chore(deps): Bump the uv group across 1 directory with 3 updates

[dependabot]

Bumps the uv group with 3 updates in the / directory: accelerate, anyio and pygments.

Updates accelerate from 1.13.0 to 1.15.0

Release notes

Sourced from accelerate's releases.

v1.15.0: FSDP2 activation memory, dtensor improvements

FSDP2

A large batch of FSDP2 work this release: two fixes that cut activation memory at long sequence lengths, tied-embedding support on torch >= 2.13, and a round of checkpointing correctness and scale fixes.

Activation checkpointing was wrapping each child of the matched layer (self_attn, mlp, the norms) instead of the layer itself, so every inter-child activation stayed saved for backward. It now wraps the layer.

There's also a new FSDP2-only activation_checkpointing_offload, which moves the remaining per-layer checkpoint inputs to pinned CPU memory. Gradients are exactly those of plain activation checkpointing:

# fsdp2.yaml
fsdp_config:
  fsdp_version: 2
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_activation_checkpointing: true
  fsdp_activation_checkpointing_offload: true

accelerate launch --config_file fsdp2.yaml train.py

• FSDP2 activation checkpointing: wrap the matched transformer layer itself, not each of its children by @​qgallouedec in #4172
• Add FSDP2 activation_checkpointing_offload: offload checkpointed layer inputs to pinned CPU memory by @​qgallouedec in #4175
• Fix FSDP2 tied-embedding models on torch >= 2.13: put the output embedding in the same fully_shard group by @​qgallouedec in #4171
• Fix FSDP2/PEFT/FULL_STATE_DICT dropping every rank's adapter shard except rank 0 by @​AmineDiro in #4206
• FSDP2: per-rank torch.save/load for SHARDED_STATE_DICT to fix 2800+ NPU checkpoint timeout by @​gygdh-001 in #4105
• Fix FSDP sharded checkpoint path resolution by @​HaomingSong in #4119
• Raise a clear error when FSDP is enabled on a mesh with no shard dimension by @​qgallouedec in #4180

DTensor

Two fixes for DTensor-sharded models, which you hit with FSDP2, tensor parallelism, or any N-D parallelism setup: gradient clipping no longer fails on the foreach op when plain tensors and DTensors are mixed, and prepare_model leaves an already-sharded model where it is:

• Clip grad norm support for dtensors by @​michaelbenayoun in #4219
• prepare_model: don't move DTensor-sharded models to the device by @​qgallouedec in #4181

Offloading & Quantization

An entire model can now be dispatched to disk, including tied weights — useful for tools like llm-compressor that compress large models on machines that can't hold them:

• [Offloading] Support full disk offloading by @​kylesayrs in #4083
• Fix disk offload crash on FP8 tensors by @​shoemoney in #4151

Trackers

Custom trackers can be registered by name and then selected from log_with= like any built-in one:

from accelerate import Accelerator
from accelerate.tracking import register_tracker_class
</tr></table> 

... (truncated)

Commits

• 6afc1e5 Release: v1.15.0
• 636a934 Fix nightly CI: tensor parallel size detection and DeepSpeed warmup steps (#4...
• 385d9fb docs: fix three dead links left by the docs restructure (#4230)
• c424d0b docs: fix garbled sentence in CONTRIBUTING.md (#4231)
• 0f7e35f Clip grad norm support for dtensors (#4219)
• b2ac509 Fix FSDP2/ PEFT/FULL_STATE_DICT dropping every rank's adapter shard except ...
• 33e8bc4 Fix load_accelerator_state only restoring one RNG backend (#4217)
• 50de3ad Treat MPS out-of-memory errors as OOM in find_executable_batch_size (#4227)
• ab5fe8e feat: add the neuron device branch in state (#4218)
• 9baf95b Fix MLFLOW_NESTED_RUN never being able to turn nesting off (#4223)
• Additional commits viewable in compare view

Updates anyio from 4.12.1 to 4.14.2

Release notes

Sourced from anyio's releases.

