Upload q4-ort-metal.html with huggingface_hub
Browse files- q4-ort-metal.html +1365 -0
q4-ort-metal.html
ADDED
|
@@ -0,0 +1,1365 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<!DOCTYPE html>
|
| 2 |
+
<html lang="en">
|
| 3 |
+
<head>
|
| 4 |
+
<meta charset="utf-8"/>
|
| 5 |
+
<meta name="viewport" content="width=device-width, initial-scale=1"/>
|
| 6 |
+
<title>Magenta RT · Live (custom WGSL temporal + custom WGSL depth)</title>
|
| 7 |
+
<style>
|
| 8 |
+
:root{ --bg:#fffdf3; --ink:#1c1917; --mut:#78716c; --amber:#f59e0b; --amber2:#fbbf24; --lime:#84cc16; --line:#e7e1c8; --card:#fffef9; }
|
| 9 |
+
*{ box-sizing:border-box; } body{ margin:0; font-family:"Segoe UI",-apple-system,sans-serif; color:var(--ink);
|
| 10 |
+
background:radial-gradient(900px 480px at 80% -5%, #fef3c7, var(--bg) 55%); min-height:100vh; }
|
| 11 |
+
.wrap{ max-width:680px; margin:0 auto; padding:30px 20px 60px; }
|
| 12 |
+
h1{ font-size:23px; margin:0 0 2px; font-weight:800; letter-spacing:-.4px; }
|
| 13 |
+
h1 .em{ background:linear-gradient(90deg,var(--amber),var(--lime)); -webkit-background-clip:text; background-clip:text; color:transparent; }
|
| 14 |
+
.sub{ color:var(--mut); font-size:13px; margin-bottom:22px; }
|
| 15 |
+
.card{ background:var(--card); border:1px solid var(--line); border-radius:16px; padding:18px; margin-bottom:14px; }
|
| 16 |
+
label{ font-size:12px; color:var(--mut); text-transform:uppercase; letter-spacing:.5px; display:block; margin-bottom:8px; }
|
| 17 |
+
.row{ display:flex; gap:18px; flex-wrap:wrap; align-items:flex-end; }
|
| 18 |
+
select,button.go{ font-family:inherit; font-weight:700; }
|
| 19 |
+
select{ padding:10px 12px; border-radius:10px; border:1.5px solid var(--line); background:#fff; font-size:15px; color:var(--ink); }
|
| 20 |
+
button.go{ padding:12px 26px; border:none; border-radius:999px; font-size:15px; color:#3a2c05;
|
| 21 |
+
background:linear-gradient(135deg,var(--amber2),var(--amber)); box-shadow:0 4px 14px #f59e0b55; cursor:pointer; }
|
| 22 |
+
button.go:disabled{ opacity:.5; cursor:default; }
|
| 23 |
+
.stat{ font-size:13px; color:var(--mut); margin-top:10px; min-height:18px; }
|
| 24 |
+
.brk{ font-size:12px; color:var(--mut); margin-top:4px; min-height:16px; font-variant-numeric:tabular-nums; }
|
| 25 |
+
canvas{ width:100%; height:70px; display:block; border-radius:10px; background:#fffdf3; margin-top:10px; border:1px solid var(--line); }
|
| 26 |
+
#prog{ display:none; height:10px; background:#fef9e7; border:1px solid var(--line); border-radius:999px; overflow:hidden; margin-top:10px; }
|
| 27 |
+
#bar{ height:100%; width:0%; border-radius:999px; background:linear-gradient(90deg,var(--amber),var(--lime)); transition:width .15s; }
|
| 28 |
+
code{ background:#fef9e7; padding:1px 6px; border-radius:6px; font-size:12px; }
|
| 29 |
+
.badge{ font-size:11px; padding:2px 8px; border-radius:999px; background:#f7fee7; color:#3f6212; border:1px solid #d9f99d; }
|
| 30 |
+
</style>
|
| 31 |
+
</head>
|
| 32 |
+
<body>
|
| 33 |
+
<div class="wrap">
|
| 34 |
+
<h1>🍋 Magenta RT · <span class="em">full custom WGSL on your GPU</span></h1>
|
| 35 |
+
<div class="sub">ORT encoder → <b>custom WGSL temporal</b> → <b>custom WGSL depth</b> (all GPU-resident, shared ORT device) → ORT streaming decoder. Both transformers run as hand-written WebGPU kernels.</div>
|
| 36 |
+
|
| 37 |
+
<div class="card">
|
| 38 |
+
<label>Style</label>
|
| 39 |
+
<div class="row">
|
| 40 |
+
<select id="prompt"></select>
|
| 41 |
+
<button class="go" id="load">Load model</button>
|
| 42 |
+
<button class="go" id="play" disabled>▶ Play</button>
|
| 43 |
+
<button class="go" id="stop" disabled style="background:#2a2e45;color:#fff">■ Stop</button>
|
| 44 |
+
<span class="badge" id="rt"></span>
|
| 45 |
+
</div>
|
| 46 |
+
<div class="stat" id="stat">Press <b>Load model</b> (downloads encoder + decoder + 157 MB q4 WGSL temporal + 66 MB q4 WGSL depth weights, cached after first run).</div>
|
| 47 |
+
<div class="brk" id="brk"></div>
|
| 48 |
+
<div id="prog"><div id="bar"></div></div>
|
| 49 |
+
<canvas id="viz" width="640" height="70"></canvas>
|
| 50 |
+
</div>
|
| 51 |
+
</div>
|
| 52 |
+
|
| 53 |
+
<script type="module">
|
| 54 |
+
import * as ort from "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.23.0/dist/ort.webgpu.min.mjs";
|
| 55 |
+
ort.env.wasm.numThreads = navigator.hardwareConcurrency ? Math.min(4, navigator.hardwareConcurrency) : 2;
|
| 56 |
+
ort.env.webgpu.powerPreference = "high-performance";
|
| 57 |
+
|
| 58 |
+
const $ = id => document.getElementById(id);
|
| 59 |
+
// fixed config: WebGPU. Temporal AND depth are custom WGSL kernels on ORT's device.
