Built in
the open.

Good infrastructure is invisible.
Making it better shouldn’t be.

Accepted contributions to the compilers, runtimes, and machine learning frameworks I build with.

53Accepted contributions

11Repositories

PyTorch / Inductor

Add global-scratch support to PyTorch's static Triton launcher

Device-side TMA kernels could not launch from cached FX graphs because the static Triton launcher rejected global-scratch metadata. The landed change preserves scratch size and alignment, allocates workspace for the full launch grid, passes it through the kernel ABI, and keeps scratch kernels off the incompatible fast launcher.

PyTorchTritonGPU runtime

Landed · Aug 14#191133

TensorFlow / XLA

Enable complex Cumsum and Cumprod in XLA

XLA rejected valid complex Cumsum and Cumprod operations before lowering. The landed change separates the real-only Logsumexp registration, enables complex64 and complex128 scans, and covers both widths across scan directions and axes.

C++XLAComplex dtypes

Landed · Aug 24#125286

Runtime correctness

Trap on synchronous cancel while inside a waitable set

Wasmtime needed cancellation to trap consistently when synchronous streams, futures, or subtasks were still registered in a waitable set. The change closes the runtime edge case and adds coverage across the component model implementation.

RustWasmAsync runtime

Merged · Jul 08#13708

Memory safety

Fix TFLite FFT buffer leak on temporary lookup failure

TFLite FFT kernels could leak heap work buffers when a fallible temporary-tensor lookup returned early. The fix moves the buffers into the interpreter-managed arena and covers dynamic FFT lengths.

C++TFLiteFFT

Merged · Aug 07#123994

Observability

Include configured guardrails in usage details

LiteLLM usage details omitted guardrails loaded from configuration, creating an incomplete view of what protected a request. The change unifies configured and request-level guardrail reporting.

Add global-scratch support to PyTorch's static Triton launcher

Enable complex Cumsum and Cumprod in XLA

Trap on synchronous cancel while inside a waitable set

Fix TFLite FFT buffer leak on temporary lookup failure

Include configured guardrails in usage details

Add global-scratch support to PyTorch's static Triton launcher

Prevent overflow when rounding BFC allocation sizes

Fix sort comparison type selection

Validate unigram sampler ranges before construction

Use safe shape validation for TensorScatter ops

Skip dense suffix expansion for empty ragged values

Fix FFN depth-scaled initialization target

Check sparse tensor map allocations

Infer zero-sized VALID AvgPool outputs

Fix zero gradients for variables with string dtypes

Preserve errors during eager uint64 conversion

Fix complex powers with zero exponent

Core: Preserve bfloat16 packet sign under FTZ and DAZ

Core: Preserve float and double sign under DAZ/FTZ

Reject out-of-range targets in XLA InTopK

Reject negative integer exponents in GPU Pow

Align subtract scalar coercion across execution modes

Validate FusedPadConv2D input rank before dimension access

Reject negative GPU NMS output sizes

Validate RandomIndexShuffle round count

Support int32 argmax in CPU MaxPool gradients

Add regression coverage for float64 softmax tail gradients

Preserve float64 log-softmax tail gradients

Use SoftplusGrad for stable higher derivatives

Handle zero-element batch slices safely

Preserve float64 SiLU second derivatives

Fix sqrt gradient for float64 subnormals

Validate dtype before capturing eager tensors

Cover XLA RandomShuffle comparisons for float vectors

Fix SELU gradient precision for negative inputs

Reject unrepresentable TensorList scatter indices

Fix ELU gradient precision for negative inputs

Enable complex Cumsum and Cumprod in XLA

Fix SiLU gradient underflow for large negative inputs

Prevent high-rank broadcast stack overflow in TFLite

Avoid eager shape ops in boolean_mask with static shapes

Preserve Keras tracing errors on empty fallback graphs

Clarify TFLite Interpreter input detail ordering

Fix TFLite FFT buffer leak on temporary lookup failure

Saturate constant StridedSlice indices when lowering to int32

Validate oneDNN depthwise input ranges before scalar access

Prevent process crashes across nine sparse optimizer kernels

Trap on synchronous cancel while inside a waitable set

Validate TensorListSetItem indices to prevent a segfault

Fix float32 erfinv precision loss near ±1

Make _uv_safe_path space-safe on macOS and Linux

Remove inter-task wakeup stream before cancellation

Force store=false for OpenRouter Responses API

Include configured guardrails in usage details

Use UTF-8 for code-execution subprocess I/O

Fix dataloader resume warning interpolation

Restore fast_inference compatibility with vLLM 0.22

Fix tokenizer compatibility after the vLLM 0.22 removal