Sep 16, 2026 – Sep 22, 2026

core/validation: keep stored ResourceQuota values out of update checksGo · 37 + / 19 −

Introduces 3 new declarations in pkg/apis/core/validation/validation.go.

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

pkg/apis/core/validation/validation.go ↗ · 2 files

@@ -8528,21 +8528,48 @@ func ValidateResourceQuotaStatus(status *core.ResourceQuotaStatus, fld *field.Pa }  func ValidateResourceQuotaSpec(resourceQuotaSpec *core.ResourceQuotaSpec, fld *field.Path) field.ErrorList {+	return validateResourceQuotaSpec(resourceQuotaSpec, nil, fld)+}++// validateResourceQuotaSpec validates a spec against the hard limits the object+// already stores, oldHard, which is nil on create. Nothing is mutated.+func validateResourceQuotaSpec(resourceQuotaSpec *core.ResourceQuotaSpec, oldHard core.ResourceList, fld *field.Path) field.ErrorList { 	allErrs := field.ErrorList{} -	fldPath := fld.Child("hard")-	for k, v := range resourceQuotaSpec.Hard {-		resPath := fldPath.Key(string(k))-		allErrs = append(allErrs, ValidateResourceQuotaResourceName(k, resPath)...)-		allErrs = append(allErrs, ValidateResourceQuantityValue(k, v, resPath)...)-	}+	allErrs = append(allErrs, validateResourceQuotaResourceList(resourceQuotaSpec.Hard, fld.Child("hard"), oldHard)...)  	allErrs = append(allErrs, validateResourceQuotaScopes(resourceQuotaSpec, fld)...) 	allErrs = append(allErrs, validateScopeSelector(resourceQuotaSpec, fld)...)  	return allErrs } +// validateResourceQuotaResourceList validates every name in values, and every+// value that none of the stored lists holds under the same key. Nothing is mutated.+func validateResourceQuotaResourceList(values core.ResourceList, fldPath *field.Path, stored ...core.ResourceList) field.ErrorList {+	allErrs := field.ErrorList{}+	for k, v := range values {+		resPath := fldPath.Key(string(k))+		allErrs = append(allErrs, ValidateResourceQuotaResourceName(k, resPath)...)+		// A value the object already holds was accepted when it was stored.

kubectl explain: add auto completionGo · 265 + / 5 −

Introduces 7 new declarations in staging/src/k8s.io/kubectl/pkg/cmd/explain/completion.go.

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

staging/src/k8s.io/kubectl/pkg/cmd/explain/completion.go ↗ · 4 files

@@ -0,0 +1,252 @@+/*+Copyright The Kubernetes Authors.++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.+*/++package explain++import (+	"slices"+	"strings"++	"github.com/spf13/cobra"++	"k8s.io/apimachinery/pkg/api/meta"+	"k8s.io/apimachinery/pkg/runtime/schema"+	"k8s.io/cli-runtime/pkg/genericclioptions"+	"k8s.io/client-go/openapi3"+	"k8s.io/kube-openapi/pkg/validation/spec"+	"k8s.io/kubectl/pkg/explain"+	"k8s.io/kubectl/pkg/util/completion"+)+

Don't set nf_conntrack_max if it's already larger than neededGo · 43 + / 30 −

Introduces 5 new declarations in pkg/proxy/conntrack/sysctls.go.

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

pkg/proxy/conntrack/sysctls.go ↗ · 2 files

@@ -128,39 +152,28 @@ func getConntrackMax(ctx context.Context, config *kubeproxyconfig.KubeProxyConnt 	return 0, nil } -// detectNumCPU returns the CPU count used to size nf_conntrack_max. That limit+type realConntrackConfigurer struct {+	sys sysctl.Interface+}++// DetectNumCPU returns the CPU count used to size nf_conntrack_max. That limit // is host-wide, so it must be based on the node's CPU count, not runtime.NumCPU(): // runtime.NumCPU() honors the process cpuset and undercounts when kube-proxy // runs under a static CPU policy. cpuset.NumCPU() reads the node's online CPU // count from sysfs instead, falling back to runtime.NumCPU() if it can't.-func detectNumCPU() int {+func (rct realConntrackConfigurer) DetectNumCPU() int { 	if n, err := cpuset.NumCPU(); err == nil && n > 0 { 		return n 	} 	return runtime.NumCPU() } -type realConntrackConfigurer struct {+func (rct realConntrackConfigurer) GetMax(_ context.Context) (int, error) {+	return rct.sys.GetSysctl("net/netfilter/nf_conntrack_max") }  func (rct realConntrackConfigurer) SetMax(ctx context.Context, max int) error {-	logger := klog.FromContext(ctx)-	logger.Info("Setting nf_conntrack_max", "nfConntrackMax", max)-	if err := rct.setIntSysCtl(ctx, "nf_conntrack_max", max); err != nil {-		return err-	}-

sqlite: add virtual table support via createModule()C++ · 826 + / 0 −

Introduces 7 new declarations in src/node_sqlite.cc.

