Add files using upload-large-folder tool

LTA_openwebtext_dualt/docs/lta_samples/metrics_20260517/selected_long20k_len256_bs512_ode128_20260517_1830long20k/rollin_p50_s4_old/step_3000/decode_token_acc.tsv

@@ -0,0 +1,2 @@
+run	ckpt_step	endpoint_softening	token_acc_mean	token_acc_min	token_acc_max	exact_acc	exact_count	exact_ref_coverage	exact_ref_count
+train8_rollin_len256_rollin_p50_s4_i32_20260517_171654	3000	none	0.9420166015625	0.12109375	1.0	0.140625	9	0.125	1

LTA_openwebtext_dualt/logs/ctx1024_sampleds_sweep_4gpu/ctx1024_sampledpath_true_20260517_224139.log

@@ -0,0 +1,331 @@
+[ctx1024-sampleds] start stamp=ctx1024_sampledpath_true_20260517_224139 len=1024 vocab=2664 out=docs/lta_samples/metrics_20260517/ctx1024_sampleds_sweep_bs512_ode128_ctx1024_sampledpath_true_20260517_224139
+[ctx1024-sampleds] config=p50_path4_unif0_0p125_outwdm1 run=train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139 p=0.50 mode=sampled_path steps=4 s_dist=uniform s_frac=0.0->0.125 beta=2.0,6.0 outwd=-1 sync_t=1
+[ctx1024-sampleds] train config=p50_path4_unif0_0p125_outwdm1 from=0 to=1000
+[ctx1024-sampleds] eval config=p50_path4_unif0_0p125_outwdm1 step=1000
+[eval-decode-acc] train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139 step=1000 soft=none
+[decode] max_len=1024 generated=64/64
+{
+  "num_rows": 1,
+  "best_by_run": {
+    "train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139::none": {
+      "run": "train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139",
+      "checkpoint": "runs/train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139/step_0001000.pt",
+      "ckpt_step": 1000,
+      "endpoint_softening": "none",
+      "decode_rule": "flowmap",
+      "steps": 128,
+      "time_schedule": "logit_normal",
+      "model_t_mode": "post",
+      "final_from": "state",
+      "n_gen": 64,
+      "n_refs": 8,
+      "token_acc_mean": 0.0288543701171875,
+      "token_acc_min": 0.01953125,
+      "token_acc_max": 0.0478515625,
+      "exact_acc": 0.0,
+      "exact_count": 0,
+      "exact_ref_coverage": 0.0,
+      "exact_ref_count": 0,
+      "exact_ref_hits": [],
+      "best_ref_idx": [
+        7,
+        2,
+        5,
+        7,
+        7,
+        7,
+        2,
+        0,
+        7,
+        3,
+        7,
+        4,
+        2,
+        7,
+        7,
+        5,
+        7,
+        2,
+        1,
+        0,
+        7,
+        2,
+        7,
+        5,
+        7,
+        0,
+        5,
+        5,
+        7,
+        7,
+        5,
+        7,
+        4,
+        7,
+        7,
+        5,
+        5,
+        4,
+        7,
+        5,
+        2,
+        7,
+        5,
+        7,
+        7,
+        7,
+        7,
+        4,
+        4,
+        5,
+        7,
+        2,
+        4,
+        7,
+        0,
+        7,
+        7,
+        5,
+        0,
+        4,
+        0,
+        7,
+        4,
+        5
+      ],
+      "best_token_acc": [
+        0.0263671875,
+        0.021484375,
+        0.029296875,
+        0.01953125,
+        0.029296875,
+        0.0283203125,
+        0.0234375,
+        0.021484375,
+        0.0263671875,
+        0.0283203125,
+        0.0234375,
+        0.025390625,
+        0.0302734375,
+        0.0263671875,
+        0.0263671875,
+        0.0263671875,
+        0.0263671875,
+        0.0234375,
+        0.0283203125,
+        0.0302734375,
+        0.033203125,
+        0.0380859375,
+        0.0263671875,
+        0.0283203125,
+        0.02734375,
+        0.03515625,
+        0.044921875,
+        0.0419921875,
+        0.03125,
+        0.0283203125,
+        0.0478515625,
+        0.0224609375,
+        0.0263671875,
+        0.0283203125,
+        0.033203125,
+        0.041015625,
+        0.03125,
+        0.0224609375,
+        0.03125,
+        0.03515625,
+        0.0244140625,
+        0.0263671875,
+        0.0302734375,
+        0.0234375,
+        0.01953125,
+        0.021484375,
+        0.0224609375,
+        0.0322265625,
+        0.025390625,
+        0.0322265625,
+        0.029296875,
+        0.033203125,
+        0.0244140625,
+        0.033203125,
+        0.02734375,
+        0.041015625,
+        0.025390625,
+        0.0302734375,
+        0.0234375,
+        0.02734375,
+        0.0244140625,
+        0.025390625,
+        0.025390625,
+        0.0439453125
+      ]
+    }
+  },
+  "first_exact_by_run": {}
+}
+RESULT config=p50_path4_unif0_0p125_outwdm1 run=train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139 ckpt_step=1000 views=512000 token_acc=0.0289 exact=0/64 exact_refs=0 hits=[]
+[ctx1024-sampleds] train config=p50_path4_unif0_0p125_outwdm1 from=1000 to=2000
+[ctx1024-sampleds] eval config=p50_path4_unif0_0p125_outwdm1 step=2000
+[eval-decode-acc] train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139 step=2000 soft=none
+[decode] max_len=1024 generated=64/64
+{
+  "num_rows": 1,
+  "best_by_run": {
+    "train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139::none": {
+      "run": "train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139",
+      "checkpoint": "runs/train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139/step_0002000.pt",
+      "ckpt_step": 2000,
+      "endpoint_softening": "none",
+      "decode_rule": "flowmap",
+      "steps": 128,
+      "time_schedule": "logit_normal",
+      "model_t_mode": "post",
+      "final_from": "state",
+      "n_gen": 64,
+      "n_refs": 8,
+      "token_acc_mean": 0.86529541015625,
+      "token_acc_min": 0.2119140625,
+      "token_acc_max": 0.9921875,
+      "exact_acc": 0.0,
+      "exact_count": 0,
+      "exact_ref_coverage": 0.0,
+      "exact_ref_count": 0,
+      "exact_ref_hits": [],
+      "best_ref_idx": [
+        4,
+        4,
+        5,
+        4,
+        1,
+        7,
+        4,
+        4,
+        5,
+        1,
+        5,
+        1,
+        4,
+        1,
+        4,
+        4,
+        1,
+        5,
+        4,
+        1,
+        4,
+        4,
+        4,
+        4,
+        4,
+        4,
+        7,
+        4,
+        4,
+        1,
+        4,
+        4,
+        4,
+        1,
+        4,
+        4,
+        4,
+        4,
+        5,
+        4,
+        1,
+        4,
+        4,
+        3,
+        4,
+        5,
+        7,
+        7,
+        5,
+        4,
+        4,
+        4,
+        5,
+        4,
+        4,
+        4,
+        5,
+        4,
+        7,
+        4,
+        5,
+        4,
+        1,
+        4
+      ],
+      "best_token_acc": [
+        0.962890625,
+        0.966796875,
+        0.9912109375,
+        0.3701171875,
+        0.78125,
+        0.7607421875,
+        0.9658203125,
+        0.9560546875,
+        0.9873046875,
+        0.376953125,
+        0.9921875,
+        0.970703125,
+        0.95703125,
+        0.72265625,
+        0.71484375,
+        0.7314453125,
+        0.9765625,
+        0.9892578125,
+        0.73828125,
+        0.861328125,
+        0.9521484375,
+        0.9716796875,
+        0.8662109375,
+        0.8330078125,
+        0.96875,
+        0.9482421875,
+        0.9853515625,
+        0.689453125,
+        0.962890625,
+        0.9775390625,
+        0.9580078125,
+        0.2119140625,
+        0.9345703125,
+        0.2138671875,
+        0.90625,
+        0.5400390625,
+        0.9736328125,
+        0.685546875,
+        0.9716796875,
+        0.8994140625,
+        0.974609375,
+        0.9736328125,
+        0.9248046875,
+        0.984375,
+        0.9677734375,
+        0.7060546875,
+        0.9345703125,
+        0.9228515625,
+        0.9912109375,
+        0.8466796875,
+        0.9599609375,
+        0.9638671875,
+        0.4013671875,
+        0.9033203125,
+        0.9736328125,
+        0.943359375,
+        0.9912109375,
+        0.9716796875,
+        0.98828125,
+        0.9716796875,
+        0.97265625,
+        0.94921875,
+        0.97265625,
+        0.9658203125
+      ]
+    }
+  },
+  "first_exact_by_run": {}
+}
+RESULT config=p50_path4_unif0_0p125_outwdm1 run=train8_ctx1024_p50_path4_unif0_0p125_outwdm1_ctx1024_sampledpath_true_20260517_224139 ckpt_step=2000 views=1024000 token_acc=0.8653 exact=0/64 exact_refs=0 hits=[]
+[ctx1024-sampleds] train config=p50_path4_unif0_0p125_outwdm1 from=2000 to=3000

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/anyio/from_thread.py

@@ -0,0 +1,578 @@
+from __future__ import annotations
+
+__all__ = (
+    "BlockingPortal",
+    "BlockingPortalProvider",
+    "check_cancelled",
+    "run",
+    "run_sync",
+    "start_blocking_portal",
+)
+
+import sys
+from collections.abc import Awaitable, Callable, Generator
+from concurrent.futures import Future
+from contextlib import (
+    AbstractAsyncContextManager,
+    AbstractContextManager,
+    contextmanager,
+)
+from dataclasses import dataclass, field
+from functools import partial
+from inspect import isawaitable
+from threading import Lock, Thread, current_thread, get_ident
+from types import TracebackType
+from typing import (
+    Any,
+    Generic,
+    TypeVar,
+    cast,
+    overload,
+)
+
+from ._core._eventloop import (
+    get_cancelled_exc_class,
+    threadlocals,
+)
+from ._core._eventloop import run as run_eventloop
+from ._core._exceptions import NoEventLoopError
+from ._core._synchronization import Event
+from ._core._tasks import CancelScope, create_task_group
+from .abc._tasks import TaskStatus
+from .lowlevel import EventLoopToken, current_token
+
+if sys.version_info >= (3, 11):
+    from typing import TypeVarTuple, Unpack
+else:
+    from typing_extensions import TypeVarTuple, Unpack
+
+T_Retval = TypeVar("T_Retval")
+T_co = TypeVar("T_co", covariant=True)
+PosArgsT = TypeVarTuple("PosArgsT")
+
+
+def _token_or_error(token: EventLoopToken | None) -> EventLoopToken:
+    if token is not None:
+        return token
+
+    try:
+        return threadlocals.current_token
+    except AttributeError:
+        raise NoEventLoopError(
+            "Not running inside an AnyIO worker thread, and no event loop token was "
+            "provided"
+        ) from None
+
+
+def run(
+    func: Callable[[Unpack[PosArgsT]], Awaitable[T_Retval]],
+    *args: Unpack[PosArgsT],
+    token: EventLoopToken | None = None,
+) -> T_Retval:
+    """
+    Call a coroutine function from a worker thread.
+
+    :param func: a coroutine function
+    :param args: positional arguments for the callable
+    :param token: an event loop token to use to get back to the event loop thread
+        (required if calling this function from outside an AnyIO worker thread)
+    :return: the return value of the coroutine function
+    :raises MissingTokenError: if no token was provided and called from outside an
+        AnyIO worker thread
+    :raises RunFinishedError: if the event loop tied to ``token`` is no longer running
+
+    .. versionchanged:: 4.11.0
+        Added the ``token`` parameter.
+
+    """
+    explicit_token = token is not None
+    token = _token_or_error(token)
+    return token.backend_class.run_async_from_thread(
+        func, args, token=token.native_token if explicit_token else None
+    )
+
+
+def run_sync(
+    func: Callable[[Unpack[PosArgsT]], T_Retval],
+    *args: Unpack[PosArgsT],
+    token: EventLoopToken | None = None,
+) -> T_Retval:
+    """
+    Call a function in the event loop thread from a worker thread.
+
+    :param func: a callable
+    :param args: positional arguments for the callable
+    :param token: an event loop token to use to get back to the event loop thread
+        (required if calling this function from outside an AnyIO worker thread)
+    :return: the return value of the callable
+    :raises MissingTokenError: if no token was provided and called from outside an
+        AnyIO worker thread
+    :raises RunFinishedError: if the event loop tied to ``token`` is no longer running
+
+    .. versionchanged:: 4.11.0
+        Added the ``token`` parameter.
+
+    """
+    explicit_token = token is not None
+    token = _token_or_error(token)
+    return token.backend_class.run_sync_from_thread(
+        func, args, token=token.native_token if explicit_token else None
+    )
+
+
+class _BlockingAsyncContextManager(Generic[T_co], AbstractContextManager):
+    _enter_future: Future[T_co]
+    _exit_future: Future[bool | None]
+    _exit_event: Event
+    _exit_exc_info: tuple[
+        type[BaseException] | None, BaseException | None, TracebackType | None
+    ] = (None, None, None)
+
+    def __init__(
+        self, async_cm: AbstractAsyncContextManager[T_co], portal: BlockingPortal
+    ):
+        self._async_cm = async_cm
+        self._portal = portal
+
+    async def run_async_cm(self) -> bool | None:
+        try:
+            self._exit_event = Event()
+            value = await self._async_cm.__aenter__()
+        except BaseException as exc:
+            self._enter_future.set_exception(exc)
+            raise
+        else:
+            self._enter_future.set_result(value)
+
+        try:
+            # Wait for the sync context manager to exit.
+            # This next statement can raise `get_cancelled_exc_class()` if
+            # something went wrong in a task group in this async context
+            # manager.
+            await self._exit_event.wait()
+        finally:
+            # In case of cancellation, it could be that we end up here before
+            # `_BlockingAsyncContextManager.__exit__` is called, and an
+            # `_exit_exc_info` has been set.
+            result = await self._async_cm.__aexit__(*self._exit_exc_info)
+
+        return result
+
+    def __enter__(self) -> T_co:
+        self._enter_future = Future()
+        self._exit_future = self._portal.start_task_soon(self.run_async_cm)
+        return self._enter_future.result()
+
+    def __exit__(
+        self,
+        __exc_type: type[BaseException] | None,
+        __exc_value: BaseException | None,
+        __traceback: TracebackType | None,
+    ) -> bool | None:
+        self._exit_exc_info = __exc_type, __exc_value, __traceback
+        self._portal.call(self._exit_event.set)
+        return self._exit_future.result()
+
+
+class _BlockingPortalTaskStatus(TaskStatus):
+    def __init__(self, future: Future):
+        self._future = future
+
+    def started(self, value: object = None) -> None:
+        self._future.set_result(value)
+
+
+class BlockingPortal:
+    """
+    An object that lets external threads run code in an asynchronous event loop.
+
+    :raises NoEventLoopError: if no supported asynchronous event loop is running in the
+        current thread
+    """
+
+    def __init__(self) -> None:
+        self._token = current_token()
+        self._event_loop_thread_id: int | None = get_ident()
+        self._stop_event = Event()
+        self._task_group = create_task_group()
+
+    async def __aenter__(self) -> BlockingPortal:
+        await self._task_group.__aenter__()
+        return self
+
+    async def __aexit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> bool:
+        await self.stop()
+        return await self._task_group.__aexit__(exc_type, exc_val, exc_tb)
+
+    def _check_running(self) -> None:
+        if self._event_loop_thread_id is None:
+            raise RuntimeError("This portal is not running")
+        if self._event_loop_thread_id == get_ident():
+            raise RuntimeError(
+                "This method cannot be called from the event loop thread"
+            )
+
+    async def sleep_until_stopped(self) -> None:
+        """Sleep until :meth:`stop` is called."""
+        await self._stop_event.wait()
+
+    async def stop(self, cancel_remaining: bool = False) -> None:
+        """
+        Signal the portal to shut down.
+
+        This marks the portal as no longer accepting new calls and exits from
+        :meth:`sleep_until_stopped`.
+
+        :param cancel_remaining: ``True`` to cancel all the remaining tasks, ``False``
+            to let them finish before returning
+
+        """
+        self._event_loop_thread_id = None
+        self._stop_event.set()
+        if cancel_remaining:
+            self._task_group.cancel_scope.cancel("the blocking portal is shutting down")
+
+    async def _call_func(
+        self,
+        func: Callable[[Unpack[PosArgsT]], Awaitable[T_Retval] | T_Retval],
+        args: tuple[Unpack[PosArgsT]],
+        kwargs: dict[str, Any],
+        future: Future[T_Retval],
+    ) -> None:
+        def callback(f: Future[T_Retval]) -> None:
+            if f.cancelled():
+                if self._event_loop_thread_id == get_ident():
+                    scope.cancel("the future was cancelled")
+                elif self._event_loop_thread_id is not None:
+                    self.call(scope.cancel, "the future was cancelled")
+
+        try:
+            retval_or_awaitable = func(*args, **kwargs)
+            if isawaitable(retval_or_awaitable):
+                with CancelScope() as scope:
+                    future.add_done_callback(callback)
+                    retval = await retval_or_awaitable
+            else:
+                retval = retval_or_awaitable
+        except get_cancelled_exc_class():
+            future.cancel()
+            future.set_running_or_notify_cancel()
+        except BaseException as exc:
+            if not future.cancelled():
+                future.set_exception(exc)
+
+            # Let base exceptions fall through
+            if not isinstance(exc, Exception):
+                raise
+        else:
+            if not future.cancelled():
+                future.set_result(retval)
+        finally:
+            scope = None  # type: ignore[assignment]
+
+    def _spawn_task_from_thread(
+        self,
+        func: Callable[[Unpack[PosArgsT]], Awaitable[T_Retval] | T_Retval],
+        args: tuple[Unpack[PosArgsT]],
+        kwargs: dict[str, Any],
+        name: object,
+        future: Future[T_Retval],
+    ) -> None:
+        """
+        Spawn a new task using the given callable.
+
+        :param func: a callable
+        :param args: positional arguments to be passed to the callable
+        :param kwargs: keyword arguments to be passed to the callable
+        :param name: name of the task (will be coerced to a string if not ``None``)
+        :param future: a future that will resolve to the return value of the callable,
+            or the exception raised during its execution
+
+        """
+        run_sync(
+            partial(self._task_group.start_soon, name=name),
+            self._call_func,
+            func,
+            args,
+            kwargs,
+            future,
+            token=self._token,
+        )
+
+    @overload
+    def call(
+        self,
+        func: Callable[[Unpack[PosArgsT]], Awaitable[T_Retval]],
+        *args: Unpack[PosArgsT],
+    ) -> T_Retval: ...
+
+    @overload
+    def call(
+        self, func: Callable[[Unpack[PosArgsT]], T_Retval], *args: Unpack[PosArgsT]
+    ) -> T_Retval: ...
+
+    def call(
+        self,
+        func: Callable[[Unpack[PosArgsT]], Awaitable[T_Retval] | T_Retval],
+        *args: Unpack[PosArgsT],
+    ) -> T_Retval:
+        """
+        Call the given function in the event loop thread.
+
+        If the callable returns a coroutine object, it is awaited on.
+
+        :param func: any callable
+        :raises RuntimeError: if the portal is not running or if this method is called
+            from within the event loop thread
+
+        """
+        return cast(T_Retval, self.start_task_soon(func, *args).result())
+
+    @overload
+    def start_task_soon(
+        self,
+        func: Callable[[Unpack[PosArgsT]], Awaitable[T_Retval]],
+        *args: Unpack[PosArgsT],
+        name: object = None,
+    ) -> Future[T_Retval]: ...
+
+    @overload
+    def start_task_soon(
+        self,
+        func: Callable[[Unpack[PosArgsT]], T_Retval],
+        *args: Unpack[PosArgsT],
+        name: object = None,
+    ) -> Future[T_Retval]: ...
+
+    def start_task_soon(
+        self,
+        func: Callable[[Unpack[PosArgsT]], Awaitable[T_Retval] | T_Retval],
+        *args: Unpack[PosArgsT],
+        name: object = None,
+    ) -> Future[T_Retval]:
+        """
+        Start a task in the portal's task group.
+
+        The task will be run inside a cancel scope which can be cancelled by cancelling
+        the returned future.
+
+        :param func: the target function
+        :param args: positional arguments passed to ``func``
+        :param name: name of the task (will be coerced to a string if not ``None``)
+        :return: a future that resolves with the return value of the callable if the
+            task completes successfully, or with the exception raised in the task
+        :raises RuntimeError: if the portal is not running or if this method is called
+            from within the event loop thread
+        :rtype: concurrent.futures.Future[T_Retval]
+
+        .. versionadded:: 3.0
+
+        """
+        self._check_running()
+        f: Future[T_Retval] = Future()
+        self._spawn_task_from_thread(func, args, {}, name, f)
+        return f
+
+    def start_task(
+        self,
+        func: Callable[..., Awaitable[T_Retval]],
+        *args: object,
+        name: object = None,
+    ) -> tuple[Future[T_Retval], Any]:
+        """
+        Start a task in the portal's task group and wait until it signals for readiness.
+
+        This method works the same way as :meth:`.abc.TaskGroup.start`.
+
+        :param func: the target function
+        :param args: positional arguments passed to ``func``
+        :param name: name of the task (will be coerced to a string if not ``None``)
+        :return: a tuple of (future, task_status_value) where the ``task_status_value``
+            is the value passed to ``task_status.started()`` from within the target
+            function
+        :rtype: tuple[concurrent.futures.Future[T_Retval], Any]
+
+        .. versionadded:: 3.0
+
+        """
+
+        def task_done(future: Future[T_Retval]) -> None:
+            if not task_status_future.done():
+                if future.cancelled():
+                    task_status_future.cancel()
+                elif future.exception():
+                    task_status_future.set_exception(future.exception())
+                else:
+                    exc = RuntimeError(
+                        "Task exited without calling task_status.started()"
+                    )
+                    task_status_future.set_exception(exc)
+
+        self._check_running()
+        task_status_future: Future = Future()
+        task_status = _BlockingPortalTaskStatus(task_status_future)
+        f: Future = Future()
+        f.add_done_callback(task_done)
+        self._spawn_task_from_thread(func, args, {"task_status": task_status}, name, f)
+        return f, task_status_future.result()
+
+    def wrap_async_context_manager(
+        self, cm: AbstractAsyncContextManager[T_co]
+    ) -> AbstractContextManager[T_co]:
+        """
+        Wrap an async context manager as a synchronous context manager via this portal.
+
+        Spawns a task that will call both ``__aenter__()`` and ``__aexit__()``, stopping
+        in the middle until the synchronous context manager exits.
+
+        :param cm: an asynchronous context manager
+        :return: a synchronous context manager
+
+        .. versionadded:: 2.1
+
+        """
+        return _BlockingAsyncContextManager(cm, self)
+
+
+@dataclass
+class BlockingPortalProvider:
+    """
+    A manager for a blocking portal. Used as a context manager. The first thread to
+    enter this context manager causes a blocking portal to be started with the specific
+    parameters, and the last thread to exit causes the portal to be shut down. Thus,
+    there will be exactly one blocking portal running in this context as long as at
+    least one thread has entered this context manager.
+
+    The parameters are the same as for :func:`~anyio.run`.
+
+    :param backend: name of the backend
+    :param backend_options: backend options
+
+    .. versionadded:: 4.4
+    """
+
+    backend: str = "asyncio"
+    backend_options: dict[str, Any] | None = None
+    _lock: Lock = field(init=False, default_factory=Lock)
+    _leases: int = field(init=False, default=0)
+    _portal: BlockingPortal = field(init=False)
+    _portal_cm: AbstractContextManager[BlockingPortal] | None = field(
+        init=False, default=None
+    )
+
+    def __enter__(self) -> BlockingPortal:
+        with self._lock:
+            if self._portal_cm is None:
+                self._portal_cm = start_blocking_portal(
+                    self.backend, self.backend_options
+                )
+                self._portal = self._portal_cm.__enter__()
+
+            self._leases += 1
+            return self._portal
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> None:
+        portal_cm: AbstractContextManager[BlockingPortal] | None = None
+        with self._lock:
+            assert self._portal_cm
+            assert self._leases > 0
+            self._leases -= 1
+            if not self._leases:
+                portal_cm = self._portal_cm
+                self._portal_cm = None
+                del self._portal
+
+        if portal_cm:
+            portal_cm.__exit__(None, None, None)
+
+
+@contextmanager
+def start_blocking_portal(
+    backend: str = "asyncio",
+    backend_options: dict[str, Any] | None = None,
+    *,
+    name: str | None = None,
+) -> Generator[BlockingPortal, Any, None]:
+    """
+    Start a new event loop in a new thread and run a blocking portal in its main task.
+
+    The parameters are the same as for :func:`~anyio.run`.
+
+    :param backend: name of the backend
+    :param backend_options: backend options
+    :param name: name of the thread
+    :return: a context manager that yields a blocking portal
+
+    .. versionchanged:: 3.0
+        Usage as a context manager is now required.
+
+    """
+
+    async def run_portal() -> None:
+        async with BlockingPortal() as portal_:
+            if name is None:
+                current_thread().name = f"{backend}-portal-{id(portal_):x}"
+
+            future.set_result(portal_)
+            await portal_.sleep_until_stopped()
+
+    def run_blocking_portal() -> None:
+        if future.set_running_or_notify_cancel():
+            try:
+                run_eventloop(
+                    run_portal, backend=backend, backend_options=backend_options
+                )
+            except BaseException as exc:
+                if not future.done():
+                    future.set_exception(exc)
+
+    future: Future[BlockingPortal] = Future()
+    thread = Thread(target=run_blocking_portal, daemon=True, name=name)
+    thread.start()
+    try:
+        cancel_remaining_tasks = False
+        portal = future.result()
+        try:
+            yield portal
+        except BaseException:
+            cancel_remaining_tasks = True
+            raise
+        finally:
+            try:
+                portal.call(portal.stop, cancel_remaining_tasks)
+            except RuntimeError:
+                pass
+    finally:
+        thread.join()
+
+
+def check_cancelled() -> None:
+    """
+    Check if the cancel scope of the host task's running the current worker thread has
+    been cancelled.
+
+    If the host task's current cancel scope has indeed been cancelled, the
+    backend-specific cancellation exception will be raised.
+
+    :raises RuntimeError: if the current thread was not spawned by
+        :func:`.to_thread.run_sync`
+
+    """
+    try:
+        token: EventLoopToken = threadlocals.current_token
+    except AttributeError:
+        raise NoEventLoopError(
+            "This function can only be called inside an AnyIO worker thread"
+        ) from None
+
+    token.backend_class.check_cancelled()

