Issues

• Llama2FunctionCalling prompt contains doubled curly braces

  Lightning-AI/litgpt#2418 · Filter by author Yupsecous opened on Nov 14, 2026

• SFT collator retains full untruncated batch storage despite max_seq_length

  Lightning-AI/litgpt#2414 · Filter by author Afloat16 opened on Nov 13, 2026

• Half-precision chunked cross entropy overflows while averaging finite token losses

  Lightning-AI/litgpt#2411 · Filter by author Afloat16 opened on Nov 12, 2026

• Optimizer runtime overrides mutate reusable init_args configurations

  Lightning-AI/litgpt#2409 · Filter by author Afloat16 opened on Nov 12, 2026

• Avoid retaining a duplicate input activation in LLaMAMoE

  Lightning-AI/litgpt#2407 · Filter by author Afloat16 opened on Nov 12, 2026

• LongForm and FLAN consume raw samples during repeated reads

  Lightning-AI/litgpt#2405 · Filter by author Afloat16 opened on Nov 12, 2026

• bias='lora_only' leaves LoRA layer biases frozen

  Lightning-AI/litgpt#2403 · Filter by author Afloat16 opened on Nov 12, 2026

• Adapter parameters ignore explicit constructor device and dtype

  Lightning-AI/litgpt#2401 · Filter by author Afloat16 opened on Nov 12, 2026

• Skipping an overlong fine-tuning generation example leaves the model in evaluation mode

  Lightning-AI/litgpt#2399 · Filter by author Afloat16 opened on Nov 12, 2026

• Disabling cross-entropy chunking turns fully masked validation batches into NaN

  Lightning-AI/litgpt#2397 · Filter by author Afloat16 opened on Nov 12, 2026

• Zero warmup prevents fine-tuning scheduler creation

  Lightning-AI/litgpt#2395 · Filter by author Afloat16 opened on Nov 12, 2026

• CodeLlama-34b-Instruct-hf resolves to the Llama-3 prompt style instead of CodeLlama

  Lightning-AI/litgpt#2393 · Filter by author selimfedakar opened on Nov 9, 2026

• Grouped MoE routing can select experts from rejected groups with negative correction scores

  Lightning-AI/litgpt#2387 · Filter by author davrens101 opened on Nov 3, 2026

• Greedy speculative decoding can accept tokens differing from target argmax

  Lightning-AI/litgpt#2385 · Filter by author davrens101 opened on Nov 2, 2026

• LLM.benchmark() crashes with return_as_token_ids=True

  Lightning-AI/litgpt#2381 · Filter by author remote-controlled-man opened on Nov 1, 2026

• LLM.save() discards the model's active prompt style

  Lightning-AI/litgpt#2378 · Filter by author remote-controlled-man opened on Nov 1, 2026

• R1Base prompt style differs from the DeepSeek-R1 distill chat template (<think>\n generation prompt, multi-turn)

  Lightning-AI/litgpt#2375 · Filter by author jkatdare opened on Oct 28, 2026

• finetune_full trains lm_head and wte separately for tied-embedding checkpoints; HF export drops the trained head

  Lightning-AI/litgpt#2365 · Filter by author jkatdare opened on Oct 26, 2026

• generate_speculatively crashes when draft and target have different padded_vocab_size (e.g. Qwen2.5-1.5B -> Qwen2.5-7B)

  Lightning-AI/litgpt#2363 · Filter by author jkatdare opened on Oct 26, 2026

• LLM.benchmark() runs generate twice per iteration without streaming

  Lightning-AI/litgpt#2359 · Filter by author RaphaelFakhri opened on Oct 22, 2026

• Quick survey: your experience as a PyTorch Ecosystem maintainer

  Lightning-AI/litgpt#2358 · Filter by author michellearoth opened on Oct 16, 2026

• Finetuning with gradient accumulation weights micro-batches, not tokens, so the objective depends on micro_batch_size

  Lightning-AI/litgpt#2346 · Filter by author jkatdare opened on Oct 1, 2026
  2

• Sliding window layers rebuild a dense (T, T) attention mask with a Python loop on every training forward pass

  Lightning-AI/litgpt#2344 · Filter by author jkatdare opened on Oct 1, 2026

• Proposal: deploy LitGPT on Nebius Serverless Endpoints

  Lightning-AI/litgpt#2341 · Filter by author SalikovAlex opened on Sep 28, 2026

• fix_and_load_json drops a comma that is inside a string

  Lightning-AI/litgpt#2339 · Filter by author SashaMIT opened on Sep 28, 2026