🤖 GitHub AI 热门报告

## 🌟 Summary

**v8.4.135 improves detection reliability by adapting `max_det` to dataset object counts and standardizing dataset fraction handling.**

## 📊 Key Changes

- 🚀 **Smarter `max_det` selection for detection, segmentation, pose, and OBB tasks**
  - Training and validation now inspect the largest number of labeled objects found in a single image.
  - If the default `max_det` is too low, it is automatically increased to match the observed dataset maximum.
  - User-specified `max_det` values are preserved, but a warning is shown when they may limit validation recall.
  - The resolved value is propagated to native end-to-end model heads before validation, improving consistency for NMS-free models.

- ⚠️ **Clearer warnings for object-count mismatches**
  - Users are notified when images contain more objects than `max_det` allows.
  - Warnings explain that a low limit can cap recall and produce misleading validation metrics.
  - Increasing `max_det` may increase validation cost, and cannot exceed the model or export format’s own capacity.

- ���� **Consistent `fraction` boundary behavior**
  - `fraction=1` and `fraction=1.0` now both mean “use the full dataset.”
  - Integers greater than `1` continue to represent an image count.

## [3.6.0](https://github.com/openai/openai-python/compare/v3.5.0...v3.6.0) (2026-08-27)


### Features

* **api:** add compute_units to Responses and Chat Completions usage ([#3749](https://github.com/openai/openai-python/issues/3749)) ([52421d1](https://github.com/openai/openai-python/commit/52421d197e96b4480fcc16e78345578996452c79))


### Bug Fixes

* **auth:** harden X.509 workload identity integration ([#3740](https://github.com/openai/openai-python/issues/3740)) ([fc3ad6c](https://github.com/openai/openai-python/commit/fc3ad6c55a1a250e16707396e86d7373f690a0fc))


### Chores

* **deps-dev:** bump @stdy/cli from 0.22.1 to 0.22.2 ([#3719](https://github.com/openai/openai-python/issues/3719)) ([4f5598c](https://github.com/openai/openai-python/commit/4f5598c8d822fcc281bb8862e424fa6ead310ca8))
* **deps-dev:** bump mypy from 1.17 to 2.3.1 ([#3747](https://github.com/openai/openai-python/issues/3747)) ([0b52c9e](https://github.com/openai/openai-python/commit/0b52c9ecbc2bfe7059a1273674dff7647d414141))
* **deps-dev:** bump pandas-stubs from 2.2.2.240807 to 2.3.3.260113 ([#3659](https://github.com/openai/openai-python/issues/3659)) ([95f0b43](https://github.com/openai/openai-python/commit/95f0b43d31599fe501f045dcc24964d1df5e196f))
* **deps-dev:** bump pyright from 1.1.399 to 1.1.413 ([#3744](https://github.com/openai/openai-python/issues/3744)) ([9917c6e](https://github.com/openai/openai-python/commit/9917c6e28e66e90e1227b3d223c06a8c5441515a))
* **deps-dev:** bump rich from 14.2.0 to 15.0.0 ([#3717](https://github.com/openai/openai-python/issues/3717)) ([7a5484d](https://github.com/openai/openai-python/commit/7a5484d8118e82ec004e964ab2ca484c7dd7c555))

# 🚀 Release `autogpt-platform-beta-v0.7.3`

**Date:** January 2025

---

## 🔥 What's New?

### New Features
- **#14003** - Replace the email system with the Briefing/Alert/Verdict/Ops design (by @Torantulino)
- **#14156** - Add selectable ask_question options and Needs-You prompting rule to Copilot (by @Abhi1992002)
- **#14143** - Show live compaction progress in the Copilot chat (by @0ubbe)
- **#14099** - Add expert team tools and copilot chat polish behind flags (by @Abhi1992002)
- **#14165** - Fire Google Ads conversions across the signup-to-paid journey (by @0ubbe)
- **#14184** - Report a DataFast `paywall_view` goal from onboarding (by @ntindle)
- **#14188** - Recommend Ornith 1.5 for local users (by @ntindle)

### UI/UX Improvements
- **#14164** - Drop the new-tool-ui flag and commit to the tool chain UX (by @0ubbe)
- **#14158** - Chain UI polish — card resize, expert skeletons, minimap and artifact tweaks (by @Abhi1992002)

Changes since sdk==0.4.3

* release(sdk-py): 0.4.4 (#8738)
* Merge commit from fork
* feat: route LangSmith traces from thread streams (#8723)

## What's Changed

* Ollama's app now follows the system appearance again, restoring dark mode support
* Fixed the macOS app to properly hand off to an already-running instance instead of starting a second one
* The Claude Desktop proxy no longer interrupts in-flight requests when the model catalog updates

**Full Changelog**: https://github.com/ollama/ollama/compare/v0.33.1...v0.33.2

## 1.2.0 (2026-08-27)

Full Changelog: [v1.1.0...v1.2.0](https://github.com/anthropics/anthropic-sdk-python/compare/v1.1.0...v1.2.0)

### Features

* **api:** beta files/skills namespaces use GA shapes; drop dated beta header pins ([9df4565](https://github.com/anthropics/anthropic-sdk-python/commit/9df4565fdfe4eec941809a0a3d1615ee11e16b68))


### Bug Fixes

* **aws,bedrock:** sign raw request bytes so binary file uploads work ([#531](https://github.com/anthropics/anthropic-sdk-python/issues/531)) ([f50e910](https://github.com/anthropics/anthropic-sdk-python/commit/f50e9106c002d71966f5f8027758b3d703999936))
* **ci:** resolve assignment aliases in detect-breaking-changes ([f2c4925](https://github.com/anthropics/anthropic-sdk-python/commit/f2c49254941b4d700fe97cbe6bb85205b06a6460))
* **sessions:** make event accumulator forward-compatible with new event types ([#533](https://github.com/anthropics/anthropic-sdk-python/issues/533)) ([cbbaf6e](https://github.com/anthropics/anthropic-sdk-python/commit/cbbaf6e46358d5c844eac01ed3c50a797daf93a7))
* **tools:** let read return a view_range of a file over the size cap ([#538](https://github.com/anthropics/anthropic-sdk-python/issues/538)) ([b68e876](https://github.com/anthropics/anthropic-sdk-python/commit/b68e876345bde1ecc099ef245e87b1319dc8d080))
* **tools:** preserve exact file bytes in the agent toolset and memory tool (no newline translation) ([#540](https://github.com/anthropics/anthropic-sdk-python/issues/540)) ([56921a8](https://github.com/anthropics/anthropic-sdk-python/commit/56921a8c04e0ec71192fcd22dd28db2a5e1306f7))
* **webhooks:** require headers to be passed to `unwrap()` ([0baa902](https://github.com/anthropics/anthropic-sdk-python/commit/0baa90225359f6b4bec50476c19bb52c2ad4250d))


### Documentation

Changes since langchain-core==1.6.0

revert: release(core): 1.6.2 (#39971)
release(core): 1.6.2 (#39967)
fix(core): shore up indexing in genai v1 streaming content (#39964)
fix(core): make `StructuredTool` JSON-serializable (#39631)
chore(deps): bump minor and patch dependencies (#39869)
release(core): 1.6.1 (#39832)
feat(core): propagate gateway information on error path (#39829)

# Release v5.16.1

This is a special release as we include GLM! (and a few small fixes)

# GLM-5.3-Flash

<img width="4239" height="2643" alt="image" src="https://github.com/user-attachments/assets/17bc9c29-758b-44c8-8230-42f945ded209" />

GLM-5.3-Flash, the first **natively multimodal model** in the GLM-5 series. With 320B total parameters and just 18B active parameters, it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks.

