vllm-cpu 0.8.5.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/v1/engine/detokenizer.py

@@ -0,0 +1,260 @@
+# SPDX-License-Identifier: Apache-2.0
+from abc import ABC, abstractmethod
+from typing import Optional
+
+import tokenizers
+from packaging import version
+from tokenizers import Tokenizer
+from tokenizers.decoders import DecodeStream
+from transformers import PreTrainedTokenizerFast
+
+from vllm.engine.output_processor.stop_checker import StopChecker
+from vllm.logger import init_logger
+from vllm.transformers_utils.detokenizer_utils import (
+    AnyTokenizer, convert_prompt_ids_to_tokens, detokenize_incrementally)
+from vllm.v1.engine import EngineCoreRequest
+
+logger = init_logger(__name__)
+
+
+class IncrementalDetokenizer:
+
+    def __init__(self):
+        self.token_ids: list[int] = []
+
+    @property
+    def output_token_ids(self) -> list[int]:
+        return self.token_ids
+
+    def update(self, new_token_ids: list[int],
+               stop_terminated: bool) -> Optional[str]:
+        self.token_ids.extend(new_token_ids)
+        return None
+
+    def get_next_output_text(self, finished: bool, delta: bool) -> str:
+        return ""
+
+    @classmethod
+    def from_new_request(
+        cls,
+        tokenizer: Optional[AnyTokenizer],
+        request: EngineCoreRequest,
+    ) -> "IncrementalDetokenizer":
+
+        if tokenizer is None:
+            # No tokenizer => skipping detokenization.
+            return IncrementalDetokenizer()
+
+        if (isinstance(tokenizer, PreTrainedTokenizerFast) and version.parse(
+                tokenizers.__version__) >= version.parse("0.21.1")):
+            # Fast tokenizer => use tokenizers library DecodeStream.
+            # And only tokenizers >= 0.21.1 supports Fast Detokenizer.
+            return FastIncrementalDetokenizer(tokenizer, request)
+
+        # Fall back to slow python-based incremental detokenization.
+        return SlowIncrementalDetokenizer(tokenizer, request)
+
+
+class BaseIncrementalDetokenizer(IncrementalDetokenizer, ABC):
+
+    def __init__(self, request: EngineCoreRequest):
+        super().__init__()
+
+        # Stop strings
+        params = request.sampling_params
+        self.stop = stop = params.stop
+        self.include_stop_str_in_output = params.include_stop_str_in_output
+
+        # Number of chars to hold back when stop strings are to be excluded
+        # from streamed output.
+        if stop and not self.include_stop_str_in_output:
+            self.stop_buffer_length = max(len(s) for s in stop) - 1
+        else:
+            self.stop_buffer_length = 0
+        self._last_output_text_offset: int = 0
+
+        # Generation data
+        self.output_text = ""
+
+    def update(self, new_token_ids: list[int],
+               stop_terminated: bool) -> Optional[str]:
+        """
+        Update RequestState for the request_id by:
+            1) Detokenize the new token ids incrementally.
+            2) Evaluate stop criteria.
+
+        Return matched stop string or None.
+        """
+        if not new_token_ids:
+            # Skip detokenization if no new token ids.
+            return None
+
+        if stop_terminated and not self.include_stop_str_in_output:
+            # If stop-terminated, exclude last token from detokenization
+            # based on include_stop_str_in_output parameter.
+            skipped_stop_token_id = new_token_ids[-1]
+            new_token_ids = new_token_ids[:-1]
+        else:
+            skipped_stop_token_id = None
+
+        # 1) Detokenize the new token ids incrementally.
+        # TODO(woosuk): This method becomes very inefficient when the number of
+        # new_token_ids is more than 1. We need to optimize this.
+        offset_before = len(self.output_text)
+        for new_token_id in new_token_ids:
+            self.token_ids.append(new_token_id)
+            self.output_text += self.decode_next(new_token_id)
+
+        if stop_terminated:
+            if skipped_stop_token_id is not None:
+                # Cleanup after skipping detokenization.
+                self.token_ids.append(skipped_stop_token_id)
+            # Stop token triggered; skip stop string check.
+            return None
+
+        # 2) Evaluate stop strings.
