ML engineer, Glasgow
Most of this starts on my NVIDIA GB10, where newer hardware exposes assumptions that older GPUs never tested. I reproduce the failure, find the cause and send the fix. As of 1 November 2026, the pull requests featured here include 8 landed (including one co-authored PR) and 7 in review.
Accepted through the ONNX Slack
The MatMul operator's description mostly pointed readers to NumPy, leaving them to work out the matrix, vector and batch-shape rules themselves. I rewrote it to state those rules directly, with a nine-row shape table and a worked numerical example, across the shared schema for versions 1, 9 and 13.
My patch also has 14 shape-inference checks for inputs the documentation says are invalid. The behaviour of the operator doesn't change.
Accepted through the ONNX Slack. The public pull request hasn't been opened yet.
ยท Co-author
Output stores in epilogue_subtile could split along the wrong dimension, causing compilation failures. My fix in #3896 maps the split tile to its actual output dimension and leaves unsupported indexers unsplit. It landed as part of #3963, alongside fixes for static slices and runtime index lowering.
Co-authored the merged commit.
Under torch.compile, an OrderedSet lost its insertion order at the compile boundary, because TorchDynamo modelled it as a plain set. I split OrderedSetVariable off SetVariable so it follows the real Python type.
Calling .add on a frozenset under torch.compile raised an internal assertion. CPython raises AttributeError. Splitting FrozensetVariable off SetVariable makes compiled code fail the same way eager code does.
A scaled_mm config tuned for H100 needed 147,456 bytes of shared memory, and a GB10 has 101,376, so the test ran a config that could never launch. The test now skips on GPUs with under 147,456 bytes of shared memory, GB10 included, and sm121 has its own heuristic file, which I checked in a separate run with the guard bypassed.
Added an sm121 tier to the 4-bit GEMM dispatch heuristic, chosen from measured timings. To show nothing else moved, I enumerated all 60,720 dispatch decisions in the test grid: 216 changed, and every one was on sm121.
Extended the same work to the wide_n_simt dispatch rule, so GB10 gets the wide kernel path too.
Cut the v2.10.0 documentation for HAMi, the CNCF GPU virtualisation project, in English and Chinese. I also corrected the memory units, resource formats, and the device registration and health-check explanations against the v2.10.0 release.
A one-line fix. pyproject.toml had no requires-python, so pip would install the package on interpreters it doesn't support.
Opened 20 Oct 2026
On a GPU with no heuristic file of its own, configs from an older GPU can exceed its limits. On GB10 with only the sm90 file, 19 of 24 decode shapes failed at first launch. This fits a config to the GPU when it first runs.
Opened 16 Oct 2026
The autotuner silently dropped seed configs with short per-loop lists. On GB10 that was all 97 scaled_mm configs, and the log gave no reason. This fills the short lists.
Opened 6 Sep 2026
Fixes FP8 weight synchronisation for colocated vLLM, so vLLM rebuilds its FP8 weights and scales from the current training weights on each sync.
Opened 6 Sep 2026
A model call finishing during a pricing lookup could make a usage summary disagree with itself. The summary now works from one snapshot of the call history.
Opened 19 Aug 2026
Models whose layers have different attention shapes crashed during attention or KV-cache quantisation. Shapes are now read from each layer's own config.
Opened 18 Aug 2026
Consumer Blackwell has no tile::scatter4, so desc.scatter() failed deep in ptxas as an internal error. It's now rejected at compile time with a message that names the hardware limit.
Opened 31 Jul 2026
The fused cross-entropy backward failed under torch.compile on an overload TorchInductor can't lower. The fast path now steps aside while compiling, and eager mode is unchanged.
A GB10 on my desk lets me test bugs on hardware many maintainers don't have.