GLM-5.3-Flash starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency. For the first time in the GLM series, we introduce a hybrid architecture combining sparse and linear attention, sharply reducing long-context serving costs while preserving precise long-context capabilities. The model also adopts Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency. Together with our latest **30T-token** multimodal pre-training corpus, these changes enable GLM-5.3-Flash to deliver more intelligence with less compute.

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/glm5_next)
* [Glm 5.3 Flash] GLM 5.3 Flash Support (#48342) by @Dovis01 in [#48342](https://github.com/huggingface/transformers/pull/48342)


## Small patch fixes

Mainly BC behavior for TP and pinning a hf kernel for security reasons :hugs: 

MLflow 3.15.2 is a patch release that includes several major features and improvements.

Features:

- [Evaluation] Support immutable evaluation dataset versions (#24845, @danielseong1)
- [Evaluation] Add scorer_ensemble primitive for combining scorer results (#24749, @alkispoly-db)

Bug fixes:

- [Evaluation] Preserve base judge invocation flow in MemAlign aligned judges (#24883, @veronicalyu320)
- [Tracking] Pre-import `databricks.sdk` in Databricks to avoid telemetry deadlock (#24841, @aaronteo-db)
- [Build / Tracking] Align `runs.status` constraint metadata (#24890, @joshuawong-db)

### Added

- 🚦 **Human in the loop tool approval.** Where an administrator has turned it on, you can switch a conversation from letting tools run freely to being asked first, so a model that wants to use a tool stops and waits for you to allow or deny it, one call at a time in a saved conversation, by button or by keyboard shortcut, with your choice remembered for this conversation and for future ones, switching back to running freely releasing anything already waiting, and automations, channel replies, and temporary chats unaffected. [Commit](https://github.com/open-webui/open-webui/commit/7d99b2716a0472b2100b3a71825d8eb3fcbbe877), [Commit](https://github.com/open-webui/open-webui/commit/ec36972c2b5a8d48713f1a240b0ed305e535b4cc), [Commit](https://github.com/open-webui/open-webui/commit/653562d660398c32a9a193450bbbee300d7195c3), [Commit](https://github.com/open-webui/open-webui/commit/fa94a5ab2431edba064150651a14dd4992ef29e8), [Commit](https://github.com/open-webui/open-webui/commit/55c202e841e76cd69679206cd0ecb4a65b039ea4), [Commit](https://github.com/open-webui/open-webui/commit/30f82788bc75e6965da3766e3bf43840dd971eed), [Commit](https://github.com/open-webui/open-webui/commit/bbfdbd59f29e246db1d8b5c2401bc7b11aa32c99), [Commit](https://github.com/open-webui/open-webui/commit/7fc5fa1ff3f6cd6f0efcbf847064ca8d21620ba6), [Commit](https://github.com/open-webui/open-webui/commit/3eb65f47151af4fb4ccfaf33d6e77154546e7259), [Commit](https://github.com/open-webui/open-webui/commit/62fc436999ad32e82d1405ac14d1f03e0f0341ec)
- 🙋‍♂️ **Models that can ask you a question.** A new built-in tool lets a model pause and put up to three multiple-choice questions to you before continuing, with room to type your own answer instead, and the question survives a reload in a saved conversation, so you can come back and answer it later rather than losing the conversation. [Commit](https://github.com/open-webui/open-webui/commit/4465f52a3eb521854cf190f91b0ea7cf3fa21830), [Commit](https://github.com/open-webui/open-webui/commit/133549a87ee371577453d897c2db2d8071223f27), [Commit](https://github.com/open-webui/open-webui/commit/b018feb7419e68314378b3cdd8b7b1b389400a98), [Commit](https://github.com/open-webui/open-webui/commit/083e35144152d6a301bed01aa6d898db3871ddcf), [Commit](https://github.com/open-webui/open-webui/commit/256cce505be8ddd4930e8cb2536faea718d3f789), [Commit](https://github.com/open-webui/open-webui/commit/14e4d72d9a21a10196e8e6efb04180cd9104b8e6), [Commit](https://github.com/open-webui/open-webui/commit/57bd08304e45f707768a898de9f50894929008dd), [Commit](https://github.com/open-webui/open-webui/commit/d9014b3483d4c6e8d99e50395afcdc538bcf05cd)
- 🖇 **Agents can now display terminal files directly.** A model can now show a file it made in a terminal directly in its reply, with a preview and a download button, instead of describing a path that led nowhere when clicked, and a new interface setting chooses whether these open in the reply or in the files pane. [Commit](https://github.com/open-webui/open-webui/commit/78f48a21eef330c0b78c33f2b5fc2169b084997c), [Commit](https://github.com/open-webui/open-webui/commit/e623c02acc70c4ee5d7f2eb2f32d9b7f39287663), [Commit](https://github.com/open-webui/open-webui/commit/f64c0c87e8d1bfdbe060ea5e5a3dee24c0323657), [#27650](https://github.com/open-webui/open-webui/issues/27650)
- 📶 **Streaming rebuilt from the ground up.** A reply now streams as small pieces of new text instead of resending the whole message so far with every update, so the data sent over a reply grows with its length rather than with its length squared, which on a server with many people chatting at once means far less processor time spent encoding, passing, and decoding those updates, far less load and memory on the shared cache that carries them between instances, and far less work in your browser, which no longer takes in the whole reply again and redraws the parts of it that have not changed on every update, cutting the data sent and the server work spent appending to a message by up to 1000x on a very long reply, and a reply still in progress is now kept aside on the server, so reopening the conversation after a refresh picks it up where it is rather than showing a blank message, on deployments backed by Redis. [Commit](https://github.com/open-webui/open-webui/commit/a1579a01ff43cacb357269707d36267ad35e01d6), [Commit](https://github.com/open-webui/open-webui/commit/c755ef60c6bd47ea25306bd898d9a6d1bd8e871d), [Commit](https://github.com/open-webui/open-webui/commit/d02b6a21fc02fb073782e25356968cfee3c45c36), [Commit](https://github.com/open-webui/open-webui/commit/3e186abdd91edee9e97e43c9b345714643a84cf5)
- 🪵 **Much faster throughout.** Hundreds of places across the application no longer assemble detailed log text that is switched off and thrown away unread, so sending messages, uploading and indexing files, running searches, signing in, and loading admin pages all get that time back, with the largest gains on busy servers, in long conversations, and on chats that draw from a large knowledge base. [#27834](https://github.com/open-webui/open-webui/pull/27834), [#27837](https://github.com/open-webui/open-webui/pull/27837)
- 🚀 **Faster model list lookups.** Title generation, tag suggestions, autocomplete, and other background steps of a chat turn now fetch the model list in one go, which keeps other people's responses flowing on busy Redis-backed instances with many models. [#27821](https://github.com/open-webui/open-webui/pull/27821)
- 🛰️ **Cheaper log export.** Deployments that export their logs to a telemetry collector, with "ENABLE_OTEL" and "ENABLE_OTEL_LOGS" both set, now prepare each exported line once instead of twice, which matters more than it used to now that log text is only assembled when something will actually read it. [#27840](https://github.com/open-webui/open-webui/pull/27840)
- 📇 **Faster permission checks on large instances.** Working out which groups you belong to is now a direct lookup rather than a scan of every membership on the server, so chats and the admin user list stay quick as an organization grows. [#27822](https://github.com/open-webui/open-webui/pull/27822)
- ⚙️ **Much faster JSON handling.** Saving and opening chats, reading settings, returning results from built-in tools, streaming replies, signing in and signing up, working out your permissions, and reading stored chunk details during knowledge base searches on Valkey and Oracle vector storage are all handled much faster across the application when the "ENABLE_ORJSON" option is turned on. [Commit](https://github.com/open-webui/open-webui/commit/bb0f898b431d5aa45efa7805956657ed9c3dd78d), [#28396](https://github.com/open-webui/open-webui/pull/28396), [#27841](https://github.com/open-webui/open-webui/pull/27841), [#27807](https://github.com/open-webui/open-webui/pull/27807), [#27805](https://github.com/open-webui/open-webui/pull/27805), [#27813](https://github.com/open-webui/open-webui/pull/27813)
- 📤 **Much faster outbound requests.** Conversations and embedding batches sent to Ollama and Anthropic models are packaged for delivery much faster, which is most noticeable in long chats when the "ENABLE_ORJSON" option is turned on. [#27811](https://github.com/open-webui/open-webui/pull/27811), [#27810](https://github.com/open-webui/open-webui/pull/27810)
- 🐍 **Much faster code interpreter output.** Printed output and generated images from code run in chat appear much faster when the "ENABLE_ORJSON" option is turned on. [#27812](https://github.com/open-webui/open-webui/pull/27812)
- 🪶 **Lighter page loads.** Several small requests the interface makes on every page load, along with a few administrative ones, no longer set up database access they never used, which took several times longer than the rest of the request put together. [#28178](https://github.com/open-webui/open-webui/pull/28178)
- ♻️ **One less read per message.** Sending a message no longer loads the whole conversation from the database twice over, which mattered most in long chats where that record is largest. [#28809](https://github.com/open-webui/open-webui/pull/28809)
- 🏁 **Faster skills on large instances.** Opening the skills list, or sending a message that uses one, no longer checks every skill on the instance one at a time, so both are far quicker where many skills exist and most of them are not yours. [#28798](https://github.com/open-webui/open-webui/pull/28798)
- 🩻 **Faster tools on large instances.** Listing or exporting tools no longer checks every tool on the instance one at a time, so the integrations menu and the tools workspace open faster where many exist. [Commit](https://github.com/open-webui/open-webui/commit/4807866a1cf47340f1b5ea76fded95f8114305f9)
- 🧊 **Faster file access checks.** Checking whether you may reach a file no longer walks every workspace model you can see looking for it, so opening a folder of files, downloading one, or retrieving from one is much quicker on instances with many models. [#28802](https://github.com/open-webui/open-webui/pull/28802)
- 🧱 **Faster folder listings.** Listing your folders now works out your group memberships once for the whole listing rather than again for every item in every folder. [#28810](https://github.com/open-webui/open-webui/pull/28810)