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/__init__.py

@@ -0,0 +1,62 @@
+# A highish-level implementation of the HTTP/1.1 wire protocol (RFC 7230),
+# containing no networking code at all, loosely modelled on hyper-h2's generic
+# implementation of HTTP/2 (and in particular the h2.connection.H2Connection
+# class). There's still a bunch of subtle details you need to get right if you
+# want to make this actually useful, because it doesn't implement all the
+# semantics to check that what you're asking to write to the wire is sensible,
+# but at least it gets you out of dealing with the wire itself.
+
+from h11._connection import Connection, NEED_DATA, PAUSED
+from h11._events import (
+    ConnectionClosed,
+    Data,
+    EndOfMessage,
+    Event,
+    InformationalResponse,
+    Request,
+    Response,
+)
+from h11._state import (
+    CLIENT,
+    CLOSED,
+    DONE,
+    ERROR,
+    IDLE,
+    MIGHT_SWITCH_PROTOCOL,
+    MUST_CLOSE,
+    SEND_BODY,
+    SEND_RESPONSE,
+    SERVER,
+    SWITCHED_PROTOCOL,
+)
+from h11._util import LocalProtocolError, ProtocolError, RemoteProtocolError
+from h11._version import __version__
+
+PRODUCT_ID = "python-h11/" + __version__
+
+
+__all__ = (
+    "Connection",
+    "NEED_DATA",
+    "PAUSED",
+    "ConnectionClosed",
+    "Data",
+    "EndOfMessage",
+    "Event",
+    "InformationalResponse",
+    "Request",
+    "Response",
+    "CLIENT",
+    "CLOSED",
+    "DONE",
+    "ERROR",
+    "IDLE",
+    "MUST_CLOSE",
+    "SEND_BODY",
+    "SEND_RESPONSE",
+    "SERVER",
+    "SWITCHED_PROTOCOL",
+    "ProtocolError",
+    "LocalProtocolError",
+    "RemoteProtocolError",
+)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_abnf.py

@@ -0,0 +1,132 @@
+# We use native strings for all the re patterns, to take advantage of string
+# formatting, and then convert to bytestrings when compiling the final re
+# objects.
+
+# https://svn.tools.ietf.org/svn/wg/httpbis/specs/rfc7230.html#whitespace
+#  OWS            = *( SP / HTAB )
+#                 ; optional whitespace
+OWS = r"[ \t]*"
+
+# https://svn.tools.ietf.org/svn/wg/httpbis/specs/rfc7230.html#rule.token.separators
+#   token          = 1*tchar
+#
+#   tchar          = "!" / "#" / "$" / "%" / "&" / "'" / "*"
+#                  / "+" / "-" / "." / "^" / "_" / "`" / "|" / "~"
+#                  / DIGIT / ALPHA
+#                  ; any VCHAR, except delimiters
+token = r"[-!#$%&'*+.^_`|~0-9a-zA-Z]+"
+
+# https://svn.tools.ietf.org/svn/wg/httpbis/specs/rfc7230.html#header.fields
+#  field-name     = token
+field_name = token
+
+# The standard says:
+#
+#  field-value    = *( field-content / obs-fold )
+#  field-content  = field-vchar [ 1*( SP / HTAB ) field-vchar ]
+#  field-vchar    = VCHAR / obs-text
+#  obs-fold       = CRLF 1*( SP / HTAB )
+#                 ; obsolete line folding
+#                 ; see Section 3.2.4
+#
+# https://tools.ietf.org/html/rfc5234#appendix-B.1
+#
+#   VCHAR          =  %x21-7E
+#                  ; visible (printing) characters
+#
+# https://svn.tools.ietf.org/svn/wg/httpbis/specs/rfc7230.html#rule.quoted-string
+#   obs-text       = %x80-FF
+#
+# However, the standard definition of field-content is WRONG! It disallows
+# fields containing a single visible character surrounded by whitespace,
+# e.g. "foo a bar".
+#
+# See: https://www.rfc-editor.org/errata_search.php?rfc=7230&eid=4189
+#
+# So our definition of field_content attempts to fix it up...
+#
+# Also, we allow lots of control characters, because apparently people assume
+# that they're legal in practice (e.g., google analytics makes cookies with
+# \x01 in them!):
+#   https://github.com/python-hyper/h11/issues/57
+# We still don't allow NUL or whitespace, because those are often treated as
+# meta-characters and letting them through can lead to nasty issues like SSRF.
+vchar = r"[\x21-\x7e]"
+vchar_or_obs_text = r"[^\x00\s]"
+field_vchar = vchar_or_obs_text
+field_content = r"{field_vchar}+(?:[ \t]+{field_vchar}+)*".format(**globals())
+
+# We handle obs-fold at a different level, and our fixed-up field_content
+# already grows to swallow the whole value, so ? instead of *
+field_value = r"({field_content})?".format(**globals())
+
+#  header-field   = field-name ":" OWS field-value OWS
+header_field = (
+    r"(?P<field_name>{field_name})"
+    r":"
+    r"{OWS}"
+    r"(?P<field_value>{field_value})"
+    r"{OWS}".format(**globals())
+)
+
+# https://svn.tools.ietf.org/svn/wg/httpbis/specs/rfc7230.html#request.line
+#
+#   request-line   = method SP request-target SP HTTP-version CRLF
+#   method         = token
+#   HTTP-version   = HTTP-name "/" DIGIT "." DIGIT
+#   HTTP-name      = %x48.54.54.50 ; "HTTP", case-sensitive
+#
+# request-target is complicated (see RFC 7230 sec 5.3) -- could be path, full
+# URL, host+port (for connect), or even "*", but in any case we are guaranteed
+# that it contists of the visible printing characters.
+method = token
+request_target = r"{vchar}+".format(**globals())
+http_version = r"HTTP/(?P<http_version>[0-9]\.[0-9])"
+request_line = (
+    r"(?P<method>{method})"
+    r" "
+    r"(?P<target>{request_target})"
+    r" "
+    r"{http_version}".format(**globals())
+)
+
+# https://svn.tools.ietf.org/svn/wg/httpbis/specs/rfc7230.html#status.line
+#
+#   status-line = HTTP-version SP status-code SP reason-phrase CRLF
+#   status-code    = 3DIGIT
+#   reason-phrase  = *( HTAB / SP / VCHAR / obs-text )
+status_code = r"[0-9]{3}"
+reason_phrase = r"([ \t]|{vchar_or_obs_text})*".format(**globals())
+status_line = (
+    r"{http_version}"
+    r" "
+    r"(?P<status_code>{status_code})"
+    # However, there are apparently a few too many servers out there that just
+    # leave out the reason phrase:
+    #   https://github.com/scrapy/scrapy/issues/345#issuecomment-281756036
+    #   https://github.com/seanmonstar/httparse/issues/29
+    # so make it optional. ?: is a non-capturing group.
+    r"(?: (?P<reason>{reason_phrase}))?".format(**globals())
+)
+
+HEXDIG = r"[0-9A-Fa-f]"
+# Actually
+#
+#      chunk-size     = 1*HEXDIG
+#
+# but we impose an upper-limit to avoid ridiculosity. len(str(2**64)) == 20
+chunk_size = r"({HEXDIG}){{1,20}}".format(**globals())
+# Actually
+#
+#     chunk-ext      = *( ";" chunk-ext-name [ "=" chunk-ext-val ] )
+#
+# but we aren't parsing the things so we don't really care.
+chunk_ext = r";.*"
+chunk_header = (
+    r"(?P<chunk_size>{chunk_size})"
+    r"(?P<chunk_ext>{chunk_ext})?"
+    r"{OWS}\r\n".format(
+        **globals()
+    )  # Even though the specification does not allow for extra whitespaces,
+    # we are lenient with trailing whitespaces because some servers on the wild use it.
+)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_connection.py