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

src/node_sqlite.cc ↗ · 5 files

@@ -1012,6 +1015,483 @@ BaseObjectPtr<DatabaseSyncLimits> DatabaseSyncLimits::Create(   return MakeBaseObject<DatabaseSyncLimits>(env, obj, std::move(database)); } +// ---------------------------------------------------------------------------+// VirtualTableModule+// ---------------------------------------------------------------------------++VirtualTableModule::VirtualTableModule(Environment* env,+                                       BaseObjectWeakPtr<DatabaseSync> db,+                                       Local<Function> rows_fn,+                                       std::string&& schema_sql,+                                       int num_columns,+                                       std::vector<int>&& hidden_col_indices,+                                       bool use_bigint_args,+                                       bool direct_only)+    : env_(env),+      db_(std::move(db)),+      rows_fn_(env->isolate(), rows_fn),+      schema_sql_(std::move(schema_sql)),+      num_columns_(num_columns),+      hidden_col_indices_(std::move(hidden_col_indices)),+      use_bigint_args_(use_bigint_args),+      direct_only_(direct_only),+      module_def_({}) {+  // Initialize the sqlite3_module definition. Each VirtualTableModule instance+  // gets its own copy to avoid thread-safety issues with worker threads.+  module_def_.iVersion = 1;+  module_def_.xCreate = VirtualTableModule::xCreate;+  module_def_.xConnect = VirtualTableModule::xCreate;+  module_def_.xBestIndex = VirtualTableModule::xBestIndex;+  module_def_.xDisconnect = VirtualTableModule::xDisconnect;+  module_def_.xDestroy = VirtualTableModule::xDestroy;+  module_def_.xOpen = VirtualTableModule::xOpen;

Fix non-deterministic encoding of syntax contextsRust · 87 + / 59 −

Introduces 4 new declarations in compiler/rustc_span/src/hygiene.rs.

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

compiler/rustc_span/src/hygiene.rs ↗ · 5 files

@@ -1337,58 +1381,51 @@ impl HygieneEncodeContext {                 h_ctxt.borrow().latest_ctxts             ); -            let mut mut_hctxt = h_ctxt.borrow_mut();-             // Consume the current round of syntax contexts.             // It's fine to iterate over a HashSet, because the serialization of the table             // that we insert data into doesn't depend on insertion order.             #[allow(rustc::potential_query_instability)]-            let latest_ctxts = { mem::take(&mut mut_hctxt.latest_ctxts) }.into_iter();+            let latest_contexts = { mem::take(&mut h_ctxt.borrow_mut().latest_ctxts) }.into_iter(); -            HygieneData::with(|data| {-                for ctxt in latest_ctxts {-                    if !mut_hctxt.serialized_ctxts.insert(ctxt) {-                        continue;-                    }--                    all_ctxt_data.push((ctxt.0, data.syntax_context_data[ctxt.0 as usize].key()));-                }-            });--            drop(mut_hctxt);--            for (idx, ctxt_key) in all_ctxt_data.drain(..) {-                encode_ctxt(encoder, idx, &ctxt_key);+            for (idx, ctxt) in latest_contexts {+                let key = HygieneData::with(|data| data.syntax_context_data[ctxt.0 as usize].key());+                encode_ctxt(encoder, idx, &key);             } -            let mut mut_hctxt = h_ctxt.borrow_mut();-

Fix startTime and shutdown delivery of forwarded Turbopack trace spansRust · 154 + / 33 −

Introduces 5 new declarations in turbopack/crates/turbo-tasks/src/message_queue.rs.

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

turbopack/crates/turbo-tasks/src/message_queue.rs ↗ · 5 files

@@ -59,58 +88,93 @@ impl CompilationEventQueue {                 history.pop_front();             }             history.push_back(message_clone.clone());+            drop(history);++            if deliver {+                // Send to all active receivers of the same message type+                if let Some(mut type_subscribers) = subscribers.get_mut(&EventChannelType::Type(+                    message_clone.type_name().to_owned(),+                )) {+                    let mut removal_indices = Vec::new();+                    for (ix, sender) in type_subscribers.iter().enumerate() {+                        if sender.send(message_clone.clone()).await.is_err() {+                            removal_indices.push(ix);+                        }+                    } -            // Send to all active receivers of the same message type-            if let Some(mut type_subscribers) = subscribers.get_mut(&EventChannelType::Type(-                message_clone.type_name().to_owned(),-            )) {-                let mut removal_indices = Vec::new();-                for (ix, sender) in type_subscribers.iter().enumerate() {-                    if sender.send(message_clone.clone()).await.is_err() {-                        removal_indices.push(ix);+                    for ix in removal_indices.iter().rev() {+                        type_subscribers.remove(*ix);                     }                 } -                for ix in removal_indices.iter().rev() {-                    type_subscribers.remove(*ix);-                }

fix(asr): warm up Numba RNNT and TDT losses before trainingPython · 116 + / 1 −

Introduces 7 new declarations in nemo/collections/asr/parts/numba/rnnt_loss/rnnt_pytorch.py.

