$ the-wire · showcase
TCPStore gets mutual TLS, NCCL 2.32.3 encryption exposed
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
The day's work is mostly plumbing and hardening: opt-in transport security for c10d, a CUDA graph allocator refcount fix, and a handful of CI and embedded build repairs.
[c10d] TCPStore: TLS encryption with mutual cert auth (#204159) pytorch/pytorch
TCPStore used to send unauthenticated plaintext; it now supports opt-in mutual TLS for the libuv server and blocking clients, configured through keyword arguments or TORCH_TCPSTORE_TLS_{CERT,KEY,CA}. OpenSSL 1.1.1+ bindings sit in TCPStoreOpenSSL.hpp and load at runtime, and the commit warns that TLS 1.2+ verifies certificate chains but not hostnames or subjects: "use a dedicated CA", since any...
[c10d][nccl2] Expose ncclSetEncryption PSK (#204291) pytorch/pytorch
NCCL 2.32.3 added TLS 1.3 PSK encryption of its TCP sockets via the process-global ncclSetEncryption, but PyTorch had no way to enable it. ProcessGroupNCCL.Options.encryption_psk is honored by nccl2, while legacy ProcessGroupNCCL (including shrink) rejects it and torchcomms raises rather than silently ignoring it.
Release the extra pool references taken by CUDA graph conditional nodes (#201639) pytorch/pytorch
Each conditional node in a CUDA graph re-routes allocations with endAllocateToPool plus beginAllocateToPool, and the extra pool references those re-routes took were never released. They now are, with a note that a CUDAPluggableAllocator's release_pool hook fires during capture once per re-route, so it should be treated as a reference-count decrement.
[c10d] TCPStore: TLS 1.3 pre-shared keys and TLS hardening (#204294) pytorch/pytorch
On top of certificate TLS, TCPStore gains TLS 1.3 PSK authentication with raw-value and owner-only-file configuration, all-empty disabling preserved and partial configurations rejected. Certificate TLS 1.2 is restricted to ECDHE/AEAD, certificate fallback is rejected, and REQUIRE-TLS mode logging makes accidental plaintext explicit.
Keep a packed constant alive while another node reads it (#25042) (#25042) pytorch/executorch
XNNCompiler::compileModel freed the named-data buffer of every packed filter and bias right after xnn_create_runtime_v4, even when another node still read that constant by raw pointer; Phonon's StreamingConvTranspose1d feeds the transposed-conv bias to both a convolution and a subtraction, so a file-loaded program crashed in its first mimi_decode call. The packed constant now stays alive for th...
[BugFix] detach() and dtype of a locked TensorDictParams follow model.double() and model.to() pytorch/tensordict
A locked TensorDictParams on which detach() had been called kept returning pre-move tensors after model.double(), model.to(...), model.half(), model.to("meta") or model.to_empty(...), and params.dtype stayed stale the same way; from_modules(..., as_module=True) produces exactly such a locked params object. This is the mechanism behind the RuntimeError that TorchRL 0.15.0 reproduces for a SACLos...
Revert "[Inductor] Filter unsupported transposed FP32 Blackwell configs (#196769)" pytorch/pytorch
Among the smaller entries: the Inductor transposed FP32 Blackwell config filter was reverted because it broke test signals internally, Get Started nightly commands now install torch alone, and the reusable Claude Code workflow takes a python-version input so callers can avoid the system Python 3.14 on ubuntu-latest.
Action items