@@ -0,0 +1,659 @@
+# This contains the main Connection class. Everything in h11 revolves around
+# this.
+from typing import (
+    Any,
+    Callable,
+    cast,
+    Dict,
+    List,
+    Optional,
+    overload,
+    Tuple,
+    Type,
+    Union,
+)
+
+from ._events import (
+    ConnectionClosed,
+    Data,
+    EndOfMessage,
+    Event,
+    InformationalResponse,
+    Request,
+    Response,
+)
+from ._headers import get_comma_header, has_expect_100_continue, set_comma_header
+from ._readers import READERS, ReadersType
+from ._receivebuffer import ReceiveBuffer
+from ._state import (
+    _SWITCH_CONNECT,
+    _SWITCH_UPGRADE,
+    CLIENT,
+    ConnectionState,
+    DONE,
+    ERROR,
+    MIGHT_SWITCH_PROTOCOL,
+    SEND_BODY,
+    SERVER,
+    SWITCHED_PROTOCOL,
+)
+from ._util import (  # Import the internal things we need
+    LocalProtocolError,
+    RemoteProtocolError,
+    Sentinel,
+)
+from ._writers import WRITERS, WritersType
+
+# Everything in __all__ gets re-exported as part of the h11 public API.
+__all__ = ["Connection", "NEED_DATA", "PAUSED"]
+
+
+class NEED_DATA(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class PAUSED(Sentinel, metaclass=Sentinel):
+    pass
+
+
+# If we ever have this much buffered without it making a complete parseable
+# event, we error out. The only time we really buffer is when reading the
+# request/response line + headers together, so this is effectively the limit on
+# the size of that.
+#
+# Some precedents for defaults:
+# - node.js: 80 * 1024
+# - tomcat: 8 * 1024
+# - IIS: 16 * 1024
+# - Apache: <8 KiB per line>
+DEFAULT_MAX_INCOMPLETE_EVENT_SIZE = 16 * 1024
+
+
+# RFC 7230's rules for connection lifecycles:
+# - If either side says they want to close the connection, then the connection
+#   must close.
+# - HTTP/1.1 defaults to keep-alive unless someone says Connection: close
+# - HTTP/1.0 defaults to close unless both sides say Connection: keep-alive
+#   (and even this is a mess -- e.g. if you're implementing a proxy then
+#   sending Connection: keep-alive is forbidden).
+#
+# We simplify life by simply not supporting keep-alive with HTTP/1.0 peers. So
+# our rule is:
+# - If someone says Connection: close, we will close
+# - If someone uses HTTP/1.0, we will close.
+def _keep_alive(event: Union[Request, Response]) -> bool:
+    connection = get_comma_header(event.headers, b"connection")
+    if b"close" in connection:
+        return False
+    if getattr(event, "http_version", b"1.1") < b"1.1":
+        return False
+    return True
+
+
+def _body_framing(
+    request_method: bytes, event: Union[Request, Response]
+) -> Tuple[str, Union[Tuple[()], Tuple[int]]]:
+    # Called when we enter SEND_BODY to figure out framing information for
+    # this body.
+    #
+    # These are the only two events that can trigger a SEND_BODY state:
+    assert type(event) in (Request, Response)
+    # Returns one of:
+    #
+    #    ("content-length", count)
+    #    ("chunked", ())
+    #    ("http/1.0", ())
+    #
+    # which are (lookup key, *args) for constructing body reader/writer
+    # objects.
+    #
+    # Reference: https://tools.ietf.org/html/rfc7230#section-3.3.3
+    #
+    # Step 1: some responses always have an empty body, regardless of what the
+    # headers say.
+    if type(event) is Response:
+        if (
+            event.status_code in (204, 304)
+            or request_method == b"HEAD"
+            or (request_method == b"CONNECT" and 200 <= event.status_code < 300)
+        ):
+            return ("content-length", (0,))
+        # Section 3.3.3 also lists another case -- responses with status_code
+        # < 200. For us these are InformationalResponses, not Responses, so
+        # they can't get into this function in the first place.
+        assert event.status_code >= 200
+
+    # Step 2: check for Transfer-Encoding (T-E beats C-L):
+    transfer_encodings = get_comma_header(event.headers, b"transfer-encoding")
+    if transfer_encodings:
+        assert transfer_encodings == [b"chunked"]
+        return ("chunked", ())
+
+    # Step 3: check for Content-Length
+    content_lengths = get_comma_header(event.headers, b"content-length")
+    if content_lengths:
+        return ("content-length", (int(content_lengths[0]),))
+
+    # Step 4: no applicable headers; fallback/default depends on type
+    if type(event) is Request:
+        return ("content-length", (0,))
+    else:
+        return ("http/1.0", ())
+
+
+################################################################
+#
+# The main Connection class
+#
+################################################################
+
+
+class Connection:
+    """An object encapsulating the state of an HTTP connection.
+
+    Args:
+        our_role: If you're implementing a client, pass :data:`h11.CLIENT`. If
+            you're implementing a server, pass :data:`h11.SERVER`.
+
+        max_incomplete_event_size (int):
+            The maximum number of bytes we're willing to buffer of an
+            incomplete event. In practice this mostly sets a limit on the
+            maximum size of the request/response line + headers. If this is
+            exceeded, then :meth:`next_event` will raise
+            :exc:`RemoteProtocolError`.
+
+    """
+
+    def __init__(
+        self,
+        our_role: Type[Sentinel],
+        max_incomplete_event_size: int = DEFAULT_MAX_INCOMPLETE_EVENT_SIZE,
+    ) -> None:
+        self._max_incomplete_event_size = max_incomplete_event_size
+        # State and role tracking
+        if our_role not in (CLIENT, SERVER):
+            raise ValueError(f"expected CLIENT or SERVER, not {our_role!r}")
+        self.our_role = our_role
+        self.their_role: Type[Sentinel]
+        if our_role is CLIENT:
+            self.their_role = SERVER
+        else:
+            self.their_role = CLIENT
+        self._cstate = ConnectionState()
+
+        # Callables for converting data->events or vice-versa given the
+        # current state
+        self._writer = self._get_io_object(self.our_role, None, WRITERS)
+        self._reader = self._get_io_object(self.their_role, None, READERS)
+
+        # Holds any unprocessed received data
+        self._receive_buffer = ReceiveBuffer()
+        # If this is true, then it indicates that the incoming connection was
+        # closed *after* the end of whatever's in self._receive_buffer:
+        self._receive_buffer_closed = False
+
+        # Extra bits of state that don't fit into the state machine.
+        #
+        # These two are only used to interpret framing headers for figuring
+        # out how to read/write response bodies. their_http_version is also
+        # made available as a convenient public API.
+        self.their_http_version: Optional[bytes] = None
+        self._request_method: Optional[bytes] = None
+        # This is pure flow-control and doesn't at all affect the set of legal
+        # transitions, so no need to bother ConnectionState with it:
+        self.client_is_waiting_for_100_continue = False
+
+    @property
+    def states(self) -> Dict[Type[Sentinel], Type[Sentinel]]:
+        """A dictionary like::
+
+           {CLIENT: <client state>, SERVER: <server state>}
+
+        See :ref:`state-machine` for details.
+
+        """
+        return dict(self._cstate.states)
+
+    @property
+    def our_state(self) -> Type[Sentinel]:
+        """The current state of whichever role we are playing. See
+        :ref:`state-machine` for details.
+        """
+        return self._cstate.states[self.our_role]
+
+    @property
+    def their_state(self) -> Type[Sentinel]:
+        """The current state of whichever role we are NOT playing. See
+        :ref:`state-machine` for details.
+        """
+        return self._cstate.states[self.their_role]
+
+    @property
+    def they_are_waiting_for_100_continue(self) -> bool:
+        return self.their_role is CLIENT and self.client_is_waiting_for_100_continue
+
+    def start_next_cycle(self) -> None:
+        """Attempt to reset our connection state for a new request/response
+        cycle.
+
+        If both client and server are in :data:`DONE` state, then resets them
+        both to :data:`IDLE` state in preparation for a new request/response
+        cycle on this same connection. Otherwise, raises a
+        :exc:`LocalProtocolError`.
+
+        See :ref:`keepalive-and-pipelining`.
+
+        """
+        old_states = dict(self._cstate.states)
+        self._cstate.start_next_cycle()
+        self._request_method = None
+        # self.their_http_version gets left alone, since it presumably lasts
+        # beyond a single request/response cycle
+        assert not self.client_is_waiting_for_100_continue
+        self._respond_to_state_changes(old_states)
+
+    def _process_error(self, role: Type[Sentinel]) -> None:
+        old_states = dict(self._cstate.states)
+        self._cstate.process_error(role)
+        self._respond_to_state_changes(old_states)
+
+    def _server_switch_event(self, event: Event) -> Optional[Type[Sentinel]]:
+        if type(event) is InformationalResponse and event.status_code == 101:
+            return _SWITCH_UPGRADE
+        if type(event) is Response:
+            if (
+                _SWITCH_CONNECT in self._cstate.pending_switch_proposals
+                and 200 <= event.status_code < 300
+            ):
+                return _SWITCH_CONNECT
+        return None
+
+    # All events go through here
+    def _process_event(self, role: Type[Sentinel], event: Event) -> None:
+        # First, pass the event through the state machine to make sure it
+        # succeeds.
+        old_states = dict(self._cstate.states)
+        if role is CLIENT and type(event) is Request:
+            if event.method == b"CONNECT":
+                self._cstate.process_client_switch_proposal(_SWITCH_CONNECT)
+            if get_comma_header(event.headers, b"upgrade"):
+                self._cstate.process_client_switch_proposal(_SWITCH_UPGRADE)
+        server_switch_event = None
+        if role is SERVER:
+            server_switch_event = self._server_switch_event(event)
+        self._cstate.process_event(role, type(event), server_switch_event)
+
+        # Then perform the updates triggered by it.
+
+        if type(event) is Request:
+            self._request_method = event.method
+
+        if role is self.their_role and type(event) in (
+            Request,
+            Response,
+            InformationalResponse,
+        ):
+            event = cast(Union[Request, Response, InformationalResponse], event)
+            self.their_http_version = event.http_version
+
+        # Keep alive handling
+        #
+        # RFC 7230 doesn't really say what one should do if Connection: close
+        # shows up on a 1xx InformationalResponse. I think the idea is that
+        # this is not supposed to happen. In any case, if it does happen, we
+        # ignore it.
+        if type(event) in (Request, Response) and not _keep_alive(
+            cast(Union[Request, Response], event)
+        ):
+            self._cstate.process_keep_alive_disabled()
+
+        # 100-continue
+        if type(event) is Request and has_expect_100_continue(event):
+            self.client_is_waiting_for_100_continue = True
+        if type(event) in (InformationalResponse, Response):
+            self.client_is_waiting_for_100_continue = False
+        if role is CLIENT and type(event) in (Data, EndOfMessage):
+            self.client_is_waiting_for_100_continue = False
+
+        self._respond_to_state_changes(old_states, event)
+
+    def _get_io_object(
+        self,
+        role: Type[Sentinel],
+        event: Optional[Event],
+        io_dict: Union[ReadersType, WritersType],
+    ) -> Optional[Callable[..., Any]]:
+        # event may be None; it's only used when entering SEND_BODY
+        state = self._cstate.states[role]
+        if state is SEND_BODY:
+            # Special case: the io_dict has a dict of reader/writer factories
+            # that depend on the request/response framing.
+            framing_type, args = _body_framing(
+                cast(bytes, self._request_method), cast(Union[Request, Response], event)
+            )
+            return io_dict[SEND_BODY][framing_type](*args)  # type: ignore[index]
+        else:
+            # General case: the io_dict just has the appropriate reader/writer
+            # for this state
+            return io_dict.get((role, state))  # type: ignore[return-value]
+
+    # This must be called after any action that might have caused
+    # self._cstate.states to change.
+    def _respond_to_state_changes(
+        self,
+        old_states: Dict[Type[Sentinel], Type[Sentinel]],
+        event: Optional[Event] = None,
+    ) -> None:
+        # Update reader/writer
+        if self.our_state != old_states[self.our_role]:
+            self._writer = self._get_io_object(self.our_role, event, WRITERS)
+        if self.their_state != old_states[self.their_role]:
+            self._reader = self._get_io_object(self.their_role, event, READERS)
+
+    @property
+    def trailing_data(self) -> Tuple[bytes, bool]:
+        """Data that has been received, but not yet processed, represented as
+        a tuple with two elements, where the first is a byte-string containing
+        the unprocessed data itself, and the second is a bool that is True if
+        the receive connection was closed.
+
+        See :ref:`switching-protocols` for discussion of why you'd want this.
+        """
+        return (bytes(self._receive_buffer), self._receive_buffer_closed)
+
+    def receive_data(self, data: bytes) -> None:
+        """Add data to our internal receive buffer.
+
+        This does not actually do any processing on the data, just stores
+        it. To trigger processing, you have to call :meth:`next_event`.
+
+        Args:
+            data (:term:`bytes-like object`):
+                The new data that was just received.
+
+                Special case: If *data* is an empty byte-string like ``b""``,
+                then this indicates that the remote side has closed the
+                connection (end of file). Normally this is convenient, because
+                standard Python APIs like :meth:`file.read` or
+                :meth:`socket.recv` use ``b""`` to indicate end-of-file, while
+                other failures to read are indicated using other mechanisms
+                like raising :exc:`TimeoutError`. When using such an API you
+                can just blindly pass through whatever you get from ``read``
+                to :meth:`receive_data`, and everything will work.
+
+                But, if you have an API where reading an empty string is a
+                valid non-EOF condition, then you need to be aware of this and
+                make sure to check for such strings and avoid passing them to
+                :meth:`receive_data`.
+
+        Returns:
+            Nothing, but after calling this you should call :meth:`next_event`
+            to parse the newly received data.
+
+        Raises:
+            RuntimeError:
+                Raised if you pass an empty *data*, indicating EOF, and then
+                pass a non-empty *data*, indicating more data that somehow
+                arrived after the EOF.
+
+                (Calling ``receive_data(b"")`` multiple times is fine,
+                and equivalent to calling it once.)
+
+        """
+        if data:
+            if self._receive_buffer_closed:
+                raise RuntimeError("received close, then received more data?")
+            self._receive_buffer += data
+        else:
+            self._receive_buffer_closed = True
+
+    def _extract_next_receive_event(
+        self,
+    ) -> Union[Event, Type[NEED_DATA], Type[PAUSED]]:
+        state = self.their_state
+        # We don't pause immediately when they enter DONE, because even in
+        # DONE state we can still process a ConnectionClosed() event. But
+        # if we have data in our buffer, then we definitely aren't getting
+        # a ConnectionClosed() immediately and we need to pause.
+        if state is DONE and self._receive_buffer:
+            return PAUSED
+        if state is MIGHT_SWITCH_PROTOCOL or state is SWITCHED_PROTOCOL:
+            return PAUSED
+        assert self._reader is not None
+        event = self._reader(self._receive_buffer)
+        if event is None:
+            if not self._receive_buffer and self._receive_buffer_closed:
+                # In some unusual cases (basically just HTTP/1.0 bodies), EOF
+                # triggers an actual protocol event; in that case, we want to
+                # return that event, and then the state will change and we'll
+                # get called again to generate the actual ConnectionClosed().
+                if hasattr(self._reader, "read_eof"):
+                    event = self._reader.read_eof()
+                else:
+                    event = ConnectionClosed()
+        if event is None:
+            event = NEED_DATA
+        return event  # type: ignore[no-any-return]
+
+    def next_event(self) -> Union[Event, Type[NEED_DATA], Type[PAUSED]]:
+        """Parse the next event out of our receive buffer, update our internal
+        state, and return it.
+
+        This is a mutating operation -- think of it like calling :func:`next`
+        on an iterator.
+
+        Returns:
+            : One of three things:
+
+            1) An event object -- see :ref:`events`.
+
+            2) The special constant :data:`NEED_DATA`, which indicates that
+               you need to read more data from your socket and pass it to
+               :meth:`receive_data` before this method will be able to return
+               any more events.
+
+            3) The special constant :data:`PAUSED`, which indicates that we
+               are not in a state where we can process incoming data (usually
+               because the peer has finished their part of the current
+               request/response cycle, and you have not yet called
+               :meth:`start_next_cycle`). See :ref:`flow-control` for details.
+
+        Raises:
+            RemoteProtocolError:
+                The peer has misbehaved. You should close the connection
+                (possibly after sending some kind of 4xx response).
+
+        Once this method returns :class:`ConnectionClosed` once, then all
+        subsequent calls will also return :class:`ConnectionClosed`.
+
+        If this method raises any exception besides :exc:`RemoteProtocolError`
+        then that's a bug -- if it happens please file a bug report!
+
+        If this method raises any exception then it also sets
+        :attr:`Connection.their_state` to :data:`ERROR` -- see
+        :ref:`error-handling` for discussion.
+
+        """
+
+        if self.their_state is ERROR:
+            raise RemoteProtocolError("Can't receive data when peer state is ERROR")
+        try:
+            event = self._extract_next_receive_event()
+            if event not in [NEED_DATA, PAUSED]:
+                self._process_event(self.their_role, cast(Event, event))
+            if event is NEED_DATA:
+                if len(self._receive_buffer) > self._max_incomplete_event_size:
+                    # 431 is "Request header fields too large" which is pretty
+                    # much the only situation where we can get here
+                    raise RemoteProtocolError(
+                        "Receive buffer too long", error_status_hint=431
+                    )
+                if self._receive_buffer_closed:
+                    # We're still trying to complete some event, but that's
+                    # never going to happen because no more data is coming
+                    raise RemoteProtocolError("peer unexpectedly closed connection")
+            return event
+        except BaseException as exc:
+            self._process_error(self.their_role)
+            if isinstance(exc, LocalProtocolError):
+                exc._reraise_as_remote_protocol_error()
+            else:
+                raise
+
+    @overload
+    def send(self, event: ConnectionClosed) -> None:
+        ...
+
+    @overload
+    def send(
+        self, event: Union[Request, InformationalResponse, Response, Data, EndOfMessage]
+    ) -> bytes:
+        ...
+
+    @overload
+    def send(self, event: Event) -> Optional[bytes]:
+        ...
+
+    def send(self, event: Event) -> Optional[bytes]:
+        """Convert a high-level event into bytes that can be sent to the peer,
+        while updating our internal state machine.
+
+        Args:
+            event: The :ref:`event <events>` to send.
+
+        Returns:
+            If ``type(event) is ConnectionClosed``, then returns
+            ``None``. Otherwise, returns a :term:`bytes-like object`.
+
+        Raises:
+            LocalProtocolError:
+                Sending this event at this time would violate our
+                understanding of the HTTP/1.1 protocol.
+
+        If this method raises any exception then it also sets
+        :attr:`Connection.our_state` to :data:`ERROR` -- see
+        :ref:`error-handling` for discussion.
+
+        """
+        data_list = self.send_with_data_passthrough(event)
+        if data_list is None:
+            return None
+        else:
+            return b"".join(data_list)
+
+    def send_with_data_passthrough(self, event: Event) -> Optional[List[bytes]]:
+        """Identical to :meth:`send`, except that in situations where
+        :meth:`send` returns a single :term:`bytes-like object`, this instead
+        returns a list of them -- and when sending a :class:`Data` event, this
+        list is guaranteed to contain the exact object you passed in as
+        :attr:`Data.data`. See :ref:`sendfile` for discussion.
+
+        """
+        if self.our_state is ERROR:
+            raise LocalProtocolError("Can't send data when our state is ERROR")
+        try:
+            if type(event) is Response:
+                event = self._clean_up_response_headers_for_sending(event)
+            # We want to call _process_event before calling the writer,
+            # because if someone tries to do something invalid then this will
+            # give a sensible error message, while our writers all just assume
+            # they will only receive valid events. But, _process_event might
+            # change self._writer. So we have to do a little dance:
+            writer = self._writer
+            self._process_event(self.our_role, event)
+            if type(event) is ConnectionClosed:
+                return None
+            else:
+                # In any situation where writer is None, process_event should
+                # have raised ProtocolError
+                assert writer is not None
+                data_list: List[bytes] = []
+                writer(event, data_list.append)
+                return data_list
+        except:
+            self._process_error(self.our_role)
+            raise
+
+    def send_failed(self) -> None:
+        """Notify the state machine that we failed to send the data it gave
+        us.
+
+        This causes :attr:`Connection.our_state` to immediately become
+        :data:`ERROR` -- see :ref:`error-handling` for discussion.
+
+        """
+        self._process_error(self.our_role)
+
+    # When sending a Response, we take responsibility for a few things:
+    #
+    # - Sometimes you MUST set Connection: close. We take care of those
+    #   times. (You can also set it yourself if you want, and if you do then
+    #   we'll respect that and close the connection at the right time. But you
+    #   don't have to worry about that unless you want to.)
+    #
+    # - The user has to set Content-Length if they want it. Otherwise, for
+    #   responses that have bodies (e.g. not HEAD), then we will automatically
+    #   select the right mechanism for streaming a body of unknown length,
+    #   which depends on depending on the peer's HTTP version.
+    #
+    # This function's *only* responsibility is making sure headers are set up
+    # right -- everything downstream just looks at the headers. There are no
+    # side channels.
+    def _clean_up_response_headers_for_sending(self, response: Response) -> Response:
+        assert type(response) is Response
+
+        headers = response.headers
+        need_close = False
+
+        # HEAD requests need some special handling: they always act like they
+        # have Content-Length: 0, and that's how _body_framing treats
+        # them. But their headers are supposed to match what we would send if
+        # the request was a GET. (Technically there is one deviation allowed:
+        # we're allowed to leave out the framing headers -- see
+        # https://tools.ietf.org/html/rfc7231#section-4.3.2 . But it's just as
+        # easy to get them right.)
+        method_for_choosing_headers = cast(bytes, self._request_method)
+        if method_for_choosing_headers == b"HEAD":
+            method_for_choosing_headers = b"GET"
+        framing_type, _ = _body_framing(method_for_choosing_headers, response)
+        if framing_type in ("chunked", "http/1.0"):
+            # This response has a body of unknown length.
+            # If our peer is HTTP/1.1, we use Transfer-Encoding: chunked
+            # If our peer is HTTP/1.0, we use no framing headers, and close the
+            # connection afterwards.
+            #
+            # Make sure to clear Content-Length (in principle user could have
+            # set both and then we ignored Content-Length b/c
+            # Transfer-Encoding overwrote it -- this would be naughty of them,
+            # but the HTTP spec says that if our peer does this then we have
+            # to fix it instead of erroring out, so we'll accord the user the
+            # same respect).
+            headers = set_comma_header(headers, b"content-length", [])
+            if self.their_http_version is None or self.their_http_version < b"1.1":
+                # Either we never got a valid request and are sending back an
+                # error (their_http_version is None), so we assume the worst;
+                # or else we did get a valid HTTP/1.0 request, so we know that
+                # they don't understand chunked encoding.
+                headers = set_comma_header(headers, b"transfer-encoding", [])
+                # This is actually redundant ATM, since currently we
+                # unconditionally disable keep-alive when talking to HTTP/1.0
+                # peers. But let's be defensive just in case we add
+                # Connection: keep-alive support later:
+                if self._request_method != b"HEAD":
+                    need_close = True
+            else:
+                headers = set_comma_header(headers, b"transfer-encoding", [b"chunked"])
+
+        if not self._cstate.keep_alive or need_close:
+            # Make sure Connection: close is set
+            connection = set(get_comma_header(headers, b"connection"))
+            connection.discard(b"keep-alive")
+            connection.add(b"close")
+            headers = set_comma_header(headers, b"connection", sorted(connection))
+
+        return Response(
+            headers=headers,
+            status_code=response.status_code,
+            http_version=response.http_version,
+            reason=response.reason,
+        )

