Rust async vtable bug and local LLM variance #117
Today's Letter
Rust 1.98, callable async service vtable bug
- Rust 1.98.0 emits a zero method entry in a compiler-generated vtable for a callable boxed async service
- Safe Rust dispatch through the vacant entry can trigger SIGSEGV at address zero
- The issue was observed on aarch64-apple-darwin and reported in CI for the Rama project
- The same source works on Rust 1.97.1 and nightly-2026-07-16
- The failure aborts an HTTP CONNECT request instead of returning the expected 502 Bad Gateway
- Issue #161441 is marked P-critical and includes a reproducible Rama-based test case
Source: github.com
Local LLM quality depends on the inference stack
- According to the report, identical model weights can produce different outputs across GPUs, inference engines, attention backends, quantization methods, and KV-cache precision
- A comparison using Qwen3.6-27B and roughly 100,000 tokens found attention backend changes affecting next-token selection later in the context
- KV-cache quantization alone introduced differences in tool-calling behavior, with INT8 recovering from an error while INT4 failed to recover
- INT8 W8A16 reportedly stayed closer to the BF16 reference than FP8 and NVIDIA NVFP4 in the tested setup
- NVFP4 changed the top token at about half of measured positions in an 88,000-token context
- Short prompts, temperature-zero checks, and standalone KLD values may not represent long-context agent workloads
- The report recommends evaluating the complete runtime with real context lengths, tool calls, and domain-specific tasks
Source: news.hada.io
Jocoletter curates AI, software, and product trends for developers and builders.
#GitHub #Qwen