Sub-notes

Image Source: https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-mixture-of-experts Basic MoE structure Experts are FFNN themselves, instead of passing input representation to only one dense FFNN we now have option to route them to more FFNNs. Since most LLMs have several decoder blocks, a given text will pass through multiple experts before the text is generated. Down the line it could use multiple experts but at different blocks i.e (layers) A routing layer is set to choose experts depending on how many experts are selected MoE are categorized into two i....

Calculation of FLOPs multiply accumulate cost: 2FLOPS i.e 1 for multiplication and 1 for accumulation (addition) if we multiply two matrices with sizes (a x b) and (b x c), the flops involved is b Multiply-add operation per the output size (a x c) i.e 2 x b x (a x c) Embedding lookup we initially have tokens with (seq_len,vocab_size) one-hot representation and embedding lookup matrix is (vocab_size, d_model), it will take...

After training, we generally perform alignment i.e teaching the model how to behave/act in desired manner. Post training mainly consists 1) Supervised Fine-tuning 2) RLHF the current consensus within the research community seems to be that the optimal approach to alignment is to i) perform SFT over a moderately-sized dataset of examples with very high quality and ii) invest remaining efforts into curating human preference data for fine-tuning via RLHF....

Pre-training Document packing while pretraining, different documents could be packed inside a sequence. For instance, a model with context_length 1024 can have 256 tokens from one doc and rest from the other. Demilited by EOS token. The samples may contaminate the attention, for which cross sample attention masking is used. But, it isn’t used by DeepSeek v3, lets not use it. while packing documents. we simply pack them as they appear in order and then add EOS token (used by GPT-2,3)....

torch.stack(tensors, dim) stacks the tensors across dim #usage # data has to be tensor torch.stack([data[i:i+some_number] for i in range(10)]) torch.from_numpy(numpy_array) shares the memory with the numpy_array but is tensor type a = np.array([1,2,3]) b = torch.tensor(a) # creates copy c = torch.from_numpy(a) # shares memory a[0] = 11 c # outputs: tensor([11, 2, 3]) torch.flatten(input, start,end=-1) flattens the input from dim start to end (-1 by default) t = torch.tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]]) torch....