### Features

-   [#13766](https://github.com/gradio-app/gradio/pull/13766) [`5824703`](https://github.com/gradio-app/gradio/commit/5824703d4cc2904adaaca4ea7b536800ccf6a358) - workflows: allow save as copy.  Thanks @hannahblair!
-   [#13773](https://github.com/gradio-app/gradio/pull/13773) [`375335e`](https://github.com/gradio-app/gradio/commit/375335e60e2b4504a57f3eef0eb6bca3c7bb765a) - Keep workflow canvas layout per-viewer, and add undo/redo.  Thanks @abidlabs!
-   [#13770](https://github.com/gradio-app/gradio/pull/13770) [`2cb02c0`](https://github.com/gradio-app/gradio/commit/2cb02c06228a4dfd509a0ac757ed0af466218478) - Upgrade vulnerable frontend dependencies.  Thanks @abidlabs!

### Fixes

-   [#13641](https://github.com/gradio-app/gradio/pull/13641) [`d9acd25`](https://github.com/gradio-app/gradio/commit/d9acd25491a87eda4e98b0f43d4ba262dc6a69e6) - Fix OAuth redirect loops caused by stale sessions.  Thanks @dawoodkhan82!
-   [#13760](https://github.com/gradio-app/gradio/pull/13760) [`7831e62`](https://github.com/gradio-app/gradio/commit/7831e62509a85a3cb38689655fe6ed98946d0d74) - Give the client's internal helper tasks their own thread pool.  Thanks @hysts!
-   [#13782](https://github.com/gradio-app/gradio/pull/13782) [`a9ce60a`](https://github.com/gradio-app/gradio/commit/a9ce60ae03123ed6374b739d862e62dcae2f451e) - Load a saved run's outputs, not just its inputs, when SSR is on.  Thanks @abidlabs!

# Highlights

* **Ray Serve LLM:** In this release we've completed KV cache and token aware request routing, which was previewed in 2.57. Tokenization now happens in-process on the `LLMRouter` ingress replica, the routing decision is made there, tokens are transmitted out-of-band so the engine does not re-tokenize, KV lifecycle events are broadcast to every ingress replica (\#64642, \#64920, \#64949, \#65010, \#65095). KV cache and token aware routing is also aware of CPU KV caches, so offloaded KV cache blocks count toward a replica's cache hit (\#65063).  
* **Ray Core:** We enabled the capability to offload task events from. With `RAY_enable_task_events_to_dashboard_head` on, the task event buffer is replaced by the ray event recorder, events are exported from the aggregator agent to a task events head that keeps an in-memory store, and the state APIs and `ray.timeline` read from it (\#64835, \#65028, \#65123, \#65160, \#65218). Enabling the feature removes task event ingestion and serving from the GCS hot path.   
* **Ray Data:** We’ve added Databricks integrations for writing to DeltaLake, with Catalog support. We’ve also shipped a new shuffle v2 backend, featuring improved performance for joins and aggregations.  
* **Sandboxing:** We've also added experimental Ray Sandbox, which runs task and actor code under gVisor and can run Docker-built images directly (\#64964, \#65570).  
* **TPU Support:** Ray Train adds support for TorchTPU backend (\#64796), and Ray Core adds `SubslicePlacementGroup` for gang scheduling on TPU subslices, single-host TPU support in `SlicePlacementGroup`, and resource accounting for `tpu7x` and multi-core chips (\#64578, \#64079, \#64058). This lets TPU slices and subslices be reserved and trained on without external gang-scheduling glue.

