Repository Issue Activity (beta)

#137 `confidence_eos_eot_inf` does not correctly suppress EOS confidence

#136 [Bug] A single padding token can trigger Math SDPA fallback and a quadratic GPU-memory spike [`flash_attn_varlen_func` not supported]

#135 Verify evals on Papers with Code

#133 Segfault loading LLaDA-8B-Instruct under transformers >= 5 (new core_model_loading path) — works with <5 pin

#132 iLLaDA-8B-Instruct: LLaDA's generate.py with mask_id=5 alone produces degenerate token-repetition output by iLLaDA

#122 How to evaluate the peft model

#121 Unconditional Generation

#131 LLaDA inference logits differ across NVIDIA GPUs (Blackwell vs 4090) with same weights/code

#114 Feature Request: Enable output_attentions=True for LLaDA Models

#130 AttributeError: 'LLaDAModelLM' object has no attribute 'all_tied_weights_keys'

#129 Typo in Algorithm 5: Masking ratio inverted?

#128 📋 Documentation Enhancement Suggestion

#125 Usage of CFG for Conditional Generation Benchmarks

#123 [Docs] Add PRISM test-time scaling recipe / reference for LLaDA-8B-Instruct in README

#119 cpu memory OOM when using lm-eval to evaluate LLADA

#120 RoPE and attention mask mismatch

#103 Problems when using lm_eval and calculating ppl

#36 Performance Optimization: Significantly Speed Up LLaDA Generation in generate.py

#62 Is float64 necessary？

#44 Inference Speed

#115 [Help Needed] mmengine.Config.fromfile() resolves wrong path for local configs

#112 Evaluation is too slow

#106 Unexpectedly high CPU memory usage when running GSM8K with LLaDA

#107 Update the WeChat QR code

#113 Failed to finetune llada moe with math competition dataset