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

nemo/collections/asr/parts/numba/rnnt_loss/rnnt_pytorch.py ↗ · 6 files

@@ -552,6 +583,51 @@ def __init__(         self.sigma = sigma         self.omega = omega +    def warmup(self, device: Union[str, torch.device]) -> bool:+        """Warm both TDT and RNNT CUDA branches with float32 contiguous inputs before allocating training activations.++        Preserves random states and the configured loss; returns False on CPU.+        """+        device = torch.device(device)+        if device.type != 'cuda':+            return False++        def warmup():+            # Non-singleton dimensions preserve the contiguous production array layout.+            labels_count = 3+            frames = max(8, (labels_count + 1) * max(self.durations))+            vocabulary_size = max(2, self.blank + 1)+            label = 1 if self.blank == 0 else 0+            for omega in (-1.0, 2.0):+                # Force each branch even when random.uniform returns an endpoint.+                loss = TDTLossNumba(+                    blank=self.blank,+                    durations=list(self.durations),+                    reduction='mean',+                    fastemit_lambda=self.fastemit_lambda,+                    clamp=self.clamp,+                    sigma=self.sigma,+                    omega=omega,+                )+                acts = torch.zeros(+                    (2, frames, labels_count + 1, vocabulary_size + len(self.durations)),+                    device=device,+                    dtype=torch.float32,

Always materialize the causal mask in Doge so sdpa stays causalPython · 85 + / 2 −

Introduces 3 new declarations in src/transformers/models/doge/modular_doge.py.

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

src/transformers/models/doge/modular_doge.py ↗ · 3 files

@@ -449,7 +450,61 @@ def _init_weights(self, module):   class DogeModel(MixtralModel):-    pass+    def forward(+        self,+        input_ids: torch.LongTensor | None = None,+        attention_mask: torch.Tensor | None = None,+        position_ids: torch.LongTensor | None = None,+        past_key_values: Cache | None = None,+        inputs_embeds: torch.FloatTensor | None = None,+        use_cache: bool | None = None,+        **kwargs: Unpack[TransformersKwargs],+    ) -> MoeModelOutputWithPast:+        if (input_ids is None) ^ (inputs_embeds is not None):+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")++        if use_cache and past_key_values is None:+            past_key_values = DynamicCache(config=self.config)++        if inputs_embeds is None:+            inputs_embeds = self.embed_tokens(input_ids)++        if position_ids is None:+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0+            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens+            position_ids = position_ids.unsqueeze(0)++        mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask+        causal_mask = mask_function(+            config=self.config,+            inputs_embeds=inputs_embeds,+            attention_mask=attention_mask,

model : add support for HrmTextForCausalLM (DFM Mimir 1B)C++ · 281 + / 1 −

Introduces 5 new declarations in src/models/hrm-text.cpp.

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

src/models/hrm-text.cpp ↗ · 16 files

@@ -0,0 +1,213 @@+#include "models.h"++// HRM-Text: alternating low/high transformer stacks over the same token stream.+// Reference: HrmTextModel in transformers, DFM Mimir 1B.++void llama_model_hrm_text::load_arch_hparams(llama_model_loader & ml) {+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);+    ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, false);++    ml.get_key(LLM_KV_HRM_LAYERS_PER_STACK, hparams.n_hrm_layers_per_stack);+    ml.get_key(LLM_KV_HRM_H_CYCLES, hparams.n_hrm_h_cycles);+    ml.get_key(LLM_KV_HRM_L_CYCLES, hparams.n_hrm_l_cycles);++    // prefix-LM prefill is not implemented (causal attention only); kept for round-trip+    ml.get_key(LLM_KV_HRM_PREFIX_LM, hparams.hrm_prefix_lm, false);++    GGML_ASSERT(hparams.n_hrm_layers_per_stack > 0);+    GGML_ASSERT(hparams.n_hrm_h_cycles > 0);+    GGML_ASSERT(hparams.n_hrm_l_cycles > 0);++    // the GGUF block count is the expanded cache-slot count+    const uint32_t n_slot = hparams.n_hrm_layers_per_stack * hparams.n_hrm_h_cycles * (hparams.n_hrm_l_cycles + 1);+    GGML_ASSERT(hparams.n_layer() == n_slot);++    switch (hparams.n_embd) {+        case 1536:+            type = LLM_TYPE_1B;+            break;+        default:+            type = LLM_TYPE_UNKNOWN;+    }+}+

[CB] Add pause mechanismPython · 178 + / 43 −

Introduces 10 new declarations in src/transformers/generation/continuous_batching/continuous_api.py.

