All Articles · SURL AI News

Transformers Community Reports Qwen4-Exp Cache Reset Gap That May Change Sparse Attention Results

A Transformers community report on October 19 says leftover position information may alter sparse attention selection when Qwen4-Exp reuses a reset cache. The report includes a reproduction and a local patch, but maintainers have not confirmed it, and evaluation with pretrained models remains incomplete.

2026-10-20

GitHub Repo

Hermes Agent Community Reports a Gap in MCP Token Refresh Timing; Idle Gateways May Need to Log In Again

A Hermes Agent community report on October 19 says idle connections may miss the refresh window when an MCP service’s access and refresh tokens expire at the same time. The reporter has proposed a draft patch, but upstream CI has not validated it and no fixed release has been confirmed.

2026-10-19

GitHub Repo

Ollama 0.40.7 Release Candidate Adds MLX Inference Support for PPLX Decider v1.2

Ollama released version 0.40.7-rc0 on October 18, adding MLX backend support for PPLX Decider v1.2. The update covers non-causal attention, decision readout, and image processing, but local performance and post-quantization probabilities still need validation.

2026-10-19

GitHub Repo

Hermes Agent Community Reports Scheduled Tool Allowlist Mismatch That Could Broaden Tool Access

A community report on October 19 says scheduled tasks using an `mcp-`-prefixed service name may receive tools from every enabled MCP service. The report includes reproduction code, but the full scope and an official fix remain unconfirmed.

2026-10-19

GitHub Repo

SGLang 0.5.22 Adds Deferred KV Cache Allocation Option to Improve Certain Mixed Short- and Long-Prompt Workloads

Released on October 19, SGLang 0.5.22 lets decode nodes in disaggregated deployments wait until prefill is complete before allocating KV space for a request. Official tests on specific workloads showed higher throughput but worse tail inter-token latency; the feature remains optional.

2026-10-19

GitHub Repo

Langflow Community Reports Missing PGVector Connection Parameters in Shareable Playground, Potentially Breaking Agent Retrieval

A Langflow community report on October 18 says the same flow retrieves successfully in the regular Playground but may fail in the Shareable Playground when PostgreSQL connection parameters are missing. The reporter provided a local workaround, but maintainers have not confirmed the issue or released an official fix.

2026-10-19

GitHub Repo

Dify Community Proposes MCP Type Fix for Arrays Converted to Strings in Nested Schemas

A report on October 19 says certain nested MCP tool schemas can cause valid arrays to be converted to strings, sending remote tools the wrong type. A fix proposal and regression tests were submitted the same day, but the change has not been merged and its impact on released versions remains unconfirmed.

2026-10-19

GitHub Repo

Hermes Agent Community Reports Smart Approval Gap: AppleScript May Automatically Close Unsaved Documents

A report on October 19 says smart approval may automatically allow AppleScript commands that close unsaved presentations, risking user data loss. Another contributor reproduced the approval gap. The community proposed a patch branch requiring human confirmation, but no official fix has been released.

2026-09-19

GitHub Repo

vLLM 0.32.0 Adds Structured Decision Endpoint; Check Python and Streaming Compatibility Before Upgrading

vLLM released version 0.32.0 on October 18, adding a structured decision endpoint that returns option probabilities. The release also drops Python 3.10 support and changes streaming event names, so existing deployments should verify compatibility.

2026-10-19

GitHub Repo

PyTorch Community Reports BF16 Gradient Error in Compiled Models; Convolution Followed by log_softmax May Affect Training

A report submitted to the PyTorch community on October 17 says that, under BF16 mixed precision, compiling certain convolution and log_softmax computation graphs with Inductor may produce incorrect gradients. The report includes a reproducer and temporary workarounds, but maintainers have not confirmed the full scope or announced a fix release.

2026-10-18

GitHub Repo

LangChain Community Reports Streaming Parsing Issue That May Treat Truncated Parameters as Valid Tool Calls

A LangChain community report on October 17 says incomplete tool parameters may be completed into valid JSON after a stream ends, then enter the tool execution flow. Another contributor reproduced the issue, but a proposed fix was automatically closed during contribution checks and has not been merged.

2026-10-19

GitHub Repo

Transformers Reports Zero MoE Expert Gradients; BF16 with Certain Distributed Configurations May Halt Updates

A Transformers report on October 17 said expert gradients may all become zero with BF16 and certain expert-parallel and FSDP2 configurations, even as training loss continues to fall. Maintainers proposed a fix and expanded tests on the 18th, but the proposal has not been merged and the release containing a fix remains unconfirmed.

2026-10-09