Oct 28, 2026 – Nov 3, 2026

Keep overrides of the deprecated gradient checkpointing methods workingPython · 36 + / 2 −

Introduces 3 new declarations in src/transformers/modeling_utils.py.

Adds new core implementation rather than adjusting what was there. Tests changed with it, with code of their own.

src/transformers/modeling_utils.py ↗ · 2 files

@@ -3129,6 +3129,19 @@ def activation_checkpointing_enable(             activation_checkpointing_kwargs (dict, *optional*):                 Additional keyword arguments passed along to the `torch.utils.checkpoint.checkpoint` function.         """+        # Subclasses and wrappers of the deprecated `gradient_checkpointing_enable` stay the entry point until v6+        if self._is_overridden("gradient_checkpointing_enable"):+            self.gradient_checkpointing_enable(+                gradient_checkpointing_kwargs=activation_checkpointing_kwargs,+                every_n_layers=every_n_layers,+                offload=offload,+            )+        else:+            self._activation_checkpointing_enable(activation_checkpointing_kwargs, every_n_layers, offload)++    def _activation_checkpointing_enable(+        self, activation_checkpointing_kwargs=None, every_n_layers: int = 1, offload: bool = False+    ):         if not self.supports_gradient_checkpointing:             raise ValueError(f"{self.__class__.__name__} does not support activation checkpointing.") 

Train adapters on top of FP8 weightsPython · 78 + / 43 −

Introduces 7 new declarations in src/transformers/integrations/finegrained_fp8.py.

Adds new core implementation rather than adjusting what was there. Tests changed with it, with code of their own.

src/transformers/integrations/finegrained_fp8.py ↗ · 3 files

@@ -282,6 +282,63 @@ def fp8_linear(     return finegrained_fp8_linear(input, weight, weight_scale_inv, block_size, bias, activation_scale)  +def _dequantize_fp8_weight(+    weight: torch.Tensor, weight_scale_inv: torch.Tensor, block_size: list[int] | None, dtype: torch.dtype+) -> torch.Tensor:+    """Dequantize `(..., N, K)` FP8 weights with their per-block (or per-tensor) scales, for the backward pass."""+    if weight.dtype == torch.int8:+        raise NotImplementedError("Backward through FP4 weights is not supported.")+    scale = weight_scale_inv.to(dtype)+    if block_size is not None:+        rows, cols = weight.shape[-2:]+        scale = scale.repeat_interleave(block_size[0], dim=-2)[..., :rows, :]+        scale = scale.repeat_interleave(block_size[1], dim=-1)[..., :cols]+    return weight.to(dtype).mul_(scale)+++class FP8LinearFunction(torch.autograd.Function):+    """+    `fp8_linear` with a backward for the input, so adapters (e.g. LoRA) can train on top of FP8 layers. The FP8 weights+    are frozen: the input gradient goes through the dequantized weight, and the weight, scales and bias get none.+    """++    @staticmethod+    def forward(ctx, input, weight, weight_scale_inv, block_size, bias, activation_scale, allow_deepgemm):+        ctx.save_for_backward(weight, weight_scale_inv)+        ctx.block_size = block_size+        return fp8_linear(input, weight, weight_scale_inv, block_size, bias, activation_scale, allow_deepgemm)++    @staticmethod+    def backward(ctx, grad_output):+        weight, weight_scale_inv = ctx.saved_tensors+        weight = _dequantize_fp8_weight(weight, weight_scale_inv, ctx.block_size, grad_output.dtype)

BUG: f2py: mask string literals in depend harvest and fix sort orderPython · 100 + / 17 −

Introduces 3 new declarations in numpy/f2py/crackfortran.py.

