$ the-wire · showcase
Torchtitan reshapes its config API, executorch repairs two broken CI lanes
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
Torchtitan's model config converters are gone in favor of a transform list and its FSDP all-gather now preserves quantized weight order, while executorch lands the fixes that turn its CUDA and ROCm jobs green again.
[config] Migrate ModelConfigConverter to ModelConfigTransform pytorch/torchtitan
Every converter is now a ModelConfigTransform: BatchInvariantFlexTransform, LMHeadFP32OutputTransform, and the MXFP8/NVFP4 linear and grouped-linear variants under the QuantizationTransform base. ModelConfigConverter, validate_converter_compatibility, and the converters argument of all 13 build_model_configs are deleted, so any recipe passing converters must move to apply_transforms.
Keep stacked quantized weights in order in the FSDP all-gather pytorch/torchtitan
Quantization plus FSDP scrambled linear layers because a _ShardedFSDPTensor weight sharded on a nonzero dim, such as a stacked [2, F, D] FFN w13, came out of the all-gather in the wrong order. H-Huang reports the fix has been validated on their jobs, with loss curves very similar to BF16.
[rl] SGLang engine: support SGLang's unified memory (KV and linear attention states in one pool) pytorch/torchtitan
For GDN/KDA models, SGLang splits memory into a fixed KV pool and a fixed state pool, so long-context and many-short-request workloads pull in opposite directions. The torchtitan PR reads SGLang's unified memory pool and, per acisseJZhong, "will delete this patch once upstream fix lands" in sgl-project/sglang.
Use torch 2.15.0 for every CUDA train (#25284) pytorch/executorch
Every CUDA train now pins torch 2.15.0, replacing the cu134 train's 2.15.0.dev20260921 nightly whose oldest wheels were days from expiring off the index, while cu132 and CPU move off 2.14.0. One set of pins across cpu, cu132, and cu134 removes that resolution cliff.
CUDA backend: create the temporary kernel library file safely (#24783) pytorch/executorch
The CUDA backend used to open its per-method temp kernel library with a plain std::ofstream under a pid-plus-counter name in a shared folder like /tmp, so it could fail the load or open a file it did not create. It now creates that file safely; the same convention-ledgers in source-repo note applies only to what the PR states.
Fix main CUDA CI: oversized conv configs in the test, test-models-cuda timeout (#25392) pytorch/executorch
Two CUDA jobs have been red on main since the #25270 to #25280 stack landed between fe24cf5777 and 6b82a162e7. The fix trims oversized conv configs behind the five test_conv_implicit_gemm failures and addresses the test-models-cuda timeout; the companion ROCm fix untangles ProbeTest skip order, a flaky SDPA pick test, and drops torchcodec from voxtral.
Action items