add tbo on v1 engine

vllm/envs.py

@@ -159,6 +159,7 @@ if TYPE_CHECKING:
     VLLM_ENABLE_TBO: bool = False
     VLLM_TBO_REQ_DELAY_MS: int = 0
     VLLM_TBO_DECODE_BS: int = 0
+    VLLM_TBO_MIN_TOKENS: int = 200
     VLLM_ZERO_OVERHEAD: bool = False
     VLLM_ENABLE_MOE_FUSED_GATE: bool = False
     VLLM_USE_FLASH_ATTN_PA: bool = False
@@ -1069,6 +1070,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_TBO_DECODE_BS":
     lambda: int(os.getenv("VLLM_TBO_DECODE_BS", "0")),
 
+    # set the minimum tokens size for each mini-batch to enable TBO on v1, default is 200.
+    "VLLM_TBO_MIN_TOKENS":
+    lambda: int(os.getenv("VLLM_TBO_MIN_TOKENS", "200")),
+
     # Enable zero overhead scheduler.
     "VLLM_ZERO_OVERHEAD":
     lambda: bool(int(os.getenv("VLLM_ZERO_OVERHEAD", "0"))),

vllm/two_batch_overlap/two_batch_overlap.py

@@ -16,6 +16,7 @@ from vllm.logger import init_logger
 from vllm.profiler.prof import profile
 from vllm import envs
 from vllm.utils import weak_ref_tensor
+from vllm.two_batch_overlap.v1.two_batch_overlap_v1 import is_enable_tbo_v1, tbo_all_reduce_v1
 
 tbo_one_stream = os.environ.get('VLLM_TBO_ONE_STREAM') == '1'
 
@@ -214,6 +215,8 @@ def init_two_batch_overlap():
     tbo_obj.init_tbo_thread()
 
 def tbo_all_reduce(obj):
+    if is_enable_tbo_v1():
+        return tbo_all_reduce_v1(obj)
     if envs.VLLM_ENABLE_TBO and tbo_obj != None and tbo_obj.tbo_running:
         tid = threading.get_ident()
         if not tbo_one_stream:

vllm/two_batch_overlap/v1/gpu_model_runner.py

+
+from typing import Any, Optional, Union
+import numpy as np
+import torch
+from vllm import envs
+from vllm.distributed.kv_transfer.kv_transfer_state import get_kv_transfer_group, has_kv_transfer_group
+from vllm.distributed.parallel_state import get_pp_group, get_tp_group
+from vllm.forward_context import set_forward_context
+from vllm.sequence import IntermediateTensors
+from vllm.two_batch_overlap.v1.two_batch_overlap_v1 import tbo_model_executable_v1
+from vllm.utils import async_tensor_h2d
+from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.core.sched.output import SchedulerOutput
+from vllm.v1.outputs import EMPTY_MODEL_RUNNER_OUTPUT, ModelRunnerOutput
+from vllm.v1.spec_decode.eagle import EagleProposer
+from vllm.v1.spec_decode.medusa import MedusaProposer
+from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
+from vllm.v1.spec_decode.ngram_proposer import NgramProposer
+from vllm.v1.worker.block_table import BlockTable
+from vllm.v1.worker.gpu_model_runner import GPUModelRunner
+
+
+class TBO_GPUModelRunner(GPUModelRunner):
+    def __init__(self, vllm_config, device):
+        super().__init__(vllm_config, device)
+        self.req_ids_left = []
+        self.req_ids_right = []
+        self.req_num_left = 0
+        self.req_num_right = 0
+        self.scheduler_output_left = None
+        self.scheduler_output_right = None
+
+    def split_scheduler_output(self, scheduler_output:SchedulerOutput):
+        split_tokens = scheduler_output.total_num_scheduled_tokens // 2
+        req_ids = self.input_batch.req_ids
+        tokens_counter = 0
+        min_idx = -1
+        min_counter = 0
+        for i, id in enumerate(req_ids):
+            tokens_counter += scheduler_output.num_scheduled_tokens[id]
+            diff = abs(tokens_counter - split_tokens)
+            if min_idx == -1 or diff < min_counter:
+                min_idx = i
+                min_counter = diff
+            if tokens_counter > split_tokens or diff == 0:
+                break
+        self.req_num_left = min_idx + 1
+        if self.req_num_left == len(req_ids):
+            self.req_num_left = self.req_num_left - 1
+        self.req_ids_left = req_ids[:self.req_num_left]
+        self.req_ids_right = req_ids[self.req_num_left:]
+        self.req_num_right = len(req_ids) - self.req_num_left
+        new_req_data_left = []
+        new_req_data_right = []
+        cached_reqs_left = []
+        cached_reqs_right = []
+        num_scheduled_tokens_left = {}
+        num_scheduled_tokens_right = {}
+        total_num_scheduled_tokens_left = 0
+        total_num_scheduled_tokens_right = 0
+        for new_req in scheduler_output.scheduled_new_reqs:
+            if new_req.req_id in self.req_ids_left:
+                new_req_data_left.append(new_req)
+            else:
+                new_req_data_right.append(new_req)
+        for cached_req in scheduler_output.scheduled_cached_reqs:
+            if cached_req.req_id in self.req_ids_left:
+                cached_reqs_left.append(cached_req)
+            else:
+                cached_reqs_right.append(cached_req)
+        for key, value in scheduler_output.num_scheduled_tokens.items():
+            if key in self.req_ids_left:
+                num_scheduled_tokens_left[key] = value
+                total_num_scheduled_tokens_left += value
+            else:
+                num_scheduled_tokens_right[key] = value
+                total_num_scheduled_tokens_right += value
+
+        self.scheduler_output_left = SchedulerOutput(
+            scheduled_new_reqs=new_req_data_left,
+            scheduled_cached_reqs=cached_reqs_left,
+            num_scheduled_tokens=num_scheduled_tokens_left,
+            total_num_scheduled_tokens=total_num_scheduled_tokens_left,
+            scheduled_spec_decode_tokens=scheduler_output.scheduled_spec_decode_tokens,
+            scheduled_encoder_inputs=scheduler_output.scheduled_encoder_inputs, ##unsupport yet
+            num_common_prefix_blocks=scheduler_output.num_common_prefix_blocks,
+            # finished_req_ids is an existing state in the scheduler,
+            # instead of being newly scheduled in this step.
+            # It contains the request IDs that are finished in between
+            # the previous and the current steps.
+            finished_req_ids=scheduler_output.finished_req_ids,
+            free_encoder_input_ids=scheduler_output.free_encoder_input_ids,
+            structured_output_request_ids=scheduler_output.structured_output_request_ids,
+            grammar_bitmask=scheduler_output.grammar_bitmask,
+        )
+        self.scheduler_output_right = SchedulerOutput(
+            scheduled_new_reqs=new_req_data_right,
+            scheduled_cached_reqs=cached_reqs_right,
+            num_scheduled_tokens=num_scheduled_tokens_right,
+            total_num_scheduled_tokens=total_num_scheduled_tokens_right,
+            scheduled_spec_decode_tokens=scheduler_output.scheduled_spec_decode_tokens,
+            scheduled_encoder_inputs=scheduler_output.scheduled_encoder_inputs, ##unsupport yet
+            num_common_prefix_blocks=scheduler_output.num_common_prefix_blocks,
+            # finished_req_ids is an existing state in the scheduler,
+            # instead of being newly scheduled in this step.
+            # It contains the request IDs that are finished in between
+            # the previous and the current steps.
+            finished_req_ids=scheduler_output.finished_req_ids,
+            free_encoder_input_ids=scheduler_output.free_encoder_input_ids,
+            structured_output_request_ids=scheduler_output.structured_output_request_ids,
+            grammar_bitmask=scheduler_output.grammar_bitmask,
+        )
+
+
+    def prepare_tbo_atten_metadata(
+        self,
+        scheduler_output: "SchedulerOutput",
+        req_ids,
+        req_offset
+    ) -> tuple[dict[str, Any], torch.Tensor, Optional[SpecDecodeMetadata]]:
+        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
+        assert total_num_scheduled_tokens > 0
+        num_reqs = len(req_ids)
+        assert num_reqs > 0
+
+        seq_len_offset = req_offset
+        if req_offset == 0: #left
+            query_start_offset = 0
+        else:
