Aug 31, 2026 – Sep 6, 2026

[generate] stop synchronizing the accelerator on every decode stepPython · 185 + / 3 −

Introduces 13 new declarations in src/transformers/generation/utils.py.

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

src/transformers/generation/utils.py ↗ · 5 files

@@ -356,6 +357,127 @@ class GenerateBeamEncoderDecoderOutput(ModelOutput): GenerateOutput = GenerateNonBeamOutput | GenerateBeamOutput  +def _undo_generation_steps(num_steps: int, input_ids: torch.LongTensor, *recorded: "tuple | None") -> tuple:+    """+    Undo the last `num_steps` decoding steps, so that they leave no trace in what `generate` returns.+    Note that the cache entries those steps wrote are dropped by `DeferredStopCheck.finish` instead.+    """+    # `[:-0]` is `[:0]`, which would empty everything rather than leave it alone+    if num_steps == 0:+        return (input_ids, *recorded)+    return (input_ids[..., :-num_steps], *(record[:-num_steps] if record else record for record in recorded))+++class StopCheck:+    """+    Decides when the decoding loop should stop, and hands each new token to a streamer.+    """++    def __init__(self, streamer: "BaseStreamer | None" = None):+        self.streamer = streamer++    def __call__(self, unfinished_sequences: torch.Tensor, tokens: torch.Tensor, length: int) -> bool:+        if self.streamer is not None:+            self.streamer.put(tokens.cpu())+        return bool(unfinished_sequences.max() == 0)++    def finish(self) -> int:+        """Flush whatever is still held back, and report how many decoding steps have to be undone."""+        return 0+++class DeferredStopCheck(StopCheck):

Model: add Tencent Hy 4 (hy_v4) preview architecture supportC++ · 734 + / 4 −

Introduces 3 new declarations in src/models/hy-v4.cpp.

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

src/models/hy-v4.cpp ↗ · 18 files

@@ -0,0 +1,601 @@+#include "models.h"++#include "llama-kv-cache.h"+#include "llama-kv-cache-dsa.h"++#include <cmath>++// iHC (independent Hyper-Connections) helpers. Same layout as the DeepSeek-V4 HC, but without+// the comb/sinkhorn term: hc_fn makes only 2*hc coefficients (pre + post). The streams mix+// through the pre-reduce / post-distribute round trip instead.++static size_t hy_v4_elem_offset(const ggml_tensor * t, int64_t i) {+    return ggml_row_size(t->type, i);+}++static ggml_tensor * hy_v4_view_1d(ggml_context * ctx, ggml_tensor * t, int64_t ne0, int64_t i0) {+    return ggml_view_1d(ctx, t, ne0, hy_v4_elem_offset(t, i0));+}++static ggml_tensor * hy_v4_view_2d(ggml_context * ctx, ggml_tensor * t, int64_t ne0, int64_t ne1, int64_t i0) {+    return ggml_view_2d(ctx, t, ne0, ne1, t->nb[1], hy_v4_elem_offset(t, i0));+}++void llama_model_hy_v4::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_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);+    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,       hparams.n_lora_q);+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,      hparams.n_lora_kv);+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,    hparams.n_embd_head_k_mla_impl);+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA,  hparams.n_embd_head_v_mla_impl);+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);

fix(npm): correct the pnpm lockfile import against real-world lockfilesRust · 848 + / 20 −

Introduces 19 new declarations in libs/resolver/pnpm_lockfile_import.rs.

Adds new language design rather than adjusting what was there. Read the linked PR for the surrounding test context.

libs/resolver/pnpm_lockfile_import.rs ↗ · 1 files

@@ -756,6 +972,618 @@ snapshots:     assert_eq!(v["npm"]["lodash@4.17.21"]["integrity"], "sha512-AAA");   } +  #[test]+  fn aliased_snapshot_dependency() {+    // `string-width-cjs: string-width@4.2.3` names a package, not a version,+    // so it must not collapse into `string-width-cjs@string-width@4.2.3`.+    let input = r#"+lockfileVersion: '9.0'++importers:+  .:+    dependencies:+      wrap-ansi:+        specifier: ^8.1.0+        version: 8.1.0++packages:+  wrap-ansi@8.1.0:+    resolution: {integrity: sha512-WRAP}+  ansi-styles@6.2.1:+    resolution: {integrity: sha512-ANSI}+  string-width@4.2.3:+    resolution: {integrity: sha512-WIDTH}+  '@scope/pkg@1.0.0':+    resolution: {integrity: sha512-SCOPED}++snapshots:+  wrap-ansi@8.1.0:+    dependencies:+      ansi-styles: 6.2.1+      string-width-cjs: string-width@4.2.3+      scoped-alias: '@scope/pkg@1.0.0'

[Model] Support for Spark2_5ForCausalLM implementationC++ · 178 + / 0 −

Introduces 2 new declarations in src/models/spark2-5.cpp.

