubatch_size"],
+        )
         parallel_group.add_argument(
             "--dbo-decode-token-threshold",
             **parallel_kwargs["dbo_decode_token_threshold"],
@@ -1557,6 +1562,7 @@ def create_engine_config(
             enable_expert_parallel=self.enable_expert_parallel,
             all2all_backend=self.all2all_backend,
             enable_dbo=self.enable_dbo,
+            ubatch_size=self.ubatch_size,
             dbo_decode_token_threshold=self.dbo_decode_token_threshold,
             dbo_prefill_token_threshold=self.dbo_prefill_token_threshold,
             disable_nccl_for_dp_synchronization=self.disable_nccl_for_dp_synchronization,

vllm/v1/attention/backends/utils.py

@@ -201,10 +201,11 @@ def _make_metadata_with_slice(
     )
     # NOTE: last token can be outside of the last request if we have CG padding.
 
-    # If the "middle" request has tokens in both ubatches, we have to split it.
-    # If ubatch_slice is the first ubatch then we will be splitting the last
-    # request. If it's the second microbatch, then we will be splitting the
-    # first request
+    # If the request is split across ubatches, we have to adjust the metadata.
+    # splits_first_request: The first request in this slice is the continuation of
+    #                       a request that started in a previous slice.
+    # splits_last_request:  The last request in this slice continues into the
+    #                       next slice.
     splits_first_request = first_tok > start_locs[first_req]
     splits_last_request = last_tok < start_locs[last_req + 1] - 1
 
@@ -225,7 +226,10 @@ def _make_metadata_with_slice(
     seq_lens_cpu = attn_metadata.seq_lens_cpu[request_slice]
 
     if splits_last_request:
-        tokens_skipped = query_start_loc_cpu[-1] - token_slice.stop
+        # NOTE: We use start_locs (the original query_start_loc_cpu) to calculate
+        # the tokens skipped because query_start_loc_cpu might have been modified
+        # if splits_first_request is True.
+        tokens_skipped = start_locs[last_req + 1] - token_slice.stop
         query_start_loc[-1] -= tokens_skipped
         query_start_loc_cpu[-1] -= tokens_skipped
 

vllm/v1/worker/dp_utils.py

@@ -11,7 +11,7 @@
 from vllm.logger import init_logger
 from vllm.v1.worker.ubatch_utils import (
     check_ubatch_thresholds,
-    is_second_ubatch_empty,
+    is_last_ubatch_empty,
 )
 
 logger = init_logger(__name__)
@@ -56,7 +56,7 @@ def _run_ar(
     return tensor
 
 
-def _post_process_ubatch(tensor: torch.Tensor) -> bool:
+def _post_process_ubatch(tensor: torch.Tensor, num_ubatches: int) -> bool:
     orig_num_tokens_tensor = tensor[0, :]
     padded_num_tokens_tensor = tensor[1, :]
 
@@ -68,7 +68,7 @@ def _post_process_ubatch(tensor: torch.Tensor) -> bool:
     # there are no "empty" second ubatches
     orig_min_num_tokens = int(orig_num_tokens_tensor.min().item())
     padded_max_num_tokens = int(padded_num_tokens_tensor.max().item())
-    if is_second_ubatch_empty(orig_min_num_tokens, padded_max_num_tokens):
+    if is_last_ubatch_empty(orig_min_num_tokens, padded_max_num_tokens, num_ubatches):
         logger.debug(
             "Aborting ubatching %s %s", orig_min_num_tokens, padded_max_num_tokens
         )
@@ -146,7 +146,7 @@ def _synchronize_dp_ranks(
     assert should_attempt_dp_padding == should_dp_pad
 
     # Check conditions for microbatching
-    should_ubatch = _post_process_ubatch(tensor)
+    should_ubatch = _post_process_ubatch(tensor, parallel_config.num_ubatches)
 
     if should_ubatch and not should_dp_pad:
         logger.debug_once(

vllm/v1/worker/gpu_model_runner.py

@@ -2987,7 +2987,7 @@ def execute_model(
 
                 cascade_attn_prefix_lens = None
                 # Disable cascade attention when using microbatching (DBO)
-                if self.cascade_attn_enabled and not self.parallel_config.enable_dbo:
+                if self.cascade_attn_enabled and not self.parallel_config.use_ubatching:
                     # Pre-compute cascade attention prefix lengths
                     cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
                         num_scheduled_tokens_np,
@@ -3028,6 +3028,13 @@ def execute_model(
                     num_scheduled_tokens_np,
                     num_tokens_padded,
                     num_reqs_padded,
+                    self.parallel_config.num_ubatches,
+                )
+
+                logger.debug(
+                    "ubatch_slices: %s, ubatch_slices_padded: %s",
+                    ubatch_slices,
+                    ubatch_slices_padded,
                 )
 