|
| 60 |
+
const sz = "small", dev = "webgpu";
|
| 61 |
+
const REPO = `https://huggingface.co/magenta-torch/magenta-rt-onnx-${sz}/resolve/main`;
|
| 62 |
+
const COMMUNITY = "https://huggingface.co/magenta-community/magenta-rt-onnx-small/resolve/main";
|
| 63 |
+
const DEC_URL = COMMUNITY + "/spectrostream_stream.onnx"; // stateful streaming decoder
|
| 64 |
+
const DEC_STATE_URL = COMMUNITY + "/stream_state.json";
|
| 65 |
+
const DEPTH_BASE = COMMUNITY + "/depth_wgsl/"; // depth_q4_ort.bin + depth_q4_ort_layout.json (q4)
|
| 66 |
+
const TEMP_BASE = COMMUNITY + "/temporal_wgsl/"; // temporal_q4_ort.bin + temporal_q4_ort_layout.json (q4)
|
| 67 |
+
let cfg=null, S={}, EMB=null, QUANT=null, presets={};
|
| 68 |
+
let DSJ=null, DSTATE=null, HIFLUSH=undefined;
|
| 69 |
+
let DEPTH=null; // custom WGSL depth engine (built on ort.env.webgpu.device)
|
| 70 |
+
let TEMPORAL=null;// custom WGSL temporal engine (same device)
|
| 71 |
+
|
| 72 |
+
function setBar(f){ $("bar").style.width=Math.min(100,Math.max(0,f*100)).toFixed(1)+"%"; }
|
| 73 |
+
|
| 74 |
+
// ── fp16 -> f32 (for EMB/QUANT bin tables) ──
|
| 75 |
+
function f16(u){ const s=(u&0x8000)>>15,e=(u&0x7c00)>>10,f=u&0x03ff;
|
| 76 |
+
if(e===0) return (s?-1:1)*Math.pow(2,-14)*(f/1024);
|
| 77 |
+
if(e===31) return f?NaN:(s?-1:1)*Infinity;
|
| 78 |
+
return (s?-1:1)*Math.pow(2,e-15)*(1+f/1024); }
|
| 79 |
+
function f16decode(b){ const u=new Uint16Array(b.buffer,b.byteOffset,b.byteLength>>1), o=new Float32Array(u.length); for(let i=0;i<u.length;i++)o[i]=f16(u[i]); return o; }
|
| 80 |
+
|
| 81 |
+
// ── cached fetch (Cache API) with streaming progress ──
|
| 82 |
+
async function fetchBuf(url, onFrac){
|
| 83 |
+
let cache=null;
|
| 84 |
+
try{ cache=await caches.open("mrt-fullcustom-q4-v1"); const hit=await cache.match(url);
|
| 85 |
+
if(hit){ onFrac&&onFrac(1); return new Uint8Array(await hit.arrayBuffer()); } }catch(e){}
|
| 86 |
+
const r=await fetch(url); if(!r.ok) throw new Error(r.status+" "+url.split("/").pop());
|
| 87 |
+
const total=+r.headers.get("content-length")||0, reader=r.body.getReader(), chunks=[]; let recv=0;
|
| 88 |
+
for(;;){ const {done,value}=await reader.read(); if(done)break; chunks.push(value); recv+=value.length; if(total&&onFrac)onFrac(recv/total); }
|
| 89 |
+
const buf=new Uint8Array(recv); let o=0; for(const c of chunks){buf.set(c,o);o+=c.length;}
|
| 90 |
+
if(cache){ try{ await cache.put(url, new Response(buf)); }catch(e){} }
|
| 91 |
+
return buf;
|
| 92 |
+
}
|
| 93 |
+
async function fetchJSON(url){ const r=await fetch(url); if(!r.ok) throw new Error(r.status+" "+url); return r.json(); }
|
| 94 |
+
|
| 95 |
+
/* ====================================================================== *
|
| 96 |
+
* CUSTOM WGSL TEMPORAL (ported verbatim from temporal_q4_gpu.ts — q4)
|
| 97 |
+
* 12 layers: self-attn (1 sink) -> cross-attn (1 sink) -> FFN. Big matmul
|
| 98 |
+
* matrices are q4 block-quantized along K (block_size=32, zero=8): packed
|
| 99 |
+
* nibbles in u32 (8/word, k-contiguous), per-block fp32 scales; the matvec
|
| 100 |
+
* unpacks (nibble-8)*scale with u32 shift/mask ONLY (NO unpack2x16float ->
|
| 101 |
+
* avoids the Metal compile hang). PACKED split P0|P1 (<120MB each), SCALES,
|
| 102 |
+
* and a small fp32 SMALL blob (embedding/norms/biases/sink/per_dim_scale).
|
| 103 |
+
* Self/cross KV are GPU buffers persisted across frames, slot-appended, grown
|
| 104 |
+
* to cap KV_CAP and reset on Play/restart. Built on the SAME GPUDevice ORT
|
| 105 |
+
* uses so the temporal `out` GPUBuffer feeds the WGSL depth with zero copy.
|
| 106 |
+
* ONE step = embed(prev[12]) -> 12 layers -> out[1024] (raw, no final norm).
|
| 107 |
+
* ====================================================================== */
|
| 108 |
+
const T_D_MODEL=1024, T_D_SRC=256, T_D_FF=4096, T_NUM_LAYERS=12, T_NUM_HEADS=8,
|
| 109 |
+
T_UPH=128, T_HD=T_NUM_HEADS*T_UPH, T_KV_CAP=42, T_EPS=1e-6, T_R_SOFTPLUS_0=1.442695041;
|
| 110 |
+
const T_BLK=32; // q4 block size
|
| 111 |
+
// bandwidth (q4 matvec) tiling
|
| 112 |
+
const T_RED=32, T_ROWS=8, T_T_WG=T_ROWS*T_RED; // 256
|
| 113 |
+
const tNWGT=N=>Math.ceil(N/T_ROWS);
|
| 114 |
+
|
| 115 |
+
// Build the whole temporal engine on ORT's device. q4 weights (what ships).
|
| 116 |
+
// Returns { runTemporal(prevToks)->outBuf, resetKV(), setSource(), kvLen(), outBuf }.
|
| 117 |
+
async function buildTemporal(device, layout, binBytes){
|
| 118 |
+
let USE_SUBGROUPS=false; // browser: layout:auto + shared-mem tree (subgroup-free)
|
| 119 |
+
|
| 120 |
+
const D_MODEL=T_D_MODEL, D_SRC=T_D_SRC, D_FF=T_D_FF, NUM_LAYERS=T_NUM_LAYERS,
|
| 121 |
+
NUM_HEADS=T_NUM_HEADS, UPH=T_UPH, HD=T_HD, KV_CAP=T_KV_CAP, EPS=T_EPS,
|
| 122 |
+
R_SOFTPLUS_0=T_R_SOFTPLUS_0, RED=T_RED, ROWS=T_ROWS, T_WG=T_T_WG, BLK=T_BLK;
|
| 123 |
+
|
| 124 |
+
const EMBED_SCALE=layout.embed_scale;
|
| 125 |
+
// q4 layout: q4[name]={K,N,packed_off,scale_off,...}; small[name]={offset,...}
|
| 126 |
+
const q4=n=>layout.q4[n];
|
| 127 |
+
const smallOff=n=>layout.small[n].offset;
|
| 128 |
+
const L=(li,blk,t)=>smallOff(`L${li}.${blk}.${t}`);
|
| 129 |
+
|
| 130 |
+
// section byte ranges within temporal_q4_ort.bin
|
| 131 |
+
const PACKED_BYTE_OFF=layout.packed_byte_off, SCALE_BYTE_OFF=layout.scale_byte_off,
|
| 132 |
+
SMALL_BYTE_OFF=layout.small_byte_off;
|
| 133 |
+
const PACKED_WORDS=layout.packed_words, SCALE_FLOATS=layout.scale_floats, SMALL_FLOATS=layout.small_floats;
|
| 134 |
+
// split PACKED across two storage buffers (binding 0/6) — each well under cap.
|
| 135 |
+
const P_SPLIT_WORDS=Math.ceil(PACKED_WORDS/2);
|
| 136 |
+
|
| 137 |
+
// ---------------- WGSL header / weight access ----------------
|
| 138 |
+
const CONSTS=`
|
| 139 |
+
const D_MODEL : u32 = ${D_MODEL}u;
|
| 140 |
+
const D_SRC : u32 = ${D_SRC}u;
|
| 141 |
+
const D_FF : u32 = ${D_FF}u;
|
| 142 |
+
const NUM_HEADS : u32 = ${NUM_HEADS}u;
|
| 143 |
+
const UPH : u32 = ${UPH}u;
|
| 144 |
+
const HD : u32 = ${HD}u;
|
| 145 |
+
const EPS : f32 = ${EPS};
|
| 146 |
+
const R_SOFTPLUS_0 : f32 = ${R_SOFTPLUS_0};
|
| 147 |
+
const RED : u32 = ${RED}u;
|
| 148 |
+
const ROWS : u32 = ${ROWS}u;
|
| 149 |
+
const BLK : u32 = ${BLK}u;
|
| 150 |
+
`;
|
| 151 |
+
// SMALL fp32 params (binding 8), float-indexed via wv().
|
| 152 |
+
const SMALLBUF=`
|
| 153 |
+
@group(0) @binding(8) var<storage, read> SMALL : array<f32>;
|
| 154 |
+
fn wv(i : u32) -> f32 { return SMALL[i]; }
|
| 155 |
+
fn softplus(x: f32) -> f32 {
|
| 156 |
+
return log(1.0 + exp(-abs(x))) + max(x, 0.0);
|
| 157 |
+
}
|
| 158 |
+
fn gelu_tanh(x: f32) -> f32 {
|
| 159 |
+
let c : f32 = 0.7978845608028654;
|
| 160 |
+
// clamp: Metal fast-math tanh overflows to 0/NaN for |arg|>~44.
|
| 161 |
+
// |t|>=12 already rounds tanh(t) to +/-1 in f32, so this is exact.
