GitHub / SafeAILab/EAGLE issues and pull requests
#283 - Hardcoded Draft Tokens Probabilities
Issue -
State: closed - Opened by maryamhgf 19 days ago
- 1 comment
#282 - Support on Qwen3
Issue -
State: closed - Opened by aladinggit 19 days ago
- 3 comments
#281 - Rollback main to pre-PR #275 state (commit 2282dc4) to resolve Issue #278 (LossKwargs import error)
Pull Request -
State: closed - Opened by y344shi 21 days ago
- 2 comments
#276 - Eagle3 训练Acc length为1.9
Issue -
State: closed - Opened by ShallowMDream 23 days ago
- 1 comment
#275 - 增加了MiniCPM4的eagle3权重,并且将MiniCPM4的在sglang中的适配同步提交了
Pull Request -
State: open - Opened by LDLINGLINGLING 26 days ago
#274 - fix: torch.checkpoint() incorrectly wraps single forward step in original codebase.
Pull Request -
State: open - Opened by COLAZERO2 28 days ago
#273 - How to training 32K ctx on llama3.1-8b model?
Issue -
State: closed - Opened by fan-niu 29 days ago
- 1 comment
#271 - add support for qwen3 with eagle3
Pull Request -
State: closed - Opened by quanfeifan about 1 month ago
#269 - The results are less optimal than expected
Issue -
State: closed - Opened by ggg-s about 1 month ago
- 3 comments
#267 - Lora Support
Issue -
State: closed - Opened by ggg-s about 1 month ago
- 1 comment
#266 - Performance Regression in Eagle3: vLLM 0.9.0 vs. vLLM 0.9.
Issue -
State: closed - Opened by ggg-s about 1 month ago
- 2 comments
#265 - About LM Head in EAGLE Model Parameters
Issue -
State: open - Opened by seohyunwoo-0407 about 1 month ago
#264 - Can EAGLE3 overfit one sample to see performance?
Issue -
State: open - Opened by ChiikawaSama about 1 month ago
#263 - Support target model init from AutoModelForCausalLM
Pull Request -
State: open - Opened by KerwinKai about 1 month ago
#262 - Clarification on How KVCache Is Utilized in EAGLE3 Training Flow
Issue -
State: closed - Opened by chriszhang1 about 1 month ago
- 1 comment
#261 - Question about EAGLE-3 Draft Model Input: All Hidden States Concatenation vs. Low/Mid/High Layer Features
Issue -
State: closed - Opened by seohyunwoo-0407 about 2 months ago
- 1 comment
#259 - why use fp16 for eagle-llama3.1-8b
Issue -
State: closed - Opened by SUDA-HLT-ywfang about 2 months ago
- 3 comments
#258 - Vocab size Issue between target model and draft model (Maybe Tokenizer Issue?)
Issue -
State: closed - Opened by seohyunwoo-0407 about 2 months ago
- 2 comments
#257 - Can eagle3 support qwen-vl?
Issue -
State: closed - Opened by ChiikawaSama about 2 months ago
- 3 comments
#256 - Discrepancy in the computation of Eagle 3 bw paper and code
Issue -
State: closed - Opened by ekagra-ranjan about 2 months ago
- 1 comment
#255 - chore: Clean eagenerate and naivegenerate
Pull Request -
State: open - Opened by tonylt about 2 months ago
#255 - chore: Clean eagenerate and naivegenerate
Pull Request -
State: open - Opened by tonylt about 2 months ago
#254 - Question about EAGLE-3 paper
Issue -
State: closed - Opened by ebubekir-pulat about 2 months ago
- 1 comment
#253 - utils_alpha.py
Issue -
State: open - Opened by Duze204 about 2 months ago
#252 - Add Eagle2-Qwen2.5-14B-Instruct weight
Pull Request -
State: open - Opened by ShiyiZheng123 2 months ago
#252 - Add Eagle2-Qwen2.5-14B-Instruct weight
Pull Request -
State: closed - Opened by ShiyiZheng123 2 months ago
#251 - EAGLE-3 Training and Test Data
Issue -
State: open - Opened by ebubekir-pulat 2 months ago
- 14 comments
#250 - Batch size >1 execution
Issue -
State: open - Opened by Lena-Jurkschat 2 months ago
#249 - Loading EAGLE-3 Model state dictionary fails
Issue -
State: open - Opened by Lena-Jurkschat 2 months ago
#248 - train eagle3: target_p'shape and out_logp'shape not equal cause error
Issue -
State: open - Opened by dongyibo 2 months ago
- 2 comments
#247 - Inquiry about attention mask used for EAGLE-3 Training
Issue -
State: open - Opened by YanzuoLu 2 months ago
#246 - d2t and t2d cache seem not work during the training phase?
