Merge branch 'v0.8.5-zero_overhead' into 'v0.8.5.post1-dev'

V0.8.5 zero overhead

See merge request dcutoolkit/deeplearing/vllm!110

vllm/attention/backends/utils.py

@@ -239,8 +239,11 @@ class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]):
             for i, block_table in enumerate(self.block_tables):
                 if block_table:
                     input_block_tables[i, :len(block_table)] = block_table
-            block_tables = torch.from_numpy(input_block_tables).to(
-                device, non_blocking=True)
+            # block_tables = torch.from_numpy(input_block_tables).to(
+            #     device, non_blocking=True)
+            block_tables = async_tensor_h2d(input_block_tables.tolist(), torch.int32,
+                                    device, self.runner.pin_memory)
+
         else:
             block_tables = make_tensor_with_pad(
                 self.block_tables,

vllm/engine/multiprocessing/engine.py

@@ -6,6 +6,8 @@ from contextlib import contextmanager
 from typing import Iterator, List, Optional, Union
 
 import cloudpickle
+from vllm.zero_overhead.llm_engine import ZeroOverheadEngine
+from vllm.zero_overhead.utils import is_zero_overhead
 import zmq
 
 from vllm import AsyncEngineArgs, SamplingParams
@@ -79,6 +81,9 @@ class MQLLMEngine:
         # the python object to be reused again.
         kwargs['use_cached_outputs'] = True
 
+        if is_zero_overhead():
+            self.engine = ZeroOverheadEngine(*args, **kwargs)
+        else:
             self.engine = LLMEngine(*args, **kwargs)
         self.log_requests = log_requests
 

vllm/entrypoints/llm.py

@@ -43,6 +43,8 @@ from vllm.transformers_utils.tokenizer import (AnyTokenizer, MistralTokenizer,
 from vllm.usage.usage_lib import UsageContext
 from vllm.utils import (Counter, Device, deprecate_args, deprecate_kwargs,
                         is_list_of)
+from vllm.zero_overhead.llm_engine import ZeroOverheadEngine
+from vllm.zero_overhead.utils import is_zero_overhead
 
 logger = init_logger(__name__)
 
@@ -244,6 +246,10 @@ class LLM:
         )
 
         # Create the Engine (autoselects V0 vs V1)
+        if is_zero_overhead():
+            self.llm_engine = ZeroOverheadEngine.from_engine_args(
+                engine_args=engine_args, usage_context=UsageContext.LLM_CLASS)
+        else:
             self.llm_engine = LLMEngine.from_engine_args(
                 engine_args=engine_args, usage_context=UsageContext.LLM_CLASS)
         self.engine_class = type(self.llm_engine)
@@ -1444,6 +1450,7 @@ class LLM:
 
         if use_tqdm:
             pbar.close()
+
         # Sort the outputs by request ID.
         # This is necessary because some requests may be finished earlier than
         # its previous requests.

vllm/model_executor/layers/sampler.py

@@ -21,6 +21,7 @@ from vllm.sequence import (VLLM_INVALID_TOKEN_ID,
                            CompletionSequenceGroupOutput, Logprob,
                            PromptLogprobs, SampleLogprobs, SequenceOutput)
 from vllm.spec_decode.metrics import SpecDecodeWorkerMetrics
+from vllm.zero_overhead.utils import is_zero_overhead
 
 if envs.VLLM_USE_FLASHINFER_SAMPLER and find_spec("flashinfer"):
     import flashinfer.sampling
@@ -38,6 +39,9 @@ def get_sampler() -> torch.nn.Module:
         # Lazy import: the v1 package isn't distributed
         from vllm.v1.sample.sampler import Sampler as V1Sampler
         return V1Sampler()
+    if is_zero_overhead():
+        from vllm.zero_overhead.sampler import ZeroOverheadSampler
+        return ZeroOverheadSampler()
     return Sampler()
 
 

vllm/profiler/prof.py

+from ctypes import *
+import os
+import time
+import threading
+
+class Prof:
+    def __init__(self):
+        self.use_nvtx = os.getenv('VLLM_PROF_NVTX') is not None
+        self.roc_tracer_flag = False
+        self.lib = None
+        if self.use_nvtx:
+            self.lib = cdll.LoadLibrary("libnvToolsExt.so")
+            self.lib.nvtxRangePushA.argtypes = [c_char_p]
+            self.lib.nvtxRangePushA.restype = c_int
+            self.lib.nvtxRangePop.restype = c_int
+
+        self.use_roctx = os.getenv('VLLM_PROF_ROCTX') is not None
+        if self.use_roctx:
+            self.lib = cdll.LoadLibrary("libroctracer64.so")
+            self.lib.roctxRangePushA.argtypes = [c_char_p]
+            self.lib.roctxRangePushA.restype = c_int
+            self.lib.roctxRangePop.restype = c_int
+        self.tm = time.perf_counter()
+        self.push_depth = {}
+
+    def StartTracer(self):
+        if self.use_roctx:
+            if self.lib is None:
+                self.lib = cdll.LoadLibrary("libroctracer64.so")
+            self.lib.roctracer_start()
+            self.roc_tracer_flag = True
+
+    def StopTracer(self):
+        if self.use_roctx:
+            if self.lib is None:
+                self.lib = cdll.LoadLibrary("libroctracer64.so")
+            self.lib.roctracer_stop()
+            self.roc_tracer_flag = False
+
+    def thread_depth_add(self, num):
+        current_thread = threading.current_thread()
+        thread_id = current_thread.ident
+        if thread_id not in self.push_depth.keys():
+            self.push_depth[thread_id] = 0
+        if num < 0 and self.push_depth[thread_id] == 0:
+            return False
+        self.push_depth[thread_id] += num
+        return True
+
+    def ProfRangePush(self, message):
+        if profile.use_nvtx:
+            profile.lib.nvtxRangePushA(message.encode('utf-8'))
+            self.thread_depth_add(1)
+        if profile.use_roctx and self.roc_tracer_flag:
+            profile.lib.roctxRangePushA(message.encode('utf-8'))
+            self.thread_depth_add(1)
+
+    def ProfRangePop(self):
+        if profile.use_nvtx:
+            if not self.thread_depth_add(-1):
+                return
+            profile.lib.nvtxRangePop()
+        if profile.use_roctx and self.roc_tracer_flag:
+            if not self.thread_depth_add(-1):
+                return
+            profile.lib.roctxRangePop()
+
+    def ProfRangeAutoPush(self, message):
+        self.ProfRangePop()
+        self.ProfRangePush(message)
+
+
+profile = Prof()

vllm/spec_decode/spec_decode_worker.py

@@ -54,6 +54,7 @@ from vllm.worker.worker_base import LoRANotSupportedWorkerBase, WorkerBase
 from vllm.worker.cache_engine import CacheEngine
 from vllm.attention.ops.paged_attn import PagedAttention
 from vllm.spec_decode.proposer_worker_base import NonLLMProposerWorkerBase
+from vllm.zero_overhead.utils import is_zero_overhead
 
 logger = init_logger(__name__)
 
@@ -206,7 +207,13 @@ class SpecDecodeWorker(LoRANotSupportedWorkerBase):
                 # Load lm_head weight for eagle in init_device
                 if draft_model_config.hf_config.model_type == "eagle":
                     enable_lm_head_weight_load = True
-
+                if is_zero_overhead():
+                    assert False, (
+                                "speculative decoding not support zero overhead scheduler yet"
+                            )
+                    from vllm.zero_overhead.spec_decode.muti_step_worker import ZeroOverheadMultiStepWorker
+                    proposer_worker = ZeroOverheadMultiStepWorker(**draft_worker_kwargs)
+                else:
                     proposer_worker = MultiStepWorker(**draft_worker_kwargs)
                 if draft_model_config.hf_config.model_type == "deepseek_mtp":
                     num_spec_prefill_steps = \
@@ -254,6 +261,20 @@ class SpecDecodeWorker(LoRANotSupportedWorkerBase):
                     "[Speculative Decoding] Disabling MQA scorer as the "
                     "target model is not running in eager mode.")
 
