| license: apache-2.0 | |
| tags: | |
| - eagle3 | |
| - speculative-decoding | |
| - pearl | |
| - llama-3.1 | |
| # Pearl EAGLE3.1 R2C best head (2026-06-19) | |
| Selected head: **R2C_t7_g115**. | |
| Selection rule: best overall 8×H100 serving throughput, ranked by aggregate best-K req/s across per-GPU concurrency C=1/4/16 and prompt-length buckets. K4/K5 vs K3 was diagnostic only; the selected checkpoint is the one with the best overall serving result. | |
| Final rank by aggregate best-K req/s: | |
| 1. R2C — 11.0288 | |
| 2. D — 11.0108 | |
| 3. I75 — 11.0029 | |
| 4. R4B — 10.9846 | |
| 5. R3B — 10.9305 | |
| Full run artifacts, training states, logs, benchmark outputs, and selection manifests were archived to: | |
| `s3://pearl-eagle31-artifacts-507286591552-us-east-2/eagle31-final-r2c-20260619T1318Z` | |
| The PR branch was also updated with the multi-replica benchmark used for final evaluation: | |
| `feat/real-prompt-eagle-mining-bench` @ `85861376 feat: add multi-replica EAGLE benchmark` | |
| This HF repo contains the selected head files needed for serving/eval (`model.safetensors`, `config.json`, `config.py`, `val_metrics.json`) plus selection metadata. Optimizer/scheduler training state is in the S3 archive, not this model repo. | |