4.14.2

• Changed ByteReceiveStream.receive() implementations to raise a ValueError when max_bytes is not a positive integer (#1191)
• Fixed CapacityLimiter.total_tokens rejecting float("inf") when the limiter was instantiated outside of an event loop. The adapter setter checked for infinity by identity (value is math.inf), so only the exact math.inf singleton was accepted, while every backend setter (using math.isinf()) accepts any positive infinity (#1189; PR by @​greymoth-jp).
• Fixed to_process.run_sync() deadlocking when the worker function writes enough data to sys.stderr to fill the (undrained) pipe buffer. The worker process now redirects sys.stderr to os.devnull as well, matching the documented behavior
• Fixed TLSStream.wrap() matching an internationalized (unicode) host name against the peer certificate using IDNA 2003 (via the standard library) instead of IDNA 2008, which could cause the host name to be matched against the wrong certificate (#1208)
• Fixed anyio.open_process() (and run_process()) ignoring the extra_groups argument, as it mistakenly passed the value of the group argument instead (#1209)
• Fixed CapacityLimiter.acquire_nowait() and CapacityLimiter.acquire_nowait_on_behalf_of() raising trio.WouldBlock instead of anyio.WouldBlock on the trio backend when there are no tokens available (#1218)
• Fixed CapacityLimiter on the asyncio backend over-granting tokens (borrowed_tokens exceeding total_tokens and available_tokens going negative) when a non-blocking acquire was made in the window between a token being released and the notified waiter resuming. The freed token is now reserved for the woken waiter right away, so the non-blocking acquire correctly raises WouldBlock (#1170; PR by @​gaoflow)
• Fixed unnecessary CPU spin when delivering cancellation from CancelScope on asyncio under certain conditions, including improper cancel scope nesting (#1111)

4.14.1

• Fixed teardown of higher-scoped async fixtures failing on asyncio with RuntimeError: Attempted to exit cancel scope in a different task than it was entered in when an async test raise an outcome exception (e.g., pytest.skip(), pytest.xfail(), or pytest.fail()) (#1179; PR by @​EmmanuelNiyonshuti)
• Fixed CapacityLimiter.total_tokens rejecting a value of 0 when the limiter was instantiated outside of an event loop, contradicting the documented behavior of allowing 0 total tokens (#1183; PR by @​nyxst4ck)

4.14.0

• Added support for Python 3.15

• Added an asynchronous implementation of the itertools module (#998; PR by @​11kkw)

• Added the local_port parameter to connect_tcp() to allow binding to a specific local port before connecting (#1067; PR by @​nullwiz)

• Added support for custom capacity limiters in async path and file I/O functions and classes

• Added the create_task() task group method for easier asyncio migration (returns a TaskHandle) (#1098)

• Changed TaskGroup.start_soon() to return a TaskHandle

• Added an option for TaskGroup.start() to return a TaskHandle (which then contains the start value in the start_value property)

• Added the cancel() convenience method to TaskGroup as a shortcut for cancelling the task group's cancel scope

• Improved the error message when a known backend is not installed to suggest the install command (#1115; PR by @​EmmanuelNiyonshuti)

• Improved anyio.Path to preserve subclass types by returning Self in methods that return path objects (#1130; PR by @​EmmanuelNiyonshuti)

• Changed the parameter type annotation in anyio.Path.write_bytes() to accept any ReadableBuffer, thus allowing it to accept bytearray and memoryview to match pathlib.Path.write_bytes() (#1135; PR by @​SAY-5)

• Changed several type annotations to only accept callables returning coroutine-like objects instead of arbitrary awaitables:

  • TaskGroup.start_soon()
  • TaskGroup.start()
  • anyio.from_thread.run()

  This reverts an earlier change from v3.7.0 which was made in error. (#1153)

• Changed anyio.run to support callables returning arbitrary awaitables at runtime on all backends. Previously, this only worked on asyncio (#1171; PR by @​gschaffner)

• Changed several classes (and their subclasses) to have __slots__ (with __weakref__):

  • anyio.CancelScope

... (truncated)

Commits

• c384f99 Bumped up the version
• dbba29d Fixed 100% CPU spin on cancel scope misuse (#1217)
• 6bbc6c3 Fix CapacityLimiter over-granting tokens on asyncio (#1172)
• 6f82b25 Refactored TestTLSStream.test_receive_invalid_max_bytes() to be less flaky
• be24b04 Relaxed timeouts to fix test flakiness
• 8113506 Fix test flakiness caused by slow callback duration logging
• 1e988b6 Fixed CapacityLimiter raising trio.WouldBlock instead of anyio.WouldBlock (#1...
• 44713f3 Pin setup-uv to a commit sha across downstream jobs (#1213)
• f1b7301 Fixed stderr writes in a worker subprocess causing a deadlock (#1207)
• 212be93 Fix flaky test_tcp_listener_same_port using a hardcoded port (#1206)
• Additional commits viewable in compare view

Updates pygments from 2.19.2 to 2.20.0

Release notes

Sourced from pygments's releases.

