Oct 16, 2026

Serve legacy module aliases through a sys.meta_path finderPython · 60 + / 51 −

Introduces 7 new declarations in src/transformers/__init__.py.

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

src/transformers/__init__.py ↗ · 3 files

@@ -821,54 +829,22 @@         extra_objects={"__version__": __version__},     ) -    def _create_module_alias(alias: str, target: str) -> None:-        """-        Lazily redirect legacy module paths to their replacements without importing heavy deps.-        """-        module = types.ModuleType(alias)-        module.__doc__ = f"Alias module for backward compatibility with `{target}`."-        # Set __file__ explicitly so that inspect.py's hasattr(module, '__file__') check-        # never falls through to __getattr__ and triggers a premature (possibly circular) import.-        module.__file__ = None--        def _get_target():-            return importlib.import_module(target, __name__)--        module.__getattr__ = lambda name: getattr(_get_target(), name)-        module.__dir__ = lambda: dir(_get_target())--        sys.modules[alias] = module-        setattr(sys.modules[__name__], alias.rsplit(".", 1)[-1], module)--    _create_module_alias(f"{__name__}.tokenization_utils_fast", ".tokenization_utils_tokenizers")-    _create_module_alias(f"{__name__}.tokenization_utils", ".tokenization_utils_sentencepiece")-    _create_module_alias(f"{__name__}.image_processing_utils_fast", ".image_processing_backends")--    for _proc_file in sorted((Path(__file__).parent / "models").rglob("image_processing_*.py")):-        _model = _proc_file.parent.name-        _module = _proc_file.stem-        _target = f".models.{_model}.{_module}"-        _create_module_alias(f"{__name__}.models.{_model}.{_module}_fast", _target)--        # Also map XImageProcessorFast -> XImageProcessor for backward compat with old class names.

Rollup of 7 pull requestsRust · 207 + / 188 −

Introduces 5 new declarations in library/std/src/sys/stdio/windows.rs.

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

library/std/src/sys/stdio/windows.rs ↗ · 38 files

@@ -135,141 +138,125 @@ fn write_console_utf16(     incomplete_utf8: &mut IncompleteUtf8,     handle: c::HANDLE, ) -> io::Result<usize> {-    if incomplete_utf8.len > 0 {-        assert!(-            incomplete_utf8.len < 4,-            "Unexpected number of bytes for incomplete UTF-8 codepoint."-        );-        if data[0] >> 6 != 0b10 {-            // not a continuation byte - reject-            incomplete_utf8.len = 0;-            return Err(io::const_error!(-                io::ErrorKind::InvalidData,-                "Windows stdio in console mode does not support writing non-UTF-8 byte sequences",-            ));-        }-        incomplete_utf8.bytes[incomplete_utf8.len as usize] = data[0];-        incomplete_utf8.len += 1;-        let char_width = utf8_char_width(incomplete_utf8.bytes[0]);-        if (incomplete_utf8.len as usize) < char_width {-            // more bytes needed-            return Ok(1);-        }-        let s = str::from_utf8(&incomplete_utf8.bytes[0..incomplete_utf8.len as usize]);-        incomplete_utf8.len = 0;-        match s {-            Ok(s) => {-                assert_eq!(char_width, s.len());-                let written = write_valid_utf8_to_console(handle, s)?;-                assert_eq!(written, s.len()); // guaranteed by write_valid_utf8_to_console() for single codepoint writes-                return Ok(1);-            }-            Err(_) => {

ENH: Expand abstract DType hierarchyC · 560 + / 239 −

Introduces 3 new declarations in numpy/_core/src/multiarray/abstractdtypes.c.

