vllm-cpu-avx512vnni 0.10.2.post2__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/inputs/registry.py

@@ -0,0 +1,251 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from collections.abc import Mapping
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, NamedTuple, Optional, Union
+
+import torch
+from transformers import BatchFeature, PretrainedConfig, ProcessorMixin
+from typing_extensions import TypeVar
+
+from vllm.logger import init_logger
+from vllm.transformers_utils.processor import cached_processor_from_config
+from vllm.utils import get_allowed_kwarg_only_overrides
+from vllm.utils.jsontree import JSONTree, json_map_leaves
+
+if TYPE_CHECKING:
+    from vllm.config import ModelConfig
+    from vllm.multimodal import (MultiModalDataDict, MultiModalPlaceholderDict,
+                                 MultiModalRegistry)
+    from vllm.sequence import SequenceData
+    from vllm.transformers_utils.tokenizer import AnyTokenizer
+else:
+    ModelConfig = Any
+    MultiModalDataDict = Any
+    MultiModalPlaceholderDict = Any
+    MultiModalRegistry = Any
+    SequenceData = Any
+    AnyTokenizer = Any
+
+_T = TypeVar("_T")
+_C = TypeVar("_C", bound=PretrainedConfig, default=PretrainedConfig)
+_P = TypeVar("_P", bound=ProcessorMixin, default=ProcessorMixin)
+
+logger = init_logger(__name__)
+
+
+@dataclass(frozen=True)
+class InputContext:
+    """
+    Contains information about the model which may be used to
+    modify the inputs.
+    """
+
+    model_config: ModelConfig
+    """The configuration of the model."""
+
+    def get_hf_config(
+        self,
+        typ: Union[type[_C], tuple[type[_C], ...]] = PretrainedConfig,
+        /,
+    ) -> _C:
+        """
+        Get the HuggingFace configuration
+        (`transformers.PretrainedConfig`) of the model,
+        additionally checking its type.
+
+        Raises:
+            TypeError: If the configuration is not of the specified type.
+        """
+        hf_config = self.model_config.hf_config
+        if not isinstance(hf_config, typ):
+            raise TypeError("Invalid type of HuggingFace config. "
+                            f"Expected type: {typ}, but "
+                            f"found type: {type(hf_config)}")
+
+        return hf_config
+
+    def get_hf_image_processor_config(self) -> dict[str, Any]:
+        """
+        Get the HuggingFace image processor configuration of the model.
+        """
+        return self.model_config.hf_image_processor_config
+
+    def get_mm_config(self):
+        """
+        Get the multimodal config of the model.
+
+        Raises:
+            RuntimeError: If the model is not a multimodal model.
+        """
+        mm_config = self.model_config.multimodal_config
+        if mm_config is None:
+            raise RuntimeError("Not a multimodal model")
+
+        return mm_config
+
+    def get_hf_processor(
+        self,
+        typ: Union[type[_P], tuple[type[_P], ...]] = ProcessorMixin,
+        /,
+        **kwargs: object,
+    ) -> _P:
+        """
+        Get the HuggingFace processor
+        (`transformers.ProcessorMixin`) of the model,
+        additionally checking its type.
+
+        Raises:
+            TypeError: If the processor is not of the specified type.
+        """
+        return cached_processor_from_config(
+            self.model_config,
+            processor_cls=typ,
+            **kwargs,
+        )
+
+    def init_processor(
+        self,
+        typ: type[_T],
+        /,
+        **kwargs: object,
+    ) -> _T:
+        """
+        Initialize a HuggingFace-like processor class, merging the
+        keyword arguments with those in the model's configuration.
+        """
+        mm_config = self.model_config.get_multimodal_config()
+        base_kwargs = mm_config.mm_processor_kwargs
+        if base_kwargs is None:
+            base_kwargs = {}
+
+        merged_kwargs = {**base_kwargs, **kwargs}
+
+        return typ(**merged_kwargs)
+
+
+@dataclass(frozen=True)
+class InputProcessingContext(InputContext):
+    tokenizer: AnyTokenizer
+    """The tokenizer used to tokenize the inputs."""
+
+    def get_hf_processor(
+        self,
+        typ: Union[type[_P], tuple[type[_P], ...]] = ProcessorMixin,
