[aws-neuron/nki-library#11] fix(scatter_add): accumulate duplicate destination indices via `unique_indices` flag

scatter_add silently drops contributions when index repeats a destination row inside the same 128-row tile. That is the common case for embedding gradients, so grad_table.scatter_add_(0, token_ids, grads) currently returns wrong gradients for any batch containing a repeated token.

This PR adds an opt-in unique_indices=False path that accumulates correctly. The default is unchanged, so existing callers and the existing fast path are untouched.

ryan-williams / marin#7470.md

Last active July 21, 2026 20:45

marin-community/marin#7470 - 2-way sync via ghpr (https://github.com/runsascoded/ghpr)

[marin-community/marin#7470] [rigging] `username_segment`: prefer launch identity over worker OS user

username_segment() now resolves the launch identity carried in MARIN_PROVENANCE (built_by, stamped into the job env by the submitting client since #7000) before falling back to the OS login. On an Iris worker the OS login is root, so users/<seg>/ checkpoints from remote runs all landed under users/root/ (~275 GB across 3 buckets in the week to 2026-07-20) even though .artifact.json built_by correctly named the submitting human. The env payload is parsed directly (no git subprocess); a malformed payload warns and falls back to the OS login, a valid payload without built_by falls back silently (the submitting host not resolving a user is absence, not an error), and the unresolvable case still raises.

When the resolved segment is a known machine login (root, runner, ubuntu, exedev), a warning is logged: a per-user namespace under a machine identity always means launch identity was n