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

numpy/_core/src/multiarray/abstractdtypes.c ↗ · 21 files

@@ -84,12 +208,144 @@ discover_descriptor_from_pycomplex( }  -NPY_NO_EXPORT PyArray_DTypeMeta *PyArray_IntAbstractDTypePtr = NULL;-NPY_NO_EXPORT PyArray_DTypeMeta *PyArray_FloatAbstractDTypePtr = NULL;-NPY_NO_EXPORT PyArray_DTypeMeta *PyArray_ComplexAbstractDTypePtr = NULL;-NPY_NO_EXPORT PyArray_DTypeMeta *PyArray_PyLongDTypePtr = NULL;-NPY_NO_EXPORT PyArray_DTypeMeta *PyArray_PyFloatDTypePtr = NULL;-NPY_NO_EXPORT PyArray_DTypeMeta *PyArray_PyComplexDTypePtr = NULL;+/* Forward declarations of the static slot tables defined further down. */+static NPY_DType_Slots pylongdtype_slots;+static NPY_DType_Slots pyfloatdtype_slots;+static NPY_DType_Slots pycomplexdtype_slots;+++NPY_NO_EXPORT int+initialize_abstract_dtypes(multiarray_umath_state *state)+{+    /*+     * Create the abstract DType classes that mirror the array-API "kind"+     * hierarchy (NumberAbstractDType -> IntegerAbstractDType ->+     * SignedIntegerAbstractDType / UnsignedIntegerAbstractDType, and+     * NumberAbstractDType -> InexactAbstractDType -> FloatingAbstractDType /+     * ComplexFloatingAbstractDType) and expose them on ``numpy.dtypes``.+     * This must run before ``set_typeinfo`` so the legacy concrete DType+     * classes can use these abstracts as ``tp_base``.+     */+    if (npy_cache_import_runtime("numpy.dtypes", "_add_dtype_helper",+                                 &state->runtime_imports._add_dtype_helper) < 0) {+        return -1;+    }+    for (size_t i = 0; i < Py_ARRAY_LENGTH(abstract_dtype_specs); ++i) {+        abstract_dtype_spec *spec = &abstract_dtype_specs[i];

Add Riva Translate v2 supportPython · 188 + / 10 −

Introduces 8 new declarations in nemo/collections/asr/inference/nmt/llm_translator.py.

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

nemo/collections/asr/inference/nmt/llm_translator.py ↗ · 3 files

@@ -263,18 +293,46 @@ def translate_batch(         for src_lang, tgt_lang, src_prefix, tgt_prefix, src_context, tgt_context in zip(             src_langs, tgt_langs, asr_transcripts, prefixes, src_contexts, tgt_contexts         ):-            text = self.prompt_template.format(src_lang, tgt_lang, src_prefix, tgt_prefix, src_context, tgt_context)+            if self.is_riva_v2:+                messages = self.prompt_template.messages(+                    src_lang, tgt_lang, src_prefix, tgt_prefix, src_context, tgt_context+                )+                has_target_prefix = bool(tgt_prefix.strip())+                text = self.prefix_tokenizer.apply_chat_template(+                    messages,+                    tokenize=False,+                    add_generation_prompt=not has_target_prefix,+                    continue_final_message=has_target_prefix,+                )+            else:+                text = self.prompt_template.format(+                    src_lang, tgt_lang, src_prefix, tgt_prefix, src_context, tgt_context+                )             input_texts.append(text)          outputs = self.nmt_model.generate(input_texts, self.sampling_params, use_tqdm=False)         translations = []-        for tgt_prefix, output in zip(prefixes, outputs):+        for tgt_prefix, tgt_lang, output in zip(prefixes, tgt_langs, outputs):             output_text = output.outputs[0].text-            output_text = self.prompt_template.extract(output_text).strip()+            output_text = self.prompt_template.extract(output_text)             if tgt_prefix:-                translations.append(f"{tgt_prefix} {output_text}")+                if self.is_riva_v2:+                    if self._is_chinese(tgt_lang):+                        separator = ""

model: add Beam (Reflection)C++ · 542 + / 11 −

Introduces 3 new declarations in src/models/beam.cpp.

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

src/models/beam.cpp ↗ · 20 files

@@ -0,0 +1,261 @@+#include "models.h"++void llama_model_beam::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);+    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_GATING_FUNC,                hparams.expert_gating_func, false);+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale);+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm);+    hparams.f_final_logit_softcapping = 0.0f; // absent key = no softcap (default is 30)+    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,           hparams.f_final_logit_softcapping, false);++    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {+        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;+    }+    if (hparams.n_layer_dense_lead > hparams.n_layer()) {+        throw std::runtime_error("beam: leading_dense_block_count exceeds the layer count");+    }++    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD; // masked iff p1 - p0 >= n_swa+    uint32_t n_pattern = 0;+    ml.get_arr_n(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, n_pattern);+    if (hparams.n_swa == 0 || n_pattern != hparams.n_layer_all) {+        throw std::runtime_error("beam: requires a positive sliding window and a sliding window pattern for each layer");+    }+    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);+    if (!hparams.is_swa_any()) {+        throw std::runtime_error("beam: sliding window pattern has no sliding window layer");+    }++    // full layers are NoPE+    hparams.rope_pattern = hparams.is_swa_impl;