+            query_start_offset = req_offset + 1
+
+        # Get the number of scheduled tokens for each request.
+        # req_ids = self.input_batch.req_ids
+        tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
+        num_scheduled_tokens = np.array(tokens, dtype=np.int32)
+        max_num_scheduled_tokens = max(tokens)
+
+        # Get request indices.
+        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
+        req_indices = np.repeat(self.arange_np[:num_reqs],
+                                num_scheduled_tokens) + req_offset
+
+        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
+        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
+        cu_num_tokens, arange = self._get_cumsum_and_arange(
+            num_scheduled_tokens)
+
+        # Get positions.
+        positions_np = self.positions_np[:total_num_scheduled_tokens]
+        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
+               arange,
+               out=positions_np)
+
+        # Calculate the slot mapping for each KV cache group.
+        for kv_cache_group_id, kv_cache_group_spec in enumerate(
+                self.kv_cache_config.kv_cache_groups):
+            block_size = kv_cache_group_spec.kv_cache_spec.block_size
+            block_table: BlockTable = self.input_batch.block_table[
+                kv_cache_group_id]
+            # E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
+            # -> [0, 0, K, K, K + 1, K + 1, K + 2, 2 * K, 2 * K, 2 * K + 1]
+            # where K is the max_num_blocks_per_req and the block size is 2.
+            # NOTE(woosuk): We can't simply use `token_indices // block_size`
+            # here because M (max_model_len) is not necessarily divisible by
+            # block_size.
+            block_table_indices = (
+                req_indices * block_table.max_num_blocks_per_req +
+                positions_np // block_size)
+            block_table_cpu = block_table.get_cpu_tensor()
+            block_numbers = block_table_cpu.flatten(
+            )[block_table_indices].numpy()
+            block_offsets = positions_np % block_size
+            np.add(
+                block_numbers * block_size,
+                block_offsets,
+                out=block_table.slot_mapping_np[:total_num_scheduled_tokens])
+
+        # Prepare the attention metadata.
+        self.query_start_loc_np[0] = 0
+        self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens
+
+        self.seq_lens_np[:num_reqs] = (
+            self.input_batch.num_computed_tokens_cpu[req_offset : req_offset + num_reqs] +
+            num_scheduled_tokens)
+        self.query_start_loc[query_start_offset: query_start_offset + num_reqs + 1].copy_(
+            self.query_start_loc_cpu[:num_reqs + 1], non_blocking=True)
+        # Note: pad query_start_loc to be non-decreasing, as kernels
+        # like FlashAttention requires that
+        if req_offset > 0: #right
+            self.query_start_loc[query_start_offset + num_reqs + 1:].fill_(
+                self.query_start_loc_cpu[num_reqs].item())
+        self.seq_lens[seq_len_offset :seq_len_offset + num_reqs].copy_(self.seq_lens_cpu[:num_reqs],
+                                       non_blocking=True)
+
+        # Fill unused with -1. Needed for reshape_and_cache
+        if req_offset > 0: #right
+            self.seq_lens[seq_len_offset + num_reqs:].fill_(0)
+
+        query_start_loc = self.query_start_loc[query_start_offset: query_start_offset + num_reqs + 1]
+        seq_lens = self.seq_lens[seq_len_offset : seq_len_offset + num_reqs]
+
+        common_attn_metadata = CommonAttentionMetadata(
+            query_start_loc=query_start_loc, seq_lens=seq_lens)
+
+        attn_metadata: dict[str, Any] = {}
+        # Prepare the attention metadata for each KV cache group and make layers
+        # in the same group share the same metadata.
+        for kv_cache_group_id, kv_cache_group_spec in enumerate(
+                self.kv_cache_config.kv_cache_groups):
+
+            # Prepare for cascade attention if enabled & beneficial.
+            common_prefix_len = 0
+            if self.cascade_attn_enabled:
+                common_prefix_len = self._compute_cascade_attn_prefix_len(
+                    num_scheduled_tokens,
+                    scheduler_output.
+                    num_common_prefix_blocks[kv_cache_group_id],
+                    kv_cache_group_spec.kv_cache_spec,
+                    self.attn_metadata_builders[kv_cache_group_id],
+                )
+            if req_offset > 0:
+                origin_block_table = self.attn_metadata_builders[kv_cache_group_id].block_table.block_table
+                self.attn_metadata_builders[kv_cache_group_id].block_table.block_table = origin_block_table[req_offset:, :]
+                origin_slot_mapping = self.attn_metadata_builders[kv_cache_group_id].block_table.slot_mapping
+                self.attn_metadata_builders[kv_cache_group_id].block_table.slot_mapping = \
+                    origin_slot_mapping[self.scheduler_output_left.total_num_scheduled_tokens:]
+
+            attn_metadata_i = (
+                self.attn_metadata_builders[kv_cache_group_id].build(
+                    num_reqs=num_reqs,
+                    num_actual_tokens=total_num_scheduled_tokens,
+                    max_query_len=max_num_scheduled_tokens,
+                    common_prefix_len=common_prefix_len,
+                    common_attn_metadata=common_attn_metadata)) # maybe FlashAttentionMetadata
+            if req_offset > 0:
+                self.attn_metadata_builders[kv_cache_group_id].block_table.block_table = origin_block_table
+                self.attn_metadata_builders[kv_cache_group_id].block_table.slot_mapping = origin_slot_mapping
+            for layer_name in kv_cache_group_spec.layer_names:
+                attn_metadata[layer_name] = attn_metadata_i
+
+        return attn_metadata
+
+    def pad_num_input_tokens(self, scheduler_output):
+        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
+        if (self.use_cuda_graph
+                and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
+            # Use piecewise CUDA graphs.
+            # Add padding to the batch size.
+            num_input_tokens = self.vllm_config.pad_for_cudagraph(