Issue -
State: closed - Opened by dongyibo 2 months ago
#245 - support: Run EAGLE on AMD ROCm
Pull Request -
State: closed - Opened by ChangLiu0709 3 months ago
#245 - support: Run EAGLE on AMD ROCm
Pull Request -
State: closed - Opened by ChangLiu0709 3 months ago
#244 - When is the Qwen3 model expected to support Eagle3?
Issue -
State: open - Opened by fanqingyu0604 3 months ago
- 12 comments
#243 - bugfix: bugfix of ui arg no_eagle3.
Pull Request -
State: open - Opened by wangzhaode 3 months ago
#243 - bugfix: bugfix of ui arg no_eagle3.
Pull Request -
State: open - Opened by wangzhaode 3 months ago
#242 - where is the training data and test data?
Issue -
State: closed - Opened by youngze0016 3 months ago
- 7 comments
#241 - 放出来的代码都不管能不能跑通的吗?
Issue -
State: open - Opened by Arcmoon-Hu 3 months ago
#240 - What is the data format of the training data?
Issue -
State: open - Opened by BucherLi 3 months ago
#239 - question about layer for eagle3 train code
Issue -
State: closed - Opened by dongyibo 3 months ago
- 2 comments
#238 - inference time increase after use eagle model
Issue -
State: closed - Opened by thuBingo 3 months ago
- 1 comment
#237 - eagel支持VLM(vision language models)模型吗?
Issue -
State: closed - Opened by sea9856 3 months ago
- 2 comments
#236 - size mismatch when load yuhuili/EAGLE3-DeepSeek-R1-Distill-LLaMA-8B model
Issue -
State: closed - Opened by thuBingo 3 months ago
- 1 comment
#235 - Calculating Average Acceptance Length on EAGLE-3
Issue -
State: open - Opened by ebubekir-pulat 3 months ago
#232 - ZeroDivisionError: division by zero
Issue -
State: open - Opened by xiaomofang 3 months ago
#231 - In getkacc(), should 'total[kk]' be updated when pre_len+kk>=seq_len?
Issue -
State: open - Opened by junghye01 3 months ago
#230 - Can EAGLE-3 be used across different machines?
Issue -
State: open - Opened by ebubekir-pulat 4 months ago
#229 - Why Does Inference Throughput Decrease When Using Eagle-V1 Draft Model with Qwen3?
Issue -
State: open - Opened by xiaomofang 4 months ago
#228 - Eagle Training Parameters and Hardware Requirements
Issue -
State: open - Opened by idankinderman 4 months ago
#227 - Acceptance rate is extremely low when runing the speculative decoding process.
Issue -
State: open - Opened by lxnlxnlxnlxnlxn 4 months ago
#226 - fix: the problem of loading local model path and downloading from hug…
Pull Request -
State: closed - Opened by Lihui-Gu 4 months ago
#226 - fix: the problem of loading local model path and downloading from hug…
Pull Request -
State: closed - Opened by Lihui-Gu 4 months ago
#225 - How to train eagle3 and support qwen2
Issue -
State: closed - Opened by skylee-01 4 months ago
- 11 comments
#224 - GPU specification used for EAGLE-3 training and inference?
Issue -
State: open - Opened by junghye01 4 months ago
#223 - when head_dim != self.hidden_size // self.num_heads
Issue -
State: open - Opened by garycaokai 4 months ago
#222 - Clarification on d2t and t2d Mapping Logic in EAGLE-3 Draft Model
Issue -
State: open - Opened by junghye01 4 months ago
#221 - Question on EAGLE tree-drafting when Temp > 0
Issue -
State: closed - Opened by luyuzhe111 4 months ago
- 10 comments
#220 - Potential Ablation Studies on Scaling Law in EAGLE3
Issue -
State: open - Opened by luyuzhe111 4 months ago
#219 - Bug in EAGLE Architecture: missing pre-attention norm.
Issue -
State: open - Opened by luyuzhe111 4 months ago
#218 - Confusion about cnets.py vs. cnets1.py when training Eagle-2 draft model
Issue -
State: closed - Opened by Yonghao-Tan 4 months ago
- 1 comment
#217 - The issue of misaligned outputs in eager mode.
Issue -
State: closed - Opened by seamoonlight-YBY 4 months ago
- 1 comment
#216 - Questions about training-time test technique & EAGLE-3 train code update dates
Issue -
State: open - Opened by junghye01 4 months ago
#215 - fix bug in top_p sample
Pull Request -
State: closed - Opened by pockers21 4 months ago
#215 - fix bug in top_p sample
Pull Request -
State: closed - Opened by pockers21 4 months ago
#214 - 'EConfig' object has no attribute 'draft_vocab_size'
Issue -
State: closed - Opened by junghye01 4 months ago
#213 - Which layers are used for low/mid/high feature fusion in EAGLE3?