+        if is_zero_overhead():
+            from vllm.zero_overhead.spec_decode.spec_decode_worker import ZeroOverheadSpecDecodeWorker
+            return ZeroOverheadSpecDecodeWorker(
+                proposer_worker,
+                scorer_worker,
+                disable_mqa_scorer=disable_mqa_scorer,
+                disable_logprobs=disable_logprobs,
+                disable_log_stats=disable_log_stats,
+                disable_by_batch_size=disable_by_batch_size,
+                spec_decode_sampler=spec_decode_sampler,
+                allow_zero_draft_token_step=allow_zero_draft_token_step,
+                enable_lm_head_weight_load=enable_lm_head_weight_load,
+                num_spec_prefill_steps=num_spec_prefill_steps)
+        else:
             return SpecDecodeWorker(
                 proposer_worker,
                 scorer_worker,

vllm/worker/model_runner.py

@@ -60,6 +60,7 @@ from vllm.worker.model_runner_base import (
     _add_sampling_metadata_broadcastable_dict,
     _init_attn_metadata_from_tensor_dict,
     _init_sampling_metadata_from_tensor_dict)
+from vllm.zero_overhead.utils import is_zero_overhead
 
 if TYPE_CHECKING:
     from vllm.attention.backends.abstract import AttentionBackend
@@ -1636,6 +1637,9 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
     _model_input_cls: Type[ModelInputForGPUWithSamplingMetadata] = (
         ModelInputForGPUWithSamplingMetadata)
     _builder_cls: Type[ModelInputForGPUBuilder] = ModelInputForGPUBuilder
+    if is_zero_overhead():
+        from vllm.zero_overhead.model_runner import ZeroOverheadModelInputForGpuBuilder
+        _builder_cls = ZeroOverheadModelInputForGpuBuilder
 
     def make_model_input_from_broadcasted_tensor_dict(
         self,

vllm/zero_overhead/model_runner.py

+
+
+import torch
+import itertools
+from typing import List, Optional, Set
+from vllm.lora.layers import LoRAMapping
+from vllm.multimodal.inputs import MultiModalKwargs
+from vllm.prompt_adapter.layers import PromptAdapterMapping
+from vllm.prompt_adapter.request import PromptAdapterRequest
+from vllm.sequence import SequenceGroupMetadata
+from vllm.utils import async_tensor_h2d, flatten_2d_lists
+from vllm.worker.model_runner import ModelInputForGPU, ModelInputForGPUBuilder
+from vllm.zero_overhead.sampler import get_last_sampler
+from vllm.zero_overhead.utils import SpecStepKind, get_accepted_token_ids, get_proposal_token_ids, get_spec_last_step, get_spec_step
+
+import triton
+import triton.language as tl
+
+@triton.jit
+def _update_input_tokens(
+    accepted_req_ids,
+    accepted_req_ids_len,
+    accepted_token_ids,
+    accepted_token_len,
+    chidren_req_ids,
+    chidren_req_ids_len,
+    input_tokens,
+    input_tokens_len,
+    input_positions,
+    seq_lens,
+    seq_lens_meta,
+    seq_lens_tensor,
+    slot_mapping,
+    seq_start_loc,
+    context_lens_tensor,
+):
+    chidren_req_ids_ = tl.load(chidren_req_ids + tl.arange(0, chidren_req_ids_len))
+    accepted_req_ids_ = tl.load(accepted_req_ids + tl.arange(0, chidren_req_ids_len))
+    
+    for seq_id_idx in range(chidren_req_ids_len / 2):
+        seq_id = chidren_req_ids_[2 * seq_id_idx]
+        for i in range(accepted_req_ids_len):
+            if seq_id == accepted_req_ids_[i]:
+                accepted_token_ids_ = tl.load(accepted_token_ids + tl.arange(i * accepted_token_len, tl.arange(0, accepted_token_len)))
+                accepted_token_counter = 0
+                for j in range(accepted_token_len):
+                    if accepted_token_ids_[j] == -1:
+                        break
+                    accepted_token_counter += 1
+                if accepted_token_counter == accepted_token_len:
+                    tl.store(input_tokens + seq_id_idx * 2 + tl.arange(0, 2), accepted_token_ids_[-2:])
+                else:
+                    tl.store(input_tokens + seq_id_idx * 2, 0)
+                    tl.store(input_tokens + seq_id_idx * 2 + 1, accepted_token_ids_[accepted_token_counter - 1])
+                    input_pos = tl.load(input_positions + seq_id_idx * 2 + tl.arange(0, 2))
+                    input_pos[0] = 0
+                    input_pos[1] = input_pos[1] - (accepted_req_ids_len - accepted_token_counter)
+                    tl.store(input_positions + seq_id_idx * 2 + tl.arange(0, 2), input_pos)
+                    tl.store(context_lens_tensor + seq_id_idx * 2 + tl.arange(0, 2), input_pos)
+                    input_pos[0] = -1
+                    tl.store(slot_mapping + seq_id_idx * 2 + tl.arange(0, 2), input_pos)
+                    input_pos[0] = 1
+                    input_pos[1] = input_pos[1] + 1
+                    tl.store(seq_lens + seq_id_idx * 2 + tl.arange(0, 2), input_pos)
+                    tl.store(seq_lens_meta + seq_id_idx * 2 + tl.arange(0, 2), input_pos)
+                    tl.store(seq_lens_tensor + seq_id_idx * 2 + tl.arange(0, 2), input_pos)
+    seq_lens_ = tl.load(seq_lens + tl.arange(0, input_tokens_len))
+    seq_start_loc_ = tl.zero_like(seq_start_loc)
+    for i in range(input_tokens_len):
+        seq_start_loc_[i + 1] = seq_start_loc_[i] + seq_lens_[i]
+    tl.store(seq_start_loc + tl.arange(0, input_tokens_len + 1), seq_start_loc_)
+
+
+
+class ZeroOverheadModelInputForGpuBuilder(ModelInputForGPUBuilder):
+    def __init__(self, runner, finished_requests_ids = None):
+        super().__init__(runner, finished_requests_ids)
+        self.req_ids = []
+
+    def prepare(self,
+                finished_requests_ids: Optional[List[str]] = None) -> None:
+        self.req_ids.clear()
+        return super().prepare(finished_requests_ids)
+    
+    def add_seq_group(self, seq_group_metadata: SequenceGroupMetadata):
+        seq_ids = seq_group_metadata.seq_data.keys()
+        n_seqs = len(seq_ids)
+        seq_ids = list(seq_ids)
+        for seq_idx in range(n_seqs):
+            self.req_ids.append(seq_ids[seq_idx])
+        return super().add_seq_group(seq_group_metadata)
+
+    def build(self) -> ModelInputForGPU:
+        model_input = super().build()
+        last_sampler = get_last_sampler()
+        spec_step = get_spec_step()
+        last_step = get_spec_last_step()
+        if last_sampler is not None:
+            if spec_step == SpecStepKind.KIND_DEFAULT:
+                update_indices = []
+                select_indices = []
+                for i, seq_id in enumerate(self.req_ids):
+                    for j, seq_id_ in enumerate(last_sampler.seq_ids):
+                        if seq_id == seq_id_:
+                            select_indices.append(j)
+                            update_indices.append(i)
+                            break      
+                if len(select_indices) > 0:   
+                    select_indices = async_tensor_h2d(select_indices, torch.long,
+                                                    self.runner.device,
+                                                    self.runner.pin_memory)
+                    update_indices = async_tensor_h2d(update_indices, torch.long,
+                                                    self.runner.device,
+                                                    self.runner.pin_memory)       
+                    model_input.input_tokens[update_indices] = last_sampler.sampled_token_ids_tensor[select_indices, 0]
+            if spec_step == SpecStepKind.OTHER_PROPOSAL:
+                if last_step ==  SpecStepKind.OTHER_PROPOSAL: # copy last sampled token ids to input tokens directly.
+                    update_indices = [i for i in range(len(self.req_ids))]
+                    update_indices = async_tensor_h2d(update_indices, torch.long,
+                                                    self.runner.device,
+                                                    self.runner.pin_memory)    
+                    model_input.input_tokens[update_indices] = last_sampler.sampled_token_ids_tensor[update_indices, 0]
+                if last_step == SpecStepKind.FIRST_PROPOSAL: # TODO: ajust input tokens number to 1 per request.
+                    update_indices = [i for i in range(len(self.req_ids))]
+                    update_indices = async_tensor_h2d(update_indices, torch.long,
+                                                    self.runner.device,
+                                                    self.runner.pin_memory)    
+                    model_input.input_tokens[update_indices] = last_sampler.sampled_token_ids_tensor[update_indices, 0]
+                    
+        if spec_step == SpecStepKind.SCORE_DECODE:
+            proposal_token_ids = get_proposal_token_ids()
+            shape = proposal_token_ids.shape
+            batch_size = shape[0]
+            proposal_len = shape[1]
+            update_indices = []
+            for i in range(batch_size):
+                for j in range(proposal_len): 
+                    update_indices.append(i * (proposal_len + 1) + j + 1)
+
+            update_indices = async_tensor_h2d(update_indices, torch.long,
+                                            self.runner.device,
+                                            self.runner.pin_memory)
+            model_input.input_tokens[update_indices] = proposal_token_ids.view(-1)
+        if spec_step == SpecStepKind.FIRST_PROPOSAL:
+            if last_step == SpecStepKind.PREFILL:# TODO: when last step is prefill, just update the input ids for last seqence_id onely.
+                pass
+            if last_step == SpecStepKind.SCORE_DECODE:# TODO: when last step is score decode, fix input ids、seq_lens、input_positions use accepte token ids
+                accept_token_ids, accept_seq_ids = get_accepted_token_ids()
+
+                chidren_req_ids = async_tensor_h2d(self.req_ids, torch.long,
+                                                self.runner.device,
+                                                self.runner.pin_memory)    
+                grid = [1, 1, 1]
+                _update_input_tokens[grid](
+                    accept_seq_ids, accept_seq_ids.shape[0], 
+                    accept_token_ids, accept_token_ids.shape[1],
+                    chidren_req_ids, chidren_req_ids.shape[0],
+                    model_input.input_tokens, model_input.input_tokens.shape[0],
+                    model_input.input_positions, 
+                    model_input.seq_lens,
+                    model_input.attn_metadata.seq_lens_tensor, 
+                    model_input.attn_metadata.seq_lens,
+                    model_input.attn_metadata.slot_mapping,
+                    model_input.attn_metadata.seq_start_loc,
+                    model_input.attn_metadata.context_lens_tensor,
+                    )
+            
+
+        return model_input

