[Speculative decoding 2/9] Multi-step worker for draft model (#2424)

vllm/engine/llm_engine.py

@@ -18,7 +18,7 @@ from vllm.sequence import (SamplerOutput, Sequence, SequenceGroup,
                            SequenceGroupOutput, SequenceOutput, SequenceStatus)
 from vllm.transformers_utils.tokenizer import (detokenize_incrementally,
                                                get_tokenizer)
-from vllm.utils import Counter, set_cuda_visible_devices, get_ip, get_open_port
+from vllm.utils import Counter, set_cuda_visible_devices, get_ip, get_open_port, get_distributed_init_method
 
 if ray:
     from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
@@ -132,7 +132,8 @@ class LLMEngine:
             "Ray is required if parallel_config.world_size > 1.")
 
         self.workers: List[Worker] = []
-        distributed_init_method = f"tcp://{get_ip()}:{get_open_port()}"
+        distributed_init_method = get_distributed_init_method(
+            get_ip(), get_open_port())
         self.driver_worker = Worker(
             self.model_config,
             self.parallel_config,
@@ -207,7 +208,8 @@ class LLMEngine:
         for worker, (node_id, _) in zip(self.workers, worker_node_and_gpu_ids):
             worker.set_cuda_visible_devices.remote(node_gpus[node_id])
 
-        distributed_init_method = f"tcp://{driver_ip}:{get_open_port()}"
+        distributed_init_method = get_distributed_init_method(
+            driver_ip, get_open_port)
 
         # Lazy import the Worker to avoid importing torch.cuda/xformers
         # before CUDA_VISIBLE_DEVICES is set in the Worker

vllm/engine/ray_utils.py

@@ -65,10 +65,9 @@ def initialize_cluster(
             the default Ray cluster address.
 
     Returns:
-        A tuple of (`distributed_init_method`, `placement_group`). The
-        `distributed_init_method` is the address for initializing the
-        distributed backend. `placement_group` includes the specification
-        of the resources for each distributed worker.
+        An optional `PlacementGroup`. It includes the specification
+        of the resources for each distributed worker. None if Ray is
+        not used.
     """
     if parallel_config.worker_use_ray or engine_use_ray:
         if ray is None:

vllm/model_executor/parallel_utils/parallel_state.py

@@ -83,6 +83,31 @@ def initialize_model_parallel(
             _PIPELINE_GLOBAL_RANKS = ranks
 
 
+def ensure_model_parallel_initialized(
+    tensor_model_parallel_size: int,
+    pipeline_model_parallel_size: int,
+) -> None:
+    """Helper to initialize model parallel groups if they are not initialized,
+    or ensure tensor-parallel and pipeline-parallel sizes are equal to expected
+    values if the model parallel groups are initialized.
+    """
+    if not model_parallel_is_initialized():
+        initialize_model_parallel(tensor_model_parallel_size,
+                                  pipeline_model_parallel_size)
+        return
+
+    assert (
+        get_tensor_model_parallel_world_size() == tensor_model_parallel_size
+    ), ("tensor parallel group already initialized, but of unexpected size: "
+        f"{get_tensor_model_parallel_world_size()=} vs. "
+        f"{tensor_model_parallel_size=}")
+    assert (get_pipeline_model_parallel_world_size(
+    ) == pipeline_model_parallel_size), (
+        "pipeline parallel group already initialized, but of unexpected size: "
+        f"{get_pipeline_model_parallel_world_size()=} vs. "
+        f"{pipeline_model_parallel_size=}")
+
+
 def model_parallel_is_initialized():
     """Check if tensor and pipeline parallel groups are initialized."""
     return (_TENSOR_MODEL_PARALLEL_GROUP is not None

vllm/utils.py

@@ -65,6 +65,10 @@ def get_ip() -> str:
     return s.getsockname()[0]
 