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_events.py

@@ -0,0 +1,369 @@
+# High level events that make up HTTP/1.1 conversations. Loosely inspired by
+# the corresponding events in hyper-h2:
+#
+#     http://python-hyper.org/h2/en/stable/api.html#events
+#
+# Don't subclass these. Stuff will break.
+
+import re
+from abc import ABC
+from dataclasses import dataclass
+from typing import List, Tuple, Union
+
+from ._abnf import method, request_target
+from ._headers import Headers, normalize_and_validate
+from ._util import bytesify, LocalProtocolError, validate
+
+# Everything in __all__ gets re-exported as part of the h11 public API.
+__all__ = [
+    "Event",
+    "Request",
+    "InformationalResponse",
+    "Response",
+    "Data",
+    "EndOfMessage",
+    "ConnectionClosed",
+]
+
+method_re = re.compile(method.encode("ascii"))
+request_target_re = re.compile(request_target.encode("ascii"))
+
+
+class Event(ABC):
+    """
+    Base class for h11 events.
+    """
+
+    __slots__ = ()
+
+
+@dataclass(init=False, frozen=True)
+class Request(Event):
+    """The beginning of an HTTP request.
+
+    Fields:
+
+    .. attribute:: method
+
+       An HTTP method, e.g. ``b"GET"`` or ``b"POST"``. Always a byte
+       string. :term:`Bytes-like objects <bytes-like object>` and native
+       strings containing only ascii characters will be automatically
+       converted to byte strings.
+
+    .. attribute:: target
+
+       The target of an HTTP request, e.g. ``b"/index.html"``, or one of the
+       more exotic formats described in `RFC 7320, section 5.3
+       <https://tools.ietf.org/html/rfc7230#section-5.3>`_. Always a byte
+       string. :term:`Bytes-like objects <bytes-like object>` and native
+       strings containing only ascii characters will be automatically
+       converted to byte strings.
+
+    .. attribute:: headers
+
+       Request headers, represented as a list of (name, value) pairs. See
+       :ref:`the header normalization rules <headers-format>` for details.
+
+    .. attribute:: http_version
+
+       The HTTP protocol version, represented as a byte string like
+       ``b"1.1"``. See :ref:`the HTTP version normalization rules
+       <http_version-format>` for details.
+
+    """
+
+    __slots__ = ("method", "headers", "target", "http_version")
+
+    method: bytes
+    headers: Headers
+    target: bytes
+    http_version: bytes
+
+    def __init__(
+        self,
+        *,
+        method: Union[bytes, str],
+        headers: Union[Headers, List[Tuple[bytes, bytes]], List[Tuple[str, str]]],
+        target: Union[bytes, str],
+        http_version: Union[bytes, str] = b"1.1",
+        _parsed: bool = False,
+    ) -> None:
+        super().__init__()
+        if isinstance(headers, Headers):
+            object.__setattr__(self, "headers", headers)
+        else:
+            object.__setattr__(
+                self, "headers", normalize_and_validate(headers, _parsed=_parsed)
+            )
+        if not _parsed:
+            object.__setattr__(self, "method", bytesify(method))
+            object.__setattr__(self, "target", bytesify(target))
+            object.__setattr__(self, "http_version", bytesify(http_version))
+        else:
+            object.__setattr__(self, "method", method)
+            object.__setattr__(self, "target", target)
+            object.__setattr__(self, "http_version", http_version)
+
+        # "A server MUST respond with a 400 (Bad Request) status code to any
+        # HTTP/1.1 request message that lacks a Host header field and to any
+        # request message that contains more than one Host header field or a
+        # Host header field with an invalid field-value."
+        # -- https://tools.ietf.org/html/rfc7230#section-5.4
+        host_count = 0
+        for name, value in self.headers:
+            if name == b"host":
+                host_count += 1
+        if self.http_version == b"1.1" and host_count == 0:
+            raise LocalProtocolError("Missing mandatory Host: header")
+        if host_count > 1:
+            raise LocalProtocolError("Found multiple Host: headers")
+
+        validate(method_re, self.method, "Illegal method characters")
+        validate(request_target_re, self.target, "Illegal target characters")
+
+    # This is an unhashable type.
+    __hash__ = None  # type: ignore
+
+
+@dataclass(init=False, frozen=True)
+class _ResponseBase(Event):
+    __slots__ = ("headers", "http_version", "reason", "status_code")
+
+    headers: Headers
+    http_version: bytes
+    reason: bytes
+    status_code: int
+
+    def __init__(
+        self,
+        *,
+        headers: Union[Headers, List[Tuple[bytes, bytes]], List[Tuple[str, str]]],
+        status_code: int,
+        http_version: Union[bytes, str] = b"1.1",
+        reason: Union[bytes, str] = b"",
+        _parsed: bool = False,
+    ) -> None:
+        super().__init__()
+        if isinstance(headers, Headers):
+            object.__setattr__(self, "headers", headers)
+        else:
+            object.__setattr__(
+                self, "headers", normalize_and_validate(headers, _parsed=_parsed)
+            )
+        if not _parsed:
+            object.__setattr__(self, "reason", bytesify(reason))
+            object.__setattr__(self, "http_version", bytesify(http_version))
+            if not isinstance(status_code, int):
+                raise LocalProtocolError("status code must be integer")
+            # Because IntEnum objects are instances of int, but aren't
+            # duck-compatible (sigh), see gh-72.
+            object.__setattr__(self, "status_code", int(status_code))
+        else:
+            object.__setattr__(self, "reason", reason)
+            object.__setattr__(self, "http_version", http_version)
+            object.__setattr__(self, "status_code", status_code)
+
+        self.__post_init__()
+
+    def __post_init__(self) -> None:
+        pass
+
+    # This is an unhashable type.
+    __hash__ = None  # type: ignore
+
+
+@dataclass(init=False, frozen=True)
+class InformationalResponse(_ResponseBase):
+    """An HTTP informational response.
+
+    Fields:
+
+    .. attribute:: status_code
+
+       The status code of this response, as an integer. For an
+       :class:`InformationalResponse`, this is always in the range [100,
+       200).
+
+    .. attribute:: headers
+
+       Request headers, represented as a list of (name, value) pairs. See
+       :ref:`the header normalization rules <headers-format>` for
+       details.
+
+    .. attribute:: http_version
+
+       The HTTP protocol version, represented as a byte string like
+       ``b"1.1"``. See :ref:`the HTTP version normalization rules
+       <http_version-format>` for details.
+
+    .. attribute:: reason
+
+       The reason phrase of this response, as a byte string. For example:
+       ``b"OK"``, or ``b"Not Found"``.
+
+    """
+
+    def __post_init__(self) -> None:
+        if not (100 <= self.status_code < 200):
+            raise LocalProtocolError(
+                "InformationalResponse status_code should be in range "
+                "[100, 200), not {}".format(self.status_code)
+            )
+
+    # This is an unhashable type.
+    __hash__ = None  # type: ignore
+
+
+@dataclass(init=False, frozen=True)
+class Response(_ResponseBase):
+    """The beginning of an HTTP response.
+
+    Fields:
+
+    .. attribute:: status_code
+
+       The status code of this response, as an integer. For an
+       :class:`Response`, this is always in the range [200,
+       1000).
+
+    .. attribute:: headers
+
+       Request headers, represented as a list of (name, value) pairs. See
+       :ref:`the header normalization rules <headers-format>` for details.
+
+    .. attribute:: http_version
+
+       The HTTP protocol version, represented as a byte string like
+       ``b"1.1"``. See :ref:`the HTTP version normalization rules
+       <http_version-format>` for details.
+
+    .. attribute:: reason
+
+       The reason phrase of this response, as a byte string. For example:
+       ``b"OK"``, or ``b"Not Found"``.
+
+    """
+
+    def __post_init__(self) -> None:
+        if not (200 <= self.status_code < 1000):
+            raise LocalProtocolError(
+                "Response status_code should be in range [200, 1000), not {}".format(
+                    self.status_code
+                )
+            )
+
+    # This is an unhashable type.
+    __hash__ = None  # type: ignore
+
+
+@dataclass(init=False, frozen=True)
+class Data(Event):
+    """Part of an HTTP message body.
+
+    Fields:
+
+    .. attribute:: data
+
+       A :term:`bytes-like object` containing part of a message body. Or, if
+       using the ``combine=False`` argument to :meth:`Connection.send`, then
+       any object that your socket writing code knows what to do with, and for
+       which calling :func:`len` returns the number of bytes that will be
+       written -- see :ref:`sendfile` for details.
+
+    .. attribute:: chunk_start
+
+       A marker that indicates whether this data object is from the start of a
+       chunked transfer encoding chunk. This field is ignored when when a Data
+       event is provided to :meth:`Connection.send`: it is only valid on
+       events emitted from :meth:`Connection.next_event`. You probably
+       shouldn't use this attribute at all; see
+       :ref:`chunk-delimiters-are-bad` for details.
+
+    .. attribute:: chunk_end
+
+       A marker that indicates whether this data object is the last for a
+       given chunked transfer encoding chunk. This field is ignored when when
+       a Data event is provided to :meth:`Connection.send`: it is only valid
+       on events emitted from :meth:`Connection.next_event`. You probably
+       shouldn't use this attribute at all; see
+       :ref:`chunk-delimiters-are-bad` for details.
+
+    """
+
+    __slots__ = ("data", "chunk_start", "chunk_end")
+
+    data: bytes
+    chunk_start: bool
+    chunk_end: bool
+
+    def __init__(
+        self, data: bytes, chunk_start: bool = False, chunk_end: bool = False
+    ) -> None:
+        object.__setattr__(self, "data", data)
+        object.__setattr__(self, "chunk_start", chunk_start)
+        object.__setattr__(self, "chunk_end", chunk_end)
+
+    # This is an unhashable type.
+    __hash__ = None  # type: ignore
+
+
+# XX FIXME: "A recipient MUST ignore (or consider as an error) any fields that
+# are forbidden to be sent in a trailer, since processing them as if they were
+# present in the header section might bypass external security filters."
+# https://svn.tools.ietf.org/svn/wg/httpbis/specs/rfc7230.html#chunked.trailer.part
+# Unfortunately, the list of forbidden fields is long and vague :-/
+@dataclass(init=False, frozen=True)
+class EndOfMessage(Event):
+    """The end of an HTTP message.
+
+    Fields:
+
+    .. attribute:: headers
+
+       Default value: ``[]``
+
+       Any trailing headers attached to this message, represented as a list of
+       (name, value) pairs. See :ref:`the header normalization rules
+       <headers-format>` for details.
+
+       Must be empty unless ``Transfer-Encoding: chunked`` is in use.
+
+    """
+
+    __slots__ = ("headers",)
+
+    headers: Headers
+
+    def __init__(
+        self,
+        *,
+        headers: Union[
+            Headers, List[Tuple[bytes, bytes]], List[Tuple[str, str]], None
+        ] = None,
+        _parsed: bool = False,
+    ) -> None:
+        super().__init__()
+        if headers is None:
+            headers = Headers([])
+        elif not isinstance(headers, Headers):
+            headers = normalize_and_validate(headers, _parsed=_parsed)
+
+        object.__setattr__(self, "headers", headers)
+
+    # This is an unhashable type.
+    __hash__ = None  # type: ignore
+
+
+@dataclass(frozen=True)
+class ConnectionClosed(Event):
+    """This event indicates that the sender has closed their outgoing
+    connection.
+
+    Note that this does not necessarily mean that they can't *receive* further
+    data, because TCP connections are composed to two one-way channels which
+    can be closed independently. See :ref:`closing` for details.
+
+    No fields.
+    """
+
+    pass