# Ray Data

### 🎉 New Features

* Add `Dataset.with_columns` for multi-column expression projection (\#63858)  
* Add `write_delta` for Delta Lake, with catalog support (\#64923, \#65079)  
* Add Torch inference API (\#65157)  
* Promote hash shuffle v2 to a selectable shuffle strategy, with aggregation support, vectorized aggregation, and block splitting during aggregation (\#64953, \#64652, \#64956, \#65329, \#64897)  
* Add `ignore_missing_paths` and `skip_paths` to `read_parquet` on DatasourceV2 (\#65118)  
* Add `delta_timestamps` (temporal windows) to `read_lerobot` (\#64877)  
* Collect cluster usage metrics by sampling in background threads during execution (\#64686)  
* Tolerate actor deaths during init via `DataContext.max_consecutive_actor_init_deaths` (\#64846)  

## Verify Docker Image Signature

All LiteLLM Docker images are signed with [cosign](https://docs.sigstore.dev/cosign/overview/). Every release is signed with the same key introduced in [commit `0112e53`](https://github.com/BerriAI/litellm/commit/0112e53046018d726492c814b3644b7d376029d0).

**Verify using the pinned commit hash (recommended):**

A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the original signing key:

```bash
cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/0112e53046018d726492c814b3644b7d376029d0/cosign.pub \
  ghcr.io/berriai/litellm:v1.98.0
```