+        stop_string = None
+        if self.stop:
+            stop = StopChecker.check_stop_strings(
+                output_text=self.output_text,
+                new_char_count=len(self.output_text) - offset_before,
+                stop=self.stop,
+                include_in_output=self.include_stop_str_in_output,
+            )
+            if stop is not None:
+                stop_string, truncate_to = stop
+                if truncate_to != -1:
+                    self.output_text = self.output_text[:truncate_to]
+
+        return stop_string
+
+    @abstractmethod
+    def decode_next(self, next_token_id: int) -> str:
+        raise NotImplementedError
+
+    def get_next_output_text(self, finished: bool, delta: bool) -> str:
+        """If delta is True, only new text since the last call to
+        this method is returned"""
+
+        # We return the full output text if the sequence is finished.
+        buffer_length = 0 if finished else self.stop_buffer_length
+        if not delta:
+            return self.output_text[:-buffer_length] if buffer_length else (
+                self.output_text)
+        length = len(self.output_text) - buffer_length
+        last_offset = self._last_output_text_offset
+        if last_offset < length:
+            self._last_output_text_offset = length
+            return self.output_text[last_offset:length]
+        return ""
+
+
+class FastIncrementalDetokenizer(BaseIncrementalDetokenizer):
+
+    def __init__(self, tokenizer: PreTrainedTokenizerFast,
+                 request: EngineCoreRequest):
+        super().__init__(request)
+
+        sampling_params = request.sampling_params
+        self.stream = DecodeStream(
+            skip_special_tokens=sampling_params.skip_special_tokens)
+
+        self.tokenizer: Tokenizer = tokenizer._tokenizer
+
+        # Find a safe place to start.
+        prompt_suffix = request.prompt_token_ids
+        prompt_len = len(prompt_suffix)
+        if prompt_len > 4:
+            for i in range(4, min(prompt_len + 1, 24)):
+                suffix = request.prompt_token_ids[-i:]
+                if '�' not in self.tokenizer.decode(suffix):
+                    prompt_suffix = suffix
+                    break
+
+        # Prime the stream.
+        for tid in prompt_suffix:
+            self.stream.step(self.tokenizer, tid)
+
+        self.spaces_between_special_tokens = (
+            sampling_params.skip_special_tokens
+            or sampling_params.spaces_between_special_tokens)
+
+        if not self.spaces_between_special_tokens:
+            # Store dict of added token ids so that we can suppress
+            # the spaces between them.
+            if (added_token_ids := getattr(self.tokenizer, "added_token_ids",
+                                           None)) is None:
+                self.tokenizer.added_token_ids = added_token_ids = {
+                    tid: tok.content
+                    for tid, tok in
+                    self.tokenizer.get_added_tokens_decoder().items()
+                }
+
+            if added_token_ids:
+                self.last_special = False
+                self.added_token_ids = added_token_ids
+            else:
+                # No added tokens.
+                self.spaces_between_special_tokens = True
+
+    def decode_next(self, next_token_id: int) -> str:
+        token = self.stream.step(self.tokenizer, next_token_id)
+
+        if not self.spaces_between_special_tokens:
+            special_token = self.added_token_ids.get(next_token_id)
+            is_special = special_token is not None
+            if is_special and self.last_special:
+                # Return raw token string without any prefixed spaces.
+                token = special_token
+            self.last_special = is_special
+
+        return token or ""
+
+
+class SlowIncrementalDetokenizer(BaseIncrementalDetokenizer):
+
+    def __init__(self, tokenizer: AnyTokenizer, request: EngineCoreRequest):
+        super().__init__(request)
+
+        self.tokenizer = tokenizer
+
+        # Metadata for incremental detokenization.
+        self.tokens, self.prefix_offset, self.read_offset = (
+            convert_prompt_ids_to_tokens(
+                tokenizer=tokenizer,
+                prompt_ids=request.prompt_token_ids,
+                skip_special_tokens=request.sampling_params.
