ExecuTorch MoE quantization, Helion runtime splits, and a broader cleanup wave

ExecuTorch landed the piece that makes quantized MoE inference practical on LLMs: a full pipeline from export to kernel. The `--use_moe_quantized_op` flag now threads through the export pipeline and LLM config [1], backed by source transformation with INT4 quantized packing [2], and the ffn was optimized by fusing w1+w3 into a single GEMM [3]. Numerical, source-transform, and Meta-kernel tests all landed [4][5], so the operator is covered at every layer of the stack.

Helion's runtime went through a month of consolidation, moving scheduler- and accelerator-specific code out of the monolithic `runtime/__init__.py` into dedicated subdirectories. Cute-specific logic moved to `runtime/cute` [6], the TPU capture and compile helpers moved to `runtime/pallas` [7], and Triton helpers moved to `runtime/triton/helpers.py` [8]. The net effect is a cleaner module boundary: importing the base runtime no longer drags in every backend's internals, which should trim import time and reduce accidental coupling.

Dynamo tightened Python semantics in two ways. `torch.compile` can now be applied directly to a `staticmethod` without wrapping [9], and nested graph breaks no longer fire when an `nn.Module` has empty hook dicts [10]. Separately, default TLS attributes for guards were added so compiled-call `try_get()` avoids a `getattr` miss [11]. Correctness came to the meta paths as well: transposed convolutions now validate `output_padding` and bias up front [12], and the clamp family gained barebones unsigned integer support [13].

PyTorch's long-running cleanup of obsolete paths hit a visible milestone and a visible wall. The legacy C extension was killed , and unused deps plus MPS dead code were cut [12]. But the Caffe2 cleanup stalled: internal dependencies forced three reverts, a reminder that vestigial code still has users that the public repo can't see [ref:12?]. On the benchmarking side, B200 bench time was slashed 5x thanks to fusing the final reduction in `matmul_reduce_scatter` into a single Triton kernel [14].

Performance and memory fixes rounded out the month. FSDP2 crash was fixed , CUDA memory leaks and checkpoint storage sharing were addressed [ref:5?], and MPS complex `std`/`var` plus `Sequential` clone naming were corrected . TensorDict 0.14 shipped a breaking API default change [15], TorchRL gained native GRPO training , and Pallas now allows HBM windows on any tensor axis . ExecuTorch also repaired its build pipeline after a commit promotion bug and cut graph replays , while Arm removed no-op resizes .

Looking ahead, the through-line is clear: inference teams are getting production MoE support on ExecuTorch, Helion's runtime is becoming modular enough to reason about, and PyTorch core is still chipping away at legacy debt, with Caffe2 being the stubborn remainder. The cleanups are routine, but the MoE work is the one to watch for anyone deploying quantized LLMs on edge devices.