Training a personal draft model for Qwen

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

I currently am using Qwen3.6 35B A3B on vLLM with MTP (num_speculative_tokens=3) for everyday coding and tasks. My workflow right now consists of using up to 2-3 subagents in the Claude Code CLI on my DGX Spark.

The performance is reasonable (~50 tok/sec sustained decode) but I want to see if I can improve the speculative decoding accuracy any further, and have a plan written up to try using DSpark from the vLLM speculators GitHub repo as a base for training my own draft model.

Have any of you tried doing something similar before? Is it worth the effort, or should I wait until the next Qwen release which may or may not support the new speculative methods natively?

Thanks!

2

Just try already existing DFlash drafter: z-lab/Qwen3.6-35B-A3B-DFlash · Hugging Face

DSpark is only marginally better

There’s no free lunch here - training will cost you couple of hundreds with uncertain benefits