Three releases on October 8 changed separate parts of AI development: Google’s LangExtract exposes failed Gemini batch items, Shepherd Model Gateway (SMG) adds an optional way to split prompt processing from token generation across AMD GPU workers, and Dora fixes input-loss and invalid-data paths in robotics dataflows.
Google’s LangExtract Python library uses a language model to pull user-defined fields from documents, links each extraction to its exact source-text span and creates an interactive HTML view for review. A matching span helps readers inspect the evidence but does not establish that the model interpreted it correctly. Version 1.7.1, released October 8, now raises an error by default for blocked or failed Gemini batch items instead of returning empty output; callers can choose to continue past them. It also passes Gemini 3 and other thinking models’ thinking_config field to the API, letting callers set a thinking level that v1.7.0 ignored. This may change output, latency or cost, or cause an API error if a model rejects the setting. The release says batch-cache keys changed, so prior v1.7.0 cache entries are not reused and requests may run again, adding provider charges. The README also documents local Ollama use; cloud models require provider access. The repository’s Apache 2.0 license does not cover provider services or their terms.
Shepherd Model Gateway (SMG) routes requests between client apps and model engines. Version 1.12.0 adds support for separate AMD GPU workers: one processes incoming prompts (“prefill”), and another generates tokens (“decode”), using vLLM and MoRI-IO. Its optional concurrent WRITE setting starts both requests together, rather than waiting for prompt processing to finish before starting token generation. In a two-node AMD MI355X test with Qwen3-0.6B, each request used about 4,350 input tokens and 64 output tokens. At four concurrent requests, sequential dispatch had a median time to first token of 71.5–72.0 milliseconds and throughput of 9.89–9.98 requests per second; concurrent WRITE reduced the latency to 46.1–46.2 milliseconds, about 36 percent, and raised throughput to 10.62–10.65 requests per second. These ranges are per-round medians from three alternating rounds. The results describe that workload, not a general serving gain. Concurrent WRITE remains opt-in because early cancellation can leave model-cache memory allocated in the tested connector. The repository is Apache 2.0; the AMD hardware and inference software are separate requirements.
Dora lets sensor, model and control programs exchange data through nodes arranged in a directed graph. Its October 8 version 1.1.0 release fixes a case where timer pressure could cause inputs to be lost and adds error handling for invalid values and output IDs that could otherwise abort a process. The project documents nodes written in Rust, Python, C and C++; developers install the command-line tool with Rust tooling, while connecting real equipment depends on its drivers and setup. The repository is Apache 2.0.
Sources
- LangExtract documentation; v1.7.1 release notes; license
- Shepherd Model Gateway documentation; v1.12.0 release notes; benchmark details; license
- Dora documentation; v1.1.0 release notes; license
- Photo: Movicontrol, “Robot industrial Kuka Movicontrol,” CC BY-SA 4.0, via Wikimedia Commons. It shows an industrial inspection cell and is illustrative; it does not depict Dora. No changes were made.