vllm/zero_overhead/sampler.py

+from importlib.util import find_spec
+from typing import Dict, List, Optional
+import torch
+
+from vllm import envs
+from vllm.model_executor.layers.sampler import MultinomialSamplesType, SampleMetadataType, \
+    SampleResultArgsType, SampleResultType, SampleResultsDictType, SampleReturnType, Sampler, \
+        SamplerOutput, _apply_min_p, _apply_min_tokens_penalty, _apply_top_k_top_p, _build_sampler_output, \
+        _modify_greedy_probs_inplace, _top_k_top_p_multinomial_with_flashinfer, get_logprobs, _multinomial
+from vllm.model_executor.layers.utils import apply_penalties
+from vllm.model_executor.sampling_metadata import SamplingMetadata, SamplingTensors, SequenceGroupToSample
+from vllm.sampling_params import SamplingType
+from vllm.sequence import VLLM_INVALID_TOKEN_ID
+if envs.VLLM_USE_FLASHINFER_SAMPLER and find_spec("flashinfer"):
+    import flashinfer.sampling
+    # yapf: disable
+    from flashinfer.sampling import (
+        top_k_top_p_sampling_from_probs as flashinfer_top_k_top_p_sampling)
+    # yapf: enable
+else:
+    flashinfer_top_k_top_p_sampling = None
+
+class SampleRecorder:
+    def __init__(self):
+        self.seq_ids:torch.Tensor = None
+        self.sampled_token_ids_tensor:torch.Tensor = None
+
+last_sampler = None
+
+def get_last_sampler():
+    return last_sampler
+
+class ZeroOverheadSampler(Sampler):
+    def __init__(self):
+        super().__init__()
+        
+    def forward(
+        self,
+        logits: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        """
+        Single-step scheduling:
+        * Perform GPU-side sampling computation & compute
+          GPU-side logprobs tensor
+        * Pythonize sampling result & logprobs tensor
+
+        Multi-step scheduling:
+        * Perform GPU-side sampling computation & compute
+          GPU-side logprobs tensor
+        * Defer Pythonization of sampling result & logprobs
+          tensor
+        * Encapsulate arguments required for deferred Pythonization
+          in the :class:`SamplerOutput` structure
+
+        Args:
+            logits: (num_tokens, vocab_size).
+            sampling_metadata: Metadata for sampling.
+        """
+        global last_sampler
+        last_sampler = SampleRecorder() 
+        assert logits is not None
+        _, vocab_size = logits.shape
+
+        # Prepare sampling tensors with pinned memory to avoid blocking.
+        if not sampling_metadata.reuse_sampling_tensors:
+            self._init_sampling_tensors(logits, sampling_metadata)
+        elif self._do_penalties:
+            # In this case, the sampling tensors logic depends on
+            # "output_tokens" of a sequence. As a result, we cannot
+            # reuse sampling tensors, since "output_tokens" changes
+            # between decode runs.
+            self._init_sampling_tensors(logits, sampling_metadata)
+
+        assert self._sampling_tensors is not None
+        sampling_tensors = self._sampling_tensors
+        do_penalties = self._do_penalties
+        do_top_p_top_k = self._do_top_p_top_k
+        do_min_p = self._do_min_p
+
+        logits = _apply_min_tokens_penalty(logits, sampling_metadata)
+
+        # Apply presence and frequency penalties.
+        if do_penalties:
+            logits = apply_penalties(logits, sampling_tensors.prompt_tokens,
+                                     sampling_tensors.output_tokens,
+                                     sampling_tensors.presence_penalties,
+                                     sampling_tensors.frequency_penalties,
+                                     sampling_tensors.repetition_penalties)
+
+        # Use float32 to apply temperature scaling.
+        # Use in-place division to avoid creating a new tensor.
+        logits = logits.to(torch.float)
+        logits.div_(sampling_tensors.temperatures.unsqueeze(dim=1))
+
+        if do_top_p_top_k and flashinfer_top_k_top_p_sampling is None:
+            logits = _apply_top_k_top_p(logits, sampling_tensors.top_ps,
+                                        sampling_tensors.top_ks)
+
+        if do_min_p:
+            logits = _apply_min_p(logits, sampling_tensors.min_ps)
+
+        # We use float32 for probabilities and log probabilities.
+        # Compute the probabilities.
+        probs = torch.softmax(logits, dim=-1, dtype=torch.float)
+        # Compute the log probabilities.
+        logprobs = torch.log_softmax(logits, dim=-1, dtype=torch.float)
+
+        # Sample the next tokens.
+        maybe_deferred_sample_results, maybe_sampled_tokens_tensor = _sample(
+            probs,
+            logprobs,
+            sampling_metadata,
+            sampling_tensors,
+            include_gpu_probs_tensor=self.include_gpu_probs_tensor,
+            modify_greedy_probs=self._should_modify_greedy_probs_inplace,
+        )
+
+        if self.include_gpu_probs_tensor:
+            # Since we will defer sampler result Pythonization,
+            # preserve GPU-side tensors in support of later
+            # deferred pythonization of logprobs
+            assert maybe_sampled_tokens_tensor is not None
+            on_device_tensors = (probs, logprobs, maybe_sampled_tokens_tensor)
+        else:
+            # Since Pythonization has already happened, don't preserve
+            # GPU-side tensors.
+            on_device_tensors = None
+
+        # Get the logprobs query results.
+        prompt_logprobs = None
+        sample_logprobs = None
+        if not sampling_metadata.skip_sampler_cpu_output:
+            # Pythonize logprobs now (GPU -> CPU); do not defer.
+            assert not isinstance(maybe_deferred_sample_results,
+                                  SampleResultArgsType)
+            prompt_logprobs, sample_logprobs = get_logprobs(
+                logprobs, sampling_metadata, maybe_deferred_sample_results)
+
+        return _build_sampler_output(
+            maybe_deferred_sample_results,
+            sampling_metadata,
+            prompt_logprobs,
+            sample_logprobs,
+            on_device_tensors=on_device_tensors,
+            skip_sampler_cpu_output=sampling_metadata.skip_sampler_cpu_output,
+            logits=logits)
+
+def _greedy_sample(
+    selected_seq_groups: List[SequenceGroupToSample],
+    samples: torch.Tensor,
+) -> SampleResultType:
+    """Run greedy sampling on a given samples.
+
+    Args:
+        selected_seq_groups: A list of sequence groups batched.
+        samples: (num_selected_samples,) A tensor of samples. The length of
+            samples could be smaller than selected_seq_groups if
+            seq_group.do_sample is False.
+    Returns:
+        Tuple of (next_token_ids, parent_ids). The length of returned list is
+        same as the length of selected_seq_groups. If the corresponding
+        seq_group has do_sample=False, tuple contains ([], [])
+    """
+    sample_idx = 0
+    results: SampleResultType = []
+    for seq_group in selected_seq_groups:
+        if not seq_group.do_sample:
+            results.append(([], []))
+            continue
+
+        seq_ids = seq_group.seq_ids
+        num_parent_seqs = len(seq_ids)
+        assert num_parent_seqs == 1, (
+            "Greedy sampling should have only one seq.")
+        parent_ids = list(range(num_parent_seqs))
+        assert num_parent_seqs == 1 # not support muti seqences in seqence group
+        next_token_ids = [0] #place holder token id
+        results.append((next_token_ids, parent_ids))
+        sample_idx += num_parent_seqs
+    return results
+
+def _random_sample(
+    selected_seq_groups: List[SequenceGroupToSample],
+    random_samples: torch.Tensor,
+) -> SampleResultType:
+    """Run random sampling on a given samples.
+
+    Args:
+        selected_seq_groups: A list of sequence groups batched.
+        random_samples: (num_selected_samples,) A tensor of samples. The
+            length of samples could be smaller than selected_seq_groups if
+            seq_group.do_sample is False.
+    Returns:
+        Tuple of (next_token_ids, parent_ids). The length of returned list is
+        same as the length of selected_seq_groups. If the corresponding
+        seq_group has do_sample=False, tuple contains ([], [])
+    """
+    # Find the maximum n value of the prompt phase requests.
+    sample_idx = 0
+    results: SampleResultType = []
+    for seq_group in selected_seq_groups:
+        if not seq_group.do_sample:
+            results.append(([], []))
+            continue
+
+        seq_ids = seq_group.seq_ids
+        sampling_params = seq_group.sampling_params
+        is_prompt = seq_group.is_prompt
+        num_parent_seqs = len(seq_ids)
+        if is_prompt:
+            # Prompt phase.
+            parent_ids = [0] * sampling_params.n
+            assert num_parent_seqs == 1 # not support muti seqences in seqence group
+            next_token_ids = [0] * sampling_params.n  #place holder token id