 
+def get_distributed_init_method(ip: str, port: int) -> str:
+    return f"tcp://{ip}:{port}"
+
+
 def get_open_port() -> int:
     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
         s.bind(("", 0))

vllm/worker/model_runner.py

@@ -277,8 +277,8 @@ class ModelRunner:
                     input_block_tables[i, :len(block_table)] = block_table
             block_tables = torch.tensor(input_block_tables, device="cuda")
         else:
-            max_block_table_len = (max_context_len + self.block_size -
-                                   1) // self.block_size
+            max_block_table_len = max(
+                len(block_table) for block_table in block_tables)
             block_tables = _make_tensor_with_pad(
                 block_tables,
                 max_len=max_block_table_len,

vllm/worker/spec_decode/multi_step_worker.py

@@ -0,0 +1,178 @@
+from typing import List, Dict
+import copy
+
+import torch
+
+from vllm.sequence import SamplerOutput, SequenceGroupMetadata
+from vllm.worker.worker import Worker
+
+
+class MultiStepWorker(Worker):
+    """The MultiStepWorker is equivalent to a Worker except that it allows
+    multiple forward passes in a single call, assuming the scheduler has
+    allocated enough space to store the additional KV. This reduces overhead
+    by invoking the scheduler less.
+
+    The MultiStepWorker does not support cache swap operations, or beam search.
+    Cache swap operations do not require large modifications. On the other hand,
+    beam search requires memory allocations during sequence forks and thus
+    requires more thought for MultiStepWorker support.
+    """
+
+    @torch.inference_mode()
+    def execute_model_multi_step(
+        self,
+        seq_group_metadata_list: List[SequenceGroupMetadata],
+        blocks_to_swap_in: Dict[int, int],
+        blocks_to_swap_out: Dict[int, int],
+        blocks_to_copy: Dict[int, List[int]],
+        num_steps: int,
+    ) -> List[SamplerOutput]:
+        """Run the model forward pass num_steps times. Returns the list of
+        sampler output, one per model forward pass.
+        """
+        self._raise_if_unsupported(seq_group_metadata_list, blocks_to_swap_in,
+                                   blocks_to_swap_out, blocks_to_copy)
+
+        # Shallow copy input data so modifications (such as appending tokens)
+        # do not cause side-effects.
+        copied_seq_group_metadata_list = self._shallow_copy_inputs(
+            seq_group_metadata_list)
+
+        # Assert enough KV space for num_steps tokens per sequence.
+        self._assert_enough_kv_space(seq_group_metadata_list, num_steps)
+
+        # Run model num_steps times.
+        model_outputs = []
+        for _ in range(num_steps):
+            model_output = super().execute_model(
+                seq_group_metadata_list=copied_seq_group_metadata_list,
+                blocks_to_swap_in=blocks_to_swap_in,
+                blocks_to_swap_out=blocks_to_swap_out,
+                blocks_to_copy=blocks_to_copy,
+            )
+
+            self._append_new_tokens(model_output,
+                                    copied_seq_group_metadata_list)
+            model_outputs.append(model_output)
+
+        return model_outputs
+
+    def _append_new_tokens(
+            self, model_output: SamplerOutput,
+            seq_group_metadata_list: SequenceGroupMetadata) -> None:
+        """Given model output from a single run, append the tokens to the
+        sequences. This is normally done outside of the worker, but it is
+        required if the worker is to perform multiple forward passes.
+        """
+        for seq_group_metadata, sequence_group_outputs in zip(
+                seq_group_metadata_list, model_output):
+            seq_group_metadata.is_prompt = False
+
+            for seq_output in sequence_group_outputs.samples:
+                # NOTE: Beam search is not supported, so we can assume that
+                # parent_seq_id == seq_id.
+                seq = seq_group_metadata.seq_data[seq_output.parent_seq_id]
+
+                token_id = seq_output.output_token
+                token_logprob = seq_output.logprobs[token_id]
+
+                seq.append_token_id(token_id, token_logprob)
+
+    def _shallow_copy_inputs(
+        self, seq_group_metadata_list: List[SequenceGroupMetadata]
+    ) -> List[SequenceGroupMetadata]:
+        """Copy input data structures to remove side-effects when input data
+        structures are shared with other modules.
+
+        The multi-step worker must be able to append tokens to sequences after
+        a forward pass. This necessitates modification of the data structures
+        used by the worker. Since these data structures are shared with other
+        parts of vLLM, like the scheduler, we must take care not to introduce
+        unexpected side-effects.
+
+        When Ray is used to orchestrate worker processes (such as when the