+                skip_special_tokens,
+            ))
+
+        self.token_ids.extend(request.prompt_token_ids)
+        self.prompt_len = len(request.prompt_token_ids)
+
+        params = request.sampling_params
+        self.skip_special_tokens = params.skip_special_tokens
+        self.spaces_between_special_tokens = (
+            params.spaces_between_special_tokens)
+
+    @property
+    def output_token_ids(self) -> list[int]:
+        return self.token_ids if not self.prompt_len else (
+            self.token_ids[self.prompt_len:])
+
+    def decode_next(self, next_token_id: int) -> str:
+        new_tokens, decoded_text, prefix_offset, read_offset = (
+            detokenize_incrementally(
+                tokenizer=self.tokenizer,
+                all_input_ids=self.token_ids,
+                prev_tokens=self.tokens,
+                prefix_offset=self.prefix_offset,
+                read_offset=self.read_offset,
+                skip_special_tokens=self.skip_special_tokens,
+                spaces_between_special_tokens=self.
+                spaces_between_special_tokens,
+            ))
+
+        self.tokens.extend(new_tokens)
+        self.prefix_offset = prefix_offset
+        self.read_offset = read_offset
+
+        return decoded_text

vllm/v1/engine/exceptions.py

@@ -0,0 +1,16 @@
+# SPDX-License-Identifier: Apache-2.0
+class EngineGenerateError(Exception):
+    """Raised when a AsyncLLM.generate() fails. Recoverable."""
+    pass
+
+
+class EngineDeadError(Exception):
+    """Raised when the EngineCore dies. Unrecoverable."""
+
+    def __init__(self, *args, suppress_context: bool = False, **kwargs):
+        ENGINE_DEAD_MESSAGE = "EngineCore encountered an issue. See stack trace (above) for the root cause."  # noqa: E501
+
+        super().__init__(ENGINE_DEAD_MESSAGE, *args, **kwargs)
+        # Make stack trace clearer when using with LLMEngine by
+        # silencing irrelevant ZMQError.
+        self.__suppress_context__ = suppress_context

vllm/v1/engine/llm_engine.py

@@ -0,0 +1,285 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from collections.abc import Mapping
+from copy import copy
+from typing import Any, Callable, Optional, Union
+
+from typing_extensions import TypeVar
+
+import vllm.envs as envs
+from vllm.config import ParallelConfig, VllmConfig
+from vllm.distributed import stateless_destroy_torch_distributed_process_group
+from vllm.engine.arg_utils import EngineArgs
+from vllm.inputs import PromptType
+from vllm.logger import init_logger
+from vllm.lora.request import LoRARequest
+from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
+from vllm.outputs import RequestOutput
+from vllm.pooling_params import PoolingParams
+from vllm.prompt_adapter.request import PromptAdapterRequest
+from vllm.sampling_params import SamplingParams
+from vllm.transformers_utils.tokenizer_group import (
+    TokenizerGroup, init_tokenizer_from_configs)
+from vllm.usage.usage_lib import UsageContext
+from vllm.utils import Device
+from vllm.v1.engine.core_client import EngineCoreClient
+from vllm.v1.engine.output_processor import OutputProcessor
+from vllm.v1.engine.parallel_sampling import ParentRequest
+from vllm.v1.engine.processor import Processor
+from vllm.v1.executor.abstract import Executor
+from vllm.v1.metrics.loggers import StatLoggerFactory
+
+logger = init_logger(__name__)
+
+_R = TypeVar("_R", default=Any)
+
+
+class LLMEngine:
+    """Legacy LLMEngine for backwards compatibility."""
+
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        executor_class: type[Executor],
+        log_stats: bool,
+        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
+        stat_loggers: Optional[list[StatLoggerFactory]] = None,
+        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
+        use_cached_outputs: bool = False,
+        multiprocess_mode: bool = False,
+    ) -> None:
+        if not envs.VLLM_USE_V1:
+            raise ValueError(
+                "Using V1 LLMEngine, but envs.VLLM_USE_V1=False. "
+                "This should not happen. As a workaround, try using "
+                "LLMEngine.from_vllm_config(...) or explicitly set "
+                "VLLM_USE_V1=0 or 1 and report this issue on Github.")
+
+        if stat_loggers is not None:
+            raise NotImplementedError(
+                "Passing StatLoggers to LLMEngine in V1 is not yet supported. "
+                "Set VLLM_USE_V1=0 and file and issue on Github.")
+
+        self.vllm_config = vllm_config
+        self.model_config = vllm_config.model_config
+        self.cache_config = vllm_config.cache_config
+
+        # important: init dp group before init the engine_core
+        # In the decoupled engine case this is handled in EngineCoreProc.