|
| 162 |
+
let t = clamp(c * (x + 0.044715 * x * x * x), -12.0, 12.0);
|
| 163 |
+
return 0.5 * x * (1.0 + tanh(t));
|
| 164 |
+
}
|
| 165 |
+
`;
|
| 166 |
+
// PACKED q4 nibbles (binding 0/6) + SCALES (binding 7). qword() routes a global
|
| 167 |
+
// u32-word index to the right packed buffer; deq() dequantizes w[col,k].
|
| 168 |
+
const QBUF=`
|
| 169 |
+
@group(0) @binding(0) var<storage, read> P0 : array<u32>;
|
| 170 |
+
@group(0) @binding(6) var<storage, read> P1 : array<u32>;
|
| 171 |
+
@group(0) @binding(7) var<storage, read> SC : array<f32>;
|
| 172 |
+
const P_SPLIT_WORDS : u32 = ${P_SPLIT_WORDS}u;
|
| 173 |
+
fn qword(wi : u32) -> u32 {
|
| 174 |
+
if (wi < P_SPLIT_WORDS) { return P0[wi]; }
|
| 175 |
+
return P1[wi - P_SPLIT_WORDS];
|
| 176 |
+
}
|
| 177 |
+
fn deq(qOff : u32, sOff : u32, K : u32, col : u32, k : u32) -> f32 {
|
| 178 |
+
let word = qword(qOff + col * (K >> 3u) + (k >> 3u));
|
| 179 |
+
let nib = (word >> ((k & 7u) * 4u)) & 0xfu;
|
| 180 |
+
let s = SC[sOff + col * (K >> 5u) + (k >> 5u)];
|
| 181 |
+
return (f32(nib) - 8.0) * s;
|
| 182 |
+
}
|
| 183 |
+
// double-single (df64) compensated summation: emulates fp64-precision reductions
|
| 184 |
+
// in fp32 so the temporal autoregressive loop stays stable indefinitely.
|
| 185 |
+
fn two_sum(a : f32, b : f32) -> vec2<f32> {
|
| 186 |
+
let s = a + b;
|
| 187 |
+
let bb = s - a;
|
| 188 |
+
return vec2<f32>(s, (a - (s - bb)) + (b - bb));
|
| 189 |
+
}
|
| 190 |
+
fn df_add_f32(acc : vec2<f32>, x : f32) -> vec2<f32> {
|
| 191 |
+
let t = two_sum(acc.x, x);
|
| 192 |
+
return vec2<f32>(t.x, acc.y + t.y);
|
| 193 |
+
}
|
| 194 |
+
fn df_add(a : vec2<f32>, b : vec2<f32>) -> vec2<f32> {
|
| 195 |
+
let t = two_sum(a.x, b.x);
|
| 196 |
+
return vec2<f32>(t.x, (a.y + b.y) + t.y);
|
| 197 |
+
}
|
| 198 |
+
`;
|
| 199 |
+
const HEADER=CONSTS+QBUF+SMALLBUF;
|
| 200 |
+
|
| 201 |
+
// ===== qMatmul (q4 bandwidth matvec; replaces gridMatmulT/gridMatmul2) =====
|
| 202 |
+
// out[n] = sum_k inEff[k] * deq(W)[n,k]. lane reads k-contiguous packed u32s.
|
| 203 |
+
// uniform: { qOff, sOff, colBase, normOff, biasOff }.
|
| 204 |
+
function qMatmul(opts){
|
| 205 |
+
const K=opts.K, sg=USE_SUBGROUPS;
|
| 206 |
+
const inDecl=`@group(0) @binding(${opts.inBinding}) var<storage, read> vin : array<f32>;`;
|
| 207 |
+
const outDecl=`@group(0) @binding(${opts.outBinding}) var<storage, read_write> vout : array<f32>;`;
|
| 208 |
+
const uniformDecl=`struct U { qOff:u32, sOff:u32, colBase:u32, normOff:u32, biasOff:u32, };
|
| 209 |
+
@group(0) @binding(5) var<uniform> u : U;`;
|
| 210 |
+
const normReduce=opts.rmsnorm?`
|
| 211 |
+
var lss : f32 = 0.0;
|
| 212 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) {
|
| 213 |
+
let xv = vin[i]; lss = lss + xv * xv;
|
| 214 |
+
}
|
| 215 |
+
red[tid] = lss;
|
| 216 |
+
workgroupBarrier();
|
| 217 |
+
for (var s : u32 = ${T_WG/2}u; s > 0u; s = s >> 1u) {
|
| 218 |
+
if (tid < s) { red[tid] = red[tid] + red[tid + s]; }
|
| 219 |
+
workgroupBarrier();
|
| 220 |
+
}
|
| 221 |
+
let rms = 1.0 / sqrt(red[0] / f32(${K}u) + EPS);
|
| 222 |
+
workgroupBarrier();
|
| 223 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) {
|
| 224 |
+
inS[i] = (vin[i] * rms) * wv(u.normOff + i);
|
| 225 |
+
}
|
| 226 |
+
workgroupBarrier();`:`
|
| 227 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) { inS[i] = vin[i]; }
|
| 228 |
+
workgroupBarrier();`;
|
| 229 |
+
let epi="";
|
| 230 |
+
if(opts.epilogue==="geluBias"){ epi=`acc = gelu_tanh(acc + wv(u.biasOff + col));`; }
|
| 231 |
+
else { epi=opts.hasBias?`acc = acc + wv(u.biasOff + col);`:``; }
|
| 232 |
+
const reduceWrite=sg?`
|
| 233 |
+
let total = subgroupAdd(acc.x) + subgroupAdd(acc.y);
|
| 234 |
+
if (lane == 0u) {
|
| 235 |
+
var accf : f32 = total;
|
| 236 |
+
${epi.replace(/\bacc\b/g,"accf")}
|
| 237 |
+
vout[outIdx] = accf;
|
| 238 |
+
}`:`
|
| 239 |
+
part[tid] = acc;
|
| 240 |
+
workgroupBarrier();
|
| 241 |
+
for (var s : u32 = ${RED/2}u; s > 0u; s = s >> 1u) {
|
| 242 |
+
if (lane < s) { part[tid] = df_add(part[tid], part[tid + s]); }
|
| 243 |
+
workgroupBarrier();
|
| 244 |
+
}
|
| 245 |
+
if (lane == 0u) {
|
| 246 |
+
var accf : f32 = part[row * RED].x;
|
| 247 |
+
${epi.replace(/\bacc\b/g,"accf")}
|
| 248 |
+
vout[outIdx] = accf;
|
| 249 |
+
}`;
|
| 250 |
+
const enableSg=sg?"enable subgroups;\n":"";
|
| 251 |
+
return enableSg+HEADER+`
|
| 252 |
+
${inDecl}
|
| 253 |
+
${outDecl}
|
| 254 |
+
${uniformDecl}
|
| 255 |
+
|
| 256 |
+
var<workgroup> inS : array<f32, ${K}>;
|
| 257 |
+
${opts.rmsnorm?`var<workgroup> red : array<f32, ${T_WG}>;`:``}
|
| 258 |
+
${sg?"":`var<workgroup> part : array<vec2<f32>, ${T_WG}>;`}
|
| 259 |
+
|
| 260 |
+
@compute @workgroup_size(${T_WG})
|
| 261 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>,
|
| 262 |
+
@builtin(workgroup_id) wid : vec3<u32>) {
|
| 263 |
+
let tid = lid.x;
|
| 264 |
+
let row = tid / RED;
|
| 265 |
+
let lane = tid % RED;
|
| 266 |
+
${normReduce}
|
| 267 |
+
let outIdx = wid.x * ROWS + row;
|
| 268 |
+
let col = u.colBase + outIdx; // global output column (== weight column)
|
| 269 |
+
var acc : vec2<f32> = vec2<f32>(0.0, 0.0); // (hi,lo) double-single accumulator
|
| 270 |
+
for (var k : u32 = lane; k < ${K}u; k = k + RED) {
|
| 271 |
+
let p = inS[k] * deq(u.qOff, u.sOff, ${K}u, col, k);
|
| 272 |
+
acc = df_add_f32(acc, p);
|
| 273 |
+
}
|
| 274 |
+
${reduceWrite}
|
| 275 |
+
}
|
| 276 |
+
`;
|
| 277 |
+
}
|
| 278 |
+
|
| 279 |
+
// ===== buildKV (combined q4 QKV self / KV cross) =====
|
| 280 |
+
// uniform: { normOff, qQ,qS, kQ,kS, vQ,vS, slot }.
