Fetching metadata from the HF Docker repository...
-
feat: add interactive plot to view rope rotation angle as a function of dimension pair index (d) and token position index (k)
-
feat: made attention logits plot more interactive and added intuitive logic fir attention logits and attention scores
-
feat: enhance attention visualization with token context and improve markdown explanations
-
feat: add formla of Q, K matrix calculation
-
fix: made the sequence length accurate for real model setup and also added information on which attention head is attending which set of dims
-
feat: add explanation on `hidden_size = num_attention_heads * dim_per_head`
-
feat: increase seq_len and auto-update of dim on UI during compute
-
fix: re-add space.GPU to comply with mandatory gpu usage
-
feat: add rope visualization only with some minor cmparison with sinusoidial positional embedding
-
fix: add spaces to comply with mandatory use og nvidia gpu on huggingface
-
fix: make numpy constraintless because of dependency on 3.10 on huggingface platform
-
feat: basic setup with docker, no real code has been added for visualization yet
-
fix: Fix Hugging Face username in workflow
DebasishDhal
-
feat: add workflow for github->hf sync (for building a hf space for visualization)
-
Added visualization for rope implementation
-
Added rope implementation code with explanation
-
feat: added bare minimum rel-position embedding
-
feat: added bare mimimum implementation of absolute sinusoidal positional encoding
-
Initial commit
DebasishDhal
