Search results

loss mean instead of dividing by loss mask.sum()

Status: Open.

#340 In SafeAILab/EAGLE;

· Manjunath679 opened on Jun 9, 2026

[Discussion] Performance characteristics of EAGLE 3.1 (Post-norm) at shallow speculative depths ($k \le 3$)

Status: Open.

#339 In SafeAILab/EAGLE;

· changerjin opened on Jun 5, 2026

Does shifting by one position when constructing the input break the causal chain? For example: input_ids_target = input_ids[:, 1:]

Status: Open.

#335 In SafeAILab/EAGLE;

· fanbinqi opened on Apr 13, 2026

Garbled text generated when using qwen3 as inference in gen_ea_answer.py

Status: Open.

#334 In SafeAILab/EAGLE;

· Rachelcoll opened on Mar 30, 2026

📋 Documentation Enhancement Suggestion

Status: Open.

#333 In SafeAILab/EAGLE;

· croviatrust opened on Mar 28, 2026

Support for multi-modal MTP？

Status: Open.

#331 In SafeAILab/EAGLE;

· 1benwu1 opened on Feb 28, 2026

Non‑greedy rejection sampling ignores draft (q); not lossless per Speculative Decoding

Status: Open.

#329 In SafeAILab/EAGLE;

· ethantsliu opened on Jan 31, 2026

Why is padding for target and input_ids performed before the for loop in traineagle3/cnets?

Status: Open.

#325 In SafeAILab/EAGLE;

· LuoKan00 opened on Jan 26, 2026

Does EAGLE3 support offloading?

Status: Open.

#322 In SafeAILab/EAGLE;

· zszdsze opened on Jan 12, 2026

loss_mask not adjusted after target shift, causing loss on invalid padding

Status: Open.

#317 In SafeAILab/EAGLE;

· oluxembourg opened on Dec 22, 2025

ImportError: transformers.utils.LossKwargs removed in transformers>=4.54 breaks import in modeling_qwen3_kv.py

Status: Open.

#306 In SafeAILab/EAGLE;

· timxu826 opened on Oct 11, 2025

Some thoughts on tree decoding

Status: Open.

#301 In SafeAILab/EAGLE;

· ggg-s opened on Sep 18, 2025