Adds new core implementation rather than adjusting what was there. Tests changed with it, with code of their own.

numpy/f2py/crackfortran.py ↗ · 8 files

@@ -2129,6 +2147,49 @@ def postcrack(block, args=None, tab=''):     return block  +def _sort_cyclic_dependents(dep, vars):+    """Order one cyclic depend group by real dependency edges.++    ``sortvarnames`` calls this only after a cycle has been reported.+    The usual f2py cycle is an input array dimensioned by hide scalars+    that are computed from that array (``n = shape(a, ...)``,+    ``lda = MAX(1, n)``). Those array-to-hide edges are dropped so+    ``getarrdims`` can blank them, and every remaining edge is honored+    by a topological sort. An initializer therefore runs after the+    names it uses, independent of declaration order.++    A cycle that is not one of those shape back-edges is left in+    ``dep`` order. The caller has already reported it. This function+    does not recurse.+    """+    remaining = set(dep)++    def blocking(name):+        var = vars[name]+        blocked_by = []+        for dep_name in var.get('depend') or []:+            if dep_name not in remaining or dep_name == name:+                continue+            target = vars.get(dep_name, {})+            # Dimension back-edge onto a hide value or an initializer.+            if (isarray(var) and isintent_nothide(var)+                    and (isintent_hide(target) or '=' in target)):+                continue+            blocked_by.append(dep_name)+        return blocked_by

ENH: Add `np.ma.assert_array_equal`Python · 125 + / 1 −

Introduces 3 new declarations in numpy/ma/core.py.

Adds new core implementation rather than adjusting what was there. Tests changed with it, with code of their own.

numpy/ma/core.py ↗ · 7 files

@@ -8744,6 +8745,129 @@ def allclose(a, b, masked_equal=True, rtol=1e-5, atol=1e-8):     return np.all(d)  +def assert_array_equal(actual, desired, err_msg='', verbose=True, *,+                       strict=False):+    """+    Raises an AssertionError if two masked arrays are not equal.++    Given two array_like objects, check that their masks are equal and that+    their data is equal wherever it is not masked. Unlike+    `numpy.testing.assert_array_equal`, which ignores masked values, a value+    that is masked in only one of the objects is reported as a mismatch.+    Objects that are not masked arrays are treated as having no masked+    values. As in `numpy.testing.assert_array_equal`, NaNs in the same+    unmasked positions are considered equal, and a scalar is compared with+    every element of an array unless `strict` is set.++    .. versionadded:: 2.6.0++    Parameters+    ----------+    actual : array_like+        The actual object to check.+    desired : array_like+        The desired, expected object.+    err_msg : str, optional+        The error message to be printed in case of failure.+    verbose : bool, optional+        If True, the conflicting values are appended to the error message.+    strict : bool, optional+        If True, raise an AssertionError when either the shape or the data+        type of the array_like objects does not match, and disable the+        special handling of scalars.

Implement `advance_back_by` on `Fuse`Rust · 41 + / 0 −

Introduces 4 new declarations in library/core/src/iter/adapters/fuse.rs.

Adds new language design rather than adjusting what was there. Tests changed with it, with code of their own.

library/core/src/iter/adapters/fuse.rs ↗ · 2 files

@@ -358,6 +366,25 @@ where         and_then_or_clear(&mut self.iter, |iter| iter.next_back())     } +    #[inline]+    default fn advance_back_by(&mut self, n: usize) -> Result<(), NonZero<usize>>+    where+        I: DoubleEndedIterator,+    {+        let Some(iter) = &mut self.iter else {+            return match NonZero::new(n) {+                Some(n) => Err(n),+                None => Ok(()),+            };+        };++        let res = iter.advance_back_by(n);+        if res.is_err() {+            self.iter = None;+        }+        res+    }+     #[inline]     default fn nth_back(&mut self, n: usize) -> Option<<I as Iterator>::Item>     where

Turbopack: fix trailing-slash package root resolutionRust · 118 + / 10 −

Introduces 3 new declarations in turbopack/crates/turbopack-core/src/resolve/mod.rs.

