Huawei Ascend AI Accellerators. DeepSeek V4 model architecture was optimized for Chinese hardware.
943 karma · joined January 11, 2012
https://github.com/python/cpython/blob/d61fcf834d197f0113a6a...
The usual pattern is abs(a - b) <= max(rel_tol * max(abs(a), abs(b)), abs_tol) to avoid both large-value and near-zero pitfalls.
Well, even Microsoft uses React Native for a lot of Windows-only apps.
https://soniox.com/docs/stt/rt/endpoint-detection
Soniox also wins the independent benchmarks done by Daily, the company behind Pipecat.
https://www.daily.co/blog/benchmarking-stt-for-voice-agents/
You can try a demo on the home page:
Disclaimer: I used to work for Soniox
Edit: I commented too soon. I only saw VAD and immediately thought of Soniox which was the first service to implement real time endpoint detection last year.
<h3>${this.getAttribute('title')}</h3>https://www.daily.co/blog/benchmarking-stt-for-voice-agents/
https://groups.google.com/g/golang-announce/search?q=form
5 fixes in 2 years related to HTTP form (url-encoded and multipart).
- Go 1.20.1 / 1.19.6: Multipart form parsing could consume excessive memory and disk (unbounded memory accounting and unlimited temp files)
- Go 1.20.3 / 1.19.8: Multipart form parsing could cause CPU and memory DoS due to undercounted memory usage and excessive allocations
- Go 1.20.3 / 1.19.8: HTTP and MIME header parsing could allocate far more memory than required from small inputs
- Go 1.22.1 / 1.21.8: Request.ParseMultipartForm did not properly limit memory usage when reading very long form lines, enabling memory exhaustion.
- Go 1.25.6 / 1.24.12: Request.ParseForm (URL-encoded forms) could allocate excessive memory when given very large numbers of key-value pairs.
Probably every HTTP server implementation in every language has similar vulnerabilities. And these are logic errors, not even memory safety bugs.
There were no good examples for how to do this but once it was set up it worked extremely well.
It uses tokio for Android/iOS/desktop and even embeds a web server for fake API for end to end testing (even on mobile)
https://github.com/guidance-ai/llguidance
Llguidance implements constrained decoding. It means that for each output token sequence you know which fixed set of tokens are allowed for decoding the next token. You prepare token masks so that in the decoding step you limit which tokens can be sampled.
So if you expect a JSON object the first token can only be whitespace or token '{'. This can be more complex because the tokenizers usually allow byte pair encoding which means they can represent any UTF-8 sequence. So if your current tokens are '{"enabled": ' and your output JSON schema requires 'enabled' field to be a boolean, the allowed tokens mask can only contain whitespace tokens, tokens 'true', 'false', 't' UTF-8 BPE token or 'f' UTF-8 BPE token ('true' and 'false' are usually a single token because they are so common)
JSON schema must first be converted into a grammar then into token masks. This takes some time to be computed and takes quite a lot of space (you need to precompute token masks) so this is usually cached for performance.