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_headers.py

@@ -0,0 +1,282 @@
+import re
+from typing import AnyStr, cast, List, overload, Sequence, Tuple, TYPE_CHECKING, Union
+
+from ._abnf import field_name, field_value
+from ._util import bytesify, LocalProtocolError, validate
+
+if TYPE_CHECKING:
+    from ._events import Request
+
+try:
+    from typing import Literal
+except ImportError:
+    from typing_extensions import Literal  # type: ignore
+
+CONTENT_LENGTH_MAX_DIGITS = 20  # allow up to 1 billion TB - 1
+
+
+# Facts
+# -----
+#
+# Headers are:
+#   keys: case-insensitive ascii
+#   values: mixture of ascii and raw bytes
+#
+# "Historically, HTTP has allowed field content with text in the ISO-8859-1
+# charset [ISO-8859-1], supporting other charsets only through use of
+# [RFC2047] encoding.  In practice, most HTTP header field values use only a
+# subset of the US-ASCII charset [USASCII]. Newly defined header fields SHOULD
+# limit their field values to US-ASCII octets.  A recipient SHOULD treat other
+# octets in field content (obs-text) as opaque data."
+# And it deprecates all non-ascii values
+#
+# Leading/trailing whitespace in header names is forbidden
+#
+# Values get leading/trailing whitespace stripped
+#
+# Content-Disposition actually needs to contain unicode semantically; to
+# accomplish this it has a terrifically weird way of encoding the filename
+# itself as ascii (and even this still has lots of cross-browser
+# incompatibilities)
+#
+# Order is important:
+# "a proxy MUST NOT change the order of these field values when forwarding a
+# message"
+# (and there are several headers where the order indicates a preference)
+#
+# Multiple occurences of the same header:
+# "A sender MUST NOT generate multiple header fields with the same field name
+# in a message unless either the entire field value for that header field is
+# defined as a comma-separated list [or the header is Set-Cookie which gets a
+# special exception]" - RFC 7230. (cookies are in RFC 6265)
+#
+# So every header aside from Set-Cookie can be merged by b", ".join if it
+# occurs repeatedly. But, of course, they can't necessarily be split by
+# .split(b","), because quoting.
+#
+# Given all this mess (case insensitive, duplicates allowed, order is
+# important, ...), there doesn't appear to be any standard way to handle
+# headers in Python -- they're almost like dicts, but... actually just
+# aren't. For now we punt and just use a super simple representation: headers
+# are a list of pairs
+#
+#   [(name1, value1), (name2, value2), ...]
+#
+# where all entries are bytestrings, names are lowercase and have no
+# leading/trailing whitespace, and values are bytestrings with no
+# leading/trailing whitespace. Searching and updating are done via naive O(n)
+# methods.
+#
+# Maybe a dict-of-lists would be better?
+
+_content_length_re = re.compile(rb"[0-9]+")
+_field_name_re = re.compile(field_name.encode("ascii"))
+_field_value_re = re.compile(field_value.encode("ascii"))
+
+
+class Headers(Sequence[Tuple[bytes, bytes]]):
+    """
+    A list-like interface that allows iterating over headers as byte-pairs
+    of (lowercased-name, value).
+
+    Internally we actually store the representation as three-tuples,
+    including both the raw original casing, in order to preserve casing
+    over-the-wire, and the lowercased name, for case-insensitive comparisions.
+
+    r = Request(
+        method="GET",
+        target="/",
+        headers=[("Host", "example.org"), ("Connection", "keep-alive")],
+        http_version="1.1",
+    )
+    assert r.headers == [
+        (b"host", b"example.org"),
+        (b"connection", b"keep-alive")
+    ]
+    assert r.headers.raw_items() == [
+        (b"Host", b"example.org"),
+        (b"Connection", b"keep-alive")
+    ]
+    """
+
+    __slots__ = "_full_items"
+
+    def __init__(self, full_items: List[Tuple[bytes, bytes, bytes]]) -> None:
+        self._full_items = full_items
+
+    def __bool__(self) -> bool:
+        return bool(self._full_items)
+
+    def __eq__(self, other: object) -> bool:
+        return list(self) == list(other)  # type: ignore
+
+    def __len__(self) -> int:
+        return len(self._full_items)
+
+    def __repr__(self) -> str:
+        return "<Headers(%s)>" % repr(list(self))
+
+    def __getitem__(self, idx: int) -> Tuple[bytes, bytes]:  # type: ignore[override]
+        _, name, value = self._full_items[idx]
+        return (name, value)
+
+    def raw_items(self) -> List[Tuple[bytes, bytes]]:
+        return [(raw_name, value) for raw_name, _, value in self._full_items]
+
+
+HeaderTypes = Union[
+    List[Tuple[bytes, bytes]],
+    List[Tuple[bytes, str]],
+    List[Tuple[str, bytes]],
+    List[Tuple[str, str]],
+]
+
+
+@overload
+def normalize_and_validate(headers: Headers, _parsed: Literal[True]) -> Headers:
+    ...
+
+
+@overload
+def normalize_and_validate(headers: HeaderTypes, _parsed: Literal[False]) -> Headers:
+    ...
+
+
+@overload
+def normalize_and_validate(
+    headers: Union[Headers, HeaderTypes], _parsed: bool = False
+) -> Headers:
+    ...
+
+
+def normalize_and_validate(
+    headers: Union[Headers, HeaderTypes], _parsed: bool = False
+) -> Headers:
+    new_headers = []
+    seen_content_length = None
+    saw_transfer_encoding = False
+    for name, value in headers:
+        # For headers coming out of the parser, we can safely skip some steps,
+        # because it always returns bytes and has already run these regexes
+        # over the data:
+        if not _parsed:
+            name = bytesify(name)
+            value = bytesify(value)
+            validate(_field_name_re, name, "Illegal header name {!r}", name)
+            validate(_field_value_re, value, "Illegal header value {!r}", value)
+        assert isinstance(name, bytes)
+        assert isinstance(value, bytes)
+
+        raw_name = name
+        name = name.lower()
+        if name == b"content-length":
+            lengths = {length.strip() for length in value.split(b",")}
+            if len(lengths) != 1:
+                raise LocalProtocolError("conflicting Content-Length headers")
+            value = lengths.pop()
+            validate(_content_length_re, value, "bad Content-Length")
+            if len(value) > CONTENT_LENGTH_MAX_DIGITS:
+                raise LocalProtocolError("bad Content-Length")
+            if seen_content_length is None:
+                seen_content_length = value
+                new_headers.append((raw_name, name, value))
+            elif seen_content_length != value:
+                raise LocalProtocolError("conflicting Content-Length headers")
+        elif name == b"transfer-encoding":
+            # "A server that receives a request message with a transfer coding
+            # it does not understand SHOULD respond with 501 (Not
+            # Implemented)."
+            # https://tools.ietf.org/html/rfc7230#section-3.3.1
+            if saw_transfer_encoding:
+                raise LocalProtocolError(
+                    "multiple Transfer-Encoding headers", error_status_hint=501
+                )
+            # "All transfer-coding names are case-insensitive"
+            # -- https://tools.ietf.org/html/rfc7230#section-4
+            value = value.lower()
+            if value != b"chunked":
+                raise LocalProtocolError(
+                    "Only Transfer-Encoding: chunked is supported",
+                    error_status_hint=501,
+                )
+            saw_transfer_encoding = True
+            new_headers.append((raw_name, name, value))
+        else:
+            new_headers.append((raw_name, name, value))
+    return Headers(new_headers)
+
+
+def get_comma_header(headers: Headers, name: bytes) -> List[bytes]:
+    # Should only be used for headers whose value is a list of
+    # comma-separated, case-insensitive values.
+    #
+    # The header name `name` is expected to be lower-case bytes.
+    #
+    # Connection: meets these criteria (including cast insensitivity).
+    #
+    # Content-Length: technically is just a single value (1*DIGIT), but the
+    # standard makes reference to implementations that do multiple values, and
+    # using this doesn't hurt. Ditto, case insensitivity doesn't things either
+    # way.
+    #
+    # Transfer-Encoding: is more complex (allows for quoted strings), so
+    # splitting on , is actually wrong. For example, this is legal:
+    #
+    #    Transfer-Encoding: foo; options="1,2", chunked
+    #
+    # and should be parsed as
+    #
+    #    foo; options="1,2"
+    #    chunked
+    #
+    # but this naive function will parse it as
+    #
+    #    foo; options="1
+    #    2"
+    #    chunked
+    #
+    # However, this is okay because the only thing we are going to do with
+    # any Transfer-Encoding is reject ones that aren't just "chunked", so
+    # both of these will be treated the same anyway.
+    #
+    # Expect: the only legal value is the literal string
+    # "100-continue". Splitting on commas is harmless. Case insensitive.
+    #
+    out: List[bytes] = []
+    for _, found_name, found_raw_value in headers._full_items:
+        if found_name == name:
+            found_raw_value = found_raw_value.lower()
+            for found_split_value in found_raw_value.split(b","):
+                found_split_value = found_split_value.strip()
+                if found_split_value:
+                    out.append(found_split_value)
+    return out
+
+
+def set_comma_header(headers: Headers, name: bytes, new_values: List[bytes]) -> Headers:
+    # The header name `name` is expected to be lower-case bytes.
+    #
+    # Note that when we store the header we use title casing for the header
+    # names, in order to match the conventional HTTP header style.
+    #
+    # Simply calling `.title()` is a blunt approach, but it's correct
+    # here given the cases where we're using `set_comma_header`...
+    #
+    # Connection, Content-Length, Transfer-Encoding.
+    new_headers: List[Tuple[bytes, bytes]] = []
+    for found_raw_name, found_name, found_raw_value in headers._full_items:
+        if found_name != name:
+            new_headers.append((found_raw_name, found_raw_value))
+    for new_value in new_values:
+        new_headers.append((name.title(), new_value))
+    return normalize_and_validate(new_headers)
+
+
+def has_expect_100_continue(request: "Request") -> bool:
+    # https://tools.ietf.org/html/rfc7231#section-5.1.1
+    # "A server that receives a 100-continue expectation in an HTTP/1.0 request
+    # MUST ignore that expectation."
+    if request.http_version < b"1.1":
+        return False
+    expect = get_comma_header(request.headers, b"expect")
+    return b"100-continue" in expect

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_readers.py

@@ -0,0 +1,250 @@
+# Code to read HTTP data
+#
+# Strategy: each reader is a callable which takes a ReceiveBuffer object, and
+# either:
+# 1) consumes some of it and returns an Event
+# 2) raises a LocalProtocolError (for consistency -- e.g. we call validate()
+#    and it might raise a LocalProtocolError, so simpler just to always use
+#    this)
+# 3) returns None, meaning "I need more data"
+#
+# If they have a .read_eof attribute, then this will be called if an EOF is
+# received -- but this is optional. Either way, the actual ConnectionClosed
+# event will be generated afterwards.
+#
+# READERS is a dict describing how to pick a reader. It maps states to either:
+# - a reader
+# - or, for body readers, a dict of per-framing reader factories
+
+import re
+from typing import Any, Callable, Dict, Iterable, NoReturn, Optional, Tuple, Type, Union
+
+from ._abnf import chunk_header, header_field, request_line, status_line
+from ._events import Data, EndOfMessage, InformationalResponse, Request, Response
+from ._receivebuffer import ReceiveBuffer
+from ._state import (
+    CLIENT,
+    CLOSED,
+    DONE,
+    IDLE,
+    MUST_CLOSE,
+    SEND_BODY,
+    SEND_RESPONSE,
+    SERVER,
+)
+from ._util import LocalProtocolError, RemoteProtocolError, Sentinel, validate
+
+__all__ = ["READERS"]
+
+header_field_re = re.compile(header_field.encode("ascii"))
+obs_fold_re = re.compile(rb"[ \t]+")
+
+
+def _obsolete_line_fold(lines: Iterable[bytes]) -> Iterable[bytes]:
+    it = iter(lines)
+    last: Optional[bytes] = None
+    for line in it:
+        match = obs_fold_re.match(line)
+        if match:
+            if last is None:
+                raise LocalProtocolError("continuation line at start of headers")
+            if not isinstance(last, bytearray):
+                # Cast to a mutable type, avoiding copy on append to ensure O(n) time
+                last = bytearray(last)
+            last += b" "
+            last += line[match.end() :]
+        else:
+            if last is not None:
+                yield last
+            last = line
+    if last is not None:
+        yield last
+
+
+def _decode_header_lines(
+    lines: Iterable[bytes],
+) -> Iterable[Tuple[bytes, bytes]]:
+    for line in _obsolete_line_fold(lines):
+        matches = validate(header_field_re, line, "illegal header line: {!r}", line)
+        yield (matches["field_name"], matches["field_value"])
+
+
+request_line_re = re.compile(request_line.encode("ascii"))
+
+
+def maybe_read_from_IDLE_client(buf: ReceiveBuffer) -> Optional[Request]:
+    lines = buf.maybe_extract_lines()
+    if lines is None:
+        if buf.is_next_line_obviously_invalid_request_line():
+            raise LocalProtocolError("illegal request line")
+        return None
+    if not lines:
+        raise LocalProtocolError("no request line received")
+    matches = validate(
+        request_line_re, lines[0], "illegal request line: {!r}", lines[0]
+    )
+    return Request(
+        headers=list(_decode_header_lines(lines[1:])), _parsed=True, **matches
+    )
+
+
+status_line_re = re.compile(status_line.encode("ascii"))
+
+
+def maybe_read_from_SEND_RESPONSE_server(
+    buf: ReceiveBuffer,
+) -> Union[InformationalResponse, Response, None]:
+    lines = buf.maybe_extract_lines()
+    if lines is None:
+        if buf.is_next_line_obviously_invalid_request_line():
+            raise LocalProtocolError("illegal request line")
+        return None
+    if not lines:
+        raise LocalProtocolError("no response line received")
+    matches = validate(status_line_re, lines[0], "illegal status line: {!r}", lines[0])
+    http_version = (
+        b"1.1" if matches["http_version"] is None else matches["http_version"]
+    )
+    reason = b"" if matches["reason"] is None else matches["reason"]
+    status_code = int(matches["status_code"])
+    class_: Union[Type[InformationalResponse], Type[Response]] = (
+        InformationalResponse if status_code < 200 else Response
+    )
+    return class_(
+        headers=list(_decode_header_lines(lines[1:])),
+        _parsed=True,
+        status_code=status_code,
+        reason=reason,
+        http_version=http_version,
+    )
+
+
+class ContentLengthReader:
+    def __init__(self, length: int) -> None:
+        self._length = length
+        self._remaining = length
+
+    def __call__(self, buf: ReceiveBuffer) -> Union[Data, EndOfMessage, None]:
+        if self._remaining == 0:
+            return EndOfMessage()
+        data = buf.maybe_extract_at_most(self._remaining)
+        if data is None:
+            return None
+        self._remaining -= len(data)
+        return Data(data=data)
+
+    def read_eof(self) -> NoReturn:
+        raise RemoteProtocolError(
+            "peer closed connection without sending complete message body "
+            "(received {} bytes, expected {})".format(
+                self._length - self._remaining, self._length
+            )
+        )
+
+
+chunk_header_re = re.compile(chunk_header.encode("ascii"))
+
+
+class ChunkedReader:
+    def __init__(self) -> None:
+        self._bytes_in_chunk = 0
+        # After reading a chunk, we have to throw away the trailing \r\n.
+        # This tracks the bytes that we need to match and throw away.
+        self._bytes_to_discard = b""
+        self._reading_trailer = False
+
+    def __call__(self, buf: ReceiveBuffer) -> Union[Data, EndOfMessage, None]:
+        if self._reading_trailer:
+            lines = buf.maybe_extract_lines()
+            if lines is None:
+                return None
+            return EndOfMessage(headers=list(_decode_header_lines(lines)))
+        if self._bytes_to_discard:
+            data = buf.maybe_extract_at_most(len(self._bytes_to_discard))
+            if data is None:
+                return None
+            if data != self._bytes_to_discard[: len(data)]:
+                raise LocalProtocolError(
+                    f"malformed chunk footer: {data!r} (expected {self._bytes_to_discard!r})"
+                )
+            self._bytes_to_discard = self._bytes_to_discard[len(data) :]
+            if self._bytes_to_discard:
+                return None
+            # else, fall through and read some more
+        assert self._bytes_to_discard == b""
+        if self._bytes_in_chunk == 0:
+            # We need to refill our chunk count
+            chunk_header = buf.maybe_extract_next_line()
+            if chunk_header is None:
+                return None
+            matches = validate(
+                chunk_header_re,
+                chunk_header,
+                "illegal chunk header: {!r}",
+                chunk_header,
+            )
+            # XX FIXME: we discard chunk extensions. Does anyone care?
+            self._bytes_in_chunk = int(matches["chunk_size"], base=16)
+            if self._bytes_in_chunk == 0:
+                self._reading_trailer = True
+                return self(buf)
+            chunk_start = True
+        else:
+            chunk_start = False
+        assert self._bytes_in_chunk > 0
+        data = buf.maybe_extract_at_most(self._bytes_in_chunk)
+        if data is None:
+            return None
+        self._bytes_in_chunk -= len(data)
+        if self._bytes_in_chunk == 0:
+            self._bytes_to_discard = b"\r\n"
+            chunk_end = True
+        else:
+            chunk_end = False
+        return Data(data=data, chunk_start=chunk_start, chunk_end=chunk_end)
+
+    def read_eof(self) -> NoReturn:
+        raise RemoteProtocolError(
+            "peer closed connection without sending complete message body "
+            "(incomplete chunked read)"
+        )
+
+
+class Http10Reader:
+    def __call__(self, buf: ReceiveBuffer) -> Optional[Data]:
+        data = buf.maybe_extract_at_most(999999999)
+        if data is None:
+            return None
+        return Data(data=data)
+
+    def read_eof(self) -> EndOfMessage:
+        return EndOfMessage()
+
+
+def expect_nothing(buf: ReceiveBuffer) -> None:
+    if buf:
+        raise LocalProtocolError("Got data when expecting EOF")
+    return None
+
+
+ReadersType = Dict[
+    Union[Type[Sentinel], Tuple[Type[Sentinel], Type[Sentinel]]],
+    Union[Callable[..., Any], Dict[str, Callable[..., Any]]],
+]
+
+READERS: ReadersType = {
+    (CLIENT, IDLE): maybe_read_from_IDLE_client,
+    (SERVER, IDLE): maybe_read_from_SEND_RESPONSE_server,
+    (SERVER, SEND_RESPONSE): maybe_read_from_SEND_RESPONSE_server,
+    (CLIENT, DONE): expect_nothing,
+    (CLIENT, MUST_CLOSE): expect_nothing,
+    (CLIENT, CLOSED): expect_nothing,
+    (SERVER, DONE): expect_nothing,
+    (SERVER, MUST_CLOSE): expect_nothing,
+    (SERVER, CLOSED): expect_nothing,
+    SEND_BODY: {
+        "chunked": ChunkedReader,
+        "content-length": ContentLengthReader,
+        "http/1.0": Http10Reader,
+    },
+}