+        /,
+        **kwargs: object,
+    ) -> _P:
+        return super().get_hf_processor(
+            typ,
+            tokenizer=self.tokenizer,
+            **kwargs,
+        )
+
+    def call_hf_processor(
+        self,
+        hf_processor: ProcessorMixin,
+        data: Mapping[str, object],
+        kwargs: Mapping[str, object] = {},
+    ) -> Union[BatchFeature, JSONTree]:
+        """
+        Call `hf_processor` on the prompt `data`
+        (text, image, audio...) with configurable options `kwargs`.
+        """
+        assert callable(hf_processor)
+
+        mm_config = self.model_config.get_multimodal_config()
+        merged_kwargs = mm_config.merge_mm_processor_kwargs(kwargs)
+
+        allowed_kwargs = get_allowed_kwarg_only_overrides(
+            hf_processor,
+            merged_kwargs,
+            requires_kw_only=False,
+            allow_var_kwargs=True,
+        )
+
+        def maybe_cast_dtype(x):
+            # This mimics the behavior of transformers.BatchFeature
+            if isinstance(x, torch.Tensor) and x.is_floating_point():
+                return x.to(dtype=self.model_config.dtype)
+            return x
+
+        try:
+            output = hf_processor(**data,
+                                  **allowed_kwargs,
+                                  return_tensors="pt")
+            # this emulates output.to(dtype=self.model_config.dtype)
+            if isinstance(output, BatchFeature):
+                cast_output = json_map_leaves(maybe_cast_dtype, output.data)
+                return BatchFeature(cast_output)
+
+            cast_output = json_map_leaves(maybe_cast_dtype, output)
+
+            logger.warning_once(
+                f"{type(hf_processor).__name__} did not return `BatchFeature`. "
+                "Make sure to match the behaviour of `ProcessorMixin` when "
+                "implementing custom processors.")
+            return cast_output
+
+        except Exception as exc:
+            msg = (f"Failed to apply {type(hf_processor).__name__} "
+                   f"on data={data} with kwargs={allowed_kwargs}")
+
+            raise ValueError(msg) from exc
+
+
+class DummyData(NamedTuple):
+    """
+    Dummy data used for profiling.
+
+    Note: This is only used in V0.
+    """
+
+    seq_data: SequenceData
+    multi_modal_data: Optional[MultiModalDataDict] = None
+    multi_modal_placeholders: Optional[MultiModalPlaceholderDict] = None
+
+
+class InputRegistry:
+    """
+    Note: This is only used in V0.
+    """
+
+    def dummy_data_for_profiling(
+        self,
+        model_config: ModelConfig,
+        seq_len: int,
+        mm_registry: MultiModalRegistry,
+        is_encoder_data: bool = False,
+    ) -> DummyData:
+        """
+        Create dummy data for profiling the memory usage of a model.
+
+        The model is identified by ``model_config``.
+        """
+        # Avoid circular import
+        from vllm.multimodal.cache import processor_only_cache_from_config
+        from vllm.sequence import SequenceData
+
+        if not model_config.is_multimodal_model:
+            seq_data = SequenceData.from_prompt_token_counts((0, seq_len))
+            return DummyData(seq_data=seq_data)
+
+        cache = processor_only_cache_from_config(model_config, mm_registry)
+
+        # Encoder dummy data does not contain multi-modal data
+        if is_encoder_data:
+            enc_data = mm_registry.get_encoder_dummy_data(model_config,
+                                                          seq_len,
+                                                          cache=cache)
+            seq_data = SequenceData.from_seqs(enc_data.prompt_token_ids)
+            return DummyData(seq_data=seq_data)
+
+        dec_data = mm_registry.get_decoder_dummy_data(model_config,
+                                                      seq_len,
+                                                      cache=cache)
+
+        return DummyData(
+            seq_data=SequenceData.from_seqs(dec_data.prompt_token_ids),
+            multi_modal_data=dec_data.multi_modal_data.get_data(),
+            multi_modal_placeholders=dec_data.multi_modal_placeholders,
+        )