Adds new framework internals rather than adjusting what was there. Tests changed with it, with code of their own.

turbopack/crates/turbopack-core/src/resolve/mod.rs ↗ · 6 files

@@ -3908,6 +3914,108 @@ mod tests {         .await;     } +    #[rstest::rstest]+    #[case::main(rcstr!("package").into(), r#"{"main":"./main.js"}"#, vec![("package", "main.js")])]+    #[case::main_trailing_slash(rcstr!("package/").into(), r#"{"main":"./main.js"}"#, vec![("package/", "main.js")])]+    #[case::index(rcstr!("package").into(), "{}", vec![("package", "index.js")])]+    #[case::index_trailing_slash(rcstr!("package/").into(), "{}", vec![("package/", "index.js")])]+    #[case::exports(rcstr!("package").into(), r#"{"exports":"./main.js"}"#, vec![("package", "main.js")])]+    #[case::exports_trailing_slash(rcstr!("package/").into(), r#"{"exports":"./main.js"}"#, vec![])]+    #[case::exports_subpath(rcstr!("package/subpath.js").into(), r#"{"exports":{"./subpath.js":"./main.js"}}"#, vec![("package/subpath.js", "main.js")])]+    #[case::dynamic_subpath(+        Pattern::concat([rcstr!("package/").into(), Pattern::Dynamic, rcstr!(".js").into()]),+        r#"{"main":"./main.js"}"#,+        vec![("package/index.js", "index.js"), ("package/main.js", "main.js"), ("package/subpath.js", "subpath.js")]+    )]+    #[case::dynamic_trailing_slash(+        Pattern::concat([rcstr!("package/").into(), Pattern::Dynamic]),+        r#"{"main":"./main.js"}"#,+        vec![+            ("package/", "main.js"),+            ("package/index", "index.js"), ("package/index.js", "index.js"),+            ("package/main", "main.js"), ("package/main.js", "main.js"),+            ("package/package", "package.json"), ("package/package.json", "package.json"),+            ("package/subpath", "subpath.js"), ("package/subpath.js", "subpath.js")+        ]+    )]+    #[case::dynamic_roots(+        Pattern::Alternatives(vec![+            rcstr!("package").into(),+            Pattern::concat([rcstr!("package/").into(), Pattern::Dynamic]),+        ]),+        r#"{"main":"./main.js"}"#,

fix: Protect math from markdown emphasis parsing in Markdown panePython · 58 + / 3 −

Introduces 10 new declarations in panel/pane/markup.py.

Adds new core implementation rather than adjusting what was there. Tests changed with it, with code of their own.

panel/pane/markup.py ↗ · 3 files

@@ -38,6 +38,54 @@ # links, definition lists) or which breaks line offsets (\r). _NON_LOCAL_SYNTAX = re.compile(r'\[\^|\]:|^ {0,3}[:~][ \t]|\r', re.MULTILINE) +def _math_plugin(md: MarkdownIt) -> None:+    """+    Keeps math delimited by $$ verbatim, since the client-side+    MathJax pass runs on the rendered HTML and would otherwise see+    markup generated from characters such as '*' inside the math.+    """+    from mdit_py_plugins.dollarmath import dollarmath_plugin+    from mdit_py_plugins.dollarmath.index import math_inline_dollar++    dollarmath_plugin(md, allow_labels=False, double_inline=True)++    # The client only typesets $$, so single $ must stay ordinary text+    # (e.g. currency) and keep its surrounding emphasis.+    inline_rule = math_inline_dollar(True, True, True)++    def double_only(state, silent):+        if state.src[state.pos:state.pos + 2] != '$$':+            return False+        return inline_rule(state, silent)++    md.inline.ruler.at('math_inline', double_only)++    def raw(delimiter, block=False):+        def render(self, tokens, idx, options, env):+            math = escape(tokens[idx].content, quote=False)+            return f'{delimiter}{math}{delimiter}\n' if block else f'{delimiter}{math}{delimiter}'+        return render++    md.add_render_rule('math_inline_double', raw('$$'))+    md.add_render_rule('math_block', raw('$$', block=True))

fix: Send updates from long running sync callbacks before they returnPython · 51 + / 2 −

Introduces 3 new declarations in panel/io/document.py.