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_receivebuffer.py

@@ -0,0 +1,153 @@
+import re
+import sys
+from typing import List, Optional, Union
+
+__all__ = ["ReceiveBuffer"]
+
+
+# Operations we want to support:
+# - find next \r\n or \r\n\r\n (\n or \n\n are also acceptable),
+#   or wait until there is one
+# - read at-most-N bytes
+# Goals:
+# - on average, do this fast
+# - worst case, do this in O(n) where n is the number of bytes processed
+# Plan:
+# - store bytearray, offset, how far we've searched for a separator token
+# - use the how-far-we've-searched data to avoid rescanning
+# - while doing a stream of uninterrupted processing, advance offset instead
+#   of constantly copying
+# WARNING:
+# - I haven't benchmarked or profiled any of this yet.
+#
+# Note that starting in Python 3.4, deleting the initial n bytes from a
+# bytearray is amortized O(n), thanks to some excellent work by Antoine
+# Martin:
+#
+#     https://bugs.python.org/issue19087
+#
+# This means that if we only supported 3.4+, we could get rid of the code here
+# involving self._start and self.compress, because it's doing exactly the same
+# thing that bytearray now does internally.
+#
+# BUT unfortunately, we still support 2.7, and reading short segments out of a
+# long buffer MUST be O(bytes read) to avoid DoS issues, so we can't actually
+# delete this code. Yet:
+#
+#     https://pythonclock.org/
+#
+# (Two things to double-check first though: make sure PyPy also has the
+# optimization, and benchmark to make sure it's a win, since we do have a
+# slightly clever thing where we delay calling compress() until we've
+# processed a whole event, which could in theory be slightly more efficient
+# than the internal bytearray support.)
+blank_line_regex = re.compile(b"\n\r?\n", re.MULTILINE)
+
+
+class ReceiveBuffer:
+    def __init__(self) -> None:
+        self._data = bytearray()
+        self._next_line_search = 0
+        self._multiple_lines_search = 0
+
+    def __iadd__(self, byteslike: Union[bytes, bytearray]) -> "ReceiveBuffer":
+        self._data += byteslike
+        return self
+
+    def __bool__(self) -> bool:
+        return bool(len(self))
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    # for @property unprocessed_data
+    def __bytes__(self) -> bytes:
+        return bytes(self._data)
+
+    def _extract(self, count: int) -> bytearray:
+        # extracting an initial slice of the data buffer and return it
+        out = self._data[:count]
+        del self._data[:count]
+
+        self._next_line_search = 0
+        self._multiple_lines_search = 0
+
+        return out
+
+    def maybe_extract_at_most(self, count: int) -> Optional[bytearray]:
+        """
+        Extract a fixed number of bytes from the buffer.
+        """
+        out = self._data[:count]
+        if not out:
+            return None
+
+        return self._extract(count)
+
+    def maybe_extract_next_line(self) -> Optional[bytearray]:
+        """
+        Extract the first line, if it is completed in the buffer.
+        """
+        # Only search in buffer space that we've not already looked at.
+        search_start_index = max(0, self._next_line_search - 1)
+        partial_idx = self._data.find(b"\r\n", search_start_index)
+
+        if partial_idx == -1:
+            self._next_line_search = len(self._data)
+            return None
+
+        # + 2 is to compensate len(b"\r\n")
+        idx = partial_idx + 2
+
+        return self._extract(idx)
+
+    def maybe_extract_lines(self) -> Optional[List[bytearray]]:
+        """
+        Extract everything up to the first blank line, and return a list of lines.
+        """
+        # Handle the case where we have an immediate empty line.
+        if self._data[:1] == b"\n":
+            self._extract(1)
+            return []
+
+        if self._data[:2] == b"\r\n":
+            self._extract(2)
+            return []
+
+        # Only search in buffer space that we've not already looked at.
+        match = blank_line_regex.search(self._data, self._multiple_lines_search)
+        if match is None:
+            self._multiple_lines_search = max(0, len(self._data) - 2)
+            return None
+
+        # Truncate the buffer and return it.
+        idx = match.span(0)[-1]
+        out = self._extract(idx)
+        lines = out.split(b"\n")
+
+        for line in lines:
+            if line.endswith(b"\r"):
+                del line[-1]
+
+        assert lines[-2] == lines[-1] == b""
+
+        del lines[-2:]
+
+        return lines
+
+    # In theory we should wait until `\r\n` before starting to validate
+    # incoming data. However it's interesting to detect (very) invalid data
+    # early given they might not even contain `\r\n` at all (hence only
+    # timeout will get rid of them).
+    # This is not a 100% effective detection but more of a cheap sanity check
+    # allowing for early abort in some useful cases.
+    # This is especially interesting when peer is messing up with HTTPS and
+    # sent us a TLS stream where we were expecting plain HTTP given all
+    # versions of TLS so far start handshake with a 0x16 message type code.
+    def is_next_line_obviously_invalid_request_line(self) -> bool:
+        try:
+            # HTTP header line must not contain non-printable characters
+            # and should not start with a space
+            return self._data[0] < 0x21
+        except IndexError:
+            return False

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_state.py

@@ -0,0 +1,365 @@
+################################################################
+# The core state machine
+################################################################
+#
+# Rule 1: everything that affects the state machine and state transitions must
+# live here in this file. As much as possible goes into the table-based
+# representation, but for the bits that don't quite fit, the actual code and
+# state must nonetheless live here.
+#
+# Rule 2: this file does not know about what role we're playing; it only knows
+# about HTTP request/response cycles in the abstract. This ensures that we
+# don't cheat and apply different rules to local and remote parties.
+#
+#
+# Theory of operation
+# ===================
+#
+# Possibly the simplest way to think about this is that we actually have 5
+# different state machines here. Yes, 5. These are:
+#
+# 1) The client state, with its complicated automaton (see the docs)
+# 2) The server state, with its complicated automaton (see the docs)
+# 3) The keep-alive state, with possible states {True, False}
+# 4) The SWITCH_CONNECT state, with possible states {False, True}
+# 5) The SWITCH_UPGRADE state, with possible states {False, True}
+#
+# For (3)-(5), the first state listed is the initial state.
+#
+# (1)-(3) are stored explicitly in member variables. The last
+# two are stored implicitly in the pending_switch_proposals set as:
+#   (state of 4) == (_SWITCH_CONNECT in pending_switch_proposals)
+#   (state of 5) == (_SWITCH_UPGRADE in pending_switch_proposals)
+#
+# And each of these machines has two different kinds of transitions:
+#
+# a) Event-triggered
+# b) State-triggered
+#
+# Event triggered is the obvious thing that you'd think it is: some event
+# happens, and if it's the right event at the right time then a transition
+# happens. But there are somewhat complicated rules for which machines can
+# "see" which events. (As a rule of thumb, if a machine "sees" an event, this
+# means two things: the event can affect the machine, and if the machine is
+# not in a state where it expects that event then it's an error.) These rules
+# are:
+#
+# 1) The client machine sees all h11.events objects emitted by the client.
+#
+# 2) The server machine sees all h11.events objects emitted by the server.
+#
+#    It also sees the client's Request event.
+#
+#    And sometimes, server events are annotated with a _SWITCH_* event. For
+#    example, we can have a (Response, _SWITCH_CONNECT) event, which is
+#    different from a regular Response event.
+#
+# 3) The keep-alive machine sees the process_keep_alive_disabled() event
+#    (which is derived from Request/Response events), and this event
+#    transitions it from True -> False, or from False -> False. There's no way
+#    to transition back.
+#
+# 4&5) The _SWITCH_* machines transition from False->True when we get a
+#    Request that proposes the relevant type of switch (via
+#    process_client_switch_proposals), and they go from True->False when we
+#    get a Response that has no _SWITCH_* annotation.
+#
+# So that's event-triggered transitions.
+#
+# State-triggered transitions are less standard. What they do here is couple
+# the machines together. The way this works is, when certain *joint*
+# configurations of states are achieved, then we automatically transition to a
+# new *joint* state. So, for example, if we're ever in a joint state with
+#
+#   client: DONE
+#   keep-alive: False
+#
+# then the client state immediately transitions to:
+#
+#   client: MUST_CLOSE
+#
+# This is fundamentally different from an event-based transition, because it
+# doesn't matter how we arrived at the {client: DONE, keep-alive: False} state
+# -- maybe the client transitioned SEND_BODY -> DONE, or keep-alive
+# transitioned True -> False. Either way, once this precondition is satisfied,
+# this transition is immediately triggered.
+#
+# What if two conflicting state-based transitions get enabled at the same
+# time?  In practice there's only one case where this arises (client DONE ->
+# MIGHT_SWITCH_PROTOCOL versus DONE -> MUST_CLOSE), and we resolve it by
+# explicitly prioritizing the DONE -> MIGHT_SWITCH_PROTOCOL transition.
+#
+# Implementation
+# --------------
+#
+# The event-triggered transitions for the server and client machines are all
+# stored explicitly in a table. Ditto for the state-triggered transitions that
+# involve just the server and client state.
+#
+# The transitions for the other machines, and the state-triggered transitions
+# that involve the other machines, are written out as explicit Python code.
+#
+# It'd be nice if there were some cleaner way to do all this. This isn't
+# *too* terrible, but I feel like it could probably be better.
+#
+# WARNING
+# -------
+#
+# The script that generates the state machine diagrams for the docs knows how
+# to read out the EVENT_TRIGGERED_TRANSITIONS and STATE_TRIGGERED_TRANSITIONS
+# tables. But it can't automatically read the transitions that are written
+# directly in Python code. So if you touch those, you need to also update the
+# script to keep it in sync!
+from typing import cast, Dict, Optional, Set, Tuple, Type, Union
+
+from ._events import *
+from ._util import LocalProtocolError, Sentinel
+
+# Everything in __all__ gets re-exported as part of the h11 public API.
+__all__ = [
+    "CLIENT",
+    "SERVER",
+    "IDLE",
+    "SEND_RESPONSE",
+    "SEND_BODY",
+    "DONE",
+    "MUST_CLOSE",
+    "CLOSED",
+    "MIGHT_SWITCH_PROTOCOL",
+    "SWITCHED_PROTOCOL",
+    "ERROR",
+]
+
+
+class CLIENT(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class SERVER(Sentinel, metaclass=Sentinel):
+    pass
+
+
+# States
+class IDLE(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class SEND_RESPONSE(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class SEND_BODY(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class DONE(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class MUST_CLOSE(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class CLOSED(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class ERROR(Sentinel, metaclass=Sentinel):
+    pass
+
+
+# Switch types
+class MIGHT_SWITCH_PROTOCOL(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class SWITCHED_PROTOCOL(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class _SWITCH_UPGRADE(Sentinel, metaclass=Sentinel):
+    pass
+
+
+class _SWITCH_CONNECT(Sentinel, metaclass=Sentinel):
+    pass
+
+
+EventTransitionType = Dict[
+    Type[Sentinel],
+    Dict[
+        Type[Sentinel],
+        Dict[Union[Type[Event], Tuple[Type[Event], Type[Sentinel]]], Type[Sentinel]],
+    ],
+]
+
+EVENT_TRIGGERED_TRANSITIONS: EventTransitionType = {
+    CLIENT: {
+        IDLE: {Request: SEND_BODY, ConnectionClosed: CLOSED},
+        SEND_BODY: {Data: SEND_BODY, EndOfMessage: DONE},
+        DONE: {ConnectionClosed: CLOSED},
+        MUST_CLOSE: {ConnectionClosed: CLOSED},
+        CLOSED: {ConnectionClosed: CLOSED},
+        MIGHT_SWITCH_PROTOCOL: {},
+        SWITCHED_PROTOCOL: {},
+        ERROR: {},
+    },
+    SERVER: {
+        IDLE: {
+            ConnectionClosed: CLOSED,
+            Response: SEND_BODY,
+            # Special case: server sees client Request events, in this form
+            (Request, CLIENT): SEND_RESPONSE,
+        },
+        SEND_RESPONSE: {
+            InformationalResponse: SEND_RESPONSE,
+            Response: SEND_BODY,
+            (InformationalResponse, _SWITCH_UPGRADE): SWITCHED_PROTOCOL,
+            (Response, _SWITCH_CONNECT): SWITCHED_PROTOCOL,
+        },
+        SEND_BODY: {Data: SEND_BODY, EndOfMessage: DONE},
+        DONE: {ConnectionClosed: CLOSED},
+        MUST_CLOSE: {ConnectionClosed: CLOSED},
+        CLOSED: {ConnectionClosed: CLOSED},
+        SWITCHED_PROTOCOL: {},
+        ERROR: {},
+    },
+}
+
+StateTransitionType = Dict[
+    Tuple[Type[Sentinel], Type[Sentinel]], Dict[Type[Sentinel], Type[Sentinel]]
+]
+
+# NB: there are also some special-case state-triggered transitions hard-coded
+# into _fire_state_triggered_transitions below.
+STATE_TRIGGERED_TRANSITIONS: StateTransitionType = {
+    # (Client state, Server state) -> new states
+    # Protocol negotiation
+    (MIGHT_SWITCH_PROTOCOL, SWITCHED_PROTOCOL): {CLIENT: SWITCHED_PROTOCOL},
+    # Socket shutdown
+    (CLOSED, DONE): {SERVER: MUST_CLOSE},
+    (CLOSED, IDLE): {SERVER: MUST_CLOSE},
+    (ERROR, DONE): {SERVER: MUST_CLOSE},
+    (DONE, CLOSED): {CLIENT: MUST_CLOSE},
+    (IDLE, CLOSED): {CLIENT: MUST_CLOSE},
+    (DONE, ERROR): {CLIENT: MUST_CLOSE},
+}
+
+
+class ConnectionState:
+    def __init__(self) -> None:
+        # Extra bits of state that don't quite fit into the state model.
+
+        # If this is False then it enables the automatic DONE -> MUST_CLOSE
+        # transition. Don't set this directly; call .keep_alive_disabled()
+        self.keep_alive = True
+
+        # This is a subset of {UPGRADE, CONNECT}, containing the proposals
+        # made by the client for switching protocols.
+        self.pending_switch_proposals: Set[Type[Sentinel]] = set()
+
+        self.states: Dict[Type[Sentinel], Type[Sentinel]] = {CLIENT: IDLE, SERVER: IDLE}
+
+    def process_error(self, role: Type[Sentinel]) -> None:
+        self.states[role] = ERROR
+        self._fire_state_triggered_transitions()
+
+    def process_keep_alive_disabled(self) -> None:
+        self.keep_alive = False
+        self._fire_state_triggered_transitions()
+
+    def process_client_switch_proposal(self, switch_event: Type[Sentinel]) -> None:
+        self.pending_switch_proposals.add(switch_event)
+        self._fire_state_triggered_transitions()
+
+    def process_event(
+        self,
+        role: Type[Sentinel],
+        event_type: Type[Event],
+        server_switch_event: Optional[Type[Sentinel]] = None,
+    ) -> None:
+        _event_type: Union[Type[Event], Tuple[Type[Event], Type[Sentinel]]] = event_type
+        if server_switch_event is not None:
+            assert role is SERVER
+            if server_switch_event not in self.pending_switch_proposals:
+                raise LocalProtocolError(
+                    "Received server _SWITCH_UPGRADE event without a pending proposal"
+                )
+            _event_type = (event_type, server_switch_event)
+        if server_switch_event is None and _event_type is Response:
+            self.pending_switch_proposals = set()
+        self._fire_event_triggered_transitions(role, _event_type)
+        # Special case: the server state does get to see Request
+        # events.
+        if _event_type is Request:
+            assert role is CLIENT
+            self._fire_event_triggered_transitions(SERVER, (Request, CLIENT))
+        self._fire_state_triggered_transitions()
+
+    def _fire_event_triggered_transitions(
+        self,
+        role: Type[Sentinel],
+        event_type: Union[Type[Event], Tuple[Type[Event], Type[Sentinel]]],
+    ) -> None:
+        state = self.states[role]
+        try:
+            new_state = EVENT_TRIGGERED_TRANSITIONS[role][state][event_type]
+        except KeyError:
+            event_type = cast(Type[Event], event_type)
+            raise LocalProtocolError(
+                "can't handle event type {} when role={} and state={}".format(
+                    event_type.__name__, role, self.states[role]
+                )
+            ) from None
+        self.states[role] = new_state
+
+    def _fire_state_triggered_transitions(self) -> None:
+        # We apply these rules repeatedly until converging on a fixed point
+        while True:
+            start_states = dict(self.states)
+
+            # It could happen that both these special-case transitions are
+            # enabled at the same time:
+            #
+            #    DONE -> MIGHT_SWITCH_PROTOCOL
+            #    DONE -> MUST_CLOSE
+            #
+            # For example, this will always be true of a HTTP/1.0 client
+            # requesting CONNECT.  If this happens, the protocol switch takes
+            # priority. From there the client will either go to
+            # SWITCHED_PROTOCOL, in which case it's none of our business when
+            # they close the connection, or else the server will deny the
+            # request, in which case the client will go back to DONE and then
+            # from there to MUST_CLOSE.
+            if self.pending_switch_proposals:
+                if self.states[CLIENT] is DONE:
+                    self.states[CLIENT] = MIGHT_SWITCH_PROTOCOL
+
+            if not self.pending_switch_proposals:
+                if self.states[CLIENT] is MIGHT_SWITCH_PROTOCOL:
+                    self.states[CLIENT] = DONE
+
+            if not self.keep_alive:
+                for role in (CLIENT, SERVER):
+                    if self.states[role] is DONE:
+                        self.states[role] = MUST_CLOSE
+
+            # Tabular state-triggered transitions
+            joint_state = (self.states[CLIENT], self.states[SERVER])
+            changes = STATE_TRIGGERED_TRANSITIONS.get(joint_state, {})
+            self.states.update(changes)
+
+            if self.states == start_states:
+                # Fixed point reached
+                return
+
+    def start_next_cycle(self) -> None:
+        if self.states != {CLIENT: DONE, SERVER: DONE}:
+            raise LocalProtocolError(
+                f"not in a reusable state. self.states={self.states}"
+            )
+        # Can't reach DONE/DONE with any of these active, but still, let's be
+        # sure.
+        assert self.keep_alive
+        assert not self.pending_switch_proposals
+        self.states = {CLIENT: IDLE, SERVER: IDLE}