+        else:
+            # Generation phase.
+            parent_ids = list(range(num_parent_seqs))
+            assert num_parent_seqs == 1 # not support muti seqences in seqence group
+            next_token_ids = [0] * num_parent_seqs  #place holder token id
+        results.append((next_token_ids, parent_ids))
+        sample_idx += num_parent_seqs
+    return results
+
+def _sample(
+    probs: torch.Tensor,
+    logprobs: torch.Tensor,
+    sampling_metadata: SamplingMetadata,
+    sampling_tensors: SamplingTensors,
+    include_gpu_probs_tensor: bool,
+    modify_greedy_probs: bool,
+) -> SampleReturnType:
+    """
+    Args:
+        probs: (num_query_tokens_in_batch, num_vocab)
+        logprobs: (num_query_tokens_in_batch, num_vocab)
+        sampling_metadata: The metadata for a batch for sampling.
+        sampling_tensors: Tensors that include sampling related metadata.
+
+    Returns:
+        (next_token_ids, parent_seq_ids) for each seq group in a batch.
+            If sampling is skipped, it returns ([], [])
+        sampled_token_ids_tensor: A tensor of sampled token ids.
+    """
+    return _sample_with_torch(
+        probs,
+        logprobs,
+        sampling_metadata,
+        sampling_tensors,
+        include_gpu_probs_tensor=include_gpu_probs_tensor,
+        modify_greedy_probs=modify_greedy_probs,
+    )
+
+def _sample_with_torch(
+    probs: torch.Tensor,
+    logprobs: torch.Tensor,
+    sampling_metadata: SamplingMetadata,
+    sampling_tensors: SamplingTensors,
+    include_gpu_probs_tensor: bool,
+    modify_greedy_probs: bool,
+) -> SampleReturnType:
+    '''Torch-oriented _sample() implementation.
+
+    Single-step scheduling:
+    * Perform GPU-side sampling computation
+    * Immediately Pythonize sampling result
+
+    Multi-step scheduling:
+    * Perform GPU-side sampling computation
+    * Defer Pythonization & preserve GPU-side
+      tensors required for Pythonization
+    '''
+
+    categorized_seq_group_ids: Dict[SamplingType, List[int]] = {
+        t: []
+        for t in SamplingType
+    }
+    last_sampler.seq_ids = []
+    categorized_sample_indices = sampling_metadata.categorized_sample_indices
+    for i, seq_group in enumerate(sampling_metadata.seq_groups):
+        last_sampler.seq_ids.append(seq_group.seq_ids[0])
+        sampling_params = seq_group.sampling_params
+        sampling_type = sampling_params.sampling_type
+        categorized_seq_group_ids[sampling_type].append(i)
+    sample_results_dict: SampleResultsDictType = {}
+    sample_metadata: SampleMetadataType = {}
+    multinomial_samples: MultinomialSamplesType = {}
+    greedy_samples: Optional[torch.Tensor] = None
+
+    # Create output tensor for sampled token ids.
+    if include_gpu_probs_tensor:
+        sampled_token_ids_tensor = torch.full((logprobs.shape[0], 1),
+                                              VLLM_INVALID_TOKEN_ID,
+                                              dtype=torch.long,
+                                              device=logprobs.device)
+    else:
+        sampled_token_ids_tensor = None
+
+    # Counterintiutively, having two loops here is actually faster.
+    # The first loop can run without waiting on GPU<->CPU sync.
+    for sampling_type in SamplingType:
+        sample_indices = categorized_sample_indices[sampling_type]
+        num_tokens = len(sample_indices)
+        if num_tokens == 0:
+            continue
+
+        seq_group_id = categorized_seq_group_ids[sampling_type]
+        seq_groups = [sampling_metadata.seq_groups[i] for i in seq_group_id]
+        sample_metadata[sampling_type] = (seq_group_id, seq_groups)
+        long_sample_indices = sample_indices.long()
+        if sampling_type == SamplingType.GREEDY:
+            greedy_samples = torch.argmax(logprobs[long_sample_indices],
+                                          dim=-1)
+            
+            last_sampler.sampled_token_ids_tensor = greedy_samples.unsqueeze(-1)
+
+            if sampled_token_ids_tensor is not None:
+                # Store sampled tokens in output tensor.
+                sampled_token_ids_tensor[
+                    long_sample_indices] = greedy_samples.unsqueeze(-1)
+
+            if modify_greedy_probs:
+                # If required, modify the probabilities such that sampling from
+                # the modified distribution would always sample the argmax
+                # token id.
+                _modify_greedy_probs_inplace(logprobs, probs,
+                                             long_sample_indices,
+                                             greedy_samples)
+
+        elif sampling_type in (SamplingType.RANDOM, SamplingType.RANDOM_SEED):
+            max_n_in_batch = 1
+            for seq_group in seq_groups:
+                if seq_group.is_prompt:
+                    sampling_params = seq_group.sampling_params
+                    max_n_in_batch = max(max_n_in_batch, sampling_params.n)
+            seq_groups_arg = (None if sampling_type == SamplingType.RANDOM else
+                              seq_groups)
+
+            if flashinfer_top_k_top_p_sampling is not None:
+                multinomial_samples[
+                    sampling_type] = _top_k_top_p_multinomial_with_flashinfer(
+                        probs[long_sample_indices],
+                        sampling_tensors.top_ks[long_sample_indices],
+                        sampling_tensors.top_ps[long_sample_indices],
+                        max_n_in_batch,
+                        seq_groups_arg,
+                    )
+            else:
+                multinomial_samples[sampling_type] = _multinomial(
+                    probs[long_sample_indices],
+                    max_n_in_batch,
+                    seq_groups=seq_groups_arg)
+                
+            last_sampler.sampled_token_ids_tensor = \
+                multinomial_samples[sampling_type].to(torch.long)
+
+            if sampled_token_ids_tensor is not None:
+                # Store sampled tokens in output tensor.
+                sampled_token_ids_tensor[long_sample_indices] = \
+                    multinomial_samples[sampling_type].to(torch.long)
+
+    # Encapsulate arguments for computing Pythonized sampler
+    # results, whether deferred or otherwise.
+    maybe_deferred_args = SampleResultArgsType(
+        sampling_metadata=sampling_metadata,
+        sample_metadata=sample_metadata,
+        multinomial_samples=multinomial_samples,
+        greedy_samples=greedy_samples,
+        sample_results_dict=sample_results_dict)
+
+    if not sampling_metadata.skip_sampler_cpu_output:
+        # GPU<->CPU sync happens here.
+        # This also converts the sampler output to a Python object.
+        # Return Pythonized sampler result & sampled token ids
+        return get_pythonized_sample_results(
+            maybe_deferred_args), sampled_token_ids_tensor
+    else:
+        # Defer sampler result Pythonization; return deferred
+        # Pythonization args & sampled token ids
+        return (
+            maybe_deferred_args,
+            sampled_token_ids_tensor,
+        )
+
+
+def get_pythonized_sample_results(
+        sample_result_args: SampleResultArgsType) -> SampleResultType:
+    '''This function consumes GPU-side sampler results and computes
+    Pythonized CPU-side sampler results (GPU -> CPU sync.)
+
+    Single-step scheduling: this function is invoked at sampling-time
+    for immediate Pythonization.
+
+    Multi-step scheduling: Pythonization is deferred until after multiple
+    GPU-side steps have been completed.
+
+    Args:
+      sample_result_args: GPU-side inputs to the Pythonization process
+
+    Returns:
+      Pythonized sampler results
+    '''
+
+    (
+        sample_metadata,
+        sampling_metadata,
+        greedy_samples,
+        multinomial_samples,
+        sample_results_dict,
+    ) = (
+        sample_result_args.sample_metadata,
+        sample_result_args.sampling_metadata,
+        sample_result_args.greedy_samples,
+        sample_result_args.multinomial_samples,
+        sample_result_args.sample_results_dict,
+    )
+
+    for sampling_type in SamplingType:
+        if sampling_type not in sample_metadata:
+            continue
+        (seq_group_id, seq_groups) = sample_metadata[sampling_type]
+        if sampling_type == SamplingType.GREEDY:
+            sample_results = _greedy_sample(seq_groups, greedy_samples)
+        elif sampling_type in (SamplingType.RANDOM, SamplingType.RANDOM_SEED):
+            sample_results = _random_sample(seq_groups,
+                                            multinomial_samples[sampling_type])
+        sample_results_dict.update(zip(seq_group_id, sample_results))
+
+    return [
+        sample_results_dict.get(i, ([], []))
+        for i in range(len(sampling_metadata.seq_groups))
+    ]
\ No newline at end of file

