10 Pull requests merged by 7 people

• TE: update test to be more stable

  #2156 merged Jun 1, 2025

• Update coverage requirement from ~=7.6.8 to ~=7.8.2

  #2158 merged Jun 1, 2025

• Update ipython[all] requirement from ~=8.36.0 to ~=8.37.0

  #2159 merged Jun 1, 2025

• Bump transformers from 4.50.3 to 4.52.4

  #2160 merged Jun 1, 2025

• Set input_quantizer.internal to True

  #2146 merged May 30, 2025

• Print repro command when test_core_vs_torch_consistency fails with sample index specified

  #2131 merged May 30, 2025

• add non-None check for torch.utils.collect_env.get_pip_packages outputs

  #2124 merged May 30, 2025

• Remove qualified access to methods from autodiff

  #2147 merged May 28, 2025

• [pre-commit.ci] pre-commit suggestions

  #2139 merged May 28, 2025

• Add custom decompositions for cross entropy loss for the nvfuser executor

  #2043 merged May 27, 2025

7 Pull requests opened by 7 people

• TE: Fix redundant compute for PEFT using transform

  #2138 opened May 26, 2025

• Diffuser test

  #2141 opened May 27, 2025

• try enabling ruff in pre-commit

  #2142 opened May 27, 2025

• Fix autodiff joint trace dataflow and in-place ops in higher order functions

  #2143 opened May 27, 2025

• add a use_hf option to benchmarking

  #2154 opened May 30, 2025

• Fix HFTransformers recipe for kv-cache

  #2155 opened May 30, 2025

• Bump pytest-timeout from 2.3.1 to 2.4.0

  #2161 opened Jun 1, 2025

2 Issues closed by 2 people

• Better to check `get_pip_packages` return values in `thunder.dynamo.utils.get_env`

  #2123 closed May 30, 2025

• `thunderfx_benchmark_report`'s output files could wrongly specify `executors`

  #2140 closed May 27, 2025

9 Issues opened by 6 people

• ThunderFX in BF16 mode shows a larger divergence from eager FP64 results compared to the divergence between eager BF16 and eager FP64

  #2153 opened May 30, 2025

• [pre-commit] Bump Python default version to 3.12?

  #2152 opened May 30, 2025

• can I use thunder to autotune and select kernel implementations based on input shapes.

  #2151 opened May 29, 2025

• register an error message to show to the user when an executor is not instantiated

  #2150 opened May 29, 2025

• don't register cudnn executor if cudnn is not available

  #2149 opened May 29, 2025

• Should thunderfx's splitter reach executor level while determining if thunder can execute an operation

  #2148 opened May 28, 2025

• Add trace dataflow correctness tests

  #2145 opened May 27, 2025

• Remove "block" and "layer" bucketing modes from FSDP implementation

  #2144 opened May 27, 2025

• order of dimensions in advanced indexing does not match PyTorch

  #2137 opened May 26, 2025

17 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Use joint trace in transform_for_execution

  #2102 commented on May 29, 2025 • 20 new comments

• `prims.where` computes broadcast shape to avoid runtime-trace mismatch

  #2135 commented on May 30, 2025 • 6 new comments

• ci: test with latest dependencies

  #2122 commented on May 27, 2025 • 2 new comments

• Deprecate `add_transform`

  #2136 commented on May 26, 2025 • 1 new comment

• Reduces the test time

  #2077 commented on May 30, 2025 • 1 new comment

• switch tests using check vjp to use PyTorch autograd

  #2134 commented on May 27, 2025 • 0 new comments

• Fixes the value_and_grad check in splitter on symbol scaled_dot_product_attention

  #2073 commented on May 28, 2025 • 0 new comments

• Add parity check of shape/dtype/device of runtime and trace

  #2069 commented on May 30, 2025 • 0 new comments

• [wip]

  #2052 commented on May 30, 2025 • 0 new comments

• Add test for Phi-3-vision-128k-instruct

  #1850 commented on May 31, 2025 • 0 new comments

• Use `f.write(...)` instead of `print(..., file=f)`

  #1831 commented on May 27, 2025 • 0 new comments

• ThunderFX fails with FP8 and Activation Checkpointing

  #1424 commented on May 28, 2025 • 0 new comments

• Cannot correctly compute cross entropy loss (backward) when the reduction used is "sum" and torch executor is used.

  #2056 commented on May 28, 2025 • 0 new comments

• cudnnex SDPA checker function doesn't handle FloatProxy when `cache="symbolic values"`

  #2120 commented on May 28, 2025 • 0 new comments

• Workaround for Adam.step's lazy initialization of its inner state

  #1057 commented on May 27, 2025 • 0 new comments

• ThunderFX reporting: The e2e performance result shows Torch is faster, but the saved FXgraph script says Thunder is faster

  #1968 commented on May 27, 2025 • 0 new comments

• TE Integration fails with TE main

  #2060 commented on May 26, 2025 • 0 new comments