**Verify using the release tag (convenience):**

Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:

```bash
cosign verify \

> [!TIP]
> This release features several new pipelines, including LTX2.5, MiniMax H3, and Wan Animate 2. We're also graduating Modular Diffusers out of the experimental phase and announcing its stable support. Additionally, this release includes minimal support for tensor-parallel. There's a lot more that went down in this release. So, please consult the notes for details. 

## New Pipelines

### MiniMax-H3

[**MiniMax-H3**](https://huggingface.co/docs/diffusers/main/api/pipelines/minimax_h3) generates video and its soundtrack together. A single transformer denoises one packed sequence containing the text conditioning, the conditioning media, and the target video *and* audio latents — there is no separate vocoder and no post-hoc audio pass. Its conditioner is a `Qwen3VLForConditionalGeneration` whose unnormalized 50th-decoder-layer hidden state is read instead of the last one.

MiniMax-H3 is integrated as [Modular Diffusers](https://huggingface.co/docs/diffusers/main/modular_diffusers/overview) blocks only — `MiniMaxH3Blocks` and their `MiniMaxH3ModularPipeline` are the whole integration. The conversion ships both checkpoint partitions in one repository and exposes three workflows (`t2va`, `fl2va`, `ref2va`) that can be pruned at `from_pretrained` time so only that task's components are declared and downloaded.

### MiniMax Music 3

[**MiniMax Music 3**](https://huggingface.co/docs/diffusers/main/api/pipelines/minimax_music3) produces complete songs up to five minutes long from lyrics and a music description, with expressive vocals and long-range structure. It is a hybrid of an autoregressive and a diffusion stage: an 8B Qwen3-based global language model predicts one semantic audio token per frame while a small depth decoder fills in seven residual RVQ codebooks, and their fused hidden states condition a 2.4B flow-matching transformer that produces Flow-VAE latents in overlapping chunks. A DAC-style decoder turns the latents into 44.1 kHz stereo audio.

### Stable Audio 3

[**Stable Audio 3**](https://huggingface.co/docs/diffusers/main/api/pipelines/stable_audio_3) is a text-to-audio model from Stability AI that generates high-quality stereo audio at 44.1 kHz. It uses a rectified-flow DiT conditioned on a frozen T5Gemma text encoder (via cross-attention) and on duration (a float embedded by `StableAudio3DurationEmbedder` and used for adaptive layer norm), and decodes with the SAME (Semantically-Aligned Music Encoder) autoencoder, `AutoencoderSAME`.

Three pipelines ship: `StableAudio3Pipeline`, `StableAudio3AudioToAudioPipeline`, and `StableAudio3InpaintPipeline`.

# Version 3.12.0 (August 20, 2026)

## General

1. [3f8db4c1](https://github.com/google-deepmind/mujoco/commit/3f8db4c1) The MJCF grammar is now defined in a single source of truth schema file, [`src/xml/mjcf.schema`](https://github.com/google-deepmind/mujoco/tree/main/src/xml/mjcf.schema). The parser's grammar table, presence constraints, keyword maps, typed attribute bindings and save policies are generated from it and gated by tests, as are the schema's enum keywords and declared defaults against the C headers and default-constructors.

> [!WARNING]
> **Breaking API changes**
> 2. [6fe04aa8](https://github.com/google-deepmind/mujoco/commit/6fe04aa8) Removed the custom binary texture format (`image/vnd.mujoco.texture`) and the automatic fallback to custom textures when loading files with unrecognized extensions. Textures can now only be loaded from PNG (`image/png`) and KTX (`image/ktx`) files.

## Actuation

3. [279df98c](https://github.com/google-deepmind/mujoco/commit/279df98c) Added the [`pid`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-pid) actuator: a PID controller with real position and velocity setpoint inputs, optional integral action ([`ki`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-pid-ki), integrating the position error with [`imax`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-pid-imax) anti-windup), setpoint rate limiting ([`slewmax`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-pid-slewmax)), and an optional feedforward input. This subsumes the functionality of the `mujoco.pid` plugin with proper activation state: correct under all integrators and visible to keyframes and sensors. With a zero velocity setpoint it is identical to [`position`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-position). The input signature is any subset of `[pos, vel, ff]`, selected by [`input`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-pid-input); absent setpoint inputs are fixed at zero, so the control vector contains no inert entries.
4. [2f1843f4](https://github.com/google-deepmind/mujoco/commit/2f1843f4) The [`dcmotor`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-dcmotor) on-board controller is redesigned: the [`input`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-dcmotor-input) attribute selects any subset of `[pos, vel, ff, voltage]`, where `pos` and `vel` are setpoints for the controller, `ff` is a torque feedforward, and `voltage` is the raw terminal voltage (the default, a plain voltage-commanded motor). Controller gains are in torque space, as for [`pid`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-pid), and the drive voltage compensates back-EMF as in a current-controlled driver: commanded torque is delivered exactly until a limit is reached. The keyword `input="none"` selects the empty signature: the actuator has no control inputs and is purely passive, so friction, cogging and back-EMF braking can be used as passive joint forces.

> [!WARNING]
> **Breaking API changes**
> 5. [2f1843f4](https://github.com/google-deepmind/mujoco/commit/2f1843f4) The mode-flag semantics of [`dcmotor/input`](https://mujoco.readthedocs.io/en/stable/XMLreference.html#actuator-dcmotor-input) ("voltage", "position", "velocity", selecting the interpretation of a single control) are replaced by input signatures, and the controller gains changed from voltage space to torque space. The old velocity mode's integral term (integrated-velocity tracking) is retired without replacement; the integrator always accumulates position error.
>
>    **Migration:** Voltage-commanded motors (the default) are unchanged. Replace `input="position"` with `input="pos"` and `input="velocity"` with `input="vel"`, and multiply the controller gains by $K/R$ (torque per volt). The motor's back-EMF damping, previously felt in addition to the controller's damping, is now compensated: to preserve behavior when the velocity setpoint is zero, add $K^2/R$ to the converted `kd`.

<!-- Release notes generated using configuration in .github/release.yml at 1.62.0 -->

## What's Changed
### Breaking Changes 🛠
* [chore] Remove deprecated st.cache API by @lukasmasuch in https://github.com/streamlit/streamlit/pull/15787
* [chore] Deprecate savefig kwargs on st.pyplot by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16450
* [chore] Remove global-figure support in `st.pyplot` by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16464
### New Features 🎉
* [feature] Provide fully-typed selection state return values by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16275
* [feature] Add client-side validation to st.text_input by @lukasmasuch in https://github.com/streamlit/streamlit/pull/15714
* [feature] Add public streamlit.typing namespace by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16295
* [feature] Add `wrap` parameter to button-like elements to control label wrapping by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16325
* [feature] Expose DataEditorState in streamlit.typing by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16351
* Add support for `50` step font weights in theme configs by @mayagbarnes in https://github.com/streamlit/streamlit/pull/16396
* Expand chart theme config support for light/dark/sidebar by @mayagbarnes in https://github.com/streamlit/streamlit/pull/16357
* [feature] Add email / url / phone / search types to `st.text_input` by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16385
* Remove `BaseWeb` for `DateInput` by @mayagbarnes in https://github.com/streamlit/streamlit/pull/16460