+                num_scheduled_tokens)
+        else:
+            # Eager mode.
+            # Pad tokens to multiple of tensor_parallel_size when
+            # enabled collective fusion for SP
+            tp_size = self.vllm_config.parallel_config.tensor_parallel_size
+            if self.vllm_config.compilation_config.pass_config. \
+                enable_sequence_parallelism and tp_size > 1:
+                from vllm.utils import round_up
+                num_input_tokens = round_up(num_scheduled_tokens, tp_size)
+            else:
+                num_input_tokens = num_scheduled_tokens
+
+        # Padding for DP
+        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
+        num_input_tokens += num_pad
+        return num_input_tokens, num_tokens_across_dp
+
+    @torch.inference_mode()
+    def execute_model(
+        self,
+        scheduler_output: "SchedulerOutput",
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+    ) -> Union[ModelRunnerOutput, IntermediateTensors]:
+        self._update_states(scheduler_output)
+        if not scheduler_output.total_num_scheduled_tokens:
+            if not has_kv_transfer_group():
+                # Return empty ModelRunnerOutput if there's no work to do.
+                return EMPTY_MODEL_RUNNER_OUTPUT
+
+            return self.kv_connector_no_forward(scheduler_output)
+
+        assert not self.use_cuda_graph, 'v1 engine with tbo do not support cuda-graph'
+        
+        # Prepare the decoder inputs.
+        attn_metadata, logits_indices, spec_decode_metadata = (
+            super()._prepare_inputs(scheduler_output))
+        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
+        if (self.use_cuda_graph
+                and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
+            # Use piecewise CUDA graphs.
+            # Add padding to the batch size.
+            num_input_tokens = self.vllm_config.pad_for_cudagraph(
+                num_scheduled_tokens)
+        else:
+            # Eager mode.
+            # Pad tokens to multiple of tensor_parallel_size when
+            # enabled collective fusion for SP
+            tp_size = self.vllm_config.parallel_config.tensor_parallel_size
+            if self.vllm_config.compilation_config.pass_config. \
+                enable_sequence_parallelism and tp_size > 1:
+                from vllm.utils import round_up
+                num_input_tokens = round_up(num_scheduled_tokens, tp_size)
+            else:
+                num_input_tokens = num_scheduled_tokens
+
+        # Padding for DP
+        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
+        num_input_tokens += num_pad
+
+
+        # _prepare_inputs may reorder the batch, so we must gather multi
+        # modal outputs after that to ensure the correct order
+        if self.is_multimodal_model:
+            # Run the multimodal encoder if any.
+            self._execute_mm_encoder(scheduler_output)
+            mm_embeds = self._gather_mm_embeddings(scheduler_output)
+        else:
+            mm_embeds = []
+
+        if self.is_multimodal_model and get_pp_group().is_first_rank:
+            # NOTE(woosuk): To unify token ids and soft tokens (vision
+            # embeddings), we always use embeddings (rather than token ids)
+            # as input to the multimodal model, even when the input is text.
+            input_ids = self.input_ids[:num_scheduled_tokens]
+            if mm_embeds:
+                inputs_embeds = self.model.get_input_embeddings(
+                    input_ids, mm_embeds)
+            else:
+                inputs_embeds = self.model.get_input_embeddings(input_ids)
+            # TODO(woosuk): Avoid the copy. Optimize.
+            self.inputs_embeds[:num_scheduled_tokens].copy_(inputs_embeds)
+            inputs_embeds = self.inputs_embeds[:num_input_tokens]
+            input_ids = None
+        else:
+            # For text-only models, we use token ids as input.
+            # While it is possible to use embeddings as input just like the
+            # multimodal models, it is not desirable for performance since
+            # then the embedding layer is not included in the CUDA graph.
+            input_ids = self.input_ids[:num_input_tokens]
+            inputs_embeds = None
+        if self.uses_mrope:
+            positions = self.mrope_positions[:, :num_input_tokens]
+        else:
+            positions = self.positions[:num_input_tokens]
+
+        if get_pp_group().is_first_rank:
+            intermediate_tensors = None
+        else:
+            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
+                num_input_tokens, intermediate_tensors, True)
+            
+        use_tbo = False
+        if len(scheduler_output.num_scheduled_tokens) > 1:
+            self.split_scheduler_output(scheduler_output)
+            if self.scheduler_output_left.total_num_scheduled_tokens >= envs.VLLM_TBO_MIN_TOKENS and \
+                self.scheduler_output_right.total_num_scheduled_tokens >= envs.VLLM_TBO_MIN_TOKENS:
+                use_tbo = True
+
+        if use_tbo:
+            num_input_tokens_left = self.scheduler_output_left.total_num_scheduled_tokens
+            num_input_tokens_right = num_input_tokens - num_input_tokens_left
+            
+            attn_metadata_left = self.prepare_tbo_atten_metadata(self.scheduler_output_left, self.req_ids_left, 0)
+            attn_metadata_right = self.prepare_tbo_atten_metadata(self.scheduler_output_right, self.req_ids_right, self.req_num_left)
+
+            model_output = tbo_model_executable_v1(
+                self,
+                attn_metadata_left, 
+                attn_metadata_right, 
+                num_input_tokens_left,
+                num_input_tokens_right,
+                num_tokens_across_dp,
+                input_ids,
+                positions,
+                intermediate_tensors,
+                inputs_embeds) 
+            finished_sending, finished_recving = None, None
+        else:
+            # Run the decoder.
+            # Use persistent buffers for CUDA graphs.
+            with set_forward_context(attn_metadata,
+                                    self.vllm_config,
+                                    num_tokens=num_input_tokens,
