0.2.4

What's Changed

• cleaning skipif for past Torch dev versions by @Borda in #2125
• fix missing images when released on PyPI by @Borda in #2130
• Add custom decompositions for cross entropy loss for the nvfuser executor by @protonu in #2043
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2139
• Remove qualified access to methods from autodiff by @riccardofelluga in #2147
• add non-None check for torch.utils.collect_env.get_pip_packages outputs by @crcrpar in #2124
• Print repro command when test_core_vs_torch_consistency fails with sample index specified by @crcrpar in #2131
• Set input_quantizer.internal to True by @beverlylytle in #2146
• Bump transformers from 4.50.3 to 4.52.4 by @dependabot in #2160
• Update ipython[all] requirement from ~=8.36.0 to ~=8.37.0 by @dependabot in #2159
• Update coverage requirement from ~=7.6.8 to ~=7.8.2 by @dependabot in #2158
• TE: update test to be more stable by @kshitij12345 in #2156
• Bump pytest-timeout from 2.3.1 to 2.4.0 by @dependabot in #2161
• Update dependabot - reviewers by @Borda in #2162
• Fix autodiff joint trace dataflow and in-place ops in higher order functions by @riccardofelluga in #2143
• Reduces the test time by @kiya00 in #2077
• add a use_hf option to benchmarking by @t-vi in #2154
• Bump pytest-xdist from 3.6.1 to 3.7.0 by @dependabot in #2164
• Update hypothesis requirement from ~=6.131.9 to ~=6.133.0 by @dependabot in #2165
• Update snowballstemmer requirement from <3 to <4 by @dependabot in #2168
• nvFuser Executor: Ensure cross-entropy loss fwd is not recomputed when computing bwd by @protonu in #2180
• Add parity check of shape/dtype/device of runtime and trace by @crcrpar in #2069
• Add docstrings for recipes by @KaelanDt in #2185
• sdpa_ex: relax test tolerances by @kshitij12345 in #2178
• removing nv_enable_embedding by @jjsjann123 in #2057
• Improve error reporting in benchmark job and add cleanup logic by @Borda in #2176
• Add mode flag to TorchCompileExecutor by @t-vi in #2188
• Use to_dtype and to_torch_dtype not _torch_to_thunder_dtype_map and _thunder_to_torch_dtype_map by @crcrpar in #2181
• use hf recipe in quickstart by @t-vi in #2191
• Remove kwarg construction from FusionDefinitionWrapper.call by @IvanYashchuk in #1871
• Add tests for HFTransformers recipe with static cache by @KaelanDt in #2179
• bump: PyTorch to be latest 2.7.1 by @Borda in #2193
• prims.where ignores shape/device of pred if it's a CPU scalar tensor by @crcrpar in #2135
• add decomposition for repeat interleave by @t-vi in #2194
• fix traceback in with / try: finally: for Python 3.10 by @t-vi in #2195
• Use joint trace in transform_for_execution by @beverlylytle in #2102
• Handle proxy objects in the cuDNN SDPA checker by @kiya00 in #2073
• default to hf recipe in thunder.compile for hf models by @KaelanDt in #2199
• add autocast lookaside to hf recipe for tracing on meta device by @t-vi in #2200
• make test_networks.py not rely on HF downloads by @KaelanDt in #2202
• Add plugins documentation by @KaelanDt in #2207
• implement partial, avoid tuple addition, test partialmethod by @t-vi in #2209
• Add bitwise_left_shift and bitwise_right_shift by @crcrpar in #2210
• [thunderfx] Avoid split at Tensor.__eq__ by registering it in thunder.torch by @crcrpar in #2211
• fixed installing NCCL for CUDA by @Borda in #2208
• Enable ruff-check in pre-commit by @crcrpar in #2192
• unxfail passing test by @t-vi in #2220
• Add #2192 to .git-blame-ignore-revs by @crcrpar in #2219
• fix cache validity issue, tighten assert by @t-vi in #2223
• Avoid negative number rhs values to bitwise shift tests by @crcrpar in #2227
• Add missing opinfo for bitwise_right_shift to elementwise_binary_ops by @crcrpar in #2214
• Enable ruff format in pre-commit by @crcrpar in #2142
• Fixes baddbmm() got an unexpected keyword argument 'batch1' by @kiya00 in #2228
• Only register cudnn executor if it is available by @KaelanDt in #2174
• Representing DTensor in thunder traces by @kshitij12345 in #1907
• install cudnn in quickstarts by @KaelanDt in #2235
• add experimental by @t-vi in #2236
• DTensor: don't error if torch.distributed is unavailable by @kshitij12345 in #2243
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2245
• bump: OS versions for CI by @Borda in #2249
• Plumbing the topk to the nvFuser executor by @protonu in #2237
• bump: bitsandbytes to its next compatible release by @Borda in #2248
• ci: test with latest dependencies by @Borda in #2122
• Adds empty_like,rand_like by @kiya00 in #2225
• Support converting SymTypes Node to input proxy by @kiya00 in #2171
• fix test_reports_benchmark timeout by @kiya00 in #2229
• Fix torch.gather function signature to accept input passed as keyword argument by @kiya00 in #2250
• Fix bitsandbytes dependency conditions to use platform_machine instead of sys_platform by @Borda in #2257
• add float exception to assertion in jit_ext by @KaelanDt in #2256
• register softmax fudge function for stacklevel by @t-vi in #2259
• include message in NotImplementedError in proxy methods by @t-vi in #2260
• add support for full with tensor input by @t-vi in #2262
• Remove W291, W293, E702, and F722 from ignore by @crcrpar in #2267
• Add #2142 of ruff format integration to .git-blame-ignore-revs by @crcrpar in #2266
• split getitem into basic and "purely" advanced indexing by @t-vi in #2258
• TE: fix related to delayed forward-backward split by @kshitij12345 in #2222
• bump version to 0.2.4 for release by @t-vi in #2273

Full Changelog: 0.2.3...0.2.4