vllm/zero_overhead/sequence.py

+
+
+from typing import Union
+from vllm.sequence import Sequence
+from typing import Sequence as GenericSequence
+
+
+class ZeroOverheadSequence(Sequence):
+    def __init__(self, seq_id, inputs, block_size, eos_token_id = None, lora_request = None, prompt_adapter_request = None):
+        super().__init__(seq_id, inputs, block_size, eos_token_id, lora_request, prompt_adapter_request)
+        self.effective_output_len : int = 0
+
+    def fix_last_token_id(self, token_id: int) -> None:
+        effect_offset = self.effective_output_len - len(self.data.output_token_ids)
+        assert effect_offset < 0
+        self.data._output_token_ids[effect_offset] = token_id
+        if len(self.data._new_appended_tokens) >= effect_offset * -1:
+            self.data._new_appended_tokens[effect_offset] = token_id
+        self.data._cached_all_token_ids[effect_offset] = token_id
+        self.effective_output_len += 1
+    
+    def remove_last_place_holder(self, count):
+        self.data._output_token_ids = self.data._output_token_ids[:-1 * count]
+        self.data._new_appended_tokens = self.data._new_appended_tokens[:-1 * count]
+        self.data._cached_all_token_ids = self.data._cached_all_token_ids[:-1 * count]
+        self.data._num_computed_tokens -= count
+
+    def zero_overhead_get_output_token_ids(self) -> tuple[int, ...]:
+        return self.data.output_token_ids[:self.effective_output_len]
+    
+    def zero_overhead_get_output_len(self) -> int:
+        return self.effective_output_len
+    
+    def zero_overhead_get_last_token_id(self) -> int:
+        if self.effective_output_len == 0:
+            return self.data._prompt_token_ids[-1]
+        return self.data._output_token_ids[self.effective_output_len - 1]
+    
+    def zero_overhead_get_len(self) -> int:
+        return self.effective_output_len + len(self.data._prompt_token_ids)
+    
+    def get_output_token_ids_to_return(
+            self, delta: bool) -> Union[GenericSequence[int], int]:
+        """If delta is True, only new tokens since the last call to
+        this method are returned"""
+        if not delta:
+            return self.zero_overhead_get_output_token_ids()
+
+        output_len = self.zero_overhead_get_output_len()
+
+        # Get the number of new tokens
+        num_new_tokens = output_len - self._last_output_token_ids_offset
+        self._last_output_token_ids_offset = output_len
+
+        # Return new tokens
+        if num_new_tokens == 1:
+            # Optimization for single decode token case
+            # (which is what we have most of the time)
+            return self.data._cached_all_token_ids[self.effective_output_len - 1]
+
+        if num_new_tokens == 0:
+            return []
+
+        effect_offset = self.effective_output_len - len(self.data.output_token_ids)
+        return self.data._cached_all_token_ids[-num_new_tokens : effect_offset]
\ No newline at end of file

