PyTorch — full archive
117 daily briefings · 9 reviews.
$ ls pytorch/reviews/ # the step-back reads
- week torch.compile dtype fix, two Arm CVEs patched
- week Inductor correctness sweep lands with nextafter and index_put fixes
- week TorchTitan drops pure TP on routed experts
- week Helion autotunes Metal, ending default 16x16x16 tiles
- week TorchRL 0.14 lands with DreamerV3 and async collectors
- month PyTorch 2.14.0 ships, TorchRL 0.14 lands
- week PyTorch 2.14.0 ships, security docs clarified
- week Executorch stabilizes its build and strips legacy backends
- month ExecuTorch MoE quantization, Helion runtime splits, and a broader cleanup wave
- Tensordict compiles masked indexing under fullgraph; ExecuTorch links WebGPU into the Wasm build
- TensorDict 0.15.0 goes pure Python, 1.77x faster
- Inductor hands Triton kernels to the worker pool, RL rollout KV cache persists across turns
- ExecuTorch drops Python 3.10, TensorRT trims compile startup, tensordict fixes silent h5 loss
- GreenLight leaves beta on merge_rules.yaml, FSDP2 numerics fixed in TorchTitan
- TorchTitan RL workers stop importing vLLM, tensordict fixes its pytree registry leak
- One FSDP module at a time, packed-atomic revert, and TensorDict 0.15 removals
- TensorDict store scalars, TorchTitan's one-line GC fix
- Silently wrong MoE gradients, and E8M0 wrappers decompose again
- Executorch closes a security hole and two silent-alias bugs
- ReinforceLoss stops resampling actions, torch_tensorrt save fixed
- Executorch patches CWE-125 in Arm VGF decoder, PyTorch cleans up snapshots and CI
- torch.compile dtype fixes, Vulkan reduction fallbacks, and a silent RL temperature bug
- AOTI fallback fixes, Vulkan reductions stop clamping, Dist-MoE gets a TorchTitan adapter
- NVFP4 kernel pipeline gets PDL, fused scaling, and tuned autotuning; torchtitan drops stale train command
- NCCL split/shrink init failures now surface, torchtitan lands an RL KV-cache fix
- TorchTitan drops graph-based EP chunking, ditching FlexAttention in DSV4
- Inductor's nextafter lowering, cuDNN RNN strides, and a memory-snapshot refcount fix
- Inductor stops reinplacing index_put, ExecuTorch fixes ConvTranspose BatchNorm fusion
- Inductor cat merge fixed, AVX512 grid_sample returned zeros
- TorchTitan pipeline and CP sharding land, ExecuTorch quantization finally holds
- Green-context SM splitting, cat-kernel DCE, and a debug map that pinned every exported module
- TorchTitan unifies graph execution and makes grouped MoE projections the default
- torch.compiler.precompile becomes a module, callable API retired
- Dynamo stops silently folding with runtime_checkable Protocols, ExecuTorch drops CUDA 12.6 wheels
- NVFP4 and MXFP4 quantization land in TensorRT, GreenLight tightens its merge bar
- TorchTitan drops pure TP on routed experts, restacks fused FFN weights
- torchtitan fixes multi-node HF export, torch._native opens to non-aten namespaces
- Fake-tensor hardening, ROCm graph fixes, and an XNNPACK use-after-free
- CPU caching allocator teardown race fixed, torchtitan shares one training engine
- ROCm test timeouts traced to duplicated runs, and distributed/test_store drops 15 minutes
- TorchRL resume lands, TorchTitan fixes a silent gradient bug
- Dynamo stops deleting globals it never created, FP8 blockwise scale fixes land
- Helion autotunes Metal, TorchRL stops cacheing constructor device
- PyTorch trims CuTeDSL known-good versions, fixes lerp on MPS
- TorchTitan reshapes gradient accumulation for CUDA graphs, TensorRT 11.3 lands
- TorchRL 0.14 lands, ExecuTorch tunes Arm64 BF16 copies and Vulkan dispatch
- TorchRL flips async collector defaults, GreenLight adds shadow state
- DreamerV3 gains compile auto-selection, background replay, and native resume
- AsyncEnvPool shared-memory fixes, CUDA 12.6 nightly removal, and stale autotune Future collision
- MPS complex solves land, Arm round() matches torch, dynamo fixes repro fidelity
- Torchtitan trains MoE routers in fp32 with bf16x9 on Blackwell
- executorch fixes dropped edge-manager passes; TorchTitan overhauls MxFP8
- Inductor fixes stale captured-tensor mutations in while_loop
- PyTorch 2.14.0 ships, import reload guarded, AI reviewer unblocks merges
- ExecuTorch and TensorRT harden CUDA edge cases
- PyTorch fixes MPS complex std/var and Sequential clone naming
- CuTe grouped GEMM physical lowering lands, torch repros emit Python
- TorchRL gains native GRPO training; TensorRT fixes NCCL rebind crash
- PyTorch Kills the Legacy C Extension
- EXECUTORCH STRIPS MPS, PYTORCH DROPS UNUSED DEPS
- EXECUTORCH CUTS GRAPH REPLAYS, ARM KILLS NO-OP RESIZES
- FSDP2 CRASH FIXED, EXECUTORCH BUILD PIPELINE REPAIRED AFTER COMMIT PROMOTION BUG
- PyTorch's Caffe2 cleanup hits a wall as internal dependencies force three reverts
- Pallas HBM Windows Open on Any Tensor Axis
- NVFP4 GroupedExperts lands on Blackwell
- Pytorch Fixes XPU Dispatch Gap and CI OOMs
- PYTORCH SLASHES B200 BENCH TIME 5X
- PYTORCH CUTS MPS DEAD CODE, DREAMER V3 FIXES GPU MEMORY LEAK
- PYTORCH CORE FIXES CUDA MEMORY LEAKS AND CHECKPOINT STORAGE SHARING
Never miss a PyTorch briefing.
Get PyTorch's week in review every Monday - and security advisories the day they land. Free, unsubscribe in one click.
One click. Every Monday: what mattered across the projects you've been reading (or our AI picks) — security news same-day. You refine anytime. Choose specific journals · Read a sample issue →