vllm/logger.py

@@ -0,0 +1,229 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Logging configuration for vLLM."""
+import datetime
+import json
+import logging
+import os
+import sys
+from collections.abc import Hashable
+from functools import lru_cache, partial
+from logging import Logger
+from logging.config import dictConfig
+from os import path
+from types import MethodType
+from typing import Any, Optional, cast
+
+import vllm.envs as envs
+
+VLLM_CONFIGURE_LOGGING = envs.VLLM_CONFIGURE_LOGGING
+VLLM_LOGGING_CONFIG_PATH = envs.VLLM_LOGGING_CONFIG_PATH
+VLLM_LOGGING_LEVEL = envs.VLLM_LOGGING_LEVEL
+VLLM_LOGGING_PREFIX = envs.VLLM_LOGGING_PREFIX
+VLLM_LOGGING_STREAM = envs.VLLM_LOGGING_STREAM
+
+_FORMAT = (f"{VLLM_LOGGING_PREFIX}%(levelname)s %(asctime)s "
+           "[%(fileinfo)s:%(lineno)d] %(message)s")
+_DATE_FORMAT = "%m-%d %H:%M:%S"
+
+DEFAULT_LOGGING_CONFIG = {
+    "formatters": {
+        "vllm": {
+            "class": "vllm.logging_utils.NewLineFormatter",
+            "datefmt": _DATE_FORMAT,
+            "format": _FORMAT,
+        },
+    },
+    "handlers": {
+        "vllm": {
+            "class": "logging.StreamHandler",
+            "formatter": "vllm",
+            "level": VLLM_LOGGING_LEVEL,
+            "stream": VLLM_LOGGING_STREAM,
+        },
+    },
+    "loggers": {
+        "vllm": {
+            "handlers": ["vllm"],
+            "level": "DEBUG",
+            "propagate": False,
+        },
+    },
+    "version": 1,
+    "disable_existing_loggers": False
+}
+
+
+@lru_cache
+def _print_debug_once(logger: Logger, msg: str, *args: Hashable) -> None:
+    # Set the stacklevel to 2 to print the original caller's line info
+    logger.debug(msg, *args, stacklevel=2)
+
+
+@lru_cache
+def _print_info_once(logger: Logger, msg: str, *args: Hashable) -> None:
+    # Set the stacklevel to 2 to print the original caller's line info
+    logger.info(msg, *args, stacklevel=2)
+
+
+@lru_cache
+def _print_warning_once(logger: Logger, msg: str, *args: Hashable) -> None:
+    # Set the stacklevel to 2 to print the original caller's line info
+    logger.warning(msg, *args, stacklevel=2)
+
+
+class _VllmLogger(Logger):
+    """
+    Note:
+        This class is just to provide type information.
+        We actually patch the methods directly on the [`logging.Logger`][]
+        instance to avoid conflicting with other libraries such as
+        `intel_extension_for_pytorch.utils._logger`.
+    """
+
+    def debug_once(self, msg: str, *args: Hashable) -> None:
+        """
+        As [`debug`][logging.Logger.debug], but subsequent calls with
+        the same message are silently dropped.
+        """
+        _print_debug_once(self, msg, *args)
+
+    def info_once(self, msg: str, *args: Hashable) -> None:
+        """
+        As [`info`][logging.Logger.info], but subsequent calls with
+        the same message are silently dropped.
+        """
+        _print_info_once(self, msg, *args)
+
+    def warning_once(self, msg: str, *args: Hashable) -> None:
+        """
+        As [`warning`][logging.Logger.warning], but subsequent calls with
+        the same message are silently dropped.
+        """
+        _print_warning_once(self, msg, *args)
+
+
+# Pre-defined methods mapping to avoid repeated dictionary creation
+_METHODS_TO_PATCH = {
+    "debug_once": _print_debug_once,
+    "info_once": _print_info_once,
+    "warning_once": _print_warning_once,
+}
+
+
+def _configure_vllm_root_logger() -> None:
+    logging_config = dict[str, Any]()
+
+    if not VLLM_CONFIGURE_LOGGING and VLLM_LOGGING_CONFIG_PATH:
+        raise RuntimeError(
+            "VLLM_CONFIGURE_LOGGING evaluated to false, but "
+            "VLLM_LOGGING_CONFIG_PATH was given. VLLM_LOGGING_CONFIG_PATH "
+            "implies VLLM_CONFIGURE_LOGGING. Please enable "
+            "VLLM_CONFIGURE_LOGGING or unset VLLM_LOGGING_CONFIG_PATH.")
+
+    if VLLM_CONFIGURE_LOGGING:
+        logging_config = DEFAULT_LOGGING_CONFIG
+
+    if VLLM_LOGGING_CONFIG_PATH:
+        if not path.exists(VLLM_LOGGING_CONFIG_PATH):
+            raise RuntimeError(
+                "Could not load logging config. File does not exist: %s",
