Will the training code be released later?

[jiahe7ay]

Thank you very much for making the paper public. I believe dFlash is an outstanding and forward-looking piece of work.

I’m writing to ask if there are plans to release the training code in the future.

My primary reasons for asking are as follows:

1. Once the training code is available, the community can adapt dFlash to more models and datasets, which will further strengthen the dFlash ecosystem.

2.Exploring 3-layer Configurations: I am particularly interested in testing the performance of a 3-layer setup. As I mentioned previously, a 5-layer configuration can impact TTFT and throughput under high concurrency. I would like to try training a model with fewer layers myself to optimize for these metrics.

3.Integration with SpecForge: After the official training code is released, we can quickly integrate it into the SpecForge project (part of sgl-project). This would enable dFlash to support training for larger-scale models and make it much easier for users to get started with dFlash training.
Thank you again for your contribution~

[mitalbert]

While I appreciate your effort to release speculative models like Qwen3-Coder-Next and GLM-4.7, the current release cadence from major labs is too fast for a single lab. By the time a speculative model is trained and released for something like GLM-4.7-Flash, the base model is often already outdated.

To solve this, we really need better automation for training and easier integration into inference engines (like the one OP mentioned). This would allow the community to train their own drafts right after a model is released rather than waiting for official releases, which understandably takes much time and resources.