2.20.0

• New lexers:

  • Rell (#2914)

• Updated lexers:

  • archetype: Fix catastrophic backtracking in GUID and ID patterns (#3064)
  • ASN.1: Recognize minus sign and fix range operator (#3014, #3060)
  • C++: Add C++26 keywords (#2955), add integer literal suffixes (#2966)
  • ComponentPascal: Fix analyse_text (#3028, #3032)
  • Coq renamed to Rocq (#2883, #2908)
  • Cython: Various improvements (#2932, #2933)
  • Debian control: Improve architecture parsing (#3052)
  • Devicetree: Add support for overlay/fragments (#3021), add bytestring support (#3022), fix catastrophic backtracking (#3057)
  • Fennel: Various improvements (#2911)
  • Haskell: Handle escape sequences in character literals (#3069, #1795)
  • Java: Add module keywords (#2955)
  • Lean4: Add operators ]', ]?, ]! (#2946)
  • LESS: Support single-line comments (#3005)
  • LilyPond: Update to 2.25.29 (#2974)
  • LLVM: Support C-style comments (#3023, #2978)
  • Lua(u): Fix catastrophic backtracking (#3047)
  • Macaulay2: Update to 1.25.05 (#2893), 1.25.11 (#2988)
  • Mathematica: Various improvements (#2957)
  • meson: Add additional operators (#2919)
  • MySQL: Update keywords (#2970)
  • org-Mode: Support both schedule and deadline (#2899)
  • PHP: Add __PROPERTY__ magic constant (#2924), add reserved keywords (#3002)
  • PostgreSQL: Add more keywords (#2985)
  • protobuf: Fix namespace tokenization (#2929)
  • Python: Add t-string support (#2973, #3009, #3010)
  • Tablegen: Fix infinite loop (#2972, #2940)
  • Tera Term macro: Add commands introduced in v5.3 through v5.6 (#2951)
  • TOML: Support TOML 1.1.0 (#3026, #3027)
  • Turtle: Allow empty comment lines (#2980)
  • XML: Added .xbrl as file ending (#2890, #2891)

• Drop Python 3.8, and add Python 3.14 as a supported version (#2987, #3012)

• Various improvements to autopygmentize (#2894)

• Update onedark style to support more token types (#2977)

• Update rtt style to support more token types (#2895)

• Cache entry points to improve performance (#2979)

• Fix xterm-256 color table (#3043)

• Fix kwargs dictionary getting mutated on each call (#3044)

Changelog

Sourced from pygments's changelog.

Version 2.20.0

(released March 29th, 2026)

• New lexers:

  • Rell (#2914)

• Updated lexers:

  • archetype: Fix catastrophic backtracking in GUID and ID patterns (#3064)
  • ASN.1: Recognize minus sign and fix range operator (#3014, #3060)
  • C++: Add C++26 keywords (#2955), add integer literal suffixes (#2966)
  • ComponentPascal: Fix analyse_text (#3028, #3032)
  • Coq renamed to Rocq (#2883, #2908)
  • Cython: Various improvements (#2932, #2933)
  • Debian control: Improve architecture parsing (#3052)
  • Devicetree: Add support for overlay/fragments (#3021), add bytestring support (#3022), fix catastrophic backtracking (#3057)
  • Fennel: Various improvements (#2911)
  • Haskell: Handle escape sequences in character literals (#3069, #1795)
  • Java: Add module keywords (#2955)
  • Lean4: Add operators ]', ]?, ]! (#2946)
  • LESS: Support single-line comments (#3005)
  • LilyPond: Update to 2.25.29 (#2974)
  • LLVM: Support C-style comments (#3023, #2978)
  • Lua(u): Fix catastrophic backtracking (#3047)
  • Macaulay2: Update to 1.25.05 (#2893), 1.25.11 (#2988)
  • Mathematica: Various improvements (#2957)
  • meson: Add additional operators (#2919)
  • MySQL: Update keywords (#2970)
  • org-Mode: Support both schedule and deadline (#2899)
  • PHP: Add __PROPERTY__ magic constant (#2924), add reserved keywords (#3002)
  • PostgreSQL: Add more keywords (#2985)
  • protobuf: Fix namespace tokenization (#2929)
  • Python: Add t-string support (#2973, #3009, #3010)
  • Tablegen: Fix infinite loop (#2972, #2940)
  • Tera Term macro: Add commands introduced in v5.3 through v5.6 (#2951)
  • TOML: Support TOML 1.1.0 (#3026, #3027)
  • Turtle: Allow empty comment lines (#2980)
  • XML: Added .xbrl as file ending (#2890, #2891)