+                                    num_tokens_across_dp=num_tokens_across_dp):
+                self.maybe_setup_kv_connector(scheduler_output)
+
+                model_output = self.model(
+                    input_ids=input_ids,
+                    positions=positions,
+                    intermediate_tensors=intermediate_tensors,
+                    inputs_embeds=inputs_embeds,
+                )
+
+                self.maybe_wait_for_kv_save()
+                finished_sending, finished_recving = (
+                    self.get_finished_kv_transfers(scheduler_output))
+
+        if self.use_aux_hidden_state_outputs:
+            hidden_states, aux_hidden_states = model_output
+        else:
+            hidden_states = model_output
+        # Broadcast PP output for external_launcher (torchrun)
+        # to make sure we are synced across pp ranks
+        # TODO: Support overlapping mirco-batches
+        # https://github.com/vllm-project/vllm/issues/18019
+        broadcast_pp_output = \
+            self.parallel_config.distributed_executor_backend \
+            == "external_launcher" and len(get_pp_group().ranks) > 0
+        if not get_pp_group().is_last_rank:
+            # For mid-pipeline stages, return the hidden states.
+            if not broadcast_pp_output:
+                return hidden_states
+            assert isinstance(hidden_states, IntermediateTensors)
+            get_pp_group().send_tensor_dict(hidden_states.tensors,
+                                            all_gather_group=get_tp_group())
+            logits = None
+        else:
+            sample_hidden_states = hidden_states[logits_indices]
+            logits = self.model.compute_logits(sample_hidden_states, None)
+        if broadcast_pp_output:
+            model_output_broadcast_data = {
+                "logits": logits.contiguous(),
+            } if logits is not None else {}
+            model_output_broadcast_data = get_pp_group().broadcast_tensor_dict(
+                model_output_broadcast_data, src=len(get_pp_group().ranks) - 1)
+            assert model_output_broadcast_data is not None
+            logits = model_output_broadcast_data["logits"]
+
+        # Apply structured output bitmasks if present
+        if scheduler_output.grammar_bitmask is not None:
+            self.apply_grammar_bitmask(scheduler_output, logits)
+
+        # Sample the next token and get logprobs if needed.
+        sampling_metadata = self.input_batch.sampling_metadata
+        if spec_decode_metadata is None:
+            sampler_output = self.sampler(
+                logits=logits,
+                sampling_metadata=sampling_metadata,
+            )
+        else:
+            # When indexing with a tensor (bonus_logits_indices), PyTorch
+            # creates a new tensor with separate storage from the original
+            # logits tensor. This means any in-place operations on bonus_logits
+            # won't affect the original logits tensor.
+            assert logits is not None
+            bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
+            sampler_output = self.sampler(
+                logits=bonus_logits,
+                sampling_metadata=sampling_metadata,
+            )
+            bonus_token_ids = sampler_output.sampled_token_ids
+
+            # Just like `bonus_logits`, `target_logits` is a new tensor with
+            # separate storage from the original `logits` tensor. Therefore,
+            # it is safe to update `target_logits` in place.
+            target_logits = logits[spec_decode_metadata.target_logits_indices]
+            output_token_ids = self.rejection_sampler(
+                spec_decode_metadata,
+                None,  # draft_probs
+                target_logits,
+                bonus_token_ids,
+                sampling_metadata,
+            )
+            sampler_output.sampled_token_ids = output_token_ids
+
+        # TODO(woosuk): The following loop can be slow since it iterates over
+        # the requests one by one. Optimize.
+        discard_sampled_tokens_req_indices = []
+        for i, req_id in enumerate(self.input_batch.req_ids):
+            req_state = self.requests[req_id]
+            seq_len = (req_state.num_computed_tokens +
+                       scheduler_output.num_scheduled_tokens[req_id])
+            if seq_len < req_state.num_tokens:
+                # Ignore the sampled token for partial prefills.
+                # Rewind the generator state as if the token was not sampled.
+                # This relies on cuda-specific torch-internal impl details
+                generator = self.input_batch.generators.get(i)
+                if generator is not None:
+                    generator.set_offset(generator.get_offset() - 4)
+                # Record the index of the request that should not be sampled,
+                # so that we could clear the sampled tokens before returning.
+                discard_sampled_tokens_req_indices.append(i)
+
+        # NOTE: GPU -> CPU Sync happens here.
+        # Move as many CPU operations as possible before this sync point.
+        logprobs_tensors = sampler_output.logprobs_tensors
+        logprobs_lists = logprobs_tensors.tolists() \
+            if logprobs_tensors is not None else None
+
+        # Compute prompt logprobs if needed.
+        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
+            hidden_states[:num_scheduled_tokens],
+            scheduler_output,
+        )
+
+        # Get the valid generated tokens.
+        sampled_token_ids = sampler_output.sampled_token_ids
+        max_gen_len = sampled_token_ids.shape[-1]
+        if max_gen_len == 1:
+            # No spec decode tokens.
+            valid_sampled_token_ids = sampled_token_ids.tolist()
+        else:
+            # Includes spec decode tokens.
+            valid_sampled_token_ids = self.rejection_sampler.parse_output(
+                sampled_token_ids,
+                self.input_batch.vocab_size,
+            )
+        # Mask out the sampled tokens that should not be sampled.
+        for i in discard_sampled_tokens_req_indices:
+            valid_sampled_token_ids[i].clear()
+
+        if not self.speculative_config:
+            # Speculative decoding is not enabled.