+                VLLM_LOGGING_CONFIG_PATH)
+        with open(VLLM_LOGGING_CONFIG_PATH, encoding="utf-8") as file:
+            custom_config = json.loads(file.read())
+
+        if not isinstance(custom_config, dict):
+            raise ValueError("Invalid logging config. Expected dict, got %s.",
+                             type(custom_config).__name__)
+        logging_config = custom_config
+
+    for formatter in logging_config.get("formatters", {}).values():
+        # This provides backwards compatibility after #10134.
+        if formatter.get("class") == "vllm.logging.NewLineFormatter":
+            formatter["class"] = "vllm.logging_utils.NewLineFormatter"
+
+    if logging_config:
+        dictConfig(logging_config)
+
+
+def init_logger(name: str) -> _VllmLogger:
+    """The main purpose of this function is to ensure that loggers are
+    retrieved in such a way that we can be sure the root vllm logger has
+    already been configured."""
+
+    logger = logging.getLogger(name)
+
+    for method_name, method in _METHODS_TO_PATCH.items():
+        setattr(logger, method_name, MethodType(method, logger))
+
+    return cast(_VllmLogger, logger)
+
+
+# The root logger is initialized when the module is imported.
+# This is thread-safe as the module is only imported once,
+# guaranteed by the Python GIL.
+_configure_vllm_root_logger()
+
+logger = init_logger(__name__)
+
+
+def _trace_calls(log_path, root_dir, frame, event, arg=None):
+    if event in ['call', 'return']:
+        # Extract the filename, line number, function name, and the code object
+        filename = frame.f_code.co_filename
+        lineno = frame.f_lineno
+        func_name = frame.f_code.co_name
+        if not filename.startswith(root_dir):
+            # only log the functions in the vllm root_dir
+            return
+        # Log every function call or return
+        try:
+            last_frame = frame.f_back
+            if last_frame is not None:
+                last_filename = last_frame.f_code.co_filename
+                last_lineno = last_frame.f_lineno
+                last_func_name = last_frame.f_code.co_name
+            else:
+                # initial frame
+                last_filename = ""
+                last_lineno = 0
+                last_func_name = ""
+            with open(log_path, 'a') as f:
+                ts = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")
+                if event == 'call':
+                    f.write(f"{ts} Call to"
+                            f" {func_name} in {filename}:{lineno}"
+                            f" from {last_func_name} in {last_filename}:"
+                            f"{last_lineno}\n")
+                else:
+                    f.write(f"{ts} Return from"
+                            f" {func_name} in {filename}:{lineno}"
+                            f" to {last_func_name} in {last_filename}:"
+                            f"{last_lineno}\n")
+        except NameError:
+            # modules are deleted during shutdown
+            pass
+    return partial(_trace_calls, log_path, root_dir)
+
+
+def enable_trace_function_call(log_file_path: str,
+                               root_dir: Optional[str] = None):
+    """
+    Enable tracing of every function call in code under `root_dir`.
+    This is useful for debugging hangs or crashes.
+    `log_file_path` is the path to the log file.
+    `root_dir` is the root directory of the code to trace. If None, it is the
+    vllm root directory.
+
+    Note that this call is thread-level, any threads calling this function
+    will have the trace enabled. Other threads will not be affected.
+    """
+    logger.warning(
+        "VLLM_TRACE_FUNCTION is enabled. It will record every"
+        " function executed by Python. This will slow down the code. It "
+        "is suggested to be used for debugging hang or crashes only.")
+    logger.info("Trace frame log is saved to %s", log_file_path)
+    if root_dir is None:
+        # by default, this is the vllm root directory
+        root_dir = os.path.dirname(os.path.dirname(__file__))
+    sys.settrace(partial(_trace_calls, log_file_path, root_dir))