• Drop Python 3.8, and add Python 3.14 as a supported version (#2987, #3012)

• Various improvements to autopygmentize (#2894)

• Update onedark style to support more token types (#2977)

• Update rtt style to support more token types (#2895)

• Cache entry points to improve performance (#2979)

• Fix xterm-256 color table (#3043)

• Fix kwargs dictionary getting mutated on each call (#3044)

Commits

• 708197d Fix underline length.
• 1d4538a Prepare 2.20 release.
• 2ceaee4 Update CHANGES.
• e3a3c54 Fix Haskell lexer: handle escape sequences in character literals (#3069)
• d7c3453 Merge pull request #3071 from pygments/harden-html-formatter
• 0f97e7c Harden the HTML formatter against CSS.
• 9f981b2 Update CHANGES.
• 1d88915 Update CHANGES.
• c3d93ad Fix ASN.1 lexer: recognize minus sign and fix range operator (#3060)
• 4f06bcf fix bad behaving backtracking regex in CommonLispLexer
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore <dependency name> major version will close this group update PR and stop Dependabot creating any more for the specific dependency's major version (unless you unignore this specific dependency's major version or upgrade to it yourself)
• @dependabot ignore <dependency name> minor version will close this group update PR and stop Dependabot creating any more for the specific dependency's minor version (unless you unignore this specific dependency's minor version or upgrade to it yourself)
• @dependabot ignore <dependency name> will close this group update PR and stop Dependabot creating any more for the specific dependency (unless you unignore this specific dependency or upgrade to it yourself)
• @dependabot unignore <dependency name> will remove all of the ignore conditions of the specified dependency
• @dependabot unignore <dependency name> <ignore condition> will remove the ignore condition of the specified dependency and ignore conditions
  You can disable automated security fix PRs for this repo from the Security Alerts page.

[timeleft--]

Hermes review findings were written locally at /Users/yianaga/git/MachineWisdomAI/fava-trails/.wise-agents/reviews/hermes/20260919T221028Z-main-findings.md.

PR 120 (head aa5adb5 vs base 10f689f) — findings from the supplied uv.lock-only diff.

1. severity: medium
   file: uv.lock (cuda-bindings optional-dependencies, ~418–459; cufile extra in particular)
   issue: NVIDIA extra markers were widened from linux (and limited win32/Python) gates to platform_machine == 'aarch64' or platform_machine == 'x86_64' with no sys_platform check. nvidia-cufile was previously linux-only and is now selected on any aarch64/x86_64 OS. Same pattern on cublas, cudart, cufft, cupti, curand, cusolver, cusparse, nvjitlink, nvrtc, nvtx. That can try to pull Linux-only NVIDIA wheels on macOS/Windows or install CUDA extras where they were previously excluded.

2. severity: medium
   file: uv.lock (cuda-bindings dependencies ~385–389; nvidia-cublas ~1258–1262; nvidia-cudnn-cu13 ~1297–1301; nvidia-cufft ~1309–1313; nvidia-cusolver ~1339–1346; nvidia-cusparse ~1353–1357)
   issue: python_full_version / sys_platform != 'win32' markers were dropped on cuda-pathfinder and several nvidia-* runtime deps, so those packages are now unconditional. Package versions themselves did not change (cuda-bindings still 13.3.1). This is lockfile graph rewrite, not an upstream version bump, and can change who gets the CUDA stack on Windows and Python 3.12+.

3. severity: low
   file: uv.lock (accelerate ~14–27)
   issue: accelerate jumps 1.13.0 -> 1.15.0 in the lockfile with no matching pyproject.toml or code change in this diff. That is a two-minor Hugging Face runtime bump (training/device placement) with no visible constraint or test update in the supplied evidence.

No other application, test, or config files appear in the supplied exact-head diff. No credentials or tokens in the diff.

[timeleft--]

Captain follow-up on the Hermes findings at exact head aa5adb5: the marker rewrites change lockfile metadata, but a frozen uv tree comparison for aarch64 macOS, x86_64 macOS, x86_64 Windows, and x86_64 Linux shows no platform dependency drift—both base and PR select no NVIDIA/CUDA packages on macOS or Windows and the same CUDA graph on Linux. The accelerate 1.13.0 to 1.15.0 change is one of the three dependency updates explicitly declared by this Dependabot PR, not an unrelated lock artifact. All repository checks pass. Approved with the initial findings superseded by this platform-resolution evidence.