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

src/models/spark2-5.cpp ↗ · 18 files

@@ -0,0 +1,146 @@+#include "models.h"++void llama_model_spark2_5::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_ATTENTION_SLIDING_WINDOW, hparams.n_swa);++    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;+    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);++    hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;+    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);++    switch (hparams.n_layer()) {+        case 28: type = LLM_TYPE_1_7B; break;+        default: type = LLM_TYPE_UNKNOWN;+    }+}++void llama_model_spark2_5::load_arch_tensors(llama_model_loader &) {+    LLAMA_LOAD_LOCALS;++    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);++    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);+    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);+    if (output == nullptr) {+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);+    }++    for (int i = 0; i < n_layer; ++i) {+        auto & layer = layers[i];+

Introduce `rustc_middle::middle::resolve`Rust · 378 + / 363 −

Introduces 28 new declarations in compiler/rustc_middle/src/middle/resolve.rs.

Adds new language design rather than adjusting what was there. Read the linked PR for the surrounding test context.

compiler/rustc_middle/src/middle/resolve.rs ↗ · 28 files

@@ -0,0 +1,309 @@+//! This module contains types that carry name resolution results from `rustc_resolve` to a+//! consumer in another crate (e.g. AST lowering, metadata, or a query).++use rustc_ast::node_id::NodeMap;+use rustc_ast::{self as ast, NodeId};+use rustc_attr_ir::StrippedCfgItem;+use rustc_data_structures::fx::{FxIndexMap, FxIndexSet};+use rustc_data_structures::steal::Steal;+use rustc_data_structures::unord::{UnordMap, UnordSet};+use rustc_errors::{ErrorGuaranteed, LintBuffer};+use rustc_hir::def::{DefKind, Namespace, PerNS, Res};+use rustc_hir::def_id::{CrateNum, DefId, LocalDefId, LocalDefIdMap, LocalModId, ModId};+use rustc_hir::definitions::PerParentDisambiguatorState;+use rustc_hir::{MissingLifetimeKind, TraitCandidate};+use rustc_macros::{StableHash, TyDecodable, TyEncodable};+use rustc_span::{ExpnId, Ident, Span, Symbol};+use smallvec::SmallVec;++use crate::middle::privacy::EffectiveVisibilities;+use crate::ty::Visibility;++/// The result of resolving a path before lowering to HIR,+/// with "module" segments resolved and associated item+/// segments deferred to type checking.+/// `base_res` is the resolution of the resolved part of the+/// path, `unresolved_segments` is the number of unresolved+/// segments.+///+/// ```text+/// module::Type::AssocX::AssocY::MethodOrAssocType+/// ^~~~~~~~~~~~  ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~+/// base_res      unresolved_segments = 3+///

crypto: fix multi-prime RSA JWKsC++ · 102 + / 8 −

Reworks 99 lines of existing logic in src/crypto/crypto_rsa.cc.

Changes how existing runtime behaves. Tests changed with it, with code of their own.

src/crypto/crypto_rsa.cc ↗ · 7 files

@@ -417,19 +448,79 @@ KeyObjectData ImportJWKRsaKey(Environment* env, Local<Object> jwk) {     ByteSource dq = ByteSource::FromEncodedString(env, dq_value.As<String>());     ByteSource qi = ByteSource::FromEncodedString(env, qi_value.As<String>()); -    if (!rsa_view.setPrivateKey(-            d.ToBN(), q.ToBN(), p.ToBN(), dp.ToBN(), dq.ToBN(), qi.ToBN())) {+    ncrypto::Rsa::OtherPrimeInfoPointers other_prime_infos;+    if (!oth_value->IsUndefined()) {+      if (!oth_value->IsArray()) {+        THROW_ERR_CRYPTO_INVALID_JWK(env, "Invalid JWK RSA key");+        return {};+      }++      Local<Array> oth = oth_value.As<Array>();+      const uint32_t length = oth->Length();+      if (length == 0 || length > kMaxRsaOtherPrimeInfos) {+        THROW_ERR_CRYPTO_INVALID_JWK(env, "Invalid JWK RSA key");+        return {};+      }+      other_prime_infos.reserve(length);+      for (uint32_t i = 0; i < length; i++) {+        Local<Value> item_value;+        Local<Value> r_value;+        Local<Value> other_d_value;+        Local<Value> t_value;+        if (!oth->Get(env->context(), i).ToLocal(&item_value) ||+            !item_value->IsObject()) {+          THROW_ERR_CRYPTO_INVALID_JWK(env, "Invalid JWK RSA key");+          return {};+        }++        Local<Object> item = item_value.As<Object>();+        if (!item->Get(env->context(), env->jwk_r_string()).ToLocal(&r_value) ||+            !item->Get(env->context(), env->jwk_d_string())

Clean up the AST visitorRust · 189 + / 230 −

Introduces 14 new declarations in compiler/rustc_ast/src/visit.rs.