* [feature] Add wrap control to horizontal st.container by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16484
* Remove `BaseWeb` for `DateTimeInput` by @mayagbarnes in https://github.com/streamlit/streamlit/pull/16501
* [feature] Add wrap parameter to st.checkbox and st.toggle by @lukasmasuch in https://github.com/streamlit/streamlit/pull/16470

* New features
  * Added an error check for trying to deserialize JAX exports that are older
    than the backwards compatibility window. Without this check the
    deserialization of expired artifacts may succeed and then result in
    obscure downstream errors.
    Added a configuration flag `--jax_export_deserialize_expired_versions` to
    temporarily bypass the error check.
    See https://docs.jax.dev/en/latest/export/export.html#compatibility-guarantees.
  * Added `jax.numpy.top_k`, which implements `numpy.top_k`, added in
    in NumPy v2.6.0 (#39729).

* Breaking changes
  * The `exec_time_optimization_effort` and `memory_fitting_effort` flags have been
    removed in favor of the `EffortLevel` enum.
  * JAX does not support anymore deserialization of Exported modules from
    before January 15th, 2026 because they are beyond the backwards compatibility
    window. On that date we added support to serialize shardings as NamedSharding,
    and now that is the only sharding serialization that is supported.
  * jnp.take_along_axis now always defaults wrap_negative_indices to True.
    It used to default to False for mode=promise_in_bounds and True otherwise.

Version 3.10.3 2026-08-12

* docs: wrap Chat-80 HOWTO output
* Sandbox Stanford JAR execution to nltk_data directories
* Harden path-traversal / file-I/O sandbox: close write-side symlink TOCTOU + shared-temp squat, lock the cluster with a living audit (CWE-22/59/377)
* Extend algorithmic-complexity DoS hardening: repo-wide sweep + two-string distances (CWE-407/CWE-400)
* Bound unbounded-work DoS in parsers and grammar transforms (CWE-407/674/835)
* fix(security): sandbox MaltParser's Java execution (CVE-2026-12252, CVE-2026-12841)
* fix(security): trust the system temp dir only when it is private (CWE-377/CWE-378)
* fix(security): validate corpus-reader roots against the data sandbox (CWE-73)
* fix(security): validate per-call java() options and replace the -XX:/-D allowlist with a minimal one (CWE-88)
* Additional security hardening (CWE-407, CWE-426, CWE-427, CWE-502, CWE-59, CWE-776, CWE-918)

Thanks to the following contributors to 3.10.3: Mohammad Favas S, leduckhuong, Ziyu Lin, dougtrainer28-cmyk, Chaitanya Kadian, 0xRenSec, Arpit Jain, Jace, nguyencanhthuong, Liling Tan, medimedi, Eric Kafe.



## What's Changed
* Sandbox Stanford JAR execution to nltk_data directories by @ekaf in https://github.com/nltk/nltk/pull/3743
* docs: wrap Chat-80 HOWTO output by @medisean in https://github.com/nltk/nltk/pull/3741

# Change log

## Features :point_up:

* https://github.com/qdrant/qdrant/milestone/50 - TurboQuant 4-bit as a datatype of primary vector storage. Only store 4-bit quantized vectors and spare disk space on original vectors. [[docs](https://qdrant.tech/documentation/manage-data/vectors/#turbo4)]
* https://github.com/qdrant/qdrant/pull/9669, https://github.com/qdrant/qdrant/pull/9684, https://github.com/qdrant/qdrant/pull/9950 - Unify definition of memory usage strategy for collection components. Use `"memory": "cold" / "cached" / "pinned"` to define memory behavior for each individual collection component. Allows for more fine-grained control over memory usage and performance. [[docs](https://qdrant.tech/documentation/ops-configuration/memory-tiers/)]
* https://github.com/qdrant/qdrant/pull/9683 - Allow `"match": {"prefix": "..."}` in `filter` to match keywords by prefix, must be enabled in keyword index. [[docs](https://qdrant.tech/documentation/search/filtering/#prefix-match)]
* https://github.com/qdrant/qdrant/pull/9661 - Per-query IDF corpus for sparse vector search [[docs](https://qdrant.tech/documentation/search/text-search/full-text-search/#per-tenant-idf-statistics)]
* https://github.com/qdrant/qdrant/pull/9899 - Slice filtering condition: sliced scroll / deterministic sampling [[docs](https://qdrant.tech/documentation/search/filtering/#slice)]
* https://github.com/qdrant/qdrant/pull/10035 - Global quota API [[docs](https://qdrant.tech/documentation/ops-configuration/quotas/)]
* https://github.com/qdrant/qdrant/pull/9338 - Add routing token for deterministic read routes [[docs](https://qdrant.tech/documentation/scaling/consistency-guarantees/#read-affinity)]

## Improvements :point_left:

* https://github.com/qdrant/qdrant/pull/9113 - Use batched reads in vector and payload storage
* https://github.com/qdrant/qdrant/pull/9409 - Utilize `io_uring` for payload storage
* https://github.com/qdrant/qdrant/pull/9448 - Lower default update queue length, reduce from 1M to 200
* https://github.com/qdrant/qdrant/pull/9500 - Use Entry API to avoid redundant map double-lookups, improving performance
* https://github.com/qdrant/qdrant/pull/9332 - Enable single file mmap vector storage by default for immutable segments
* https://github.com/qdrant/qdrant/pull/9376 - Add option to explicitly disable BM25 stemmer; deprecate "none" hack

# Keras 3.12.4

Keras 3.12.4 is a security patch release that hardens dataset loading and model file handling against insecure deserialization and decompression-bomb attacks.

## Security Fixes

* **Restrict unpickling when loading IMDB and Reuters datasets** — Replaces `np.load(allow_pickle=True)` with a restricted unpickler that only permits numpy array reconstruction, preventing arbitrary code execution via crafted `.npz` files (CWE-502). ([#23047](https://github.com/keras-team/keras/pull/23047)) by @LinZiyuu
* **Verify all intermediary H5 groups when navigating H5 files** — Manually resolves nested H5 group paths to verify group types at each step, preventing potential path traversal. ([#23168](https://github.com/keras-team/keras/pull/23168)) by @hertschuh
* **Reject decompression-bomb members on the `.keras` asset extraction path** — Adds per-member decompression-ratio checks before extracting `.keras` archives to disk, preventing disk-exhaustion attacks via crafted archives. ([#23101](https://github.com/keras-team/keras/pull/23101)) by @LinZiyuu
* **Restrict unpickling when loading CIFAR datasets** — Replaces bare `cPickle.load` in CIFAR-10/100 batch loading with the numpy-only `RestrictedUnpickler`, blocking arbitrary code execution via pickle gadgets. ([#23252](https://github.com/keras-team/keras/pull/23252)) by @SABITHSAHEB

---

## Contributors

Thank you to all the contributors who made this release possible! 🎉

* @LinZiyuu — Security hardening for IMDB, Reuters, and `.keras` asset extraction ([#23047](https://github.com/keras-team/keras/pull/23047), [#23101](https://github.com/keras-team/keras/pull/23101))
* @hertschuh — H5 group verification ([#23168](https://github.com/keras-team/keras/pull/23168))
* @SABITHSAHEB — CIFAR dataset pickle restriction ([#23252](https://github.com/keras-team/keras/pull/23252))

## v3.0.0

Release date: July 29, 2026

| Milvus Version | Python SDK Version | Node.js SDK Version | Java SDK Version | Go SDK Version |
| -------------- | ------------------ | ------------------- | ---------------- | -------------- |
| 3.0.0          | 3.0.1              | 3.0.3               | 3.0.5            | 3.0.0          |

Milvus 3.0.0 is officially released! Building on the lake-native architecture introduced in [3.0-beta](https://milvus.io/docs/release_notes.md#v30-beta), this release completes what the beta started: External Collection covers more lakehouse workflows; schema supports online add / backfill / drop; the sparse index is rebuilt around SINDI; StructArray and faceted search round out the retrieval engine; FAISS passthrough, and TEXT extend index and modality choices; and Woodpecker runs as a standalone service.

If you are new to the 3.0 line, the Core 3.0 features recall section below summarizes the capabilities introduced in 3.0-beta; the [3.0-beta release notes](https://milvus.io/docs/release_notes.md#v30-beta) have the full write-ups.