+            spec_token_ids = None
+        elif self.speculative_config.method == "ngram":
+            assert isinstance(self.drafter, NgramProposer)
+            spec_token_ids = self.generate_draft_token_ids(
+                valid_sampled_token_ids, sampling_metadata)
+        elif self.speculative_config.method == "medusa":
+            assert isinstance(self.drafter, MedusaProposer)
+            if max_gen_len == 1:
+                hidden_states = sample_hidden_states
+            else:
+                indices = []
+                offset = 0
+                for num_draft, tokens in zip(
+                        spec_decode_metadata.num_draft_tokens,
+                        valid_sampled_token_ids):
+                    indices.append(offset + len(tokens) - 1)
+                    offset += num_draft + 1
+
+                indices = torch.tensor(indices,
+                                       device=sample_hidden_states.device)
+                hidden_states = sample_hidden_states[indices]
+
+            spec_token_ids = self.drafter.propose(
+                target_hidden_states=hidden_states,
+                sampling_metadata=sampling_metadata,
+            )
+        elif self.speculative_config.use_eagle():
+            assert isinstance(self.drafter, EagleProposer)
+            # TODO(woosuk): Refactor the loop.
+            next_token_ids: list[int] = []
+            for i, token_ids in enumerate(valid_sampled_token_ids):
+                if token_ids:
+                    # Common case.
+                    next_token_id = token_ids[-1]
+                else:
+                    # Partial prefill (rare case).
+                    # Get the next token id from the request state.
+                    req_id = self.input_batch.req_ids[i]
+                    req_state = self.requests[req_id]
+                    seq_len = (req_state.num_computed_tokens +
+                               scheduler_output.num_scheduled_tokens[req_id])
+                    next_token_id = req_state.get_token_id(seq_len)
+                next_token_ids.append(next_token_id)
+            next_token_ids = torch.tensor(next_token_ids,
+                                          dtype=torch.int32,
+                                          device=self.device)
+            # At this moment, we assume all eagle layers belong to the same KV
+            # cache group, thus using the same attention metadata.
+            eagle_attn_metadata = attn_metadata[
+                self.drafter.attn_layer_names[0]]
+
+            # NOTE: deepseek_mtp uses MLA which does not have `block_table`
+            if hasattr(eagle_attn_metadata, "block_table"):
+                block_table = eagle_attn_metadata.block_table
+            else:
+                block_table = None
+
+            num_rejected_tokens_tuple = None
+            if spec_decode_metadata is None:
+                # input_ids can be None for multimodal models.
+                target_token_ids = self.input_ids[:num_scheduled_tokens]
+                target_positions = positions[:num_scheduled_tokens]
+                if self.use_aux_hidden_state_outputs:
+                    target_hidden_states = torch.cat(
+                        [h[:num_scheduled_tokens] for h in aux_hidden_states],
+                        dim=-1)
+                else:
+                    target_hidden_states = hidden_states[:num_scheduled_tokens]
+                target_slot_mapping = eagle_attn_metadata.slot_mapping
+                cu_num_tokens = eagle_attn_metadata.query_start_loc
+            else:
+                # TODO(woosuk): Refactor this.
+                num_draft_tokens = spec_decode_metadata.num_draft_tokens
+                num_rejected_tokens = [
+                    n + 1 - len(valid_sampled_token_ids[i]) if n > 0 else 0
+                    for i, n in enumerate(num_draft_tokens)
+                ]
+                num_rejected_tokens_tensor = async_tensor_h2d(
+                    num_rejected_tokens,
+                    dtype=torch.int32,
+                    target_device=self.device,
+                    pin_memory=True)
+                num_tokens = num_scheduled_tokens - sum(num_rejected_tokens)
+                cu_num_tokens, token_indices = self.drafter.prepare_inputs(
+                    eagle_attn_metadata.query_start_loc,
+                    num_rejected_tokens_tensor,
+                    num_tokens,
+                )
+                target_token_ids = self.input_ids[token_indices]
+                target_positions = positions[token_indices]
+                if self.use_aux_hidden_state_outputs:
+                    target_hidden_states = torch.cat(
+                        [h[token_indices] for h in aux_hidden_states], dim=-1)
+                else:
+                    target_hidden_states = hidden_states[token_indices]
+                target_slot_mapping = eagle_attn_metadata.slot_mapping[
+                    token_indices]
+                num_rejected_tokens_tuple = (num_rejected_tokens, num_rejected_tokens_tensor)
+            draft_token_ids = self.drafter.propose(
+                target_token_ids=target_token_ids,
+                target_positions=target_positions,
+                target_hidden_states=target_hidden_states,
+                target_slot_mapping=target_slot_mapping,
+                next_token_ids=next_token_ids,
+                cu_num_tokens=cu_num_tokens,
+                block_table=block_table,
+                sampling_metadata=sampling_metadata,
+                num_rejected_tokens_tuple=num_rejected_tokens_tuple
+            )
+            spec_token_ids = draft_token_ids.tolist()
+
+        # Clear KVConnector state after all KVs are generated.
+        if has_kv_transfer_group():
+            get_kv_transfer_group().clear_connector_metadata()
+
+        return ModelRunnerOutput(
+            req_ids=self.input_batch.req_ids,
+            req_id_to_index=self.input_batch.req_id_to_index,
+            sampled_token_ids=valid_sampled_token_ids,
+            spec_token_ids=spec_token_ids,
+            logprobs=logprobs_lists,
+            prompt_logprobs_dict=prompt_logprobs_dict,
+            finished_sending=finished_sending,
+            finished_recving=finished_recving,
+        )
\ No newline at end of file