Adds new language design rather than adjusting what was there. Read the linked PR for the surrounding test context.

compiler/rustc_ast/src/visit.rs ↗ · 2 files

@@ -929,98 +916,110 @@ macro_rules! common_visitor_and_walkers {             ) -> V::Result {                 match self {                     ForeignItemKind::Static(item) =>-                        visit_visitable!($($mut)? vis, item),+                        visit_visitable!(vis, item),                     ForeignItemKind::Fn(func) => {-                        let kind = FnKind::Fn(FnCtxt::Foreign, visibility, &$($mut)?*func);+                        let kind = FnKind::Fn(FnCtxt::Foreign, visibility, &$($mut)? *func);                         try_visit!(vis.visit_fn(kind, attrs, span, id))                     }                     ForeignItemKind::TyAlias(alias) =>-                        visit_visitable!($($mut)? vis, alias),+                        visit_visitable!(vis, alias),                     ForeignItemKind::MacCall(mac) =>-                        visit_visitable!($($mut)? vis, mac),+                        visit_visitable!(vis, mac),                 }                 V::Result::output()             }         } -        pub fn walk_fn<$($lt,)? V: $Visitor$(<$lt>)?>(vis: &mut V, kind: FnKind<$($lt)? $(${ignore($mut)} '_)?>) -> V::Result {+        pub fn walk_fn<$($lt,)? V: $Visitor$(<$lt>)?>(+            vis: &mut V,+            kind: FnKind<$($lt)? $(${ignore($mut)} '_)?>,+        ) -> V::Result {             match kind {                 FnKind::Fn(                     _ctxt,                     // Visibility is visited as a part of the item.                     _vis,-                    Fn { defaultness, ident, sig, generics, contract, body, define_opaque, eii_impl },+                    Fn {

[rust-compiler] Propagate block lowering errorsRust · 38 + / 27 −

Introduces 1 new declaration in compiler/crates/react_compiler_lowering/src/build_hir.rs.

Adds new language design rather than adjusting what was there. Read the linked PR for the surrounding test context.

compiler/crates/react_compiler_lowering/src/build_hir.rs ↗ · 2 files

@@ -2589,17 +2588,25 @@ fn lower_block_statement(     block: &react_compiler_ast::statements::BlockStatement,     parent_scope: Option<react_compiler_ast::scope::ScopeId>, ) -> Result<(), CompilerError> {-    let _ = lower_block_statement_inner(builder, block, None, parent_scope);-    Ok(())+    Ok(lower_block_statement_inner(+        builder,+        block,+        None,+        parent_scope,+    )?) }  fn lower_block_statement_with_scope(     builder: &mut HirBuilder,     block: &react_compiler_ast::statements::BlockStatement,     scope_override: react_compiler_ast::scope::ScopeId, ) -> Result<(), CompilerError> {-    let _ = lower_block_statement_inner(builder, block, Some(scope_override), None);-    Ok(())+    Ok(lower_block_statement_inner(+        builder,+        block,+        Some(scope_override),+        None,+    )?) }  fn lower_block_statement_inner(

quic: stop guarding ngtcp2_recv_stop_sending callback fieldC++ · 16 + / 20 −

Reworks 32 lines of existing logic in src/quic/session.cc.

Changes how existing framework internals behaves. Read the linked PR for the surrounding test context.

src/quic/session.cc ↗ · 2 files

@@ -1720,44 +1718,42 @@ struct Session::Impl final : public MemoryRetainer {       on_stream_open,       on_stream_close,       nullptr,  // recv_stateless_reset (deprecated, use v2 below)-      nullptr,+      nullptr,  // recv_retry       on_extend_max_streams_bidi,       on_extend_max_streams_uni,       on_rand,       nullptr,  // get_new_connection_id (deprecated, use v2 below)       on_remove_connection_id,       ngtcp2_crypto_update_key_cb,       on_path_validation,-      nullptr,+      nullptr,  // select_preferred_addr       on_stream_reset,       on_extend_max_remote_streams_bidi,       on_extend_max_remote_streams_uni,       on_extend_max_stream_data,       nullptr,  // dcid_status (deprecated, use v2 below)-      nullptr,-      nullptr,+      nullptr,  // handshake_confirmed+      nullptr,  // recv_new_token       ngtcp2_crypto_delete_crypto_aead_ctx_cb,       ngtcp2_crypto_delete_crypto_cipher_ctx_cb,       on_receive_datagram,       on_acknowledge_datagram,       on_lost_datagram,       nullptr,  // get_path_challenge_data (deprecated, use v2 below)-      nullptr,  // stream_stop_sending+      nullptr,  // stream_stop_sending (deprecated, use v2 below)       ngtcp2_crypto_version_negotiation_cb,-      nullptr,