硬體

AI 晶片、GPU、資料中心、邊緣裝置與硬體安全。

Transformers Proposes Unified Quantization Architecture; MXFP4 Migration Could Change Loading Behavior on Older GPUs

On October 20, Transformers maintainers proposed a shared block quantization architecture covering formats such as FP8 and MXFP4, along with mixed-precision loading. The MXFP4 draft also raises the hardware requirements for quantization kernels, potentially changing execution on A100, T4, and CPU; it has not yet been merged.

2026-10-21

GitHub Repo

Dify Community Reports Accidental Deletion of Attachments Shared Across Knowledge Bases

A Dify community report filed on October 20 says deleting a knowledge base dataset may remove attachments still used by another dataset. A proposed fix adds cross-dataset reference checks, but has not been merged; the impact on released versions and the cloud service remains unclear.

2026-10-21

GitHub Repo

LangChain Community Reports Counting Gap: Tool Limit May End Run Early and Allow Model Calls Past Cap

A LangChain community report on October 19 said that the ordering of two built-in middleware components may leave completed model calls uncounted, disabling cumulative conversation limits. Another contributor said they reproduced the issue and prepared a fix, but no maintainer confirmation or official fix release was available.

2026-09-27

GitHub Repo

Dify Community Reports an Index Deletion Race That May Use Up RAG Retrieval Slots

A Dify community report on October 19 says deleting a document while it is being indexed may leave orphaned entries, causing retrieval to return fewer results than configured. A patch proposal includes tests but has not been merged and will not clean up existing orphaned data.

2026-10-10

GitHub Repo

Transformers 5.20.2 Fixes Lost Input Types in Generation and Includes BF16 Expert Gradient Patch

Hugging Face released Transformers 5.20.2 on October 19, fixing a bug that could discard input container types when moving inputs during generation and affect multimodal models. The release also formally includes a BF16 gradient patch for a specific expert-parallel configuration, providing a deployable version for an issue previously reported by the community.

2026-10-20

GitHub Repo

Dify Community Reports Inconsistent Date Formats, Potentially Causing Entire Knowledge Base Queries to Fail

A Dify community report on October 16 says inconsistent write formats for built-in metadata may cause time filters to fail or source filters to miss documents. A proposed fix has been submitted but not merged, and does not include migration of existing data.

2026-10-18

Kyutai Releases REKI, a 2.4B Robotics Model That Uses Cached Memory to Reduce Repetitive Oscillations

On October 16, Kyutai announced REKI and released its pretrained and fine-tuned weights, along with data processing, training, and deployment code. The model adds short-term memory to its action expert; the team’s tests show fewer repetitive motions, but the weights are limited to non-commercial use and the results await independent replication.

2026-10-18

GitHub Repo

Transformers Community Reports Missing Mask2Former Parameter Registration; Projection Layers May Not Move to GPU with the Model

A Transformers community report on October 17 said some Mask2Former projection layers are stored in a plain Python list and may be omitted from optimizer and device-transfer workflows. A patch has been proposed, but the official fix release and any effect on segmentation accuracy remain unconfirmed.

2026-10-18

GitHub Repo

Ollama 0.40.6 fixes lingering Windows processes by cleaning up inference subprocesses with a Job Object

Ollama released version 0.40.6 on October 16, fixing an issue where inference subprocesses could continue running after its Windows command-line service was forcibly terminated. The update uses operating-system process lifecycle management and covers llama-server and the MLX runner.

2026-09-28

NVIDIA Releases Dory 4B Recurrent Model Weights, Using Adjustable Loops to Increase Inference Compute Depth

On October 16, NVIDIA released Dory 1.0, which runs shared layers repeatedly so the same 4B weights can support different inference compute budgets. The model offers a final-loop KV cache option, but performance results are still based on NVIDIA’s own tests, and integration with major inference frameworks remains pending.

2026-10-18

GitHub Repo

ROCm maintainers confirm PyTorch package mismatch issue mitigated, while stable-channel docs still list alpha versions

On October 16, ROCm TheRock maintainers marked a PyTorch release workflow package mismatch as fixed or mitigated. The patches added version pinning and adjusted version numbers on the release branch, but official installation docs still list some alpha packages, so deployers should verify the actual combinations.

2026-10-18

GitHub Repo

Dify Community Proposes a Knowledge Chunking Fix: Character-Level Splitting May Remove Chinese Text and List Line Breaks

On October 15, the Dify community reported that a fallback to character-by-character splitting in the knowledge base can remove single line breaks, joining rows that were originally separate. On October 16, a contributor reported reproducing the issue and proposed a fix, but the proposal has not been merged, and its impact on retrieval has yet to be measured.

2026-10-16