My take on batched Medusa: Simple Framework for Accelerating LLM Generation with Multiple Decoding Heads
Jupyter Notebook · active 2024-07-17 → 2026-04-20 (UTC)
Activity over time
Daily event counts in the loaded window
Line chart, 643 days from 2024-07-17 to 2026-04-20. Pushes: 0 total, peak 0 in a day. Pull requests: 9 total, peak 2 in a day. Issues: 10 total, peak 1 in a day. Comments: 11 total, peak 2 in a day. Stars: 384 total, peak 12 in a day.
- Pushes
- Pull requests
- Issues
- Comments
- Stars
Stars, PRs, issues and forks are under-captured in the later part of this window. GH Archive progressively stopped capturing non-push events during 2026 — −95% or worse by the end of the window. Every series here except Pushes fades for that reason, so a decline above reflects the archive, not this repository. Pushes stay reliable throughout, so read them, and the contributor counts derived from them, as the real signal. Data health has the measurements.
Top contributors
Pushes, PRs, issues, reviews and comments — stars and forks excluded, so this is contribution rather than popularity
| Contributor | Contributions | Pushes | PRs | Comments |
|---|---|---|---|---|
| JackCharlesZhang | 4 | 0 | 4 | 0 |
| ksajan | 3 | 0 | 0 | 2 |
| wittycheng | 3 | 0 | 0 | 2 |
| alex4321 | 2 | 0 | 2 | 0 |
| ByronHsu | 2 | 0 | 0 | 1 |
| simrathanspal | 2 | 0 | 2 | 0 |
| Achazwl | 1 | 0 | 0 | 0 |
| buvnswrn | 1 | 0 | 1 | 0 |
| Mewo518 | 1 | 0 | 0 | 0 |
| xhjcxxl | 1 | 0 | 0 | 1 |
| userljz | 1 | 0 | 0 | 0 |
| shrango | 1 | 0 | 0 | 0 |
| chuangzhidan | 1 | 0 | 0 | 1 |
| PineTreeWss | 1 | 0 | 0 | 0 |
| zhoujian-z | 1 | 0 | 0 | 1 |
| YunSeoHwan | 1 | 0 | 0 | 1 |
| kimjoohyungsd | 1 | 0 | 0 | 0 |
| PKUfreshman | 1 | 0 | 0 | 1 |
| Yi-Yu-Yvonne | 1 | 0 | 0 | 0 |
| huangzl19 | 1 | 0 | 0 | 1 |
Recent activity
Latest issues, pull requests and releases
- Issue#139wittycheng2025-08-22 14:58loss value nan
- Issue comment#95wittycheng2025-08-22 01:57Medusa Training Loss
- Issue comment#134wittycheng2025-08-22 01:52why my train medusa head result is {'medusa0_top1': nan, 'medusa0_loss': nan, 'medusa1_top1': nan, 'medusa1_loss': nan, 'medusa2_top1': nan, 'medusa2_loss': nan, 'epoch': 0}
- Issue comment#55PKUfreshman2025-07-27 07:29TypeError: __init__() got an unexpected keyword argument 'medusa_num_heads'
- Issue#138Yi-Yu-Yvonne2025-07-25 20:30Support for Dynamic Cache?
- Pull request#135JackCharlesZhang2025-04-14 00:37
- Pull request#135JackCharlesZhang2025-04-14 00:37
- Pull request#135JackCharlesZhang2025-04-13 23:37
- Pull request#135JackCharlesZhang2025-04-13 23:36
- Issue#134Mewo5182025-03-22 02:19why my train medusa head result is {'medusa0_top1': nan, 'medusa0_loss': nan, 'medusa1_top1': nan, 'medusa1_loss': nan, 'medusa2_top1': nan, 'medusa2_loss': nan, 'epoch': 0}
- Pull request#131buvnswrn2024-12-28 01:32
- Pull request#130alex43212024-12-15 04:10
- Pull request#130alex43212024-12-15 04:10
- Issue#126kimjoohyungsd2024-11-01 10:34About Code compatability
- Issue comment#55YunSeoHwan2024-10-28 03:46TypeError: __init__() got an unexpected keyword argument 'medusa_num_heads'
- Issue#125Achazwl2024-10-24 05:29Ask for data recipe to reproduce Medusa-2
- Issue#124shrango2024-10-10 21:20[report bug] Encountered when inferencing with Mistral models
- Issue comment#41chuangzhidan2024-09-29 08:28vLLM support
- Issue comment#98huangzl192024-09-21 09:21ImportError: cannot import name 'is_flash_attn_available' from 'transformers.utils'
- Issue comment#115xhjcxxl2024-09-14 03:01jinja2.exceptions.UndefinedError: dict object has no element 0
- Issue#122PineTreeWss2024-09-13 11:47About the Tree Sparsity
- Pull request#120simrathanspal2024-09-06 11:35
- Pull request#120simrathanspal2024-09-06 11:20
- Issue comment#119ByronHsu2024-08-26 05:22[Retraining] Use Liger Kernel to avoid multi-head logits materialization and scale the context length by N times
- Issue#119ByronHsu2024-08-26 05:22[Retraining] Use Liger Kernel to avoid multi-head logits materialization and scale the context length by N times
Totals cover only the window loaded into ClickHouse and count events, not GitHub's lifetime totals — 384 stars here means stars gained during the window, not the repo's star count.