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

src/transformers/generation/continuous_batching/continuous_api.py ↗ · 4 files

@@ -137,58 +139,77 @@ def fail_and_deliver(self, state: RequestState, error: Exception) -> None:  class BackgroundThreadStatus:     """Tracks the status of the background thread locally and in its TP group. The status is an int that can only-    increase, representing how soon the thread should stop."""+    increase, representing how soon the thread should stop.+    Through this object, threads sharing the model can also pause the generation loop. Any number of threads may ask at+    once, and the loop resumes once they are all done.+    """ +    # Stop statuses, in increasing order of urgency. The local status can only ever increase.     DONT_STOP = 0     FLUSH_AND_STOP = 1     HARD_STOP = 2     STOPPED = 3      def __init__(self) -> None:-        self._local_status_lock = threading.Lock()+        self.fatal_error = None+        self._condition = threading.Condition()         self._local_status = self.DONT_STOP         self._tp_status = self.DONT_STOP-        self.fatal_error = None+        self._pauses_requested = 0+        self._local_pauses_requested = ThreadLocalCounter()+        self._paused = False      def clear(self) -> None:-        """Clear the local and TP statuses. This method should ONLY be called by the main thread itself BEFORE starting-        the background thread."""-        self._tp_status = self.DONT_STOP-        with self._local_status_lock:+        """Clear the statuses and the pause state. This method should ONLY be called by the main thread itself BEFORE+        starting the background thread."""

BUG: named agg result depended on the output name matching the column name (GH#63743)Python · 40 + / 4 −

Introduces 3 new declarations in pandas/core/apply.py.

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

pandas/core/apply.py ↗ · 4 files

@@ -1902,10 +1902,22 @@ def reconstruct_func(     names, and the reconstructed order of columns.     If relabeling is False, the columns and order will be None. +    Named aggregation is the one exception: when ``allow_skip_normalization`` is+    True, every output name equals its source column name, and every aggfunc+    reduces to a scalar, relabeling is reported as False (and columns/order as+    None) even though named aggregation was used, because the caller can consume+    the un-normalized func directly.+     Parameters     ----------     func: agg function (e.g. 'min' or Callable) or list of agg functions         (e.g. ['min', np.max]) or dictionary (e.g. {'A': ['min', np.max]}).+    allow_skip_normalization: bool, default False+        Whether the caller can handle the un-normalized ``{column: aggfunc}`` form+        that named aggregation reduces to when every output name equals its source+        column name and every aggfunc is a scalar reduction. Callers that rely on+        ``columns``/``order`` being returned whenever named aggregation was used+        must leave this False.     **kwargs: dict, kwargs used in is_multi_agg_with_relabel and         normalize_keyword_aggregation function for relabelling 

Staticlib rename internal symbols: add COFF supportRust · 183 + / 17 −

Introduces 3 new declarations in compiler/rustc_codegen_ssa/src/back/symbol_edit.rs.

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

compiler/rustc_codegen_ssa/src/back/symbol_edit.rs ↗ · 4 files

@@ -441,6 +476,125 @@ fn macho_rebuild_strtab(     Some(result) } +// ---------------------------------------------------------------------------+// COFF: single-pass collection + apply+// ---------------------------------------------------------------------------++fn coff_collect_impl<'data, Coff: CoffHeader>(+    data: &'data [u8],+    header: &'data Coff,+    exported: &FxHashSet<String>,+    out: &mut FxHashSet<String>,+    strip_underscore: bool,+) {+    let Ok(symbols) = header.symbols(data) else { return };+    let strings = symbols.strings();++    for (_index, sym) in symbols.iter() {+        let sclass = sym.storage_class();+        if sclass != pe::IMAGE_SYM_CLASS_EXTERNAL && sclass != pe::IMAGE_SYM_CLASS_WEAK_EXTERNAL {+            continue;+        }+        if sym.section_number() <= 0 {+            continue;+        }+        let Ok(name_bytes) = sym.name(strings) else { continue };+        let Ok(mut name) = str::from_utf8(name_bytes).map(String::from) else { continue };+        if strip_underscore {+            name = name.strip_prefix('_').unwrap_or(&name).to_string();+        }+        if !exported.contains(&name) {+            out.insert(name);+        }