|
| 281 |
+
function buildKV(opts){
|
| 282 |
+
const K=opts.K, sg=USE_SUBGROUPS, enableSg=sg?"enable subgroups;\n":"";
|
| 283 |
+
const qPart=opts.writeQ?`var qa : vec2<f32> = vec2<f32>(0.0, 0.0);`:``;
|
| 284 |
+
const qAccum=opts.writeQ?`qa = df_add_f32(qa, hv * deq(u.qQ, u.qS, ${K}u, oi, k));`:``;
|
| 285 |
+
const reduceWrite=sg?`
|
| 286 |
+
let sk = subgroupAdd(ka.x) + subgroupAdd(ka.y);
|
| 287 |
+
let sv = subgroupAdd(va.x) + subgroupAdd(va.y);
|
| 288 |
+
${opts.writeQ?"let sq = subgroupAdd(qa.x) + subgroupAdd(qa.y);":""}
|
| 289 |
+
if (lane == 0u) {
|
| 290 |
+
let cbase = u.slot * HD;
|
| 291 |
+
kcache[cbase + oi] = sk;
|
| 292 |
+
vcache[cbase + oi] = sv;
|
| 293 |
+
${opts.writeQ?"qbuf[oi] = sq;":""}
|
| 294 |
+
}`:`
|
| 295 |
+
pk[tid] = ka; pv[tid] = va;
|
| 296 |
+
${opts.writeQ?"pq[tid] = qa;":""}
|
| 297 |
+
workgroupBarrier();
|
| 298 |
+
for (var s : u32 = ${RED/2}u; s > 0u; s = s >> 1u) {
|
| 299 |
+
if (lane < s) {
|
| 300 |
+
pk[tid] = df_add(pk[tid], pk[tid + s]);
|
| 301 |
+
pv[tid] = df_add(pv[tid], pv[tid + s]);
|
| 302 |
+
${opts.writeQ?"pq[tid] = df_add(pq[tid], pq[tid + s]);":""}
|
| 303 |
+
}
|
| 304 |
+
workgroupBarrier();
|
| 305 |
+
}
|
| 306 |
+
if (lane == 0u) {
|
| 307 |
+
let cbase = u.slot * HD;
|
| 308 |
+
kcache[cbase + oi] = pk[row * RED].x;
|
| 309 |
+
vcache[cbase + oi] = pv[row * RED].x;
|
| 310 |
+
${opts.writeQ?"qbuf[oi] = pq[row * RED].x;":""}
|
| 311 |
+
}`;
|
| 312 |
+
const normReduce=opts.rmsnorm?`
|
| 313 |
+
var lss : f32 = 0.0;
|
| 314 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) {
|
| 315 |
+
let xv = vin[i]; lss = lss + xv * xv;
|
| 316 |
+
}
|
| 317 |
+
red[tid] = lss;
|
| 318 |
+
workgroupBarrier();
|
| 319 |
+
for (var s : u32 = ${T_WG/2}u; s > 0u; s = s >> 1u) {
|
| 320 |
+
if (tid < s) { red[tid] = red[tid] + red[tid + s]; }
|
| 321 |
+
workgroupBarrier();
|
| 322 |
+
}
|
| 323 |
+
let rms = 1.0 / sqrt(red[0] / f32(${K}u) + EPS);
|
| 324 |
+
workgroupBarrier();
|
| 325 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) {
|
| 326 |
+
inS[i] = (vin[i] * rms) * wv(u.normOff + i);
|
| 327 |
+
}
|
| 328 |
+
workgroupBarrier();`:`
|
| 329 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) { inS[i] = vin[i]; }
|
| 330 |
+
workgroupBarrier();`;
|
| 331 |
+
return enableSg+HEADER+`
|
| 332 |
+
@group(0) @binding(1) var<storage, read> vin : array<f32>;
|
| 333 |
+
${opts.writeQ?`@group(0) @binding(2) var<storage, read_write> qbuf : array<f32>;`:``}
|
| 334 |
+
@group(0) @binding(3) var<storage, read_write> kcache : array<f32>;
|
| 335 |
+
@group(0) @binding(4) var<storage, read_write> vcache : array<f32>;
|
| 336 |
+
struct U {
|
| 337 |
+
normOff:u32,
|
| 338 |
+
qQ:u32, qS:u32,
|
| 339 |
+
kQ:u32, kS:u32,
|
| 340 |
+
vQ:u32, vS:u32,
|
| 341 |
+
slot:u32,
|
| 342 |
+
};
|
| 343 |
+
@group(0) @binding(5) var<uniform> u : U;
|
| 344 |
+
|
| 345 |
+
var<workgroup> inS : array<f32, ${K}>;
|
| 346 |
+
var<workgroup> red : array<f32, ${T_WG}>;
|
| 347 |
+
${sg?"":`var<workgroup> pk : array<vec2<f32>, ${T_WG}>;
|
| 348 |
+
var<workgroup> pv : array<vec2<f32>, ${T_WG}>;
|
| 349 |
+
${opts.writeQ?`var<workgroup> pq : array<vec2<f32>, ${T_WG}>;`:``}`}
|
| 350 |
+
|
| 351 |
+
@compute @workgroup_size(${T_WG})
|
| 352 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>,
|
| 353 |
+
@builtin(workgroup_id) wid : vec3<u32>) {
|
| 354 |
+
let tid = lid.x;
|
| 355 |
+
let row = tid / RED;
|
| 356 |
+
let lane = tid % RED;
|
| 357 |
+
${normReduce}
|
| 358 |
+
let oi = wid.x * ROWS + row; // output column in [0,HD)
|
| 359 |
+
var ka : vec2<f32> = vec2<f32>(0.0, 0.0); var va : vec2<f32> = vec2<f32>(0.0, 0.0);
|
| 360 |
+
${qPart}
|
| 361 |
+
for (var k : u32 = lane; k < ${K}u; k = k + RED) {
|
| 362 |
+
let hv = inS[k];
|
| 363 |
+
ka = df_add_f32(ka, hv * deq(u.kQ, u.kS, ${K}u, oi, k));
|
| 364 |
+
va = df_add_f32(va, hv * deq(u.vQ, u.vS, ${K}u, oi, k));
|
| 365 |
+
${qAccum}
|
| 366 |
+
}
|
| 367 |
+
${reduceWrite}
|
| 368 |
+
}
|
| 369 |
+
`;
|
| 370 |
+
}
|
| 371 |
+
const buildQonly=()=>qMatmul({K:D_MODEL,inBinding:1,outBinding:2,rmsnorm:true,epilogue:"none",hasBias:false});
|
| 372 |
+
|
| 373 |
+
// ===== attention core (1 sink + per-dim scale) =====
|
| 374 |
+
const ATTN_SINK_WGSL=HEADER+`
|
| 375 |
+
@group(0) @binding(1) var<storage, read> qbuf : array<f32>; // [HD]
|
| 376 |
+
@group(0) @binding(2) var<storage, read> kcache : array<f32>; // [cap,HD]
|
| 377 |
+
@group(0) @binding(3) var<storage, read> vcache : array<f32>; // [cap,HD]
|
| 378 |
+
@group(0) @binding(4) var<storage, read_write> ctxbuf : array<f32>; // [HD]
|
| 379 |
+
struct U { per_dim_scale:u32, sink_k:u32, sink_v:u32, cacheLen:u32, start:u32, };
|
| 380 |
+
@group(0) @binding(5) var<uniform> u : U;
|
| 381 |
+
|
| 382 |
+
var<workgroup> logitsW : array<f32, ${KV_CAP + 1}>;
|
| 383 |
+
var<workgroup> red : array<f32, 256>;
|
| 384 |
+
var<workgroup> mden : array<f32, 2>;
|
| 385 |
+
|
| 386 |
+
@compute @workgroup_size(256)
|
| 387 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>) {
|
| 388 |
+
let t = lid.x;
|
| 389 |
+
let T = u.cacheLen; // number of real keys
|
| 390 |
+
let nkeys = T + 1u; // + sink
|
| 391 |
+
let qscale = R_SOFTPLUS_0 * (1.0 / sqrt(f32(UPH)));
|
| 392 |
+
|
| 393 |
+
for (var h : u32 = 0u; h < NUM_HEADS; h = h + 1u) {
|
| 394 |
+
let qbase = h * UPH;
|
| 395 |
+
for (var ki : u32 = t; ki < nkeys; ki = ki + 256u) {
|
| 396 |
+
var acc : vec2<f32> = vec2<f32>(0.0, 0.0); // (hi,lo) double-single
|
| 397 |
+
if (ki == 0u) {
|
| 398 |
+
let sbase = u.sink_k + qbase;
|
| 399 |
+
for (var d : u32 = 0u; d < UPH; d = d + 1u) {
|
| 400 |
+
acc = df_add_f32(acc, qbuf[qbase + d] * wv(sbase + d));
|
| 401 |
+
}
|
| 402 |
+
} else {
|
| 403 |
+
let kk = ki - 1u;
|
| 404 |
+
// read KV in temporal order (oldest->newest): physical ring slot.
