[WIP] Eagle3 Training Implementation

[fynnsu]

WORK IN PROGRESS

Todos/Missing:

• Eagle3Draft Model definition with TTT steps and loss calculations
• Patched Decoder layer definitions
• Simple data loading from sample files
• FlexAttention masking and implementation
• Loss Masking
• Training loop
  • Train data loader
  • loss.backward() + optimizer steps
  • Distributed loss reduction
  • Val data loader
  • Metric collection/reporting
  • Model checkpointing
• Data batching
  • Collate fn
  • Batch sampler (dynamic batch size through sample packing)
  • Distributed (rank) aware sampling
• Distributed support
• Better Verifier lm_head, embed_tokens and config values loading
• Eagle3DraftModel.__init__ signature cleanup/better configuration
• Code relocation / merging with existing definitions (Currently just have everything under speculators/train but this will need to change)
• Verify correctness of key components (attention masking, data token alignment, etc).
• General testing
• All of the stuff we don't know we need yet but will find as we go
• + all the stuff I forgot when writing this list

[github-actions]

📦 Build Artifacts Available
The build artifacts (`.whl` and `.tar.gz`) have been successfully generated and are available for download: https://github.com/vllm-project/speculators/actions/runs/18233863608/artifacts/4179292557.
They will be retained for up to 30 days.
Commit: 76aa8a1