### What's new in 3.0.0 (since 3.0-beta)

#### External Collection: more complete lakehouse workflows

3.0-beta introduced External Collection: reference lake files in place, build indexes, and search them without copying data into Milvus. This release extends it toward complete lakehouse retrieval workflows. External fields can now feed function output fields such as BM25 sparse vectors, MinHash signatures, and text embeddings, so text and model-derived retrieval fields are built inside Milvus without copying the source table. Refresh also supports additive schema evolution: when the external table gains new columns, Milvus patches the affected segments instead of rebuilding the collection.

This release also adds a `milvus-table` external format that treats Milvus Snapshot metadata and Storage V3 manifests as an external source, so a collection snapshot can itself be served as an external table — batch and serving systems get a shared, manifest-backed view of the same data.

# Highlights

<img width="1250" height="560" alt="peft-v0 20 0" src="https://github.com/user-attachments/assets/6194fae9-f3c3-48ca-b43b-40c774bfe6eb" />

This release adds no less than nine new PEFT methods and puts a lot of work into the surrounding infrastructure, for example adding a new image generation benchmark for the method comparison suite and greatly improving the documentation structure.

## New Methods

### HiRA

@hqsiswiliam added ["HiRA: Parameter-Efficient Hadamard High-Rank Adaptation for Large Language Models"](https://openreview.net/forum?id=TwJrTz9cRS) to PEFT (#2668). Instead of adding the low-rank product `BA` to the base weight, HiRA multiplies it elementwise (Hadamard product) with the frozen base weight. Because the base weight itself is full rank, the resulting update is no longer constrained to be low rank, while the trainable parameter count stays the same as LoRA's.

### GLoRA

@not-lain contributed GLoRA: ["One-for-All: Generalized LoRA for Parameter-Efficient Fine-Tuning"](https://arxiv.org/abs/2306.07967) in #3098. It is a flexible PEFT method that extends LoRA with configurable weight, activation, and bias adaptation, delivering richer fine-tuning with no extra inference cost. Use it when you need per-layer flexibility or stronger adaptation than vanilla LoRA. Skip it for non-Linear layers (e.g. Conv/Embedding) or when standard LoRA is already sufficient and simplicity matters.

### BEFT

@whubaichuan added ["BEFT: Bias-Efficient Fine-Tuning of Language Models"](https://arxiv.org/abs/2509.15974v2) in #3195. BEFT builds on the observation that fine-tuning bias terms alone can be competitive in low-data regimes, but goes further: rather than training *all* biases, it targets the value projection by default, as the authors found this to be most efficient. This brings the trainable parameter count down to roughly 0.01% of the total parameters.

## Migration

**Settings and credentials now live in a backend-managed store.**

Provider settings and credentials previously persisted in the webview's `localStorage`. In 0.8.4, Jan reads and writes them from a backend-managed store instead (secrets go to the OS keyring). Your existing settings and API keys are copied over automatically on first launch - no manual action is needed, and from this version on all new changes are saved to the backend store.

Your old `localStorage` data is **not deleted** - it's kept as a snapshot so you can downgrade to a pre-0.8.4 build and still have your previous settings. Any changes you make in 0.8.4+ live only in the backend store, so a downgrade sees the older snapshot, not your latest settings.

---

## What's Changed
* chore(flatpak): add 0.8.3 release notes to appdata by @qnixsynapse in https://github.com/janhq/jan/pull/8343
* docs: update changelogs for v0.8.3 and update documentation by @qnixsynapse in https://github.com/janhq/jan/pull/8336
* Merge release/v0.8.3 to main by @qnixsynapse in https://github.com/janhq/jan/pull/8344
* i18n(ja): update Japanese translations (#8264) by @mahirhir in https://github.com/janhq/jan/pull/8348
* i18n(ja): complete common.json translations and fix terminology (#8264) by @mahirhir in https://github.com/janhq/jan/pull/8349
* i18n(ja): complete settings.json translations (#8264) by @mahirhir in https://github.com/janhq/jan/pull/8352
* fix(markdown): force code blocks LTR in RTL language context by @qnixsynapse in https://github.com/janhq/jan/pull/8350
* feat(chat): add toggle for folding interim text into reasoning trace by @qnixsynapse in https://github.com/janhq/jan/pull/8363
* fix(rag): repair embeddings, stop chat-model reloads, and parse more PDFs by @qnixsynapse in https://github.com/janhq/jan/pull/8360

# PyTorch 2.13.0 Release Notes

- [Highlights](#highlights)
- [Backwards Incompatible Changes](#backwards-incompatible-changes)
- [Deprecations](#deprecations)
- [New Features](#new-features)
- [Improvements](#improvements)
- [Bug fixes](#bug-fixes)
- [Performance](#performance)
- [Documentation](#documentation)
- [Developers](#developers)

# Highlights

<table>
  <tr><td><strong>FlexAttention</strong> lands on Apple Silicon (MPS), with up to ~12x speedup over SDPA on sparse patterns, and gains a deterministic backward path on CUDA for reproducible gradient computation.</td></tr>
  <tr><td><strong>CuTeDSL "Native DSL" backend</strong> gives Inductor a second high-performance code path (alongside Triton) for key GPU operations, with faster compilation. [Prototype]</td></tr>
  <tr><td><strong><code>nn.LinearCrossEntropyLoss</code></strong> combines the final prediction and loss computation to cut peak GPU memory by up to 4x for large-vocabulary language model training.</td></tr>
  <tr><td><strong>torchcomms</strong>, a new communications backend for PyTorch Distributed, improves fault tolerance, scalability, and debuggability for large-cluster training.</td></tr>
  <tr><td><strong>FSDP2</strong> now overlaps reduce-scatter and all-gather communications via a dedicated process group (opt-in), increasing distributed training throughput.</td></tr>

### Breaking Changes:
- Relaxed Gymnasium version range (from `"gymnasium>=0.29.1,<1.3.0"` to `"gymnasium>=0.29.1,<2.0"`)
- `pandas` and `matplotlib` are no longer core dependencies; they are now optional and only required for loading results and plotting (moved to `stable-baselines3[extra]`).
- Moved `read_json` and `read_csv` helper functions to test files
- Raised `torch` minimum version from 2.3 to 2.8 to mitigate https://github.com/advisories/GHSA-887c-mr87-cxwp

### Bug Fixes:
- Fixed deprecated error Taxi-v3 from gymnasium v1.3.0 in tests

### [SB3-Contrib]
- Optimized tests (faster to run)
- Fixed dead link for `RecurrentPPO`.

### [RL Zoo]


### [SBX] (SB3 + Jax)

- Added support for `rollout_buffer_class` and `rollout_buffer_kwargs` arguments in `PPO` and `OnPolicyAlgorithmJax` constructors, as in Stable Baselines3. (@Trenza1ore)
- Updated Jax dependency

## FSDP2 Improvements

This release brings a large batch of FSDP2 fixes and quality-of-life improvements: correct dtype handling on load, sharding of embeddings/norms, QLoRA crash prevention, and a more robust auto-wrap policy.

  - Fsdp2 fully_shard embedding and norm by @SunMarc in #4015
  - Fix fsdp2 load full state dict dtype mismatch by @SunMarc in #4021
  - Fix region compilation fsdpv2 by @SunMarc in #4022
  - [FSDP2] Cast model to uniform dtype before fully_shard to fix mixed-dtype AssertionError by @roycho96 in #3985
  - [FSDP2] Auto-exclude non-floating frozen Params4bit from fully_shard to prevent QLoRA crash by @roycho96 in #3987
  - fix(FSDP2): auto-wrap policy ignoring _no_split_modules fallback by @JohnGiorgi in #3999
  - fix: use key-based matching in fsdp2_load_full_state_dict by @roycho96 in #3982
  - fix: add missing model_has_params4bit guard to fsdp2_load_full_state_dict call by @roycho96 in #3981
  - Fix to-fsdp2: drop REMOVED / NOT_YET_IMPLEMENTED FSDP1 keys instead of leaking them by @lollinng in #4065
  - Prevent double-wrapping models in prepare_model() by @joshuaswanson in #3977

## AMD ROCm support

Accelerate now works end-to-end on AMD ROCm devices. Thanks @Abdennacer-Badaoui!

- Make accelerate work end-to-end on AMD ROCm by @Abdennacer-Badaoui in #4025

- Fix `dialog` dependency requirement to be `>= 1.1.0`.

Version: `1.5.9`
Git ref: `refs/tags/1.5.9`