+        parallel_config = vllm_config.parallel_config
+        if not multiprocess_mode and parallel_config.data_parallel_size > 1:
+            self.dp_group = parallel_config.stateless_init_dp_group()
+        else:
+            self.dp_group = None
+        self.should_execute_dummy_batch = False
+
+        # Tokenizer (+ ensure liveness if running in another process).
+        self.tokenizer = init_tokenizer_from_configs(
+            model_config=vllm_config.model_config,
+            scheduler_config=vllm_config.scheduler_config,
+            lora_config=vllm_config.lora_config)
+
+        # Processor (convert Inputs --> EngineCoreRequests)
+        self.processor = Processor(vllm_config=vllm_config,
+                                   tokenizer=self.tokenizer,
+                                   mm_registry=mm_registry)
+
+        # OutputProcessor (convert EngineCoreOutputs --> RequestOutput).
+        self.output_processor = OutputProcessor(self.tokenizer,
+                                                log_stats=False)
+
+        # EngineCore (gets EngineCoreRequests and gives EngineCoreOutputs)
+        self.engine_core = EngineCoreClient.make_client(
+            multiprocess_mode=multiprocess_mode,
+            asyncio_mode=False,
+            vllm_config=vllm_config,
+            executor_class=executor_class,
+            log_stats=False,  # FIXME: implement
+        )
+
+        if not multiprocess_mode:
+            # for v0 compatibility
+            self.model_executor = self.engine_core.engine_core.model_executor  # type: ignore
+
+    @classmethod
+    def from_vllm_config(
+        cls,
+        vllm_config: VllmConfig,
+        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
+        stat_loggers: Optional[list[StatLoggerFactory]] = None,
+        disable_log_stats: bool = False,
+    ) -> "LLMEngine":
+        return cls(vllm_config=vllm_config,
+                   executor_class=Executor.get_class(vllm_config),
+                   log_stats=(not disable_log_stats),
+                   usage_context=usage_context,
+                   stat_loggers=stat_loggers,
+                   multiprocess_mode=envs.VLLM_ENABLE_V1_MULTIPROCESSING)
+
+    @classmethod
+    def from_engine_args(
+        cls,
+        engine_args: EngineArgs,
+        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
+        stat_loggers: Optional[list[StatLoggerFactory]] = None,
+        enable_multiprocessing: bool = False,
+    ) -> "LLMEngine":
+        """Creates an LLM engine from the engine arguments."""
+
+        # Create the engine configs.
+        vllm_config = engine_args.create_engine_config(usage_context)
+        executor_class = Executor.get_class(vllm_config)
+
+        if envs.VLLM_ENABLE_V1_MULTIPROCESSING:
+            logger.debug("Enabling multiprocessing for LLMEngine.")
+            enable_multiprocessing = True
+
+        # Create the LLMEngine.
+        return cls(vllm_config=vllm_config,
+                   executor_class=executor_class,
+                   log_stats=not engine_args.disable_log_stats,
+                   usage_context=usage_context,
+                   stat_loggers=stat_loggers,
+                   multiprocess_mode=enable_multiprocessing)
+
+    def get_num_unfinished_requests(self) -> int:
+        return self.output_processor.get_num_unfinished_requests()
+
+    def has_unfinished_requests(self) -> bool:
+        has_unfinished = self.output_processor.has_unfinished_requests()
+        if self.dp_group is None:
+            return has_unfinished
+        return self.has_unfinished_requests_dp(has_unfinished)
+
+    def has_unfinished_requests_dp(self, has_unfinished: bool) -> bool:
+        aggregated_has_unfinished = ParallelConfig.has_unfinished_dp(
+            self.dp_group, has_unfinished)
+        if not has_unfinished and aggregated_has_unfinished:
+            self.should_execute_dummy_batch = True
+        return aggregated_has_unfinished
+
+    @classmethod
+    def validate_outputs(cls, outputs, output_type):
+        return outputs
+
+    def abort_request(self, request_ids: list[str]) -> None:
+        """Remove request_ids from EngineCore and Detokenizer."""