vllm/zero_overhead/spec_decode/batch_expansion.py

+from array import array
+import numpy as np
+from itertools import chain, count
+from typing import Iterator, List, Optional, Tuple
+
+import torch
+
+from vllm import SamplingParams
+from vllm.model_executor.layers.sampler import SamplerOutput
+from vllm.sequence import (VLLM_INVALID_TOKEN_ID, VLLM_TOKEN_ID_ARRAY_TYPE,
+                           ExecuteModelRequest, SequenceData,
+                           SequenceGroupMetadata, get_all_seq_ids)
+from vllm.spec_decode.batch_expansion import BatchExpansionTop1Scorer
+from vllm.spec_decode.interfaces import (SpeculativeProposals,
+                                         SpeculativeScorer, SpeculativeScores)
+from vllm.spec_decode.util import nvtx_range, split_batch_by_proposal_len
+from vllm.utils import async_tensor_h2d
+from vllm.zero_overhead.utils import get_proposal_lens_list, record_proposal_token_ids
+
+SeqId = int
+TargetSeqId = int
+TokenId = int
+
+DEFAULT_SIMPLE_SAMPLING_PARAMS = SamplingParams()
+
+
+class ZeroOverheadBatchExpansionTop1Scorer(BatchExpansionTop1Scorer):
+    
+    @nvtx_range("BatchExpansionTop1Scorer.score_proposals")
+    def score_proposals(
+        self,
+        execute_model_req: ExecuteModelRequest,
+        proposals: SpeculativeProposals,
+    ) -> SpeculativeScores:
+        """Score the proposed tokens via the scorer model.
+
+        This converts each input sequence to a set of k+1 target sequences. The
+        target sequences have the unique continuations to be scored and a
+        unique sequence ID that is different from all input sequence ids.
+
+        If a speculative sequence length would exceed the max model length, then
+        no speculation is produced for that sequence.
+
+        Args:
+            execute_model_req: The execution request.
+            proposals: The speculative proposals to score.
+        Returns:
+            SpeculativeScores: The scores of each speculative token, along with
+                which sequences were ignored during scoring.
+        """
+
+        proposal_lens_list = get_proposal_lens_list()
+        record_proposal_token_ids(proposals.proposal_token_ids)
+        proposal_token_ids_list = np.zeros(proposals.proposal_token_ids.shape, dtype=int).tolist() # place holder tokens
+
+        # Filter the list to ignore invalid proposals.
+        proposal_token_ids_list_without_skips = [
+            proposals for proposals in proposal_token_ids_list
+            if VLLM_INVALID_TOKEN_ID not in proposals
+        ]
+
+        (spec_indices, non_spec_indices, target_seq_group_metadata_list,
+         num_scoring_tokens) = self._expand_batch(
+             seq_group_metadata_list=execute_model_req.seq_group_metadata_list,
+             proposal_token_ids_list=proposal_token_ids_list_without_skips,
+             proposal_lens_list=proposal_lens_list,
+         )
+
+        target_sampler_output = self._scorer_worker.execute_model(
+            execute_model_req=execute_model_req.clone(
+                seq_group_metadata_list=target_seq_group_metadata_list))
+        assert len(target_sampler_output) == 1, "expected single-step output"
+        target_sampler_output = target_sampler_output[0]
+        if not non_spec_indices:
+            # All sequence groups in batch have spec decoding enabled
+            return self._contract_batch_all_spec(
+                target_sampler_output=target_sampler_output,
+                proposals=proposals,
+            )
+        else:
+            # Batch has a mix of spec decode enabled and disabled seq groups
+            return self._contract_batch(
+                execute_model_req.seq_group_metadata_list,
+                target_sampler_output=target_sampler_output,
+                proposals=proposals,
+                num_scoring_tokens=num_scoring_tokens,
+                non_spec_indices=non_spec_indices,
+                spec_indices=spec_indices,
+                k=execute_model_req.num_lookahead_slots,
+            )
+        
+    def _contract_non_speculative(
+            self, scores: SpeculativeScores,
+            seq_group_metadata_list: List[SequenceGroupMetadata],
+            non_spec_indices: List[int], non_spec_outputs: SpeculativeScores,
+            has_prompt_log: bool) -> SpeculativeScores:
+        """
+            Augment input `scores` with non-speculative requests outputs. 
+            This includes decode requests with speculation turned off, as well
+            as prefill requests when `enable_chunked_prefill` is set.
+            For the latter, prefills are further separated into terminal and 
+            non-terminal chunks (from which no token is sampled).
+        """
+        if not non_spec_indices:
+            return scores
+
+        if has_prompt_log:
+            # When prompt_logprobs is enabled, prefills yield output token
+            # (and respective prob) in the last entry (prompt|out):
+            # [.|.|.|prefill0_out|.|prefill1_out|decode0_out|..].
+            # With chunked prefill, non-terminal chunks have -1 on each
+            # position: they're still picked, but they're discarded later.
+            seq_meta = seq_group_metadata_list
+            nospec_sizes = torch.tensor([
+                seq_meta[i].token_chunk_size if seq_meta[i].is_prompt else 1
+                for i in non_spec_indices
+            ])
+            nospec_sampled_token_idxs = torch.cumsum(nospec_sizes, 0).add_(-1)
+        else:
+            # In this case only sampled tokens are returned, select all.
+            nospec_sampled_token_idxs = list(
+                range(len(non_spec_outputs.token_ids)))
+        
+        nospec_sampled_token_idxs = async_tensor_h2d(nospec_sampled_token_idxs, torch.int32,
+                                            self._device,
+                                            True)     
+        non_spec_indices = async_tensor_h2d(non_spec_indices, torch.int32,
+                                            self._device,
+                                            True)
+    
+        scores.token_ids[non_spec_indices, :1] = \
+            non_spec_outputs.token_ids[nospec_sampled_token_idxs].unsqueeze(1)
+        scores.probs[non_spec_indices, :1, :] = \
+            non_spec_outputs.probs[nospec_sampled_token_idxs].unsqueeze(1)
+        scores.logprobs[non_spec_indices, :1, :] = \
+            non_spec_outputs.logprobs[nospec_sampled_token_idxs].unsqueeze(1)
+        if scores.hidden_states is not None:
+            assert non_spec_outputs.hidden_states is not None
+            scores.hidden_states[non_spec_indices, :1, :] = \
+                non_spec_outputs.hidden_states[nospec_sampled_token_idxs].unsqueeze(1)
+        return scores
\ No newline at end of file

