Segment 23

In this sub-session, the assistant completed the entire EAGLE-3 training pipeline: fixing the synthetic data generation script to properly wrap reasoning with think tokens, running a successful 10K inference run in ~5.3 hours, extracting 828 GB of hidden states at 3,165 tok/s, and finetuning from the AQ-MedAI checkpoint in 2.6 hours. However, testing with vLLM EAGLE-3 revealed a fundamental integration issue with MLA attention—both the trained drafter and the pre-trained baseline achieved only ~15% acceptance rate, resulting in 0.66x throughput (worse than no speculation). The user directed pivoting to SGLang, which has first-class EAGLE-3 support. The assistant built sgl-kernel for SM120 (48 minutes) and verified it works; SGLang loads the 547GB model in just 22 seconds (vs 25 minutes in vLLM), but the server processes deadlock after weight loading with zero CPU/GPU utilization. Debugging is ongoing with verbose logging and CUDA graph disabling to resolve this SM120 compatibility issue.

Fix synthetic data generation scriptRun 10K inference for synthetic dataExtract hidden statesFinetune EAGLE-3 drafterTest vLLM EAGLE-3 integrationDiagnose low acceptance ratePivot to SGLangBuild sgl-kernel for SM120Debug SGLang server deadlock

The EAGLE-3 Integration Odyssey: From Training Triumph to Framework Pivot on Blackwell GPUs 3954 words

Chunks