|
| 405 |
+
let phys = (u.start + kk) % ${KV_CAP}u;
|
| 406 |
+
let kbase = phys * HD + qbase;
|
| 407 |
+
for (var d : u32 = 0u; d < UPH; d = d + 1u) {
|
| 408 |
+
let sc = qscale * softplus(wv(u.per_dim_scale + d));
|
| 409 |
+
acc = df_add_f32(acc, (qbuf[qbase + d] * sc) * kcache[kbase + d]);
|
| 410 |
+
}
|
| 411 |
+
}
|
| 412 |
+
logitsW[ki] = acc.x;
|
| 413 |
+
}
|
| 414 |
+
workgroupBarrier();
|
| 415 |
+
if (t == 0u) {
|
| 416 |
+
var m : f32 = logitsW[0];
|
| 417 |
+
for (var ki : u32 = 1u; ki < nkeys; ki = ki + 1u) { m = max(m, logitsW[ki]); }
|
| 418 |
+
var s : f32 = 0.0;
|
| 419 |
+
for (var ki : u32 = 0u; ki < nkeys; ki = ki + 1u) {
|
| 420 |
+
let e = exp(logitsW[ki] - m);
|
| 421 |
+
logitsW[ki] = e; s = s + e;
|
| 422 |
+
}
|
| 423 |
+
mden[0] = s;
|
| 424 |
+
}
|
| 425 |
+
workgroupBarrier();
|
| 426 |
+
let denom = mden[0];
|
| 427 |
+
for (var d : u32 = t; d < UPH; d = d + 256u) {
|
| 428 |
+
// ctx = sum_keys w * V, accumulated in double-single (hi,lo).
|
| 429 |
+
var acc : vec2<f32> = vec2<f32>((logitsW[0] / denom) * wv(u.sink_v + qbase + d), 0.0);
|
| 430 |
+
for (var kk : u32 = 0u; kk < T; kk = kk + 1u) {
|
| 431 |
+
let w = logitsW[kk + 1u] / denom;
|
| 432 |
+
// read V in temporal order (oldest->newest): physical ring slot.
|
| 433 |
+
let phys = (u.start + kk) % ${KV_CAP}u;
|
| 434 |
+
acc = df_add_f32(acc, w * vcache[phys * HD + qbase + d]);
|
| 435 |
+
}
|
| 436 |
+
ctxbuf[qbase + d] = acc.x;
|
| 437 |
+
}
|
| 438 |
+
workgroupBarrier();
|
| 439 |
+
}
|
| 440 |
+
}
|
| 441 |
+
`;
|
| 442 |
+
|
| 443 |
+
// ===== RESID =====
|
| 444 |
+
const RESID_WGSL=HEADER+`
|
| 445 |
+
@group(0) @binding(1) var<storage, read_write> x : array<f32>;
|
| 446 |
+
@group(0) @binding(2) var<storage, read> y : array<f32>;
|
| 447 |
+
struct U { post:u32, };
|
| 448 |
+
@group(0) @binding(5) var<uniform> u : U;
|
| 449 |
+
|
| 450 |
+
var<workgroup> red : array<f32, 256>;
|
| 451 |
+
@compute @workgroup_size(256)
|
| 452 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>) {
|
| 453 |
+
let t = lid.x;
|
| 454 |
+
var ss : f32 = 0.0;
|
| 455 |
+
for (var i : u32 = t; i < D_MODEL; i = i + 256u) { let v = y[i]; ss = ss + v*v; }
|
| 456 |
+
red[t] = ss;
|
| 457 |
+
workgroupBarrier();
|
| 458 |
+
for (var s : u32 = 128u; s > 0u; s = s >> 1u) {
|
| 459 |
+
if (t < s) { red[t] = red[t] + red[t + s]; }
|
| 460 |
+
workgroupBarrier();
|
| 461 |
+
}
|
| 462 |
+
let rms = 1.0 / sqrt(red[0] / f32(D_MODEL) + EPS);
|
| 463 |
+
workgroupBarrier();
|
| 464 |
+
for (var i : u32 = t; i < D_MODEL; i = i + 256u) {
|
| 465 |
+
x[i] = x[i] + (y[i] * rms) * wv(u.post + i);
|
| 466 |
+
}
|
| 467 |
+
}
|
| 468 |
+
`;
|
| 469 |
+
|
| 470 |
+
// ===== EMBED prev =====
|
| 471 |
+
const EMBED_WGSL=HEADER+`
|
| 472 |
+
@group(0) @binding(1) var<storage, read> prevBuf : array<u32>; // [12]
|
| 473 |
+
@group(0) @binding(2) var<storage, read_write> xout : array<f32>; // [1024]
|
| 474 |
+
struct EU { embeddingOff:u32, embed_scale:f32, };
|
| 475 |
+
@group(0) @binding(5) var<uniform> u : EU;
|
| 476 |
+
|
| 477 |
+
@compute @workgroup_size(256)
|
| 478 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>) {
|
| 479 |
+
let i = lid.x;
|
| 480 |
+
for (var di : u32 = i; di < D_MODEL; di = di + 256u) {
|
| 481 |
+
var acc : f32 = 0.0;
|
| 482 |
+
for (var c : u32 = 0u; c < 12u; c = c + 1u) {
|
| 483 |
+
let tok = prevBuf[c];
|
| 484 |
+
acc = acc + wv(u.embeddingOff + tok * D_MODEL + di);
|
| 485 |
+
}
|
| 486 |
+
xout[di] = (acc / 12.0) * u.embed_scale;
|
| 487 |
+
}
|
| 488 |
+
}
|
| 489 |
+
`;
|
| 490 |
+
|
| 491 |
+
// ---------------- q4 weight storage buffers (PACKED P0|P1, SCALES, SMALL) ----
|
| 492 |
+
const P0_BYTES=P_SPLIT_WORDS*4, P1_BYTES=PACKED_WORDS*4-P0_BYTES;
|
| 493 |
+
const p0Buf=device.createBuffer({ size:Math.max(4,P0_BYTES), usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 494 |
+
const p1Buf=device.createBuffer({ size:Math.max(4,P1_BYTES), usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 495 |
+
device.queue.writeBuffer(p0Buf,0,binBytes,PACKED_BYTE_OFF,P0_BYTES);
|
| 496 |
+
device.queue.writeBuffer(p1Buf,0,binBytes,PACKED_BYTE_OFF+P0_BYTES,P1_BYTES);
|
| 497 |
+
const scBuf=device.createBuffer({ size:Math.max(4,SCALE_FLOATS*4), usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 498 |
+
device.queue.writeBuffer(scBuf,0,binBytes,SCALE_BYTE_OFF,SCALE_FLOATS*4);
|
| 499 |
+
const smBuf=device.createBuffer({ size:Math.max(4,SMALL_FLOATS*4), usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 500 |
+
device.queue.writeBuffer(smBuf,0,binBytes,SMALL_BYTE_OFF,SMALL_FLOATS*4);
|
| 501 |
+
console.log(`temporal q4: PACKED ${(PACKED_WORDS*4/1048576).toFixed(1)}MB (P0 ${(P0_BYTES/1048576).toFixed(1)} / P1 ${(P1_BYTES/1048576).toFixed(1)}), SCALES ${(SCALE_FLOATS*4/1048576).toFixed(1)}MB, SMALL ${(SMALL_FLOATS*4/1048576).toFixed(1)}MB`);
|
| 502 |
+
|
| 503 |
+
// ---------------- activation buffers (GPU-resident) ----------------
|
| 504 |
+
const makeStorage=(floats,extra=0)=>device.createBuffer({
|
| 505 |
+
size:Math.max(4,floats*4),
|
| 506 |
+
usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST|GPUBufferUsage.COPY_SRC|extra });
|
| 507 |
+
const xBuf=makeStorage(D_MODEL); // residual stream x[1024] == temporal `out`
|
| 508 |
+
const srcBuf=makeStorage(D_SRC); // source[256] (uploaded per style)
|
| 509 |
+
const qBuf=makeStorage(HD);
|
| 510 |
+
const ctxBuf=makeStorage(HD);
|
| 511 |
+
const blkOutBuf=makeStorage(D_MODEL);
|
| 512 |
+
const ffBuf=makeStorage(D_FF);
|
| 513 |
+
// KV caches: per layer, capacity KV_CAP+1 rows of HD.
