* sd: sync with master-692-9b0fceb
* sd: sync with master-694-276025e
* sd: sync with master-697-5db680c
* sd: sync to master-700-c2df4e1
* sd: sync with master-704-6e66a1a
* sd: sync with master-707-5a34bc7
* sd: sync with master-690-3a54597
* sd: drop Chroma workaround for Flash Attention
Fixed in master-686-1b702a5 .
* sd: skip setting cpu flags (soon to be removed)
* Experimental Multi-thread MP3
I used Deepseek V4 Flash to come up with a method of making the MP3 conversion with Ace-step multi-threaded.
I built and tested. Resulting MP3 doesnt have any glitches and generates almost instantly (compared to dozens of seconds previously)
* Implement WBRUNA suggestion
Updated as suggested
* further tidyup
Both issues fixed:
fprintf removed from encode_chunk (line 651) — it now just returns "" silently on failure. Error reporting moved to call sites:
Single-threaded (line 689-692): checks mp3_data.empty() and logs once
Multi-threaded (line 722-742): uses std::atomic<int> init_failures{0}, each thread does fetch_add(1) on failure, then after join a single fprintf reports the count. No interleaved output.
No leftover encode call outside threads — all chunks are encoded inside threads (lines 724-732). No results[0] = encode_chunk(...) exists between splitting and joining.
* Overlap by 1 frame
Overlapping by 1 frame from the previous chunk and discarding appears to have resolved the audible boundary glitch.
* Improve multi-threading logic in audio encoding
Refactor thread count calculation and chunk partitioning logic for audio encoding.
---------
Co-authored-by: LostRuins Concedo <39025047+LostRuins@users.noreply.github.com>
Concedo
Wagner Bruna
askmyteapot
Wagner Bruna