vllm/two_batch_overlap/v1/two_batch_overlap_v1.py

+import os
+import queue
+import threading
+import torch
+from vllm.distributed.communication_op import tensor_model_parallel_all_reduce
+from vllm.distributed.parallel_state import get_tp_group
+from vllm.forward_context import set_forward_context
+from vllm.multimodal.inputs import MultiModalKwargs
+from vllm.sequence import IntermediateTensors
+from vllm.two_batch_overlap.forward_context import init_tbo_forward_context
+from vllm.logger import init_logger
+from vllm.profiler.prof import profile
+from vllm import envs
+
+logger = init_logger(__name__)
+
+tbo_step_stream = None
+all_reduce_stream = None
+
+class TwoBatchOverlap():
+    def __init__(self):
+        global tbo_step_stream
+        global all_reduce_stream
+        self.model_input_left_queue = queue.Queue()
+        self.model_input_right_queue = queue.Queue()
+        self.states_left_queue = queue.Queue()
+        self.states_right_queue = queue.Queue()
+        self.left_thread = None
+        self.right_thread = None
+        self.left_tid = 0
+        self.right_tid = 0
+        self.sem_left = threading.Semaphore(0)
+        self.sem_right = threading.Semaphore(0)
+        self.left_first = False
+        self.tbo_running = False
+        self.tbo_in_capture = False
+        if tbo_step_stream == None:
+            tbo_step_stream = torch.cuda.Stream()
+            all_reduce_stream = torch.cuda.Stream()
+        self.step_event = torch.cuda.Event(enable_timing=False)
+        self.event_left_c2t = torch.cuda.Event(enable_timing=False)
+        self.event_right_c2t = torch.cuda.Event(enable_timing=False)
+        self.event_left_t2c = torch.cuda.Event(enable_timing=False)
+        self.event_right_t2c = torch.cuda.Event(enable_timing=False)
+
+    def init_tbo_thread(self):
+        self.model_input_left_queue.empty()
+        self.model_input_right_queue.empty()
+        self.left_thread = threading.Thread(target=self.thread_two_batch_overlap, args=(self.model_input_left_queue,))
+        self.left_thread.start()
+        self.right_thread = threading.Thread(target=self.thread_two_batch_overlap, args=(self.model_input_right_queue,))
+        self.right_thread.start()
+        logger.info('tbo:two batch overlap start')
+
+    def finish_thread(self):
+        self.left_thread.join()
+        self.left_thread = None
+        self.right_thread.join()
+        self.right_thread = None
+        
+    @torch.inference_mode()
+    def thread_two_batch_overlap(self, queue):
+        is_left_thread = False
+        tid = threading.get_ident()
+        if queue == self.model_input_left_queue:
+            self.left_tid = tid
+            is_left_thread = True
+            init_tbo_forward_context(True, self.left_tid)
+        else:
+            self.right_tid = tid
+            init_tbo_forward_context(False, self.right_tid)
+        with torch.cuda.stream(tbo_step_stream):
+            queue.get()
+            profile.ProfRangePush('start')
+            self.tbo_thread_synchronize(tid)
+            if is_left_thread:
+                attn_metadata = self.attn_metadata_left
+                num_input_tokens = self.num_input_tokens_left
+                input_ids = self.input_ids_left
+                positions = self.positions_left
+            else:
+                attn_metadata = self.attn_metadata_right
+                num_input_tokens = self.num_input_tokens_right
+                input_ids = self.input_ids_right
+                positions = self.positions_right
+
+            model_output = None
+            # Run the decoder.
+            # Use persistent buffers for CUDA graphs.
+            with set_forward_context(attn_metadata,
+                                    self.model_runner.vllm_config,
+                                    num_tokens=num_input_tokens,
+                                    num_tokens_across_dp=self.num_tokens_across_dp):
+                model_output = self.model_runner.model(
+                    input_ids=input_ids,
+                    positions=positions,
+                    intermediate_tensors=self.intermediate_tensors,
+                    inputs_embeds=self.inputs_embeds,
+                )
+            if is_left_thread:
+                self.sem_right.release()
+                self.states_left_queue.put(model_output)
+            else:
+                self.states_right_queue.put(model_output)
+            profile.ProfRangePop()
+
+    def tbo_thread_synchronize(self, tid):
+        if tid == self.left_tid:
+            if not self.left_first:
+                self.sem_right.release()
+            self.left_first = False
+            profile.ProfRangePop()
+            self.sem_left.acquire()
+            profile.ProfRangePush('left')
+            return self.event_left_c2t, self.event_left_t2c
+        else:
+            self.sem_left.release()
+            profile.ProfRangePop()
+            self.sem_right.acquire()
+            profile.ProfRangePush('right')
+            return self.event_right_c2t, self.event_right_t2c
+
+    def set_model_input(self,
+                        model_runner, 
+                        attn_metadata_left, 
+                        attn_metadata_right, 
+                        num_input_tokens_left,
+                        num_input_tokens_right,
+                        input_ids_left, 
+                        input_ids_right, 
+                        positions_left, 
+                        positions_right,