|
| 514 |
+
const selfK=[],selfV=[],crossK=[],crossV=[];
|
| 515 |
+
for(let li=0;li<NUM_LAYERS;li++){
|
| 516 |
+
selfK.push(makeStorage((KV_CAP+1)*HD));
|
| 517 |
+
selfV.push(makeStorage((KV_CAP+1)*HD));
|
| 518 |
+
crossK.push(makeStorage((KV_CAP+1)*HD));
|
| 519 |
+
crossV.push(makeStorage((KV_CAP+1)*HD));
|
| 520 |
+
}
|
| 521 |
+
const prevBuf=device.createBuffer({ size:12*4, usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 522 |
+
// debug-only: map-read copy of the temporal out (xBuf). Headless verification.
|
| 523 |
+
const dbgOutBuf=device.createBuffer({ size:D_MODEL*4, usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST });
|
| 524 |
+
|
| 525 |
+
const makeUniform=bytes=>device.createBuffer({ size:bytes, usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST });
|
| 526 |
+
const pipeline=code=>{ const m=device.createShaderModule({code}); return device.createComputePipeline({ layout:"auto", compute:{module:m,entryPoint:"main"} }); };
|
| 527 |
+
|
| 528 |
+
// ---------------- pipelines ----------------
|
| 529 |
+
const pEmbed=pipeline(EMBED_WGSL);
|
| 530 |
+
const pSelfKV=pipeline(buildKV({K:D_MODEL,writeQ:true,rmsnorm:true}));
|
| 531 |
+
const pCrossQ=pipeline(buildQonly());
|
| 532 |
+
const pCrossKV=pipeline(buildKV({K:D_SRC,writeQ:false,rmsnorm:false}));
|
| 533 |
+
const pAttn=pipeline(ATTN_SINK_WGSL);
|
| 534 |
+
const pWo=pipeline(qMatmul({K:HD,inBinding:1,outBinding:2,rmsnorm:false,epilogue:"none",hasBias:false}));
|
| 535 |
+
const pResid=pipeline(RESID_WGSL);
|
| 536 |
+
const pWi=pipeline(qMatmul({K:D_MODEL,inBinding:1,outBinding:2,rmsnorm:true,epilogue:"geluBias",hasBias:true}));
|
| 537 |
+
const pWoFfn=pipeline(qMatmul({K:D_FF,inBinding:1,outBinding:2,rmsnorm:false,epilogue:"none",hasBias:true}));
|
| 538 |
+
|
| 539 |
+
// ---------------- uniform buffers (per layer) ----------------
|
| 540 |
+
// qMatmul U { qOff,sOff,colBase,normOff,biasOff } = 20 -> pad 32.
|
| 541 |
+
const uEmbed=makeUniform(16);
|
| 542 |
+
const uSelfKV=[],uCrossQ=[],uCrossKV=[],uAttnSelf=[],uAttnCross=[],uWoSelf=[],uWoCross=[],
|
| 543 |
+
uResidSAttn=[],uResidCAttn=[],uResidFfn=[],uWi=[],uWoFfn=[];
|
| 544 |
+
for(let li=0;li<NUM_LAYERS;li++){
|
| 545 |
+
uSelfKV.push(makeUniform(32)); uCrossQ.push(makeUniform(32)); uCrossKV.push(makeUniform(32));
|
| 546 |
+
uAttnSelf.push(makeUniform(32)); uAttnCross.push(makeUniform(32)); // {per_dim_scale,sink_k,sink_v,cacheLen,start} (5 u32 -> pad 32)
|
| 547 |
+
uWoSelf.push(makeUniform(32)); uWoCross.push(makeUniform(32));
|
| 548 |
+
uResidSAttn.push(makeUniform(16)); uResidCAttn.push(makeUniform(16)); uResidFfn.push(makeUniform(16));
|
| 549 |
+
uWi.push(makeUniform(32)); uWoFfn.push(makeUniform(32));
|
| 550 |
+
}
|
| 551 |
+
// helper to fill a 5-u32 qMatmul uniform { qOff, sOff, colBase, normOff, biasOff }
|
| 552 |
+
function fillMM(buf,qOff,sOff,colBase,normOff,biasOff){
|
| 553 |
+
const a=new Uint32Array(8); a[0]=qOff; a[1]=sOff; a[2]=colBase; a[3]=normOff; a[4]=biasOff;
|
| 554 |
+
device.queue.writeBuffer(buf,0,a);
|
| 555 |
+
}
|
| 556 |
+
// embed uniform
|
| 557 |
+
{ const e=new ArrayBuffer(16); new Uint32Array(e,0,1)[0]=smallOff("embedding"); new Float32Array(e,4,1)[0]=EMBED_SCALE;
|
| 558 |
+
device.queue.writeBuffer(uEmbed,0,e); }
|
| 559 |
+
// per-layer static uniforms (slot/cacheLen filled per frame)
|
| 560 |
+
for(let li=0;li<NUM_LAYERS;li++){
|
| 561 |
+
const saWq=q4(`L${li}.sa.Wq`), saWk=q4(`L${li}.sa.Wk`), saWv=q4(`L${li}.sa.Wv`);
|
| 562 |
+
const caWk=q4(`L${li}.ca.Wk`), caWv=q4(`L${li}.ca.Wv`);
|
| 563 |
+
// self KV: {normOff=sa.pre_norm, qQ,qS, kQ,kS, vQ,vS, slot}
|
| 564 |
+
{ const a=new Uint32Array(8); a[0]=L(li,"sa","pre_norm");
|
| 565 |
+
a[1]=saWq.packed_off; a[2]=saWq.scale_off; a[3]=saWk.packed_off; a[4]=saWk.scale_off;
|
| 566 |
+
a[5]=saWv.packed_off; a[6]=saWv.scale_off; device.queue.writeBuffer(uSelfKV[li],0,a); } // slot @ word 7 per frame
|
| 567 |
+
// cross Q (qMatmul): {qOff=ca.Wq, sOff, colBase=0, normOff=ca.pre_norm, bias=0}
|
| 568 |
+
{ const caWq=q4(`L${li}.ca.Wq`); fillMM(uCrossQ[li], caWq.packed_off, caWq.scale_off, 0, L(li,"ca","pre_norm"), 0); }
|
| 569 |
+
// cross KV: {normOff(unused), q*(unused), kQ,kS, vQ,vS, slot}
|
| 570 |
+
{ const a=new Uint32Array(8); a[0]=0; a[1]=0; a[2]=0;
|
| 571 |
+
a[3]=caWk.packed_off; a[4]=caWk.scale_off; a[5]=caWv.packed_off; a[6]=caWv.scale_off;
|
| 572 |
+
device.queue.writeBuffer(uCrossKV[li],0,a); } // slot @ word 7 per frame
|
| 573 |
+
{ const a=new Uint32Array(4); a[0]=L(li,"sa","per_dim_scale"); a[1]=L(li,"sa","sink_k"); a[2]=L(li,"sa","sink_v");
|
| 574 |
+
device.queue.writeBuffer(uAttnSelf[li],0,a); }
|
| 575 |
+
{ const a=new Uint32Array(4); a[0]=L(li,"ca","per_dim_scale"); a[1]=L(li,"ca","sink_k"); a[2]=L(li,"ca","sink_v");
|
| 576 |
+
device.queue.writeBuffer(uAttnCross[li],0,a); }
|
| 577 |
+
{ const saWo=q4(`L${li}.sa.Wo`); fillMM(uWoSelf[li], saWo.packed_off, saWo.scale_off, 0, 0, 0); }
|
| 578 |
+
{ const caWo=q4(`L${li}.ca.Wo`); fillMM(uWoCross[li], caWo.packed_off, caWo.scale_off, 0, 0, 0); }
|
| 579 |
+
{ const a=new Uint32Array(4); a[0]=L(li,"sa","post_norm"); device.queue.writeBuffer(uResidSAttn[li],0,a); }
|
| 580 |
+
{ const a=new Uint32Array(4); a[0]=L(li,"ca","post_norm"); device.queue.writeBuffer(uResidCAttn[li],0,a); }
|
| 581 |
+
{ const a=new Uint32Array(4); a[0]=L(li,"ffn","post_norm"); device.queue.writeBuffer(uResidFfn[li],0,a); }
|
| 582 |
+
{ const wi=q4(`L${li}.ffn.wi`); fillMM(uWi[li], wi.packed_off, wi.scale_off, 0, L(li,"ffn","pre_norm"), L(li,"ffn","bi")); }
|
| 583 |
+
{ const wo=q4(`L${li}.ffn.wo`); fillMM(uWoFfn[li], wo.packed_off, wo.scale_off, 0, 0, L(li,"ffn","bo")); }
|
| 584 |
+
}
|
| 585 |
+
|
| 586 |
+
// ---------------- bind groups ----------------
|
| 587 |
+
// q4 binding scheme: P0@0, P1@6, SC@7, SMALL@8, in@1, out@2, uniform@5.