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_util.py

@@ -0,0 +1,135 @@
+from typing import Any, Dict, NoReturn, Pattern, Tuple, Type, TypeVar, Union
+
+__all__ = [
+    "ProtocolError",
+    "LocalProtocolError",
+    "RemoteProtocolError",
+    "validate",
+    "bytesify",
+]
+
+
+class ProtocolError(Exception):
+    """Exception indicating a violation of the HTTP/1.1 protocol.
+
+    This as an abstract base class, with two concrete base classes:
+    :exc:`LocalProtocolError`, which indicates that you tried to do something
+    that HTTP/1.1 says is illegal, and :exc:`RemoteProtocolError`, which
+    indicates that the remote peer tried to do something that HTTP/1.1 says is
+    illegal. See :ref:`error-handling` for details.
+
+    In addition to the normal :exc:`Exception` features, it has one attribute:
+
+    .. attribute:: error_status_hint
+
+       This gives a suggestion as to what status code a server might use if
+       this error occurred as part of a request.
+
+       For a :exc:`RemoteProtocolError`, this is useful as a suggestion for
+       how you might want to respond to a misbehaving peer, if you're
+       implementing a server.
+
+       For a :exc:`LocalProtocolError`, this can be taken as a suggestion for
+       how your peer might have responded to *you* if h11 had allowed you to
+       continue.
+
+       The default is 400 Bad Request, a generic catch-all for protocol
+       violations.
+
+    """
+
+    def __init__(self, msg: str, error_status_hint: int = 400) -> None:
+        if type(self) is ProtocolError:
+            raise TypeError("tried to directly instantiate ProtocolError")
+        Exception.__init__(self, msg)
+        self.error_status_hint = error_status_hint
+
+
+# Strategy: there are a number of public APIs where a LocalProtocolError can
+# be raised (send(), all the different event constructors, ...), and only one
+# public API where RemoteProtocolError can be raised
+# (receive_data()). Therefore we always raise LocalProtocolError internally,
+# and then receive_data will translate this into a RemoteProtocolError.
+#
+# Internally:
+#   LocalProtocolError is the generic "ProtocolError".
+# Externally:
+#   LocalProtocolError is for local errors and RemoteProtocolError is for
+#   remote errors.
+class LocalProtocolError(ProtocolError):
+    def _reraise_as_remote_protocol_error(self) -> NoReturn:
+        # After catching a LocalProtocolError, use this method to re-raise it
+        # as a RemoteProtocolError. This method must be called from inside an
+        # except: block.
+        #
+        # An easy way to get an equivalent RemoteProtocolError is just to
+        # modify 'self' in place.
+        self.__class__ = RemoteProtocolError  # type: ignore
+        # But the re-raising is somewhat non-trivial -- you might think that
+        # now that we've modified the in-flight exception object, that just
+        # doing 'raise' to re-raise it would be enough. But it turns out that
+        # this doesn't work, because Python tracks the exception type
+        # (exc_info[0]) separately from the exception object (exc_info[1]),
+        # and we only modified the latter. So we really do need to re-raise
+        # the new type explicitly.
+        # On py3, the traceback is part of the exception object, so our
+        # in-place modification preserved it and we can just re-raise:
+        raise self
+
+
+class RemoteProtocolError(ProtocolError):
+    pass
+
+
+def validate(
+    regex: Pattern[bytes], data: bytes, msg: str = "malformed data", *format_args: Any
+) -> Dict[str, bytes]:
+    match = regex.fullmatch(data)
+    if not match:
+        if format_args:
+            msg = msg.format(*format_args)
+        raise LocalProtocolError(msg)
+    return match.groupdict()
+
+
+# Sentinel values
+#
+# - Inherit identity-based comparison and hashing from object
+# - Have a nice repr
+# - Have a *bonus property*: type(sentinel) is sentinel
+#
+# The bonus property is useful if you want to take the return value from
+# next_event() and do some sort of dispatch based on type(event).
+
+_T_Sentinel = TypeVar("_T_Sentinel", bound="Sentinel")
+
+
+class Sentinel(type):
+    def __new__(
+        cls: Type[_T_Sentinel],
+        name: str,
+        bases: Tuple[type, ...],
+        namespace: Dict[str, Any],
+        **kwds: Any
+    ) -> _T_Sentinel:
+        assert bases == (Sentinel,)
+        v = super().__new__(cls, name, bases, namespace, **kwds)
+        v.__class__ = v  # type: ignore
+        return v
+
+    def __repr__(self) -> str:
+        return self.__name__
+
+
+# Used for methods, request targets, HTTP versions, header names, and header
+# values. Accepts ascii-strings, or bytes/bytearray/memoryview/..., and always
+# returns bytes.
+def bytesify(s: Union[bytes, bytearray, memoryview, int, str]) -> bytes:
+    # Fast-path:
+    if type(s) is bytes:
+        return s
+    if isinstance(s, str):
+        s = s.encode("ascii")
+    if isinstance(s, int):
+        raise TypeError("expected bytes-like object, not int")
+    return bytes(s)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_version.py

@@ -0,0 +1,16 @@
+# This file must be kept very simple, because it is consumed from several
+# places -- it is imported by h11/__init__.py, execfile'd by setup.py, etc.
+
+# We use a simple scheme:
+#   1.0.0 -> 1.0.0+dev -> 1.1.0 -> 1.1.0+dev
+# where the +dev versions are never released into the wild, they're just what
+# we stick into the VCS in between releases.
+#
+# This is compatible with PEP 440:
+#   http://legacy.python.org/dev/peps/pep-0440/
+# via the use of the "local suffix" "+dev", which is disallowed on index
+# servers and causes 1.0.0+dev to sort after plain 1.0.0, which is what we
+# want. (Contrast with the special suffix 1.0.0.dev, which sorts *before*
+# 1.0.0.)
+
+__version__ = "0.16.0"

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/_writers.py

@@ -0,0 +1,145 @@
+# Code to read HTTP data
+#
+# Strategy: each writer takes an event + a write-some-bytes function, which is
+# calls.
+#
+# WRITERS is a dict describing how to pick a reader. It maps states to either:
+# - a writer
+# - or, for body writers, a dict of framin-dependent writer factories
+
+from typing import Any, Callable, Dict, List, Tuple, Type, Union
+
+from ._events import Data, EndOfMessage, Event, InformationalResponse, Request, Response
+from ._headers import Headers
+from ._state import CLIENT, IDLE, SEND_BODY, SEND_RESPONSE, SERVER
+from ._util import LocalProtocolError, Sentinel
+
+__all__ = ["WRITERS"]
+
+Writer = Callable[[bytes], Any]
+
+
+def write_headers(headers: Headers, write: Writer) -> None:
+    # "Since the Host field-value is critical information for handling a
+    # request, a user agent SHOULD generate Host as the first header field
+    # following the request-line." - RFC 7230
+    raw_items = headers._full_items
+    for raw_name, name, value in raw_items:
+        if name == b"host":
+            write(b"%s: %s\r\n" % (raw_name, value))
+    for raw_name, name, value in raw_items:
+        if name != b"host":
+            write(b"%s: %s\r\n" % (raw_name, value))
+    write(b"\r\n")
+
+
+def write_request(request: Request, write: Writer) -> None:
+    if request.http_version != b"1.1":
+        raise LocalProtocolError("I only send HTTP/1.1")
+    write(b"%s %s HTTP/1.1\r\n" % (request.method, request.target))
+    write_headers(request.headers, write)
+
+
+# Shared between InformationalResponse and Response
+def write_any_response(
+    response: Union[InformationalResponse, Response], write: Writer
+) -> None:
+    if response.http_version != b"1.1":
+        raise LocalProtocolError("I only send HTTP/1.1")
+    status_bytes = str(response.status_code).encode("ascii")
+    # We don't bother sending ascii status messages like "OK"; they're
+    # optional and ignored by the protocol. (But the space after the numeric
+    # status code is mandatory.)
+    #
+    # XX FIXME: could at least make an effort to pull out the status message
+    # from stdlib's http.HTTPStatus table. Or maybe just steal their enums
+    # (either by import or copy/paste). We already accept them as status codes
+    # since they're of type IntEnum < int.
+    write(b"HTTP/1.1 %s %s\r\n" % (status_bytes, response.reason))
+    write_headers(response.headers, write)
+
+
+class BodyWriter:
+    def __call__(self, event: Event, write: Writer) -> None:
+        if type(event) is Data:
+            self.send_data(event.data, write)
+        elif type(event) is EndOfMessage:
+            self.send_eom(event.headers, write)
+        else:  # pragma: no cover
+            assert False
+
+    def send_data(self, data: bytes, write: Writer) -> None:
+        pass
+
+    def send_eom(self, headers: Headers, write: Writer) -> None:
+        pass
+
+
+#
+# These are all careful not to do anything to 'data' except call len(data) and
+# write(data). This allows us to transparently pass-through funny objects,
+# like placeholder objects referring to files on disk that will be sent via
+# sendfile(2).
+#
+class ContentLengthWriter(BodyWriter):
+    def __init__(self, length: int) -> None:
+        self._length = length
+
+    def send_data(self, data: bytes, write: Writer) -> None:
+        self._length -= len(data)
+        if self._length < 0:
+            raise LocalProtocolError("Too much data for declared Content-Length")
+        write(data)
+
+    def send_eom(self, headers: Headers, write: Writer) -> None:
+        if self._length != 0:
+            raise LocalProtocolError("Too little data for declared Content-Length")
+        if headers:
+            raise LocalProtocolError("Content-Length and trailers don't mix")
+
+
+class ChunkedWriter(BodyWriter):
+    def send_data(self, data: bytes, write: Writer) -> None:
+        # if we encoded 0-length data in the naive way, it would look like an
+        # end-of-message.
+        if not data:
+            return
+        write(b"%x\r\n" % len(data))
+        write(data)
+        write(b"\r\n")
+
+    def send_eom(self, headers: Headers, write: Writer) -> None:
+        write(b"0\r\n")
+        write_headers(headers, write)
+
+
+class Http10Writer(BodyWriter):
+    def send_data(self, data: bytes, write: Writer) -> None:
+        write(data)
+
+    def send_eom(self, headers: Headers, write: Writer) -> None:
+        if headers:
+            raise LocalProtocolError("can't send trailers to HTTP/1.0 client")
+        # no need to close the socket ourselves, that will be taken care of by
+        # Connection: close machinery
+
+
+WritersType = Dict[
+    Union[Tuple[Type[Sentinel], Type[Sentinel]], Type[Sentinel]],
+    Union[
+        Dict[str, Type[BodyWriter]],
+        Callable[[Union[InformationalResponse, Response], Writer], None],
+        Callable[[Request, Writer], None],
+    ],
+]
+
+WRITERS: WritersType = {
+    (CLIENT, IDLE): write_request,
+    (SERVER, IDLE): write_any_response,
+    (SERVER, SEND_RESPONSE): write_any_response,
+    SEND_BODY: {
+        "chunked": ChunkedWriter,
+        "content-length": ContentLengthWriter,
+        "http/1.0": Http10Writer,
+    },
+}

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/h11/py.typed

@@ -0,0 +1 @@
+Marker

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/rwkv/__init__.py

@@ -0,0 +1,27 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_rwkv import *
+    from .modeling_rwkv import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/rwkv/configuration_rwkv.py