Issue -
State: open - Opened by junghye01 4 months ago
#212 - EAGLE3-LLaMA3.1-Instruct-8B outputs are inconsistent with autoregressive (naive) decoding with temperature=0
Issue -
State: closed - Opened by taras-sereda 4 months ago
- 2 comments
#211 - Eagle-3 for LLAMA4
Issue -
State: open - Opened by tchaton 4 months ago
#210 - GPU Specs for EAGLE-2 inference on LLAMA2-13B
Issue -
State: open - Opened by junghye01 4 months ago
#209 - Unable to reproduce speed up results from the paper
Issue -
State: closed - Opened by taras-sereda 4 months ago
- 2 comments
#208 - Taras/sync fork
Pull Request -
State: closed - Opened by taras-sereda 4 months ago
#208 - Taras/sync fork
Pull Request -
State: closed - Opened by taras-sereda 4 months ago
#207 - Inconsistent Outputs of Qwen model standalone vs. EAGLE using temperature=0
Issue -
State: open - Opened by Lena-Jurkschat 4 months ago
#206 - eagl3/eagle switch, updates requirements
Pull Request -
State: closed - Opened by taras-sereda 5 months ago
#206 - eagl3/eagle switch, updates requirements
Pull Request -
State: closed - Opened by taras-sereda 5 months ago
#205 - No module named 'torch.sparse._triton_ops'
Issue -
State: open - Opened by seven1122 5 months ago
#204 - How do I train the EAGLE3 model myself?
Issue -
State: open - Opened by LiuzRush 5 months ago
#203 - When will EAGLE3 support Qwen- QWQ model?
Issue -
State: open - Opened by Ximingwang-09 5 months ago
- 2 comments
#202 - feat: added data-gen for LLama3.2-instruct-3B
Pull Request -
State: closed - Opened by zfyre 5 months ago
#202 - feat: added data-gen for LLama3.2-instruct-3B
Pull Request -
State: closed - Opened by zfyre 5 months ago
#201 - error
Issue -
State: closed - Opened by jiahe7ay 5 months ago
- 3 comments
#200 - chore: update `setup.py` with the latest changes
Pull Request -
State: closed - Opened by b8zhong 5 months ago
- 1 comment
#200 - chore: update `setup.py` with the latest changes
Pull Request -
State: closed - Opened by b8zhong 5 months ago
- 1 comment
#199 - fix qwen2 decode bug
Pull Request -
State: closed - Opened by pockers21 5 months ago
#199 - fix qwen2 decode bug
Pull Request -
State: closed - Opened by pockers21 5 months ago
#198 - When will the EAGLE3 support for Qwen2
Issue -
State: open - Opened by garycaokai 5 months ago
#197 - No module named 'eagle.ge_data'
Issue -
State: open - Opened by Lzhang-hub 5 months ago
- 3 comments
#196 - Can not reproduce the results of EAGLE3-DeepSeek-R1-Distill-LLaMA-8B
Issue -
State: open - Opened by jsttlgdkycy 5 months ago
#195 - How do you measure the acceptance rate?
Issue -
State: closed - Opened by scj0709 5 months ago
- 11 comments
#194 - How to train eagle3 with the new loss?
Issue -
State: open - Opened by carlbunny 5 months ago
#193 - Is the data generation and training pipeline different for EAGLE-3 compared with EAGLE-2?
Issue -
State: open - Opened by DrXuQian 6 months ago
#192 - do you have plan to support deepseek r1 with eagle3?
Issue -
State: closed - Opened by kingkingleeljj 6 months ago
- 1 comment
#191 - Hello!! Acceptance rage question!
Issue -
State: closed - Opened by scj0709 6 months ago
- 1 comment
#190 - Release EAGLE-3 on Hugging Face
Issue -
State: open - Opened by NielsRogge 6 months ago
- 1 comment
#189 - Could you please support QwQ-32B? I think CoT requires to speed up. Thank a lot if you can support that.
Issue -
State: closed - Opened by YF-T 6 months ago
- 1 comment
#188 - Using standard LLMs as draft models
Issue -
State: closed - Opened by sunnyc98 6 months ago
#187 - About Evaluation Details on DeepSeek-R1-LLaMA-8B
Issue -
State: closed - Opened by bingps 6 months ago
- 1 comment
#186 - Why does training eagle with my own data perform worse than medusa
Issue -
State: open - Opened by skylee-01 6 months ago
#185 - Can't wait to try out EAGLE-3
Issue -
State: closed - Opened by Shuai-Xie 6 months ago
- 7 comments