Skip to content

[WIP] Eagle3 Training Implementation #143

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Draft
wants to merge 18 commits into
base: main
Choose a base branch
from
Draft

[WIP] Eagle3 Training Implementation #143

wants to merge 18 commits into from

Conversation

fynnsu
Copy link
Collaborator

@fynnsu fynnsu commented Oct 3, 2025 •
edited
Loading

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

@fynnsu
Add Eagle3DraftModel implementation
0e0d02e 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@github-actions GitHub Actions
Copy link

github-actions bot commented Oct 3, 2025 •
edited
Loading

📦 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/18414158883/artifacts/4239947857.
They will be retained for up to 30 days.
Commit: d9a304c

fynnsu added 3 commits October 3, 2025 16:36
@fynnsu
Add loss masking
20cd35f 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Disable training on verifier_lm_head
08c8848 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Add Distributed Data Parallel (DDP)
76aa8a1 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
fynnsu added 10 commits October 8, 2025 16:55
@fynnsu
Add distributed batch sampler
aaa471a 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Refactor training loop into Trainer class
b157e69 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Rename training_loop.py to trainer.py
2e4cf4d 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Move train script to scripts/ dir
4088ec7 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Add validation dataset
976fd9d 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Small fixes
d30a0c6 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Switch to FSDP2 and add Checkpointer class
595d9a2 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Add general logging implementation with support for multiple metric b…
3476e9f 
…ackends

Supports logging to tensorboard, wandb, trackio, and console, including multiple backends at once



Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Add flex attention mask tests
7ff6186 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Move Checkpointer to own file and restore single-device support
3d12f28 
The Checkpointer class was specific to FSDP DTensor style sharded model checkpointing.
I instead converted it to an abstract-ish base class and created two subclasses for single 
device and multi-device checkpointing respectively.

Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
fynnsu added 2 commits October 9, 2025 16:29
@fynnsu
Correct data shifting
8e22439 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
@fynnsu
Add verifier embedding loading
d9a304c 
Signed-off-by: Fynn Schmitt-Ulms <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@fynnsu