vllm/zero_overhead/spec_decode/muti_step_worker.py

+import copy
+import weakref
+from typing import Dict, List, Set, Tuple
+
+import torch
+
+from vllm.model_executor.layers.sampler import SamplerOutput
+from vllm.platforms import current_platform
+from vllm.sequence import (ExecuteModelRequest, HiddenStates, SequenceData,
+                           SequenceGroupMetadata)
+from vllm.spec_decode.multi_step_worker import MultiStepWorker
+from vllm.utils import async_tensor_h2d
+from vllm.zero_overhead.spec_decode.top1_proproser import ZeroOverheadTop1Proposer
+from vllm.zero_overhead.utils import SpecStepKind, set_spec_step
+
+if current_platform.is_cuda_alike():
+    from vllm.spec_decode.draft_model_runner import TP1DraftModelRunner
+
+from vllm.spec_decode.interfaces import (SpeculativeProposals,
+                                         SpeculativeProposer)
+from vllm.spec_decode.proposer_worker_base import ProposerWorkerBase
+from vllm.worker.worker_base import DelegateWorkerBase
+
+class ZeroOverheadMultiStepWorker(MultiStepWorker):
+
+    def init_device(self) -> None:
+        self.worker.init_device()
+        self._proposer = ZeroOverheadTop1Proposer(
+            weakref.proxy(self),  # type: ignore[arg-type]
+            self.device,
+            self.vocab_size,
+            max_proposal_len=self.max_model_len,
+        )
+
+    @torch.inference_mode()
+    def sampler_output(
+        self,
+        execute_model_req: ExecuteModelRequest,
+        sample_len: int,
+        seq_ids_with_bonus_token_in_last_step: Set[int],
+    ) -> Tuple[List[SamplerOutput], bool]:
+        """Run the model forward pass sample_len times. Returns the list of
+        sampler output, one per model forward pass, along with indicator of
+        whether torch tensor in sampler output need to be transposed in latter
+        sampler_output_to_torch logic.
+
+        For multi step worker, this indicator shall be True.
+        """
+        self._raise_if_unsupported(execute_model_req)
+        # Expand the batch for sequences with a bonus token.
+        # Perform a forward pass on the expanded batch and filter the
+        # response to retain only the original sequences' responses.
+        expanded_request, indices_of_seq_with_bonus_tokens =\
+            self._expand_execute_model_request(
+                execute_model_req, seq_ids_with_bonus_token_in_last_step)
+        # Run model sample_len times.
+        model_outputs: List[SamplerOutput] = []
+        if current_platform.is_cuda_alike() and isinstance(
+                self.model_runner, TP1DraftModelRunner
+        ) and self.model_runner.supports_gpu_multi_step(expanded_request):
+            # Here we run the draft_model_runner with multi-step prepare
+            # on the GPU directly
+            expanded_request.num_steps = sample_len
+            self.model_runner.set_indices_of_seq_with_bonus_tokens(
+                indices_of_seq_with_bonus_tokens)
+            model_outputs = self.execute_model(
+                execute_model_req=expanded_request)
+        else:
+            # Here we run multi-step directly, with every step prepared
+            # on the CPU.
+            # TODO: Remove this branch once DraftModelRunner supports TP>1
+            # and other restrictions that are part of DraftModelRunner's
+            # supports_gpu_multi_step(..)
+
+            set_spec_step(SpecStepKind.FIRST_PROPOSAL)
+            for _ in range(sample_len):
+                model_output: List[SamplerOutput] = self.worker.execute_model(
+                    execute_model_req=expanded_request)
+                assert (len(model_output) == 1
+                        ), "composing multistep workers not supported"
+                model_output = model_output[0]
+                set_spec_step(SpecStepKind.OTHER_PROPOSAL)
+                self._append_new_tokens(
+                    model_output, expanded_request.seq_group_metadata_list,
+                    indices_of_seq_with_bonus_tokens)
+                model_outputs.append(model_output)
+            set_spec_step(SpecStepKind.SCORE_DECODE)
+
+        filtered_model_outputs = self._filter_model_output_zero_overhead(
+            model_outputs, indices_of_seq_with_bonus_tokens)
+        
+        return filtered_model_outputs, True
+    
+    def _filter_model_output_zero_overhead(self,
+            expanded_batch_outputs: List[SamplerOutput],
+            output_indices_to_retain: List[int]) -> List[SamplerOutput]:
+        """
+        Filters the model output to include only the specified sequence
+        outputs. This method contracts the expanded batch output from the
+        model to retain the outputs of only those sequences indicated by the
+        provided indices.
+
+        Args:
+            expanded_batch_output (List[SamplerOutput]): The expanded output
+                batch from the model.
+            output_indices_to_retain (torch.Tensor): Indices of the model
+                outputs to retain.
+
+        Returns:
+            List[SamplerOutput]: A list containing the filtered model 
+            outputs for the specified indices.
+        """
+        
+        indices_of_seq_with_bonus_tokens = async_tensor_h2d(output_indices_to_retain, torch.int32,
+                                            self.device,
+                                            True)
+        
+        return [
+            SamplerOutput(
+                outputs=[
+                    expanded_batch_output.outputs[i]
+                    for i in output_indices_to_retain
+                ] if len(expanded_batch_output.outputs) > 0 else [],
+                sampled_token_probs=(
+                    expanded_batch_output.
+                    sampled_token_probs[indices_of_seq_with_bonus_tokens]
+                    if expanded_batch_output.sampled_token_probs is not None
+                    else None),
+                logprobs=(
+                    expanded_batch_output.logprobs[indices_of_seq_with_bonus_tokens]
+                    if expanded_batch_output.logprobs is not None else None),
+                sampled_token_ids=(expanded_batch_output.
+                                   sampled_token_ids[indices_of_seq_with_bonus_tokens]
+                                   if expanded_batch_output.sampled_token_ids
+                                   is not None else None))
+            for expanded_batch_output in expanded_batch_outputs
+        ]
\ No newline at end of file

vllm/zero_overhead/spec_decode/top1_proproser.py

+import os
+from typing import List, Optional, Set, Tuple
+
+import torch
+
+from vllm.model_executor.layers.sampler import SamplerOutput
+from vllm.sequence import ExecuteModelRequest, SequenceGroupMetadata
+from vllm.spec_decode.interfaces import (SpeculativeProposals,
+                                         SpeculativeProposer)
+from vllm.spec_decode.proposer_worker_base import ProposerWorkerBase
+from vllm.spec_decode.top1_proposer import Top1Proposer
+from vllm.spec_decode.util import sampler_output_to_torch
+from vllm.utils import async_tensor_h2d
+from vllm.zero_overhead.utils import record_proposal_lens_list
+
+class ZeroOverheadTop1Proposer(Top1Proposer):
+
+    def _merge_outputs(
+        self,
+        batch_size: int,
+        proposal_len: int,
+        maybe_sampler_output: Optional[List[SamplerOutput]],
+        proposal_lens: List[int],
+        nonzero_proposal_len_indices: List[int],
+        sampler_transposed: bool,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """After speculations are produced, merge the speculation results with
+        the skipped sequences.
+        """
+        if maybe_sampler_output is None:
+            # If no speculative tokens, the sampler output will be None.
+            # In this case we return empty proposals.
+            proposal_tokens = torch.tensor(-1,
+                                           dtype=torch.long,
+                                           device=self._device).expand(
+                                               batch_size, proposal_len)
+            proposal_probs = torch.tensor(0,
+                                          dtype=torch.float32,
+                                          device=self._device).expand(
+                                              batch_size, proposal_len,
+                                              self._vocab_size)
+            proposal_lens_tensor = torch.tensor(0,
+                                                dtype=torch.long,
+                                                device=self._device).expand(
+                                                    len(proposal_lens))
+            return proposal_tokens, proposal_probs, proposal_lens_tensor
+
+        sampler_output = maybe_sampler_output
+        proposal_tokens, proposal_probs, *_ = sampler_output_to_torch(
+            sampler_output, sampler_transposed)
+
+        proposal_lens_list = [0 for i in range(batch_size)]
+        for indices in nonzero_proposal_len_indices:
+            proposal_lens_list[indices] = proposal_len
+        record_proposal_lens_list(proposal_lens_list)
+
+        nonzero_proposal_len_indices = async_tensor_h2d(nonzero_proposal_len_indices, torch.int32,
+                                            self._device,
+                                            True)
+
+        # Now, reformat the output GPU tensors such that each sequence has
+        # a proposal. the proposal can be empty, e.g. [-1, -1, -1]
+
+        entire_proposal_tokens = proposal_tokens.new_full(
+            size=(batch_size, *proposal_tokens.shape[1:]),
+            fill_value=-1,
+        )
+        entire_proposal_tokens[nonzero_proposal_len_indices] = proposal_tokens
+        entire_proposal_probs = proposal_probs.new_zeros(
+            batch_size,
+            *proposal_probs.shape[1:],
+        )
+        entire_proposal_probs[nonzero_proposal_len_indices] = proposal_probs
+
+        proposal_tokens, proposal_probs = (
+            entire_proposal_tokens,
+            entire_proposal_probs,
+        )
+        
+        proposal_lens_tensor = async_tensor_h2d(proposal_lens_list, torch.long,
+                                            self._device,
+                                            True)
+
+        return proposal_tokens, proposal_probs, proposal_lens_tensor
\ No newline at end of file