@@ -0,0 +1,76 @@
+# Copyright 2023 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""RWKV configuration"""
+
+from huggingface_hub.dataclasses import strict
+
+from ...configuration_utils import PreTrainedConfig
+from ...utils import auto_docstring
+
+
+@auto_docstring(checkpoint="RWKV/rwkv-4-169m-pile")
+@strict
+class RwkvConfig(PreTrainedConfig):
+    r"""
+    context_length (`int`, *optional*, defaults to 1024):
+        The maximum sequence length that this model can be used with in a single forward (using it in RNN mode
+        lets use any sequence length).
+    attention_hidden_size (`int`, *optional*):
+        Dimensionality of the attention hidden states. Will default to `hidden_size` if unset.
+    rescale_every (`int`, *optional*, defaults to 6):
+        At inference, the hidden states (and weights of the corresponding output layers) are divided by 2 every
+        `rescale_every` layer. If set to 0 or a negative number, no rescale is done.
+
+    Example:
+
+    ```python
+    >>> from transformers import RwkvConfig, RwkvModel
+
+    >>> # Initializing a Rwkv configuration
+    >>> configuration = RwkvConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = RwkvModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "rwkv"
+    attribute_map = {"max_position_embeddings": "context_length"}
+
+    vocab_size: int = 50277
+    context_length: int = 1024
+    hidden_size: int = 4096
+    num_hidden_layers: int = 32
+    attention_hidden_size: int | None = None
+    intermediate_size: int | None = None
+    layer_norm_epsilon: float = 1e-5
+    bos_token_id: int | None = 0
+    eos_token_id: int | list[int] | None = 0
+    rescale_every: int = 6
+    tie_word_embeddings: bool = False
+    use_cache: bool = True
+
+    def __post_init__(self, **kwargs):
+        self.attention_hidden_size = (
+            self.attention_hidden_size if self.attention_hidden_size is not None else self.hidden_size
+        )
+        self.intermediate_size = self.intermediate_size if self.intermediate_size is not None else 4 * self.hidden_size
+
+        super().__post_init__(**kwargs)
+
+
+__all__ = ["RwkvConfig"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/rwkv/modeling_rwkv.py

@@ -0,0 +1,758 @@
+# Copyright 2023 Bo Peng and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch RWKV model."""
+
+import math
+from dataclasses import dataclass
+
+import torch
+from torch import nn
+
+from ... import initialization as init
+from ...generation import GenerationMixin
+from ...modeling_layers import GradientCheckpointingLayer
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    ModelOutput,
+    auto_docstring,
+    is_bitsandbytes_available,
+    is_kernels_available,
+    is_ninja_available,
+    is_torch_cuda_available,
+    logging,
+)
+from .configuration_rwkv import RwkvConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+rwkv_cuda_kernel = None
+
+
+def load_wkv_cuda_kernel(context_length):
+    global rwkv_cuda_kernel
+    if not is_kernels_available():
+        raise ImportError("kernels is not installed, please install it with `pip install kernels`")
+
+    from ...integrations.hub_kernels import get_kernel
+
+    rwkv_cuda_kernel = get_kernel("kernels-community/rwkv")
+    rwkv_cuda_kernel.max_seq_length = context_length
+
+
+class RwkvLinearAttention(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, time_decay, time_first, key, value, state=None, return_state=False):
+        batch_size, seq_len, hidden_size = key.size()
+        if seq_len > rwkv_cuda_kernel.max_seq_length:
+            raise ValueError(
+                f"Cannot process a batch with {seq_len} tokens at the same time, use a maximum of "
+                f"{rwkv_cuda_kernel.max_seq_length} with this model."
+            )
+        if batch_size * hidden_size % min(hidden_size, 32) != 0:
+            raise ValueError(
+                f"The product of batch size ({batch_size}) and hidden size ({hidden_size}) needs to be a round "
+                f"multiple of {min(hidden_size, 32)}."
+            )
+
+        ctx.input_dtype = key.dtype
+
+        if (
+            time_decay.device.type != "cuda"
+            or time_first.device.type != "cuda"
+            or key.device.type != "cuda"
+            or value.device.type != "cuda"
+        ):
+            raise ValueError("Calling the CUDA kernel for wkv attention requires all tensors to be on CUDA devices.")
+
+        time_decay = -torch.exp(time_decay.float().contiguous())
+        if key.dtype == torch.float16:
+            time_first = time_first.float()
+            key = key.float()
+            value = value.float()
+        time_first = time_first.contiguous()
+        key = key.contiguous()
+        value = value.contiguous()
+        # The CUDA kernel will fill this tensor.
+        output = torch.empty_like(key, memory_format=torch.contiguous_format)
+        if return_state or state is not None:
+            if state is None:
+                state = torch.zeros(
+                    batch_size,
+                    hidden_size,
+                    3,
+                    dtype=torch.float32,
+                    device=key.device,
+                    memory_format=torch.contiguous_format,
+                )
+                state[:, :, 2] -= 1e38
+            else:
+                state = torch.cat([s.unsqueeze(2) for s in state], dim=2).contiguous()
+            if key.dtype == torch.bfloat16:
+                forward_func = rwkv_cuda_kernel.forward_with_state_bf16
+            else:
+                forward_func = rwkv_cuda_kernel.forward_with_state
+            forward_func(time_decay, time_first, key, value, output, state)
+        else:
+            forward_func = rwkv_cuda_kernel.forward_bf16 if key.dtype == torch.bfloat16 else rwkv_cuda_kernel.forward
+            forward_func(time_decay, time_first, key, value, output)
+
+        ctx.save_for_backward(time_decay, time_first, key, value, output)
+
+        if state is not None:
+            state = [s.squeeze(2) for s in torch.chunk(state, 3, dim=2)]
+
+        return output.to(ctx.input_dtype), state
+
+    @staticmethod
+    # g stands for grad
+    def backward(ctx, g_output, g_state=None):
+        input_dtype = ctx.input_dtype
+
+        time_decay, time_first, key, value, output = ctx.saved_tensors
+        # The CUDA kernel will fill those tensors.
+        g_time_decay = torch.empty_like(
+            time_decay,
+            memory_format=torch.contiguous_format,
+            dtype=torch.bfloat16 if input_dtype == torch.bfloat16 else torch.float32,
+        )
+        g_time_first = torch.empty_like(time_first, memory_format=torch.contiguous_format)
+        g_key = torch.empty_like(key, memory_format=torch.contiguous_format)
+        g_value = torch.empty_like(value, memory_format=torch.contiguous_format)
+
+        if input_dtype == torch.float16:
+            g_output = g_output.float()
+        backward_func = rwkv_cuda_kernel.backward_bf16 if input_dtype == torch.bfloat16 else rwkv_cuda_kernel.backward
+        backward_func(
+            time_decay,
+            time_first,
+            key,
+            value,
+            output,
+            g_output.contiguous(),
+            g_time_decay,
+            g_time_first,
+            g_key,
+            g_value,
+        )
+
+        return (
+            g_time_decay.to(input_dtype),
+            g_time_first.to(input_dtype),
+            g_key.to(input_dtype),
+            g_value.to(input_dtype),
+            None,
+            None,
+        )
+
+
+def rwkv_linear_attention_cpu(time_decay, time_first, key, value, state=None, return_state=False):
+    # For CPU fallback. Will be slower and probably take more memory than the custom CUDA kernel if not executed
+    # within a torch.no_grad.
+    _, seq_length, _ = key.size()
+    output = torch.zeros_like(key)
+
+    if state is None:
+        num_state = torch.zeros_like(key[:, 0], dtype=torch.float32)
+        den_state = torch.zeros_like(key[:, 0], dtype=torch.float32)
+        max_state = torch.zeros_like(key[:, 0], dtype=torch.float32) - 1e38
+    else:
+        num_state, den_state, max_state = state
+    # For numerical stability
+    #    real_numerator_state = num_state * torch.exp(max_state)
+    #    real_denominator_state = den_state * torch.exp(max_state)
+
+    time_decay = -torch.exp(time_decay)
+
+    for current_index in range(seq_length):
+        current_key = key[:, current_index].float()
+        current_value = value[:, current_index]
+
+        # wkv computation at time t
+        max_for_output = torch.maximum(max_state, current_key + time_first)
+        e1 = torch.exp(max_state - max_for_output)
+        e2 = torch.exp(current_key + time_first - max_for_output)
+        numerator = e1 * num_state + e2 * current_value
+        denominator = e1 * den_state + e2
+        output[:, current_index] = (numerator / denominator).to(output.dtype)
+
+        # Update state for next iteration
+        max_for_state = torch.maximum(max_state + time_decay, current_key)
+        e1 = torch.exp(max_state + time_decay - max_for_state)
+        e2 = torch.exp(current_key - max_for_state)
+        num_state = e1 * num_state + e2 * current_value
+        den_state = e1 * den_state + e2
+        max_state = max_for_state
+
+    if return_state or state is not None:
+        state = [num_state, den_state, max_state]
+
+    return output, state
+
+
+def rwkv_linear_attention(time_decay, time_first, key, value, state=None, return_state=False):
+    no_cuda = any(t.device.type != "cuda" for t in [time_decay, time_first, key, value])
+    # Launching the CUDA kernel for just one token will actually be slower (there is no for loop in the CPU version
+    # in this case).
+    one_token = key.size(1) == 1
+    if rwkv_cuda_kernel is None or no_cuda or one_token:
+        return rwkv_linear_attention_cpu(time_decay, time_first, key, value, state=state, return_state=return_state)
+    else:
+        return RwkvLinearAttention.apply(time_decay, time_first, key, value, state, return_state)
+
+
+class RwkvSelfAttention(nn.Module):
+    def __init__(self, config, layer_id=0):
+        super().__init__()
+        self.config = config
+        kernel_loaded = rwkv_cuda_kernel is not None and rwkv_cuda_kernel.max_seq_length == config.context_length
+        if is_ninja_available() and is_torch_cuda_available() and not kernel_loaded:
+            try:
+                load_wkv_cuda_kernel(config.context_length)
+            except Exception:
+                logger.info("Could not load the custom CUDA kernel for RWKV attention.")
+        self.layer_id = layer_id
+        hidden_size = config.hidden_size
+        attention_hidden_size = (
+            config.attention_hidden_size if config.attention_hidden_size is not None else hidden_size
+        )
+        self.attention_hidden_size = attention_hidden_size
+
+        self.time_decay = nn.Parameter(torch.empty(attention_hidden_size))
+        self.time_first = nn.Parameter(torch.empty(attention_hidden_size))
+
+        self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
+        self.time_mix_value = nn.Parameter(torch.empty(1, 1, hidden_size))
+        self.time_mix_receptance = nn.Parameter(torch.empty(1, 1, hidden_size))
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+        self.key = nn.Linear(hidden_size, attention_hidden_size, bias=False)
+        self.value = nn.Linear(hidden_size, attention_hidden_size, bias=False)
+        self.receptance = nn.Linear(hidden_size, attention_hidden_size, bias=False)
+        self.output = nn.Linear(attention_hidden_size, hidden_size, bias=False)
+
+    # TODO: maybe jit, otherwise move inside forward
+    def extract_key_value(self, hidden, state=None):
+        # Mix hidden with the previous timestep to produce key, value, receptance
+        if hidden.size(1) == 1 and state is not None:
+            shifted = state[1][:, :, self.layer_id]
+        else:
+            shifted = self.time_shift(hidden)
+            if state is not None:
+                shifted[:, 0] = state[1][:, :, self.layer_id]
+        key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key)
+        value = hidden * self.time_mix_value + shifted * (1 - self.time_mix_value)
+        receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance)
+
+        key = self.key(key)
+        value = self.value(value)
+        receptance = torch.sigmoid(self.receptance(receptance))
+        if state is not None:
+            state[1][:, :, self.layer_id] = hidden[:, -1]
+        return receptance, key, value, state
+
+    def forward(self, hidden, state=None, use_cache=False):
+        receptance, key, value, state = self.extract_key_value(hidden, state=state)
+        layer_state = tuple(s[:, :, self.layer_id] for s in state[2:]) if state is not None else None
+        rwkv, layer_state = rwkv_linear_attention(
+            self.time_decay,
+            self.time_first,
+            key,
+            value,
+            state=layer_state,
+            return_state=use_cache,
+        )
+
+        if layer_state is not None:
+            state[2][:, :, self.layer_id] = layer_state[0]
+            state[3][:, :, self.layer_id] = layer_state[1]
+            state[4][:, :, self.layer_id] = layer_state[2]
+
+        return self.output(receptance * rwkv), state
+
+
+class RwkvFeedForward(nn.Module):
+    def __init__(self, config, layer_id=0):
+        super().__init__()
+        self.config = config
+        self.layer_id = layer_id
+        hidden_size = config.hidden_size
+        intermediate_size = (
+            config.intermediate_size if config.intermediate_size is not None else 4 * config.hidden_size
+        )
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+        self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
+        self.time_mix_receptance = nn.Parameter(torch.empty(1, 1, hidden_size))
+
+        self.key = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.receptance = nn.Linear(hidden_size, hidden_size, bias=False)
+        self.value = nn.Linear(intermediate_size, hidden_size, bias=False)
+
+    def forward(self, hidden, state=None):
+        if hidden.size(1) == 1 and state is not None:
+            shifted = state[0][:, :, self.layer_id]
+        else:
+            shifted = self.time_shift(hidden)
+            if state is not None:
+                shifted[:, 0] = state[0][:, :, self.layer_id]
+        key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key)
+        receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance)
+
+        key = torch.square(torch.relu(self.key(key)))
+        value = self.value(key)
+        receptance = torch.sigmoid(self.receptance(receptance))
+
+        if state is not None:
+            state[0][:, :, self.layer_id] = hidden[:, -1]
+
+        return receptance * value, state
+
+
+class RwkvBlock(GradientCheckpointingLayer):
+    def __init__(self, config, layer_id):
+        super().__init__()
+        self.config = config
+        self.layer_id = layer_id
+
+        if layer_id == 0:
+            self.pre_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+        self.ln1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        self.ln2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+        self.attention = RwkvSelfAttention(config, layer_id)
+        self.feed_forward = RwkvFeedForward(config, layer_id)
+
+    def forward(self, hidden, state=None, use_cache=False, output_attentions=False):
+        if self.layer_id == 0:
+            hidden = self.pre_ln(hidden)
+
+        attention, state = self.attention(self.ln1(hidden), state=state, use_cache=use_cache)
+        hidden = hidden + attention
+
+        feed_forward, state = self.feed_forward(self.ln2(hidden), state=state)
+        hidden = hidden + feed_forward
+
+        outputs = (hidden, state)
+        if output_attentions:
+            outputs += (attention,)
+        else:
+            outputs += (None,)
+
+        return outputs
+
+
+@auto_docstring
+class RwkvPreTrainedModel(PreTrainedModel):
+    config: RwkvConfig
+    base_model_prefix = "rwkv"
+    _no_split_modules = ["RwkvBlock"]
+    _keep_in_fp32_modules = ["time_decay", "time_first"]
+    supports_gradient_checkpointing = True
+    _is_stateful = True
+
+    @torch.no_grad()
+    def _init_weights(self, module: nn.Module):
+        """Initialize the weights."""
+        if isinstance(module, RwkvSelfAttention):
+            layer_id = module.layer_id
+            num_hidden_layers = module.config.num_hidden_layers
+            hidden_size = module.config.hidden_size
+            attention_hidden_size = module.attention_hidden_size
+
+            ratio_0_to_1 = layer_id / (num_hidden_layers - 1)  # 0 to 1
+            ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers)  # 1 to ~0
+
+            time_weight = torch.tensor(
+                [i / hidden_size for i in range(hidden_size)],
+                dtype=module.time_mix_key.dtype,
+                device=module.time_mix_key.device,
+            )
+            time_weight = time_weight[None, None, :]
+
+            decay_speed = [
+                -5 + 8 * (h / (attention_hidden_size - 1)) ** (0.7 + 1.3 * ratio_0_to_1)
+                for h in range(attention_hidden_size)
+            ]
+            decay_speed = torch.tensor(decay_speed, dtype=module.time_decay.dtype, device=module.time_decay.device)
+            zigzag = (
+                torch.tensor(
+                    [(i + 1) % 3 - 1 for i in range(attention_hidden_size)],
+                    dtype=module.time_first.dtype,
+                    device=module.time_first.device,
+                )
+                * 0.5
+            )
+
+            init.copy_(module.time_decay, decay_speed)
+            init.copy_(module.time_first, torch.ones_like(module.time_first * math.log(0.3) + zigzag))
+
+            init.copy_(module.time_mix_key, torch.pow(time_weight, ratio_1_to_almost0))
+            init.copy_(module.time_mix_value, torch.pow(time_weight, ratio_1_to_almost0) + 0.3 * ratio_0_to_1)
+            init.copy_(module.time_mix_receptance, torch.pow(time_weight, 0.5 * ratio_1_to_almost0))
+        elif isinstance(module, RwkvFeedForward):
+            layer_id = module.layer_id
+            num_hidden_layers = module.config.num_hidden_layers
+            hidden_size = module.config.hidden_size
+
+            ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers)  # 1 to ~0
+
+            time_weight = torch.tensor(
+                [i / hidden_size for i in range(hidden_size)],
+                dtype=module.time_mix_key.dtype,
+                device=module.time_mix_key.device,
+            )
+            time_weight = time_weight[None, None, :]
+
+            init.copy_(module.time_mix_key, torch.pow(time_weight, ratio_1_to_almost0))
+            init.copy_(module.time_mix_receptance, torch.pow(time_weight, ratio_1_to_almost0))
+        elif isinstance(module, nn.Linear):
+            shape = module.weight.shape
+            gain = 1.0
+            scale = 1.0  # extra scale for gain
+            if module.bias is not None:
+                init.zeros_(module.bias)
+            if shape[0] > shape[1]:
+                gain = math.sqrt(shape[0] / shape[1])
+            if shape[0] == self.config.vocab_size and shape[1] == self.config.hidden_size:  # final projection?
+                scale = 0.5
+
+            gain *= scale
+            init.orthogonal_(module.weight, gain=gain)
+        elif isinstance(module, nn.Embedding):
+            shape = module.weight.shape
+            gain = 1e-4 * math.sqrt(max(shape[0], shape[1]))
+            init.orthogonal_(module.weight, gain=gain)
+        elif isinstance(module, nn.LayerNorm):
+            init.ones_(module.weight)
+            init.zeros_(module.bias)
+
+
+@auto_docstring(
+    custom_intro="""
+    Class for the RWKV model outputs.
+    """
+)
+@dataclass
+class RwkvOutput(ModelOutput):
+    r"""
+    state (list of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`):
+        The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+        avoid providing the old `input_ids`.
+    """
+
+    last_hidden_state: torch.FloatTensor | None = None
+    state: list[torch.FloatTensor] | None = None
+    hidden_states: tuple[torch.FloatTensor, ...] | None = None
+    attentions: tuple[torch.FloatTensor, ...] | None = None
+
+
+@auto_docstring(
+    custom_intro="""
+    Base class for causal language model (or autoregressive) outputs.
+    """
+)
+@dataclass
+class RwkvCausalLMOutput(ModelOutput):
+    r"""
+    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+        Language modeling loss (for next-token prediction).
+    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+    state (list of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`):
+        The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+        avoid providing the old `input_ids`.
+    """
+
+    loss: torch.FloatTensor | None = None
+    logits: torch.FloatTensor | None = None
+    state: list[torch.FloatTensor] | None = None
+    hidden_states: tuple[torch.FloatTensor, ...] | None = None
+    attentions: tuple[torch.FloatTensor, ...] | None = None
+
+
+@auto_docstring
+class RwkvModel(RwkvPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.blocks = nn.ModuleList([RwkvBlock(config, layer_id=idx) for idx in range(config.num_hidden_layers)])
+        self.ln_out = nn.LayerNorm(config.hidden_size)
+
+        self.layers_are_rescaled = False
+
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.LongTensor | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        state: list[torch.FloatTensor] | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs,
+    ) -> tuple | RwkvOutput:
+        r"""
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
+            `past_key_values.get_seq_length()` (`sequence_length` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        state (tuple of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`, *optional*):
+            If passed along, the model uses the previous state in all the blocks (which will give the output for the
+            `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
+        use_cache (`bool`, *optional*):
+            If set to `True`, the last state is returned and can be used to quickly generate the next logits.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        if attention_mask is not None:
+            logger.warning_once("`attention_mask` was passed, but it is unused in this model.")
+
+        if self.training == self.layers_are_rescaled:
+            self._rescale_layers()
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        if use_cache and state is None:
+            shape = (inputs_embeds.size(0), self.config.hidden_size, self.config.num_hidden_layers)
+            state = [
+                torch.zeros(
+                    *shape, dtype=inputs_embeds.dtype if i <= 1 else torch.float32, device=inputs_embeds.device
+                )
+                for i in range(5)
+            ]
+            state[4] -= 1e30
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        hidden_states = inputs_embeds
+
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+        for idx, block in enumerate(self.blocks):
+            hidden_states, state, attentions = block(
+                hidden_states, state=state, use_cache=use_cache, output_attentions=output_attentions
+            )
+
+            if (
+                self.layers_are_rescaled
+                and self.config.rescale_every > 0
+                and (idx + 1) % self.config.rescale_every == 0
+            ):
+                hidden_states = hidden_states / 2
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (attentions,)
+
+        hidden_states = self.ln_out(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(x for x in [hidden_states, state, all_hidden_states, all_self_attentions] if x is not None)
+
+        return RwkvOutput(
+            last_hidden_state=hidden_states,
+            state=state,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+    def _rescale_layers(self):
+        # Layers should be rescaled for inference only.
+        if self.layers_are_rescaled == (not self.training):
+            return
+        if self.config.rescale_every > 0:
+            with torch.no_grad():
+                for block_id, block in enumerate(self.blocks):
+                    if self.training:
+                        block.attention.output.weight.mul_(2 ** int(block_id // self.config.rescale_every))
+                        block.feed_forward.value.weight.mul_(2 ** int(block_id // self.config.rescale_every))
+                    else:
+                        # Deal with quantization statistics
+                        if hasattr(block.attention.output.weight, "SCB"):
+                            block.attention.output.weight.SCB.div_(2 ** int(block_id // self.config.rescale_every))
+                            block.feed_forward.value.weight.SCB.div_(2 ** int(block_id // self.config.rescale_every))
+                        elif hasattr(block.attention.output.weight, "quant_state"):
+                            self._bnb_4bit_dequantize_and_rescale(block.attention.output, block_id)
+                            self._bnb_4bit_dequantize_and_rescale(block.feed_forward.value, block_id)
+                        else:
+                            block.attention.output.weight.div_(2 ** int(block_id // self.config.rescale_every))
+                            block.feed_forward.value.weight.div_(2 ** int(block_id // self.config.rescale_every))
+
+        self.layers_are_rescaled = not self.training
+
+    def _bnb_4bit_dequantize_and_rescale(self, target_layer, block_id):
+        r"""
+        Perform the dequantization and rescaling of the weights of a given layer. After that operation the layer will
+        be quantized again.
+        """
+        if not is_bitsandbytes_available():
+            raise ImportError("Please install bitsandbytes to use this method.")
+        import bitsandbytes as bnb
+
+        dequant_weights = bnb.functional.dequantize_4bit(target_layer.weight.data, target_layer.weight.quant_state)
+
+        dequant_weights.div_(2 ** int(block_id // self.config.rescale_every))
+
+        # re-quantize the model:
+        # we need to put it first on CPU then back to the device
+        # this will create an overhead :/
+        # We set requires_grad=False as we cannot compute gradients on top of 4bit parameters anyway and to avoid
+        # bugs with bnb
+        quant_weight = bnb.nn.Params4bit(dequant_weights.to("cpu"), requires_grad=False).to(dequant_weights.device)
+        setattr(target_layer, "weight", quant_weight)
+
+
+@auto_docstring(
+    custom_intro="""
+    The RWKV Model transformer with a language modeling head on top (linear layer with weights tied to the input
+    embeddings).
+    """
+)
+class RwkvForCausalLM(RwkvPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"head.weight": "rwkv.embeddings.weight"}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.rwkv = RwkvModel(config)
+        self.head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.head = new_embeddings
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.LongTensor | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        state: list[torch.FloatTensor] | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs,
+    ) -> tuple | RwkvCausalLMOutput:
+        r"""
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
+            `past_key_values.get_seq_length()` (`sequence_length` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        state (tuple of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`, *optional*):
+            If passed along, the model uses the previous state in all the blocks (which will give the output for the
+            `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        use_cache (`bool`, *optional*):
+            If set to `True`, the last state is returned and can be used to quickly generate the next logits.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        rwkv_outputs = self.rwkv(
+            input_ids,
+            inputs_embeds=inputs_embeds,
+            state=state,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = rwkv_outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        if not return_dict:
+            output = (logits,) + rwkv_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return RwkvCausalLMOutput(
+            loss=loss,
+            logits=logits,
+            state=rwkv_outputs.state,
+            hidden_states=rwkv_outputs.hidden_states,
+            attentions=rwkv_outputs.attentions,
+        )
+
+
+__all__ = ["RwkvForCausalLM", "RwkvModel", "RwkvPreTrainedModel"]