+
+        request_ids = self.output_processor.abort_requests(request_ids)
+        self.engine_core.abort_requests(request_ids)
+
+    def add_request(
+        self,
+        request_id: str,
+        prompt: PromptType,
+        params: Union[SamplingParams, PoolingParams],
+        arrival_time: Optional[float] = None,
+        lora_request: Optional[LoRARequest] = None,
+        trace_headers: Optional[Mapping[str, str]] = None,
+        prompt_adapter_request: Optional[PromptAdapterRequest] = None,
+        priority: int = 0,
+    ) -> None:
+        # Process raw inputs into the request.
+        prompt_str, request = self.processor.process_inputs(
+            request_id, prompt, params, arrival_time, lora_request,
+            trace_headers, prompt_adapter_request, priority)
+
+        n = params.n if isinstance(params, SamplingParams) else 1
+
+        if n == 1:
+            # Make a new RequestState and queue.
+            self.output_processor.add_request(request, prompt_str, None, 0)
+            # Add the request to EngineCore.
+            self.engine_core.add_request(request)
+            return
+
+        # Fan out child requests (for n>1).
+        parent_req = ParentRequest(request_id, params)
+        for idx in range(n):
+            request_id, params = parent_req.get_child_info(idx)
+            child_request = request if idx == n - 1 else copy(request)
+            child_request.request_id = request_id
+            child_request.sampling_params = params
+
+            # Make a new RequestState and queue.
+            self.output_processor.add_request(child_request, prompt_str,
+                                              parent_req, idx)
+            # Add the request to EngineCore.
+            self.engine_core.add_request(child_request)
+
+    def step(self) -> list[RequestOutput]:
+
+        if self.should_execute_dummy_batch:
+            self.should_execute_dummy_batch = False
+            self.engine_core.execute_dummy_batch()
+            return []
+
+        # 1) Get EngineCoreOutput from the EngineCore.
+        outputs = self.engine_core.get_output()
+
+        # 2) Process EngineCoreOutputs.
+        processed_outputs = self.output_processor.process_outputs(
+            outputs.outputs)
+
+        # 3) Abort any reqs that finished due to stop strings.
+        self.engine_core.abort_requests(processed_outputs.reqs_to_abort)
+
+        return processed_outputs.request_outputs
+
+    def get_vllm_config(self):
+        return self.vllm_config
+
+    def get_model_config(self):
+        return self.model_config
+
+    def start_profile(self):
+        self.engine_core.profile(True)
+
+    def stop_profile(self):
+        self.engine_core.profile(False)
+
+    def reset_prefix_cache(self, device: Optional[Device] = None):
+        self.engine_core.reset_prefix_cache()
+
+    def sleep(self, level: int = 1):
+        self.engine_core.sleep(level)
+
+    def wake_up(self, tags: Optional[list[str]] = None):
+        self.engine_core.wake_up(tags)
+
+    def is_sleeping(self) -> bool:
+        return self.engine_core.is_sleeping()
+
+    def get_tokenizer_group(self) -> TokenizerGroup:
+        if self.tokenizer is None:
+            raise ValueError("Unable to get tokenizer because "
+                             "skip_tokenizer_init is True")
+
+        return self.tokenizer
+
+    def add_lora(self, lora_request: LoRARequest) -> bool:
+        """Load a new LoRA adapter into the engine for future requests."""
+        return self.engine_core.add_lora(lora_request)
+
+    def remove_lora(self, lora_id: int) -> bool:
+        """Remove an already loaded LoRA adapter."""
+        return self.engine_core.remove_lora(lora_id)
+
+    def list_loras(self) -> set[int]:
+        """List all registered adapters."""
+        return self.engine_core.list_loras()
+
+    def pin_lora(self, lora_id: int) -> bool:
+        """Prevent an adapter from being evicted."""
+        return self.engine_core.pin_lora(lora_id)
+
+    def collective_rpc(self,
+                       method: Union[str, Callable[..., _R]],
+                       timeout: Optional[float] = None,
+                       args: tuple = (),
+                       kwargs: Optional[dict[str, Any]] = None) -> list[_R]:
+        return self.engine_core.collective_rpc(method, timeout, args, kwargs)
+
+    def __del__(self):
+        if dp_group := getattr(self, "dp_group", None):
+            stateless_destroy_torch_distributed_process_group(dp_group)