$ the-wire · showcase
tensordict excludes CVE-affected PyArrow, clears its 0.15.3 print regressions
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
tensordict did the heaviest lifting today, blocking PyArrow releases hit by CVE-2023-47248 and walking back four HDF5/memmap printing and loading regressions introduced since 0.15.3, while test-infra made downstream triage verdicts visible in Dr. CI.
[BugFix] Exclude PyArrow versions affected by CVE-2023-47248 (#3400) pytorch/tensordict
tensordict now excludes the PyArrow versions affected by CVE-2023-47248, so a fresh install will no longer resolve to an arrow build carrying that vulnerability. The trade-off is that any environment previously pinned to an excluded arrow version will need a newer one before tensordict is installed.
[BugFix] Print h5 enum datasets, uint64 element views and lazy stacks as v0.15.3 did pytorch/tensordict
Printing an h5 or zarr file and its views had drifted from 0.15.3 in three cases, and the detailed print of an in-memory lazy stack had gotten roughly 22% slower; this restores both the text and the speed. The printing page now also lists the prints that still read leaves.
[BugFix] append_memmap appends to a directory saved again after an append was killed pytorch/tensordict
A killed append left `.meta.json.append.tmp` files behind, and after the directory was saved again with different dtypes or nested keys the next `append_memmap` raised while `load_memmap` silently read new data against old dtypes. `from_h5` had a narrower version of the same story: a `__tensordict__` attribute that is a JSON object with wrongly typed values raised on files 0.15.3 loads, and now...
[CRCR] Show downstream triage verdicts in the Dr. CI comment in PRs for L3+ repo pytorch/test-infra
Downstream repos that attach a triage verdict to their CRCR callback now see it as an "AI verdict:" line under each CRCR job in the Dr. CI comment, matching the in-tree AI CI Advisor format. The verdict is advisory only: L4 failures stay under NEW FAILURES and L3 failures stay under the non-blocking CRCR section.
[inductor] Fix test hangs from a fresh_cache() finalized by the garbage collector (#202834) pytorch/pytorch
inductor's `test_flex_decoding.py` was hanging at a random early test on the Python 3.12 and 3.13 CPU lanes because `TestFlexDecoding.setUp` entered `fresh_cache()` on the `InductorTestCase` ExitStack and then called `skipTest`, so unittest never ran `tearDown` and the GC finalized the cache later. The fix removes the intermittent hangs rather than the test coverage.
Row writes into the DRAM KV enrichment cache (#6675) pytorch/FBGEMM
by Tongxin Wang
The quiet remainder: XPU test enablement across quantization and fake-tensor suites, FBGEMM shipping `hash_zch_runtime_meta` and `hash_zch_identities` on the HBM lane so MPZCH tables arrive at serving with their metadata, and test-infra collapsing the GreenLight policy replay reads into a single `collect` job.
Action items
- → Upgrade pyarrow to a version not affected by CVE-2023-47248 before installing or updating tensordict pytorch/tensordict [plan]