                 pad_attn = cudagraph_mode == CUDAGraphMode.FULL
@@ -3710,11 +3717,14 @@ def load_model(self, eep_scale_up: bool = False) -> None:
         # wrap the model with full cudagraph wrapper if needed.
         cudagraph_mode = self.compilation_config.cudagraph_mode
         assert cudagraph_mode is not None
-        if cudagraph_mode.has_full_cudagraphs() and not self.parallel_config.enable_dbo:
+        if (
+            cudagraph_mode.has_full_cudagraphs()
+            and not self.parallel_config.use_ubatching
+        ):
             self.model = CUDAGraphWrapper(
                 self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
             )
-        elif self.parallel_config.enable_dbo:
+        elif self.parallel_config.use_ubatching:
             if cudagraph_mode.has_full_cudagraphs():
                 self.model = UBatchWrapper(
                     self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
@@ -4095,7 +4105,16 @@ def _dummy_run(
             batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
         )
         ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
-            should_ubatch, num_scheduled_tokens, num_tokens_padded, num_reqs_padded
+            should_ubatch,
+            num_scheduled_tokens,
+            num_tokens_padded,
+            num_reqs_padded,
+            self.vllm_config.parallel_config.num_ubatches,
+        )
+        logger.debug(
+            "ubatch_slices: %s, ubatch_slices_padded: %s",
+            ubatch_slices,
+            ubatch_slices_padded,
         )
 
         attn_metadata: PerLayerAttnMetadata | None = None
@@ -4644,7 +4663,7 @@ def _capture_cudagraphs(
             # is above the threshold. Otherwise we just capture a non-ubatched
             # version of the graph
             allow_microbatching = (
-                self.parallel_config.enable_dbo
+                self.parallel_config.use_ubatching
                 and cudagraph_runtime_mode == CUDAGraphMode.FULL
                 and uniform_decode
                 and check_ubatch_thresholds(
@@ -4779,8 +4798,8 @@ def initialize_metadata_builders(
                     if kv_cache_group_id < len(kernel_block_sizes)
                     else None,
                     num_metadata_builders=1
-                    if not self.parallel_config.enable_dbo
-                    else 2,
+                    if not self.parallel_config.use_ubatching
+                    else self.parallel_config.num_ubatches,
                 )
         # Calculate reorder batch threshold (if needed)
         # Note (tdoublep): do this *after* constructing builders,

vllm/v1/worker/gpu_ubatch_wrapper.py

@@ -103,8 +103,10 @@ def __init__(
         self.vllm_config = vllm_config
         self.compilation_config = vllm_config.compilation_config
         self.comm_stream = torch.cuda.Stream(device=device)
-        # Two ubatch threads plus the main thread
-        self.ready_barrier = threading.Barrier(3)
+        # Ubatch threads plus the main thread
+        self.ready_barrier = threading.Barrier(
+            self.vllm_config.parallel_config.num_ubatches + 1
+        )
 
         self.cudagraphs: dict[int, CUDAGraphMetaData] = {}
 
@@ -309,7 +311,7 @@ def _make_ubatch_metadata(
                 create_forward_context(
                     attn_metadata[i] if attn_metadata is not None else None,
                     self.vllm_config,
-                    dp_metadata=dp_metadata,
+                    dp_metadata=dp_metadata[i],
                     batch_descriptor=batch_descriptor,
                     cudagraph_runtime_mode=cudagraph_runtime_mode,
                 )
@@ -417,18 +419,19 @@ def __call__(self, *args, **kwargs):
 
         # We shouldn't be here unless we are running with multiple DP ranks
         assert dp_metadata is not None
-        num_tokens_per_ubatch = (
-            ubatch_slices[0].token_slice.stop - ubatch_slices[0].token_slice.start
-        )
-        dp_size = self.vllm_config.parallel_config.data_parallel_size
-        ubatch_num_tokens_across_dp = torch.tensor(
-            [num_tokens_per_ubatch] * dp_size, device="cpu", dtype=torch.int32
-        )
-        ubatch_dp_metadata = DPMetadata.make(
-            self.vllm_config.parallel_config,
-            num_tokens_per_ubatch,
-            ubatch_num_tokens_across_dp,
-        )
+        ubatch_dp_metadata = []
+        for ubatch_slice in ubatch_slices:
+            dp_size = self.vllm_config.parallel_config.data_parallel_size
+            ubatch_num_tokens_across_dp = torch.tensor(
+                [ubatch_slice.num_tokens] * dp_size, device="cpu", dtype=torch.int32
+            )
+            ubatch_dp_metadata.append(
+                DPMetadata.make(
+                    self.vllm_config.parallel_config,
+                    ubatch_slice.num_tokens,
+                    ubatch_num_tokens_across_dp,
+                )
+            )
 
         if (
             num_tokens not in self.cudagraphs
@@ -464,7 +467,7 @@ def __call__(self, *args, **kwargs):
                 intermediate_tensors=intermediate_tensors,
                 inputs_embeds=inputs_embeds,
                 compute_stream=compute_stream,
-                dp_metadata=dp_metadata,
+                dp_metadata=ubatch_dp_metadata,
                 batch_descriptor=batch_descriptor,
                 cudagraph_runtime_mode=CUDAGraphMode.NONE,
             )

vllm/v1/worker/ubatch_utils.py

@@ -27,36 +27,34 @@ def num_tokens(self) -> int:
 UBatchSlices: TypeAlias = list[UBatchSlice]
 