+                        num_tokens_across_dp,
+                        intermediate_tensors,
+                        inputs_embeds):
+        self.model_runner = model_runner
+        self.attn_metadata_left = attn_metadata_left
+        self.attn_metadata_right = attn_metadata_right
+        self.num_input_tokens_left = num_input_tokens_left
+        self.num_input_tokens_right = num_input_tokens_right
+        self.input_ids_left = input_ids_left
+        self.input_ids_right = input_ids_right
+        self.positions_left = positions_left
+        self.positions_right = positions_right
+        self.num_tokens_across_dp = num_tokens_across_dp
+        self.intermediate_tensors = intermediate_tensors
+        self.inputs_embeds = inputs_embeds
+
+        self.model_input_left_queue.put(None)
+        self.model_input_right_queue.put(None)
+        
+    def get_model_output(self):
+        states_left = self.states_left_queue.get()
+        states_right = self.states_right_queue.get()
+        return states_left, states_right
+    
+tbo_obj_v1 = None
+
+def is_enable_tbo_v1():
+    global tbo_obj_v1
+    return tbo_obj_v1 != None
+
+def init_two_batch_overlap():
+    global tbo_obj_v1
+    if tbo_obj_v1 == None:
+        tbo_obj_v1 = TwoBatchOverlap()
+    tbo_obj_v1.init_tbo_thread()
+
+def tbo_all_reduce_v1(obj):
+    if envs.VLLM_ENABLE_TBO and tbo_obj_v1 != None and tbo_obj_v1.tbo_running:
+        tid = threading.get_ident()
+        if tid == tbo_obj_v1.left_tid:
+            event_c2t, event_t2c = tbo_obj_v1.event_left_c2t, tbo_obj_v1.event_left_t2c
+        else:
+            event_c2t, event_t2c = tbo_obj_v1.event_right_c2t, tbo_obj_v1.event_right_t2c
+        event_c2t.record()
+        with torch.cuda.stream(all_reduce_stream):
+            all_reduce_stream.wait_event(event_c2t)
+            output = tensor_model_parallel_all_reduce(obj)
+            event_t2c.record()
+        tbo_obj_v1.tbo_thread_synchronize(tid)
+        tbo_step_stream.wait_event(event_t2c)
+        return output
+    return tensor_model_parallel_all_reduce(obj) 
+
+def merge_model_output(states_left, states_right):
+    if isinstance(states_left, IntermediateTensors):
+        output_map = {}
+        for key in states_left.tensors:
+            output_map[key] = torch.concat([states_left.tensors[key], states_right.tensors[key]], dim=0)
+        output = IntermediateTensors(output_map)
+    else:
+        output = torch.concat([states_left, states_right], dim=0)
+    return output
+
+def tbo_model_executable_v1(
+        model_runner,
+        attn_metadata_left, 
+        attn_metadata_right, 
+        num_input_tokens_left,
+        num_input_tokens_right,
+        num_tokens_across_dp,
+        input_ids,
+        positions,
+        intermediate_tensors,
+        inputs_embeds
+    ):
+
+    init_two_batch_overlap()
+    tbo_obj_v1.tbo_running = True
+    tbo_obj_v1.left_first = True
+    tbo_obj_v1.step_event.record()
+    current_stream = torch.cuda.current_stream()
+    with torch.cuda.stream(tbo_step_stream):
+        tbo_step_stream.wait_event(tbo_obj_v1.step_event)
+        tokens_split = [num_input_tokens_left, num_input_tokens_right]
+        input_ids_left, input_ids_right = torch.split(input_ids, tokens_split, dim=0)
+        positions_left, positions_right = torch.split(positions, tokens_split, dim=0)
+        tbo_obj_v1.set_model_input(model_runner, 
+                                attn_metadata_left, 
+                                attn_metadata_right, 
+                                num_input_tokens_left,
+                                num_input_tokens_right,
+                                input_ids_left, 
+                                input_ids_right, 
+                                positions_left, 
+                                positions_right,
+                                num_tokens_across_dp,
+                                intermediate_tensors,
+                                inputs_embeds)
+        model_output_left, model_output_right = tbo_obj_v1.get_model_output()
+
+        hidden_or_intermediate_states = merge_model_output(model_output_left, model_output_right)
+        tbo_obj_v1.tbo_running = False
+        tbo_obj_v1.step_event.record()
+        tbo_obj_v1.finish_thread()
+    current_stream.wait_event(tbo_obj_v1.step_event)
+    
+    return hidden_or_intermediate_states
\ No newline at end of file

vllm/v1/worker/gpu_worker.py

@@ -22,6 +22,7 @@ from vllm.lora.request import LoRARequest
 from vllm.model_executor import set_random_seed
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
+from vllm.two_batch_overlap.v1.gpu_model_runner import TBO_GPUModelRunner
 from vllm.utils import GiB_bytes, MemorySnapshot, memory_profiling
 from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
 from vllm.v1.outputs import ModelRunnerOutput
@@ -162,8 +163,12 @@ class Worker(WorkerBase):
         set_random_seed(self.model_config.seed)
 
         # Construct the model runner
-        self.model_runner: GPUModelRunner = GPUModelRunner(
-            self.vllm_config, self.device)
+        if envs.VLLM_ENABLE_TBO:
+            self.model_runner: TBO_GPUModelRunner = TBO_GPUModelRunner(
+                self.vllm_config, self.device)
+        else:
+            self.model_runner: GPUModelRunner = GPUModelRunner(
+                self.vllm_config, self.device)
 
         if self.rank == 0:
             # If usage stat is enabled, collect relevant info.