vllm/zero_overhead/stop_check.py

+
+
+from typing import Optional
+from vllm.engine.output_processor.stop_checker import StopChecker
+from vllm.lora.request import LoRARequest
+from vllm.sampling_params import SamplingParams
+from vllm.sequence import SequenceStatus
+from vllm.zero_overhead.sequence import ZeroOverheadSequence
+
+
+class ZeroOverheadStopChecker(StopChecker):
+    def __init__(self, max_model_len, get_tokenizer_for_seq):
+        super().__init__(max_model_len, get_tokenizer_for_seq)
+    
+    
+    def maybe_stop_sequence(
+        self,
+        seq: ZeroOverheadSequence,
+        new_char_count: int,
+        sampling_params: SamplingParams,
+        lora_req: Optional[LoRARequest] = None,
+    ) -> None:
+        """Stop the finished sequences.
+
+       new_char_count is the number of chars added to the
+           sequence's output text for the newly generated token
+        """
+
+        # Check if the minimum number of tokens has been generated yet;
+        # skip the stop string/token checks if not
+        if seq.zero_overhead_get_output_len() < sampling_params.min_tokens:
+            return
+
+        # Check if the sequence has generated the EOS token.
+        if ((not sampling_params.ignore_eos)
+                and seq.zero_overhead_get_last_token_id() == seq.eos_token_id):
+            # Remove the last EOS token unless explicitly specified
+            # This prevents unintended exposure of the EOS token
+            if new_char_count and (
+                    not sampling_params.include_stop_str_in_output):
+                seq.output_text = seq.output_text[:-new_char_count]
+            seq.status = SequenceStatus.FINISHED_STOPPED
+            return
+
+        # Check if a stop token was encountered.
+        # This assumes a single token produced per step.
+        last_token_id = seq.zero_overhead_get_last_token_id()
+        if last_token_id in (sampling_params.stop_token_ids or ()):
+            if new_char_count and (
+                    not sampling_params.include_stop_str_in_output):
+                # Remove last token
+                seq.output_text = seq.output_text[:-new_char_count]
+            seq.status = SequenceStatus.FINISHED_STOPPED
+            seq.stop_reason = last_token_id
+            return
+
+        # Check if any stop strings are matched.
+        stop = self.check_stop_strings(
+            seq.output_text, new_char_count, sampling_params.stop,
+            sampling_params.include_stop_str_in_output)
+        if stop is not None:
+            stop_str, truncate_to = stop
+            if truncate_to != -1:
+                seq.output_text = seq.output_text[:truncate_to]
+            seq.status = SequenceStatus.FINISHED_STOPPED
+            seq.stop_reason = stop_str
+            return
+
+        # Check if the sequence has reached max_model_len.
+        if seq.zero_overhead_get_len() > self._get_max_model_len(lora_req):
+            seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
+            return
+
+        # Check if the sequence has reached max_tokens.
+        if seq.zero_overhead_get_output_len() == sampling_params.max_tokens:
+            seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
+            return
\ No newline at end of file

vllm/zero_overhead/tokenizer.py

+
+
+from vllm.sampling_params import SamplingParams
+from vllm.sequence import VLLM_INVALID_TOKEN_ID
+from vllm.transformers_utils.detokenizer import Detokenizer
+from vllm.transformers_utils.detokenizer_utils import convert_prompt_ids_to_tokens, detokenize_incrementally
+from vllm.zero_overhead.sequence import ZeroOverheadSequence
+
+
+class ZeroOverheadDetokenizer(Detokenizer):
+    def __init__(self, tokenizer_group):
+        super().__init__(tokenizer_group)
+    
+    def decode_sequence_inplace(self, seq: ZeroOverheadSequence,
+                                prms: SamplingParams) -> int:
+        """Decodes the new token for a sequence. In-place operation.
+
+        Args:
+            seq: The sequence to decode.
+            prms: The sampling parameters used to generate the sequence.
+
+        Returns:
+            The number of characters added to the output text.
+        """   
+        eff_length = seq.get_prompt_len() + seq.effective_output_len
+        all_input_ids = seq.get_token_ids()[ : eff_length]
+
+        token_id_generated_this_iteration = all_input_ids[-1]
+        tokenizer = self.get_tokenizer_for_seq(seq)
+
+        # Convert prompt token IDs to tokens if necessary.
+        # Do it here so that we don't have to repeat this
+        # computation for each logprob.
+        if seq.tokens is None:
+            (seq.tokens, seq.prefix_offset,
+             seq.read_offset) = convert_prompt_ids_to_tokens(
+                 tokenizer=tokenizer,
+                 prompt_ids=all_input_ids[:-1],
+                 skip_special_tokens=prms.skip_special_tokens,
+             )
+
+        (new_tokens, new_decoded_token_text, prefix_offset,
+         read_offset) = detokenize_incrementally(
+             tokenizer=tokenizer,
+             all_input_ids=all_input_ids,
+             prev_tokens=seq.tokens,
+             prefix_offset=seq.prefix_offset,
+             read_offset=seq.read_offset,
+             skip_special_tokens=prms.skip_special_tokens,
+             spaces_between_special_tokens=prms.spaces_between_special_tokens,
+         )
+
+        # Decode logprobs
+        logprobs = seq.output_logprobs[-1]
+        if logprobs:
+            previous_tokens = all_input_ids[:-1]
+            for token_id, sample_logprob in logprobs.items():
+                # If the token was generated this iteration,
+                # use the provided text.
+                if token_id == token_id_generated_this_iteration:
+                    sample_logprob.decoded_token = new_decoded_token_text
+                    continue
+
+                if (sample_logprob.decoded_token is None
+                        and token_id != VLLM_INVALID_TOKEN_ID):
+                    all_input_ids_with_logprob = previous_tokens + [token_id]
+                    (_, new_text, _, _) = detokenize_incrementally(
+                        tokenizer=tokenizer,
+                        all_input_ids=all_input_ids_with_logprob,
+                        prev_tokens=seq.tokens,
+                        prefix_offset=seq.prefix_offset,
+                        read_offset=seq.read_offset,
+                        skip_special_tokens=prms.skip_special_tokens,
+                        spaces_between_special_tokens=prms.
+                        spaces_between_special_tokens,
+                    )
+                    sample_logprob.decoded_token = new_text
+
+        seq.tokens.extend(new_tokens)
+        seq.prefix_offset = prefix_offset
+        seq.read_offset = read_offset
+        seq.output_text += new_decoded_token_text
+
+        return len(new_decoded_token_text)
\ No newline at end of file

vllm/zero_overhead/utils.py

+
+
+from enum import Enum
+import os
+
+zero_overhead = os.environ.get('VLLM_ZERO_OVERHEAD') == '1'
+zero_no_thread = os.environ.get('VLLM_ZERO_NO_THREAD') == '1'
+
+def is_zero_overhead():
+    return zero_overhead
+
+def is_zero_no_thread():
+    return zero_no_thread and zero_overhead
+
+class SpecStepKind(Enum):
+    KIND_DEFAULT = 0
+    PREFILL = 1
+    FIRST_PROPOSAL = 2
+    OTHER_PROPOSAL = 3
+    SCORE_DECODE = 4
+
+class ZeroOverheadSpecContext():
+    def __init__(self):
+        self.step_kind = SpecStepKind.KIND_DEFAULT
+        self.last_step = SpecStepKind.KIND_DEFAULT
+        self.proposal_lens_list = None
+        self.proposal_token_ids = None
+        self.accepted_token_ids = None
+        self.accepted_seq_ids = None
+
+spec_context = ZeroOverheadSpecContext()  
+
+def set_spec_step(_step):
+    global spec_context
+    spec_context.last_step = spec_context.step_kind
+    spec_context.step_kind = _step
+
+def get_spec_step():
+    return spec_context.step_kind
+
+def get_spec_last_step():
+    return spec_context.last_step
+
+def record_proposal_lens_list(list):
+    global spec_context
+    spec_context.proposal_lens_list = list
+
+def get_proposal_lens_list():
+    return spec_context.proposal_lens_list
+
+def record_proposal_token_ids(tensor):
+    global spec_context
+    spec_context.proposal_token_ids = tensor
+
+def get_proposal_token_ids():
+    return spec_context.proposal_token_ids
+
+def record_accepted_token_ids(tensor, seq_ids):
+    global spec_context
+    spec_context.accepted_token_ids = tensor
+    spec_context.accepted_seq_ids = seq_ids
+
+def get_accepted_token_ids():
+    return spec_context.accepted_token_ids, spec_context.accepted_seq_ids