 
-def is_second_ubatch_empty(orig_num_tokens: int, padded_num_tokens: int) -> bool:
-    return (padded_num_tokens // 2) >= orig_num_tokens
+def is_last_ubatch_empty(
+    orig_num_tokens: int, padded_num_tokens: int, num_ubatches: int
+) -> bool:
+    return (padded_num_tokens // num_ubatches) * (num_ubatches - 1) >= orig_num_tokens
 
 
 def check_ubatch_thresholds(
     config: ParallelConfig, num_tokens: int, uniform_decode: bool
 ) -> bool:
-    if not config.enable_dbo:
+    if not config.use_ubatching:
         return False
     if uniform_decode:
         return num_tokens >= config.dbo_decode_token_threshold
     else:
         return num_tokens >= config.dbo_prefill_token_threshold
 
 
-# This just pads the second ubatch slice out to the total number of tokens
+# This pads the last ubatch slice out to the total number of tokens
 # (num_tokens + padding) since we do `create_ubatch_slices` before applying DP padding.
 def _pad_out_ubatch_slices(
     ubatch_slices: UBatchSlices, num_total_tokens: int, num_reqs_padded: int
 ) -> UBatchSlices:
-    # TODO(lucas): handle empty second ubatch
-    padded_second_request_slice = slice(
-        ubatch_slices[1].request_slice.start, num_reqs_padded
-    )
-    padded_second_token_slice = slice(
-        ubatch_slices[1].token_slice.start, num_total_tokens
-    )
-    return [
-        ubatch_slices[0],
-        UBatchSlice(padded_second_request_slice, padded_second_token_slice),
+    last_slice = ubatch_slices[-1]
+    padded_last_request_slice = slice(last_slice.request_slice.start, num_reqs_padded)
+    padded_last_token_slice = slice(last_slice.token_slice.start, num_total_tokens)
+
+    return ubatch_slices[:-1] + [
+        UBatchSlice(padded_last_request_slice, padded_last_token_slice)
     ]
 
 
@@ -65,40 +63,45 @@ def maybe_create_ubatch_slices(
     num_scheduled_tokens: np.ndarray,
     num_tokens_padded: int,
     num_reqs_padded: int,
-    split_point: int | None = None,
+    num_ubatches: int,
+    split_point: list[int] | int | None = None,
 ) -> tuple[UBatchSlices | None, UBatchSlices | None]:
     if not should_ubatch:
         return None, None
 
     if split_point is None:
-        split_point = int(num_tokens_padded) // 2
+        split_point = int(num_tokens_padded) // num_ubatches
+
+    token_split_points = [split_point * i for i in range(1, num_ubatches)]
 
     # TODO(lucas): Refactor the gpu_model_runner.py so we can pass
     # in cu_num_tokens directly (i.e. query_start_loc)
     cu_num_tokens = np.zeros(len(num_scheduled_tokens) + 1, dtype=np.int32)
     np.cumsum(num_scheduled_tokens, dtype=np.int32, out=cu_num_tokens[1:])
 
-    first_ubatch_token_slice = slice(0, split_point)
-    second_ubatch_token_slice = slice(split_point, cu_num_tokens[-1])
+    ubatch_slices = []
+    start_token = 0
 
-    # Determine request slices using exclusive stop semantics
-    # First ubatch includes requests whose tokens overlap [0, split_point)
-    first_ubatch_req_stop = int(
-        np.searchsorted(cu_num_tokens, split_point, side="left")
-    )
-    first_ubatch_req_slice = slice(0, first_ubatch_req_stop)
+    # Add the end point to the split points to make iteration easier
+    all_points = token_split_points + [cu_num_tokens[-1]]
 
-    # Second ubatch starts at the request that contains the split_point
-    # or the request starting exactly at split_point (if on boundary)
-    second_ubatch_req_start = int(
-        np.searchsorted(cu_num_tokens, split_point, side="right") - 1
-    )
-    second_ubatch_req_slice = slice(second_ubatch_req_start, len(cu_num_tokens) - 1)
+    for end_token in all_points:
+        token_slice = slice(start_token, end_token)
 
-    ubatch_slices = [
-        UBatchSlice(first_ubatch_req_slice, first_ubatch_token_slice),
-        UBatchSlice(second_ubatch_req_slice, second_ubatch_token_slice),
-    ]
+        # Determine request slices using exclusive stop semantics
+        # Ubatch includes requests whose tokens overlap [start_token, end_token)
+
+        # Start at the request that contains the start_token
+        # or the request starting exactly at start_token (if on boundary)
+        req_start = int(np.searchsorted(cu_num_tokens, start_token, side="right") - 1)
+
+        # Stop at the request that starts at or after end_token
+        req_stop = int(np.searchsorted(cu_num_tokens, end_token, side="left"))
+
+        req_slice = slice(req_start, req_stop)
+        ubatch_slices.append(UBatchSlice(req_slice, token_slice))
+
+        start_token = end_token
 
     ubatch_slices_padded = _pad_out_ubatch_slices(
         ubatch_slices, num_tokens_padded, num_reqs_padded