vllm/logging_utils/__init__.py

@@ -0,0 +1,8 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from vllm.logging_utils.formatter import NewLineFormatter
+
+__all__ = [
+    "NewLineFormatter",
+]

vllm/logging_utils/dump_input.py

@@ -0,0 +1,81 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import contextlib
+import enum
+import json
+from typing import Optional
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.v1.core.sched.output import SchedulerOutput
+from vllm.v1.metrics.stats import SchedulerStats
+from vllm.version import __version__ as VLLM_VERSION
+
+logger = init_logger(__name__)
+
+
+def prepare_object_to_dump(obj) -> str:
+    if isinstance(obj, str):
+        return f"'{obj}'"  # Double quotes
+    elif isinstance(obj, dict):
+        dict_str = ', '.join({f'{str(k)}: {prepare_object_to_dump(v)}' \
+            for k, v in obj.items()})
+        return f'{{{dict_str}}}'
+    elif isinstance(obj, list):
+        return f"[{', '.join([prepare_object_to_dump(v) for v in obj])}]"
+    elif isinstance(obj, set):
+        return f"[{', '.join([prepare_object_to_dump(v) for v in list(obj)])}]"
+        # return [prepare_object_to_dump(v) for v in list(obj)]
+    elif isinstance(obj, tuple):
+        return f"[{', '.join([prepare_object_to_dump(v) for v in obj])}]"
+    elif isinstance(obj, enum.Enum):
+        return repr(obj)
+    elif isinstance(obj, torch.Tensor):
+        # We only print the 'draft' of the tensor to not expose sensitive data
+        # and to get some metadata in case of CUDA runtime crashed
+        return (f"Tensor(shape={obj.shape}, "
+                f"device={obj.device},"
+                f"dtype={obj.dtype})")
+    elif hasattr(obj, 'anon_repr'):
+        return obj.anon_repr()
+    elif hasattr(obj, '__dict__'):
+        items = obj.__dict__.items()
+        dict_str = ', '.join([f'{str(k)}={prepare_object_to_dump(v)}' \
+            for k, v in items])
+        return f"{type(obj).__name__}({dict_str})"
+    else:
+        # Hacky way to make sure we can serialize the object in JSON format
+        try:
+            return json.dumps(obj)
+        except (TypeError, OverflowError):
+            return repr(obj)
+
+
+def dump_engine_exception(config: VllmConfig,
+                          scheduler_output: SchedulerOutput,
+                          scheduler_stats: Optional[SchedulerStats]):
+    # NOTE: ensure we can log extra info without risking raises
+    # unexpected errors during logging
+    with contextlib.suppress(Exception):
+        _dump_engine_exception(config, scheduler_output, scheduler_stats)
+
+
+def _dump_engine_exception(config: VllmConfig,
+                           scheduler_output: SchedulerOutput,
+                           scheduler_stats: Optional[SchedulerStats]):
+    logger.error(
+        "Dumping input data for V1 LLM engine (v%s) with config: %s, ",
+        VLLM_VERSION,
+        config,
+    )
+    try:
+        dump_obj = prepare_object_to_dump(scheduler_output)
+        logger.error("Dumping scheduler output for model execution: %s",
+                     dump_obj)
+        if scheduler_stats:
+            logger.error("Dumping scheduler stats: %s", scheduler_stats)
+    except Exception:
+        logger.exception("Error preparing object to dump")

vllm/logging_utils/formatter.py

@@ -0,0 +1,79 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import logging
+from pathlib import Path
+
+from vllm import envs
+
+
+class NewLineFormatter(logging.Formatter):
+    """Adds logging prefix to newlines to align multi-line messages."""
+
+    def __init__(self, fmt, datefmt=None, style="%"):
+        super().__init__(fmt, datefmt, style)
+
+        self.use_relpath = envs.VLLM_LOGGING_LEVEL == "DEBUG"
+        if self.use_relpath:
+            self.root_dir = Path(__file__).resolve().parent.parent.parent
+
+    def format(self, record):
+
+        def shrink_path(relpath: Path) -> str:
+            """
+            Shortens a file path for logging display:
+            - Removes leading 'vllm' folder if present.
+            - If path starts with 'v1',
+            keeps the first two and last two levels,
+            collapsing the middle as '...'.
+            - Otherwise, keeps the first and last two levels,
+            collapsing the middle as '...'.
+            - If the path is short, returns it as-is.
+            - Examples:
+            vllm/model_executor/layers/quantization/utils/fp8_utils.py ->
+            model_executor/.../quantization/utils/fp8_utils.py
+            vllm/model_executor/layers/quantization/awq.py ->
+            model_executor/layers/quantization/awq.py
+            vllm/v1/attention/backends/mla/common.py ->
+            v1/attention/backends/mla/common.py
+
+            Args:
+                relpath (Path): The relative path to be shortened.
+            Returns:
+                str: The shortened path string for display.
+            """
+            parts = list(relpath.parts)
+            new_parts = []
+            if parts and parts[0] == "vllm":
+                parts = parts[1:]
+            if parts and parts[0] == "v1":
+                new_parts += parts[:2]
+                parts = parts[2:]
+            elif parts:
+                new_parts += parts[:1]
+                parts = parts[1:]
+            if len(parts) > 2:
+                new_parts += ["..."] + parts[-2:]
+            else:
+                new_parts += parts
+            return "/".join(new_parts)
+
+        if self.use_relpath:
+            abs_path = getattr(record, "pathname", None)
+            if abs_path:
+                try:
+                    relpath = Path(abs_path).resolve().relative_to(
+                        self.root_dir)
+                except Exception:
+                    relpath = Path(record.filename)
+            else:
+                relpath = Path(record.filename)
+            record.fileinfo = shrink_path(relpath)
+        else:
+            record.fileinfo = record.filename
+
+        msg = super().format(record)
+        if record.message != "":
+            parts = msg.split(record.message)
+            msg = msg.replace("\n", "\r\n" + parts[0])
+        return msg