Adds new core implementation rather than adjusting what was there. Tests changed with it, with code of their own.

panel/io/document.py ↗ · 2 files

@@ -651,6 +653,46 @@ def _suppress_property_callbacks(             event.callback_invoker = None     return events +async def _write_message(+    doc: Document, connections: list[ServerConnection], message: Message+) -> None:+    for conn in connections:+        lock = getattr(conn._socket, 'write_lock', None)+        with (await lock.acquire()) if lock is not None else nullcontext():+            await _await_writes(_socket_dispatcher(conn._socket)(conn, msg=message))++def _write_from_worker_thread(+    doc: Document,+    connections: Iterable[ServerConnection],+    events: list[DocumentChangedEvent]+) -> bool:+    """+    Writes events generated by a locked callback that bokeh runs on a+    worker thread right away, rather than leaving them to bokeh, which+    only writes them once the callback returns, so that a long running+    callback can still report progress.++    Only safe while bokeh has no pending writes of its own: the message+    is serialized here, while the callback holds the Document lock, and+    the write is scheduled on the event loop ahead of bokeh's, so the+    client always receives models before they are referenced.+    """+    patches = [e for e in events if isinstance(e, DocumentPatchedEvent)]+    if state._on_loop_thread or not patches or len(patches) != len(events):+        return False+    loop = state._document_loop(doc)+    asyncio_loop = t.cast(+        'asyncio.AbstractEventLoop | None', getattr(loop, 'asyncio_loop', loop)

[`Moe`] Fix grouped mm fallbackPython · 36 + / 4 −

Introduces 3 new declarations in src/transformers/integrations/moe.py.

Adds new core implementation rather than adjusting what was there. Tests changed with it, with code of their own.

src/transformers/integrations/moe.py ↗ · 2 files

@@ -233,22 +233,43 @@ def _grouped_mm_fallback_setup_context(ctx, inputs, output):     ctx.offs = inputs[2]  -def _grouped_mm_fallback_backward(ctx, grad_output):-    """Backward pass for `_grouped_mm_fallback`. Computes grad_input and grad_weight per expert group; offs has no gradient."""-    input, weight = ctx.saved_tensors+def _grouped_mm_fallback_bwd(+    input: torch.Tensor,+    weight: torch.Tensor,+    offs: torch.Tensor,+    grad_output: torch.Tensor,+) -> tuple[torch.Tensor, torch.Tensor]:+    """Backward implementation for the grouped matmul fallback."""     grad_input = torch.zeros_like(input)     grad_weight = torch.zeros_like(weight)      start = 0     # single cpu<->gpu sync point here,     # avoids multiple syncs inside the loop-    for i, end in enumerate(ctx.offs.tolist()):+    for i, end in enumerate(offs.tolist()):         if start == end:             continue         torch.mm(grad_output[start:end], weight[i].T, out=grad_input[start:end])         torch.mm(input[start:end].T, grad_output[start:end], out=grad_weight[i])         start = end +    return grad_input, grad_weight+++def _grouped_mm_fallback_backward_fake(+    input: torch.Tensor,+    weight: torch.Tensor,

zlib: skip the stream in sync convenience calls without optionsJavaScript · 162 + / 19 −

Introduces 8 new declarations in lib/zlib.js.

Adds new runtime rather than adjusting what was there. Tests changed with it, with code of their own.

lib/zlib.js ↗ · 2 files

@@ -176,43 +177,128 @@ function zlibBufferOnEnd() {     this.cb(null, buf); } -function zlibBufferSync(engine, buffer) {+function toSyncInput(buffer) {   if (typeof buffer === 'string') {-    buffer = Buffer.from(buffer);-  } else if (!isArrayBufferView(buffer)) {+    return Buffer.from(buffer);+  }+  if (!isArrayBufferView(buffer)) {     if (isAnyArrayBuffer(buffer)) {-      buffer = Buffer.from(buffer);-    } else {-      throw new ERR_INVALID_ARG_TYPE(-        'buffer',-        ['string', 'Buffer', 'TypedArray', 'DataView', 'ArrayBuffer'],-        buffer,-      );+      return Buffer.from(buffer);     }+    throw new ERR_INVALID_ARG_TYPE(+      'buffer',+      ['string', 'Buffer', 'TypedArray', 'DataView', 'ArrayBuffer'],+      buffer,+    );   }+  return buffer;+}++function zlibBufferSync(engine, buffer) {+  buffer = toSyncInput(buffer);   buffer = processChunkSync(engine, buffer, engine._finishFlushFlag);

Refactor box downcasting methodsRust · 52 + / 43 −

Introduces 6 new declarations in library/alloc/src/boxed/convert.rs.