|
| 588 |
+
const bgEmbed=device.createBindGroup({ layout:pEmbed.getBindGroupLayout(0), entries:[
|
| 589 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 590 |
+
{binding:1,resource:{buffer:prevBuf}},{binding:2,resource:{buffer:xBuf}},{binding:5,resource:{buffer:uEmbed}}] });
|
| 591 |
+
const bgSelfKV=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pSelfKV.getBindGroupLayout(0), entries:[
|
| 592 |
+
{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},{binding:8,resource:{buffer:smBuf}},
|
| 593 |
+
{binding:1,resource:{buffer:xBuf}},{binding:2,resource:{buffer:qBuf}},
|
| 594 |
+
{binding:3,resource:{buffer:selfK[li]}},{binding:4,resource:{buffer:selfV[li]}},{binding:5,resource:{buffer:uSelfKV[li]}}] }));
|
| 595 |
+
const bgCrossQ=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pCrossQ.getBindGroupLayout(0), entries:[
|
| 596 |
+
{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},{binding:8,resource:{buffer:smBuf}},
|
| 597 |
+
{binding:1,resource:{buffer:xBuf}},{binding:2,resource:{buffer:qBuf}},{binding:5,resource:{buffer:uCrossQ[li]}}] }));
|
| 598 |
+
const bgCrossKV=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pCrossKV.getBindGroupLayout(0), entries:[
|
| 599 |
+
// pCrossKV (rmsnorm=false) never calls wv() -> SMALL@8 pruned by layout:auto.
|
| 600 |
+
{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},
|
| 601 |
+
{binding:1,resource:{buffer:srcBuf}},
|
| 602 |
+
{binding:3,resource:{buffer:crossK[li]}},{binding:4,resource:{buffer:crossV[li]}},{binding:5,resource:{buffer:uCrossKV[li]}}] }));
|
| 603 |
+
const bgAttnSelf=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pAttn.getBindGroupLayout(0), entries:[
|
| 604 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 605 |
+
{binding:1,resource:{buffer:qBuf}},{binding:2,resource:{buffer:selfK[li]}},
|
| 606 |
+
{binding:3,resource:{buffer:selfV[li]}},{binding:4,resource:{buffer:ctxBuf}},{binding:5,resource:{buffer:uAttnSelf[li]}}] }));
|
| 607 |
+
const bgAttnCross=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pAttn.getBindGroupLayout(0), entries:[
|
| 608 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 609 |
+
{binding:1,resource:{buffer:qBuf}},{binding:2,resource:{buffer:crossK[li]}},
|
| 610 |
+
{binding:3,resource:{buffer:crossV[li]}},{binding:4,resource:{buffer:ctxBuf}},{binding:5,resource:{buffer:uAttnCross[li]}}] }));
|
| 611 |
+
const bgWoSelf=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pWo.getBindGroupLayout(0), entries:[
|
| 612 |
+
// Wo qMatmul: no rmsnorm/bias -> wv() never called -> SMALL@8 pruned.
|
| 613 |
+
{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},
|
| 614 |
+
{binding:1,resource:{buffer:ctxBuf}},{binding:2,resource:{buffer:blkOutBuf}},{binding:5,resource:{buffer:uWoSelf[li]}}] }));
|
| 615 |
+
const bgWoCross=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pWo.getBindGroupLayout(0), entries:[
|
| 616 |
+
{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},
|
| 617 |
+
{binding:1,resource:{buffer:ctxBuf}},{binding:2,resource:{buffer:blkOutBuf}},{binding:5,resource:{buffer:uWoCross[li]}}] }));
|
| 618 |
+
const bgResidSAttn=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pResid.getBindGroupLayout(0), entries:[
|
| 619 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 620 |
+
{binding:1,resource:{buffer:xBuf}},{binding:2,resource:{buffer:blkOutBuf}},{binding:5,resource:{buffer:uResidSAttn[li]}}] }));
|
| 621 |
+
const bgResidCAttn=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pResid.getBindGroupLayout(0), entries:[
|
| 622 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 623 |
+
{binding:1,resource:{buffer:xBuf}},{binding:2,resource:{buffer:blkOutBuf}},{binding:5,resource:{buffer:uResidCAttn[li]}}] }));
|
| 624 |
+
const bgWi=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pWi.getBindGroupLayout(0), entries:[
|
| 625 |
+
{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},{binding:8,resource:{buffer:smBuf}},
|
| 626 |
+
{binding:1,resource:{buffer:xBuf}},{binding:2,resource:{buffer:ffBuf}},{binding:5,resource:{buffer:uWi[li]}}] }));
|
| 627 |
+
const bgWoFfn=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pWoFfn.getBindGroupLayout(0), entries:[
|
| 628 |
+
{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},{binding:8,resource:{buffer:smBuf}},
|
| 629 |
+
{binding:1,resource:{buffer:ffBuf}},{binding:2,resource:{buffer:blkOutBuf}},{binding:5,resource:{buffer:uWoFfn[li]}}] }));
|
| 630 |
+
const bgResidFfn=Array.from({length:NUM_LAYERS},(_,li)=>device.createBindGroup({ layout:pResid.getBindGroupLayout(0), entries:[
|
| 631 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 632 |
+
{binding:1,resource:{buffer:xBuf}},{binding:2,resource:{buffer:blkOutBuf}},{binding:5,resource:{buffer:uResidFfn[li]}}] }));
|
| 633 |
+
|
| 634 |
+
const WG_KV=tNWGT(HD), WG_Q=tNWGT(HD), WG_WO=tNWGT(D_MODEL), WG_WI=tNWGT(D_FF), WG_WOFFN=tNWGT(D_MODEL);
|
| 635 |
+
|
| 636 |
+
// ---------------- KV growth state (persists across frames) ----------------
|
| 637 |
+
// T = number of accumulated real KV rows. Next frame appends at slot=T, then
|
| 638 |
+
// (after this frame) the new length is min(T+1, KV_CAP). The temporal_gpu.ts
|
| 639 |
+
// reference grows by writing the appended row at slot T (no eviction below cap;
|
| 640 |
+
// at cap the model trims oldest — here we cap T at KV_CAP, appending at slot T
|
| 641 |
+
// until T==KV_CAP, then keep appending at the last slot like a ring's head).
