硬體

AI 晶片、GPU、資料中心、邊緣裝置與硬體安全。

NVIDIA Releases Nemotron 3.5 Super VL, Supporting Multimodal Agents with a Hybrid Architecture and Speculative Decoding

NVIDIA announced Nemotron 3.5 Super VL on October 15, offering BF16 and NVFP4 weights and combining text, image, and video understanding with tool calling. The company published throughput benchmarks for long-running agent tasks, though practical results still depend on hardware, cache reuse, and concurrency settings.

2026-10-16

GitHub Repo

Dify Community Proposes Node.js SDK Streaming Fix; Replacement Events May Preserve Original Text

A report on October 15 found that Node.js SDK’s `toText()` may miss workflow text and append content moderation replacement messages to the original answer. The fix includes regression tests but has not been merged; an official release version is still unconfirmed.

2026-10-16

GitHub Repo

Ollama Community Reports ROCm Graph Initialization Running Out of Memory; Partial CPU Offloading May Still Crash

An October 15 report in the Ollama community says an AMD GPU may run out of memory during HIP graph initialization when running a partially CPU-offloaded Gemma 4 model. The reporter temporarily avoided crashes by disabling graph execution, but maintainers have not confirmed the root cause or issued an official fix.

2026-10-16

GitHub Repo

Transformers Proposes Native Cellpose Integration with SAM and DINOv3 Cell Segmentation Models

An integration proposal submitted on October 15 adds Cellpose-SAM and Cellpose-DINO, with PyTorch post-processing that includes flow error checks. The proposal is still under review, and the authors’ claim of mask consistency cannot yet be generalized to all datasets and hardware.

2026-10-15

GitHub Repo

Hermes Agent v0.22.2 Fixes Incorrect Build, Improves Local Inference and Gateway Restart

NousResearch released Hermes Agent v0.22.2 on October 15, correcting a build that mistakenly included unreleased gateway code. The update also addresses false timeouts on long prompts with local models and gateway startup failures after reboot, though some sessions created by the previous release still cannot be deleted.

2026-09-27

GitHub Repo

Ollama Pre-release Ports MLX Custom Kernels to CUDA; Limited Tests Show Nemotron Prompt Throughput More Than Doubling

Ollama released v0.40.5-rc0 on October 14, adding CUDA ports of three groups of MLX custom kernels and reducing fallback to computation graph operations. The author measured more than a twofold increase in prompt throughput for two Nemotron models on an RTX 6000 Ada, but results have yet to be validated across hardware and full workloads.

2026-10-11

GitHub Repo

PyTorch Community Proposes NCCL2 Memory Pool Fix After Later Allocations May Miss Symmetric Communication Kernels

An October 13 PyTorch report says the NCCL2 backend may fail to register new segments added to a symmetric memory pool, causing collective communication to fall back to ring kernels. The proposed fix includes tests on two GB200 GPUs and kernel-path comparisons, but has not been merged and includes no end-to-end performance data.

2026-10-14

GitHub Repo

Hermes Agent v0.22.0 Integrates Local Model Management and Defaults to Stable Updates

Hermes Agent v0.22.0 adds llama.cpp downloads, hardware adaptation, and model memory management. Source installs now track stable releases by default, and a package management process that verifies a new environment before switching aims to reduce update interruptions.

2026-10-12

GitHub Repo

Langflow 1.12.6 fixes tool discovery path so listing MCP tools no longer executes stored component code

Langflow released 1.12.6 on October 13, fixing an issue where generating input schemas for MCP tool lists and A2A agent cards could execute component code. The fix includes a live reproduction and regression tests; operators of publicly deployed flows should still review custom components and anonymous access settings.

2026-10-14

GitHub Repo

Dify Community Proposes Structured Output Fix After Reasoning-Block JSON Is Mistaken for the Final Answer

A Dify community report on October 13 found that an LLM node may treat JSON in a reasoning block as structured output, even when the final answer matches the configured format. The proposed fix includes streaming and non-streaming regression tests, but has not been merged and does not add full JSON Schema validation.

2026-10-14

GitHub Repo

Transformers Community Reports Gemma Tensor Parallelism Bug; Mixing Tensor and DTensor May Interrupt Inference

Reports from the October 12–13 Transformers development builds indicate that Gemma 3 and 4 multimodal classes may fail to complete a forward pass under tensor parallelism because of inconsistent tensor types. The report includes a minimal reproduction, and its author points to a proposed fix, but the impact on stable releases and the completeness of the fix remain unconfirmed.

2026-10-09

GitHub Repo

vLLM Community Proposes MoE LoRA Backend Fix; Explicit Triton Selection May Still Fail to Start on Blackwell

Reports from the vLLM community on October 12–13 say MoE LoRA backend selection may ignore an explicit setting and trigger a startup error in fully sharded mode. The proposed fix adds configuration checks and unit tests, but has not been merged and lacks full GPU validation.

2026-10-13