Adds new language design rather than adjusting what was there. Tests changed with it, with code of their own.

library/alloc/src/boxed/convert.rs ↗ · 6 files

@@ -711,21 +717,24 @@ impl<'a, 'b> From<Cow<'b, str>> for Box<dyn Error + 'a> {     } } +impl<T: Error + ?Sized + 'static, A: Allocator> Box<T, A> {+    fn downcast_error<Dst: 'static>(this: Self) -> Result<Box<Dst, A>, Self> {+        if core::error::error_type_id::<T>(&this) == core::any::TypeId::of::<Dst>() {+            // SAFETY: The type is correct+            Ok(unsafe { Self::cast_unchecked(this) })+        } else {+            Err(this)+        }+    }+}+ impl dyn Error {     /// Attempts to downcast the box to a concrete type.     #[inline]     #[stable(feature = "error_downcast", since = "1.3.0")]     #[rustc_allow_incoherent_impl]     pub fn downcast<T: Error + 'static>(self: Box<Self>) -> Result<Box<T>, Box<dyn Error>> {-        if self.is::<T>() {-            // SAFETY: Check ensures the type is correct.-            unsafe {-                let raw: *mut dyn Error = Box::into_raw(self);-                Ok(Box::from_raw(raw as *mut T))-            }-        } else {-            Err(self)-        }+        Box::downcast_error(self)     } } 

parse inline and rustc_force_inline with same parserRust · 136 + / 106 −

Introduces 4 new declarations in compiler/rustc_attr_parsing/src/attributes/inline.rs.

Adds new language design rather than adjusting what was there. Tests changed with it, with code of their own.

compiler/rustc_attr_parsing/src/attributes/inline.rs ↗ · 3 files

@@ -3,125 +3,156 @@ use rustc_feature::AttributeStability; use rustc_lint_defs::builtin::{ILL_FORMED_ATTRIBUTE_INPUT, UNUSED_ATTRIBUTES};  use super::prelude::*;-use crate::diagnostics::{InlineForceInlineConflict, InlineIgnoredForExported};+use crate::diagnostics::{InlineForceInlineConflict, InlineIgnoredForExported, UnusedMultiple}; -pub(crate) struct InlineParser;+#[derive(Default)]+pub(crate) struct InlineParser(Option<(InlineAttr, Span)>); -impl SingleAttributeParser for InlineParser {-    const PATH: &[Symbol] = &[sym::inline];-    const ON_DUPLICATE: OnDuplicate = OnDuplicate::WarnButFutureError;-    const ALLOWED_TARGETS: AllowedTargets<'_> = AllowedTargets::AllowList(&[-        Allow(Target::Fn),-        Allow(Target::Method(MethodKind::Inherent)),-        Allow(Target::Method(MethodKind::Trait { body: true })),-        Allow(Target::Method(MethodKind::TraitImpl)),-        Allow(Target::Closure),-        Allow(Target::Delegation { mac: false }),-        Warn(Target::Method(MethodKind::Trait { body: false })),-        Warn(Target::ForeignFn),-        Warn(Target::Field),-        Warn(Target::MacroDef),-        Warn(Target::Arm),-        Warn(Target::AssocConst(AssocCtxt::Impl { of_trait: false })),-        Warn(Target::AssocConst(AssocCtxt::Trait)),-        Warn(Target::AssocConst(AssocCtxt::Impl { of_trait: true })),-        Warn(Target::MacroCall),-    ]);-    const TEMPLATE: AttributeTemplate = template!(-        Word,-        List: &["always", "never"],