Luckily, codebases like NanoGPT or the speedrun completely reset people's expectations. But I couldn't find the same for RL; just training Qwen3-4b locally requires hundreds of entreprise-y packages, many of which glitch out when you get slightly off the beaten path…
SOFTWARE
-
Unsloth Monkey-Patching Causes Distributed Training Deadlock Issues
By
–
Eventually tracked down the problem to the Unsloth monkey-patching. Distributed runs even with one process fail, but if you use the regular models from transformers and `pef`, trained with `trl` then it doesn't deadlock… Hmm.
-
Firecrawl: Lightning-Fast Website Scraping in One Line
By
–
Firecrawl scrapes websites with one sentence and it’s shockingly fast
-
Google Veo 3 Vertical Videos Launch, API Prices Cut 50%
By
–
It's been a big week on the Veo front. Here are the highlights: —You can now generate vertical format videos in @FlowbyGoogle and via the Gemini API. —We reduced prices in the Gemini API by ~50%, so now developers can build with Veo 3 at $0.40/second and Veo 3 Fast at
-
Free experiment tracking library supports images videos tables metrics
By
–
Our 𝒻𝓇ℯℯ new experiment tracking library now supports logging images, videos, tables, and of course metrics. https://t.co/gEPAE46BBS
— Hugging Face (@huggingface) 10 septembre 2025Our 𝒻𝓇ℯℯ new experiment tracking library now supports logging images, videos, tables, and of course metrics.
-
Tool Call Outputs as Container Files: Data Processing Possibilities
By
–
Can tool call outputs directly become files in the container? I’d really want that – open up a ton of data processing possibilities.
-
NCCL Timeout Issue with Qwen3 LLM Engine Initialization
By
–
`accelerate test` works with the default config… But the script above hangs for 10 minutes likely due to the NCCL timeout. Somewhere between these two log lines: INFO [core(.)py:74] Initializing a V1 LLM engine (v0.10.1.1) with config: model='unsloth/Qwen3-4B-Base'
INFO -
Training Qwen3 Multi-GPU Setup with Accelerate Library
By
–
Anyone manage to get Qwen3 to train on multiple GPUs using `accelerate`? With a fresh install, I can't get even SFT to work with two 3090s.@UnslothAI The demo notebooks are nice, but a reproducible script known to work would be great, ideally with version dependencies pinned! https://t.co/8FkFpC3pdO
— Alex J. Champandard 🌱 (@alexjc) 10 septembre 2025Anyone manage to get Qwen3 to train on multiple GPUs using `accelerate`? With a fresh install, I can't get even SFT to work with two 3090s. @UnslothAI The demo notebooks are nice, but a reproducible script known to work would be great, ideally with version dependencies pinned!
-
FFmpeg container installation and available binaries documentation
By
–
Also could you install ffmpeg into the container? And provide a documented list somewhere of all of the available binaries – I just had to create one myself https://
claude.ai/share/c9bd85ba
-71d2-4c4b-a49d-07924b657f61
… -
Claude iPhone App: Persistent Code Execution Despite Connection Loss
By
–
Here is a big one: I switched tabs on my iPhone and it looks like the code execution was interrupted – I would very much like it if the code and reasoning kept on running even if my connection drops (I also really want this in the Claude iPhone app)