# Sources and scope

The original teaching text was the user-provided handbook_with_exercises(1).html. The accompanying code is an independently written teaching implementation. It does not copy the earlier numerical records attributed to Qwen3-0.6B.

Original mechanism-check date: 2026-10-09. These are references for mechanisms, not code directly imported by the laboratory:

- GQA paper: https://arxiv.org/abs/2305.13245
- Pinned Transformers Llama implementation: https://github.com/huggingface/transformers/blob/b1463f5b871aeb20d58384a98aad57905a690799/src/transformers/models/llama/modeling_llama.py
- vLLM paged attention: https://docs.vllm.ai/en/latest/design/paged_attention/
- FlashAttention paper: https://arxiv.org/abs/2205.14135
- FlashAttention implementation and interfaces: https://github.com/Dao-AILab/flash-attention
- LMCache non-prefix blending implementation: https://github.com/LMCache/LMCache/blob/dev/lmcache/v1/compute/blend/blender.py

Online softmax checks mathematical accumulator merging rather than simulating a GPU memory hierarchy. Quantization is a custom uniform INT4 format with groupwise FP32 scales, distinct from NVFP4's E2M1 codebook, E4M3 local scales and global scale. Random-model logit differences demonstrate computational changes, not natural-language accuracy changes.