PyTorch
Inductor fixes stale captured-tensor mutations in while_loop
PyTorch's Inductor now correctly replays cond_fn mutations of captured tensors in while_loop, fixing silent wrong results in later compiled calls.
read --wire →
$ tail -f topics/python.log
Python news, every morning.
Daily updates from across the Python ecosystem - Django, FastAPI, PyTorch, Hugging Face, and the data-and-AI stack.
34 updates across 5 projects this week.
Your week in Python, every Monday.
One calm review of what shipped across Python - the commits, releases, and security advisories that matter. Every Monday, with security advisories same-day. Free, unsubscribe in one click.
We'll start you on the top Python projects - refine anytime. · Read a sample issue →
Recent wires
Hugging Face
TRL documents 1M-token training, diffusers fixes group offloading, kernels fix Intel heuristics
TRL ships a runnable million-token SFT example, diffusers repairs a group-offloading device mismatch, and triton-kernels corrects Intel autotuning constants that cost roughly 10x on Battlemage.
read --wire →
JAX removes deprecated Pallas APIs, adds barrier-skip for multi-device calls
JAX purged a set of deprecated Pallas APIs and gave Mosaic GPU callers a way to skip multi-device barriers.
read --wire →
Django
Django deprecation warnings switch to calendar-based names
Django renamed its deprecation warning classes to match the 2028 calendar-based versioning, with old names kept as aliases to avoid breaking third-party code.
read --wire →
Hugging Face
Jinja engine aligns with Jinja2; hub-kernel fallbacks now warn once
The huggingface.js Jinja engine now supports Jinja2-style namespaces, `**` unpacking, and `**` exponentiation, while transformers warns once when a hub-kernel function silently falls back to its reference PyTorch path.
read --wire →
FastAPI & Pydantic
Pydantic hardens JSON mode serialization error handling
Pydantic's JSON mode serialization now catches any exception when computing JSON Schema defaults, reducing spurious errors that leaked through.
read --wire →
PyTorch
PyTorch 2.14.0 ships, import reload guarded, AI reviewer unblocks merges
PyTorch 2.14.0 is out with NVGEMM on Inductor, while a safety guard stops reload(torch) segfaults and the AI CI Advisor now clears non-blocking failures.
read --wire →
Bigtable reroutes CheckAndMutateRows and ReadModifyWriteRows
Bigtable's conditional and append row commits now route through the data client instead of raw protos.
read --wire →
Django
Django fixes two query regressions in ORM
Two regressions in the Django ORM, one in distinct() on PostgreSQL and one in in_bulk(), are fixed in pull requests submitted this week.
read --wire →
PyTorch
ExecuTorch and TensorRT harden CUDA edge cases
ExecuTorch and TensorRT shipped fixes for CUDA load failures, annotation leaks, and graph-lifting bugs, while torchdistributed and torchtitan clarified security and checkpoint handling.
read --wire →
FastAPI & Pydantic
Dependency churn across FastAPI and SQLModel, decimal validation hardening in Pydantic
Dependency bumps dominate FastAPI and SQLModel this period, while Pydantic changes how it validates decimals.
read --wire →
Hugging Face
kernels gains minver checks, diffusers deprecates ONNX
huggingface/kernels now validates kernel metadata against a minimum version, and diffusers starts deprecating its ONNX export path.
read --wire →
JAX refactors array construction for free-threading safety
JAX's array construction is being refactored to prepare for Python free threading, removing unsafe internals and moving validation out of the constructor.
read --wire →
Django
Django fixes equality and XML serialization for unsaved composite primary keys
Django's Model.__eq__ and XML serializer now use _is_pk_set() to correctly handle unsaved instances with composite or db_default primary keys.
read --wire →
Hugging Face
Transformers drops deprecated code, adds NeoMME and gradient checkpointing offload
Transformers, TRL, and Kernels-Community ship fixes and features that touch long-standing workflows, from deprecated API removal to VLM training and flash-attn2 pointer bugs.
read --wire →
python-genai 2.21.0 ships streaming downloads; JAX hardens deserialization and locking
python-genai 2.21.0 lands streaming file downloads and new translation_config support, while JAX hardens PyTree deserialization and thread-safety.
read --wire →
Django
Django fixes changelist search crashes and boosts djangoproject.com search speed
Django fixed a changelist search crash and over-matching bug, while djangoproject.com cut search page load time by 600ms.
read --wire →
Hugging Face
Transformers tightens integration tests, TRL pins hybrid attention coverage
Hugging Face's transformers repo fixed four integration tests, TRL pinned layer types so tiny hybrid models actually cover both attention paths, and diffusers shipped a major new LTX-2.5 pipeline.
read --wire →
PyTorch
PyTorch fixes MPS complex std/var and Sequential clone naming
PyTorch core landed fixes for a wrong MPS complex variance result and a silent state_dict key rewrite on Sequential clone, plus torchtitan extended its CI to B200 for Kimi K3.
read --wire →
JAX fixes arg_exclude_mask initialization and updates XLA dependency
JAX corrects arg_exclude_mask value initialization when exclude_argnames is None and moves to a new XLA revision.
read --wire →