Build Date: `2026-05-05T05:55`
PIP Package: `chroma-1.5.9.tar.gz`
Github Container Registry Image: `:1.5.9`
DockerHub Image: `:1.5.9`

## What's Changed
* [ENH](frontend): block functions on topology dbs by @rescrv in https://github.com/chroma-core/chroma/pull/6836
* [ENH](faults): Add Tilt fault injection CLI by @rescrv in https://github.com/chroma-core/chroma/pull/6881
* [CHORE]  Debug TimeoutError in test_add.py by @rescrv in https://github.com/chroma-core/chroma/pull/6905
* [ENH]: Enable rebuilds for sharded collections by @tanujnay112 in https://github.com/chroma-core/chroma/pull/6916
* [ENH]: Group by support with sharding by @sanketkedia in https://github.com/chroma-core/chroma/pull/6909
* [CHORE]: Denormalize tenant and database into collection_compaction_cursors table by @tanujnay112 in https://github.com/chroma-core/chroma/pull/6940
* [CHORE]  Use normalized record sets for test add by @rescrv in https://github.com/chroma-core/chroma/pull/6935
* [ENH]: Add workflow to build and publish service container images by @jasonvigil in https://github.com/chroma-core/chroma/pull/6944
* [ENH] - Updates language around Chroma Cloud to be more representative. by @tjkrusinskichroma in https://github.com/chroma-core/chroma/pull/6952
* [ENH]: Add change stream to collection compaction cursors by @tanujnay112 in https://github.com/chroma-core/chroma/pull/6955
* [BUG] Switch to storing DOCKERHUB_USERNAME as var by @jasonvigil in https://github.com/chroma-core/chroma/pull/6962
* [CHORE]: Standardize Tilt CI image build on root docker-bake.hcl by @jasonvigil in https://github.com/chroma-core/chroma/pull/6958

# Release 2.21.0

## TensorFlow

### Breaking Changes

* Support for Python 3.9 has been removed starting with TF 2.21.
* The TensorBoard (TB) dependency has been removed starting with TF 2.21.

### Major Features and Improvements

* `tf.lite`
    * Adds int8 and int16x8 support for SQRT operator.
    * Adds int16x8 support for EQUAL and NOT_EQUAL operators.
    * Adds support for int2 type.
    * Adds support for int2/int4 in tfl.cast .
    * Adds support for SRQ int2 in tfl.fully_connected.
    * Adds support for int4 in tfl.slice.
    * Adds support for uint4 type.

## What's Changed
* misc(gha): expose action cache url and runtime as secrets by @mfuntowicz in https://github.com/huggingface/text-generation-inference/pull/2964
* feat: support max_image_fetch_size to limit by @drbh in https://github.com/huggingface/text-generation-inference/pull/3339
* Maintenance mode by @LysandreJik in https://github.com/huggingface/text-generation-inference/pull/3344
* Maintenance mode by @LysandreJik in https://github.com/huggingface/text-generation-inference/pull/3345
* fix(num_devices): fix num_shard/num device auto compute when NVIDIA_VISIBLE_DEVICES == "all" or "void" by @oOraph in https://github.com/huggingface/text-generation-inference/pull/3346


**Full Changelog**: https://github.com/huggingface/text-generation-inference/compare/v3.3.6...v3.3.7

## What's Changed
* Fix docs dotnet core typo by @lach-g in https://github.com/microsoft/autogen/pull/6950
* Fix loading streaming Bedrock response with tool usage with empty argument by @pawel-dabro in https://github.com/microsoft/autogen/pull/6979
* Support linear memory in RedisMemory by @justin-cechmanek in https://github.com/microsoft/autogen/pull/6972
* Fix message ID for correlation between streaming chunks and final mes… by @smalltalkman in https://github.com/microsoft/autogen/pull/6969
* fix: extra args not work to disable thinking by @liuyunrui123 in https://github.com/microsoft/autogen/pull/7006
* Add thinking mode support for anthropic client by @SrikarMannepalli in https://github.com/microsoft/autogen/pull/7002
* Fix spurious </think> tags caused by empty string reasoning_content in streaming by @Copilot in https://github.com/microsoft/autogen/pull/7025
* Fix GraphFlow cycle detection to properly clean up recursion state by @Copilot in https://github.com/microsoft/autogen/pull/7026
* Add comprehensive GitHub Copilot instructions for AutoGen development by @Copilot in https://github.com/microsoft/autogen/pull/7029
* Fix Redis caching always returning False due to unhandled string values by @Copilot in https://github.com/microsoft/autogen/pull/7022
* Fix OllamaChatCompletionClient load_component() error by adding to WELL_KNOWN_PROVIDERS by @Copilot in https://github.com/microsoft/autogen/pull/7030
* Fix finish_reason logic in Azure AI client streaming response by @litterzhang in https://github.com/microsoft/autogen/pull/6963
* Add security warnings and default to DockerCommandLineCodeExecutor by @ekzhu in https://github.com/microsoft/autogen/pull/7035
* Fix: Handle nested objects in array items for JSON schema conversion by @kkutrowski in https://github.com/microsoft/autogen/pull/6993
* Fix not supported field warnings in count_tokens_openai by @seunggil1 in https://github.com/microsoft/autogen/pull/6987
* Fix(mcp): drain pending command futures on McpSessionActor failure by @withsmilo in https://github.com/microsoft/autogen/pull/7045
* Add missing reasoning_effort parameter support for OpenAI GPT-5 models by @Copilot in https://github.com/microsoft/autogen/pull/7054
* Update version to 0.7.5 by @ekzhu in https://github.com/microsoft/autogen/pull/7058

- Added support for all GPT-5 models.
- Added support for Grok-4 via `xai/grok-4` and `openrouter/x-ai/grok-4` model names.
- Added support for `gemini/gemini-2.5-flash-lite-preview-06-17` model, by Tamir Zahavi-Brunner.
- `/clear` now prints “All chat history cleared.” so you know it worked, by Zexin Yuan.
- `/undo` output now shows only the first line of each commit message, making it easier to read.
- Added support for `openrouter/moonshotai/kimi-k2` model, by Jack Harrington.
- Display model announcements with no-arg `/model` command.
- Fixed an issue where new settings for an existing model didn't replace the old ones, by Andrew Grigorev.
- Fixed analytics to support the latest PostHog SDK event-capture API.
- Bumped dependencies to pick up latest litellm==1.75.0.

- Aider wrote 88% of the code in this release.

## 1.10.1

### Bug Fixes:
* fix image upscale on cpu ([#16275](https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/16275))

## What's Changed
* fix: resolve colab unsupported image type issue by @mashb1t in https://github.com/lllyasviel/Fooocus/pull/3506

**Full Changelog**: https://github.com/lllyasviel/Fooocus/compare/v2.5.4...v2.5.5

## Highlights
- Added SGLang worker for vision language models, lower latency and higher throughput https://github.com/lm-sys/FastChat/pull/2928
- Vision langauge WebUI https://github.com/lm-sys/FastChat/pull/2960
- OpenAI-compatible API server now supports image input https://github.com/lm-sys/FastChat/pull/2928
- Added LightLLM worker for higher throughput https://github.com/lm-sys/FastChat/blob/main/docs/lightllm_integration.md
- Added Apple MLX worker https://github.com/lm-sys/FastChat/pull/2940

## What's Changed
* fix specify local path issue use model from www.modelscope.cn by @liuyhwangyh in https://github.com/lm-sys/FastChat/pull/2934
* support openai embedding for topic clustering by @CodingWithTim in https://github.com/lm-sys/FastChat/pull/2729
* Remove duplicate API endpoint by @surak in https://github.com/lm-sys/FastChat/pull/2949
* Update Hermes Mixtral by @teknium1 in https://github.com/lm-sys/FastChat/pull/2938
* Enablement of REST API Usage within Google Colab Free Tier by @ggcr in https://github.com/lm-sys/FastChat/pull/2940
* Create a new worker implementation for Apple MLX by @aliasaria in https://github.com/lm-sys/FastChat/pull/2937
* feat: support Model Yuan2.0, a new generation Fundamental Large Language Model developed by IEIT System by @cauwulixuan in https://github.com/lm-sys/FastChat/pull/2936
* Fix the pooling method of BGE embedding model by @staoxiao in https://github.com/lm-sys/FastChat/pull/2926
* SGLang Worker by @BabyChouSr in https://github.com/lm-sys/FastChat/pull/2928
* Update mlx_worker to be async by @aliasaria in https://github.com/lm-sys/FastChat/pull/2958
* Integrate LightLLM into serve worker by @zeyugao in https://github.com/lm-sys/FastChat/pull/2888
* Copy button by @surak in https://github.com/lm-sys/FastChat/pull/2963