+        tensor-parallel degree is >1), this is not a problem because the input
+        datastructures will be serialized and created anew in the worker
+        process.
+
+        However, when Ray is not used to orchestrate the worker processes (such
+        as when the tensor-parallel degree is 1), this is a problem. We avoid
+        the problem by shallow-copying the input datastructures (specifically,
+        the parts that will change in multiple steps).
+        """
+
+        # Shallow-copy the list of SequenceGroupMetadata. This allows us to
+        # append tokens and change is_prompt without external side-effects.
+        new_seq_group_metadata_list = []
+
+        for old_seq_group_metadata in seq_group_metadata_list:
+            # We must shallow-copy seq_group_metadata as is_prompt could change.
+            seq_group_metadata = copy.copy(old_seq_group_metadata)
+            new_seq_group_metadata_list.append(seq_group_metadata)
+
+            # We must shallow-copy seq_data as we will append token ids
+            new_seq_data = {}
+            for seq_id, old_seq_data in seq_group_metadata.seq_data.items():
+                new_seq_data[seq_id] = copy.copy(old_seq_data)
+                new_seq_data[
+                    seq_id].output_token_ids = old_seq_data.output_token_ids[:]
+
+            seq_group_metadata.seq_data = new_seq_data
+
+        return new_seq_group_metadata_list
+
+    def _assert_enough_kv_space(
+            self, seq_group_metadata_list: List[SequenceGroupMetadata],
+            num_steps: int) -> None:
+        """Assert there are enough physical blocks per sequence to store the
+        current KV plus additional KV from num_steps tokens.
+        """
+        assert self.model_runner.block_size is not None
+        for seq_group_metadata in seq_group_metadata_list:
+            # Only one seq_id is guaranteed because there is no beam search.
+            seq_id = list(seq_group_metadata.seq_data.keys())[0]
+            seq = seq_group_metadata.seq_data[seq_id]
+
+            # After num_steps, the seq len will be the current seq len
+            # plus one token per step.
+            final_seq_len = seq.get_len() + num_steps
+
+            # We will have final_seq_len - 1 KV because vLLM saves KV for a
+            # token in the iteration after the token was generated.
+            required_num_kv_slots = final_seq_len - 1
+
+            # The allocated number of kv slots is the number of allocated blocks
+            # times the number of slots of block.
+            number_physical_blocks = len(
+                seq_group_metadata.block_tables[seq_id])
+            allocated_kv_slots = (number_physical_blocks *
+                                  self.model_runner.block_size)
+
+            if required_num_kv_slots > allocated_kv_slots:
+                request_id = seq_group_metadata.request_id
+                raise ValueError(
+                    "The worker attempted to run "
+                    f"{num_steps} times but found insufficient KV space for "
+                    f"{request_id=} {seq_id=}. ({allocated_kv_slots=} "
+                    f"{required_num_kv_slots=}).")
+
+    def _raise_if_unsupported(
+        self,
+        seq_group_metadata_list: List[SequenceGroupMetadata],
+        blocks_to_swap_in: Dict[int, int],
+        blocks_to_swap_out: Dict[int, int],
+        blocks_to_copy: Dict[int, List[int]],
+    ) -> None:
+        """MultiStepWorker does not yet implement support for cache swap
+        operations or beam search.
+        """
+        if any([blocks_to_swap_in, blocks_to_swap_out, blocks_to_copy]):
+            raise NotImplementedError(
+                "MultiStepWorker does not support cache operations")
+
+        if any(
+                len(seq_group_metadata.seq_data.keys()) != 1
+                for seq_group_metadata in seq_group_metadata_list):
+            raise NotImplementedError(
+                "MultiStepWorker does not support beam search.")

vllm/worker/worker.py

@@ -11,7 +11,7 @@ from vllm.model_executor import set_random_seed
 from vllm.model_executor.parallel_utils.communication_op import (
     broadcast_tensor_dict)
 from vllm.model_executor.parallel_utils.parallel_state import (
-    initialize_model_parallel)
+    ensure_model_parallel_initialized)
 from vllm.sequence import SamplerOutput, SequenceGroupMetadata
 from vllm.worker.cache_engine import CacheEngine
 from vllm.worker.model_runner import ModelRunner
@@ -227,8 +227,8 @@ def _init_distributed_environment(
 
     # A small all_reduce for warmup.
     torch.distributed.all_reduce(torch.zeros(1).cuda())
-    initialize_model_parallel(parallel_config.tensor_parallel_size,
-                              parallel_config.pipeline_parallel_size)
+    ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
+                                      parallel_config.pipeline_parallel_size)
 
 
 def _check_if_gpu_supports_dtype(torch_dtype: torch.dtype):