|
| 642 |
+
// To match the ORT temporal_step semantics (sliding window, trim oldest) the
|
| 643 |
+
// KV is treated as append-then-clip-to-cap. We implement the simple grow-to-cap
|
| 644 |
+
// policy used by the validated reference (slot = min(T, KV_CAP)).
|
| 645 |
+
let T_LEN=0; // current accumulated real KV rows (before this frame's append)
|
| 646 |
+
|
| 647 |
+
function resetKV(){ T_LEN=0; }
|
| 648 |
+
|
| 649 |
+
// upload `source` (256 floats) once per style change.
|
| 650 |
+
function setSource(srcF32){ device.queue.writeBuffer(srcBuf,0, srcF32 instanceof Float32Array?srcF32:Float32Array.from(srcF32)); }
|
| 651 |
+
|
| 652 |
+
// runTemporal: prevToks[12] (global token ids) -> writes out into xBuf[1024]
|
| 653 |
+
// (GPU-resident) and appends this frame's self/cross K,V at slot T_LEN. Returns
|
| 654 |
+
// the xBuf so the depth adapter can read it directly (zero copy). No readback.
|
| 655 |
+
function runTemporal(prevToks){
|
| 656 |
+
const slot=T_LEN % KV_CAP; // RING buffer write: overwrite oldest when full
|
| 657 |
+
const cacheLen=Math.min(T_LEN+1, KV_CAP);// recent-window length; ≤KV_CAP so sink+cacheLen fits logitsW[KV_CAP+1]
|
| 658 |
+
// start = physical index of the OLDEST key currently in the ring (read in temporal order oldest->newest).
|
| 659 |
+
const start=(T_LEN+1 <= KV_CAP) ? 0 : (T_LEN+1) % KV_CAP;
|
| 660 |
+
|
| 661 |
+
device.queue.writeBuffer(prevBuf,0,new Uint32Array(prevToks));
|
| 662 |
+
for(let li=0;li<NUM_LAYERS;li++){
|
| 663 |
+
// slot lives at word 7 (byte 28) of the 32-byte q4 selfKV/crossKV uniforms.
|
| 664 |
+
{ const a=new Uint32Array(1); a[0]=slot; device.queue.writeBuffer(uSelfKV[li],28,a); device.queue.writeBuffer(uCrossKV[li],28,a); }
|
| 665 |
+
// cacheLen (word 3 / byte 12) + start (word 4 / byte 16) of the attn uniforms.
|
| 666 |
+
{ const a=new Uint32Array(2); a[0]=cacheLen; a[1]=start; device.queue.writeBuffer(uAttnSelf[li],12,a); device.queue.writeBuffer(uAttnCross[li],12,a); }
|
| 667 |
+
}
|
| 668 |
+
|
| 669 |
+
const enc=device.createCommandEncoder();
|
| 670 |
+
const pass=enc.beginComputePass();
|
| 671 |
+
pass.setPipeline(pEmbed); pass.setBindGroup(0,bgEmbed); pass.dispatchWorkgroups(1);
|
| 672 |
+
for(let li=0;li<NUM_LAYERS;li++){
|
| 673 |
+
// self attention
|
| 674 |
+
pass.setPipeline(pSelfKV); pass.setBindGroup(0,bgSelfKV[li]); pass.dispatchWorkgroups(WG_KV);
|
| 675 |
+
pass.setPipeline(pAttn); pass.setBindGroup(0,bgAttnSelf[li]); pass.dispatchWorkgroups(1);
|
| 676 |
+
pass.setPipeline(pWo); pass.setBindGroup(0,bgWoSelf[li]); pass.dispatchWorkgroups(WG_WO);
|
| 677 |
+
pass.setPipeline(pResid); pass.setBindGroup(0,bgResidSAttn[li]);pass.dispatchWorkgroups(1);
|
| 678 |
+
// cross attention
|
| 679 |
+
pass.setPipeline(pCrossQ); pass.setBindGroup(0,bgCrossQ[li]); pass.dispatchWorkgroups(WG_Q);
|
| 680 |
+
pass.setPipeline(pCrossKV); pass.setBindGroup(0,bgCrossKV[li]); pass.dispatchWorkgroups(WG_KV);
|
| 681 |
+
pass.setPipeline(pAttn); pass.setBindGroup(0,bgAttnCross[li]); pass.dispatchWorkgroups(1);
|
| 682 |
+
pass.setPipeline(pWo); pass.setBindGroup(0,bgWoCross[li]); pass.dispatchWorkgroups(WG_WO);
|
| 683 |
+
pass.setPipeline(pResid); pass.setBindGroup(0,bgResidCAttn[li]);pass.dispatchWorkgroups(1);
|
| 684 |
+
// FFN
|
| 685 |
+
pass.setPipeline(pWi); pass.setBindGroup(0,bgWi[li]); pass.dispatchWorkgroups(WG_WI);
|
| 686 |
+
pass.setPipeline(pWoFfn); pass.setBindGroup(0,bgWoFfn[li]); pass.dispatchWorkgroups(WG_WOFFN);
|
| 687 |
+
pass.setPipeline(pResid); pass.setBindGroup(0,bgResidFfn[li]); pass.dispatchWorkgroups(1);
|
| 688 |
+
}
|
| 689 |
+
pass.end();
|
| 690 |
+
device.queue.submit([enc.finish()]);
|
| 691 |
+
|
| 692 |
+
// advance KV length — grow indefinitely; the ring wraps at KV_CAP (slot=T_LEN%KV_CAP)
|
| 693 |
+
T_LEN++;
|
| 694 |
+
return xBuf;
|
| 695 |
+
}
|
| 696 |
+
|
| 697 |
+
// debug-only: read back the current temporal out (xBuf[1024]) to CPU.
|
| 698 |
+
async function debugReadOut(){
|
| 699 |
+
const enc=device.createCommandEncoder();
|
| 700 |
+
enc.copyBufferToBuffer(xBuf,0,dbgOutBuf,0,D_MODEL*4);
|
| 701 |
+
device.queue.submit([enc.finish()]);
|
| 702 |
+
await dbgOutBuf.mapAsync(GPUMapMode.READ);
|
| 703 |
+
const out=new Float32Array(dbgOutBuf.getMappedRange().slice(0));
|
| 704 |
+
dbgOutBuf.unmap();
|
| 705 |
+
let absmax=0, hasNaN=false, sum=0, sumsq=0;
|
| 706 |
+
for(let i=0;i<out.length;i++){ const v=out[i]; if(Number.isNaN(v)||!Number.isFinite(v))hasNaN=true; const a=Math.abs(v); if(a>absmax)absmax=a; sum+=v; sumsq+=v*v; }
|
| 707 |
+
const mean=sum/out.length, std=Math.sqrt(Math.max(0,sumsq/out.length-mean*mean));
|
| 708 |
+
return { absmax, hasNaN, mean, std, first4:[out[0],out[1],out[2],out[3]] };
|
| 709 |
+
}
|
| 710 |
+
return { runTemporal, resetKV, setSource, kvLen:()=>T_LEN, outBuf:xBuf, debugReadOut };
|
| 711 |
+
}
|
| 712 |
+
|
| 713 |
+
/* ====================================================================== *
|
| 714 |
+
* CUSTOM WGSL DEPTH (ported verbatim from depth_q4_gpu.ts — q4)
|
| 715 |
+
* Big matmuls (adapter, Wq/Wk/Wv, Wo, Wi, Wo_ffn, to_logits) are q4 block-
|
| 716 |
+
* quantized along K (block_size=32, zero=8): packed nibbles in u32 (8/word,
|
| 717 |
+
* k-contiguous) + per-block fp32 scales; unpack via u32 shift/mask only (NO
|
| 718 |
+
* unpack2x16float). Small params (norms/biases/per_dim_scale/embedding) stay
|
| 719 |
+
* fp32 in SMALL. Built on the SAME GPUDevice ORT uses so the temporal-out
|
| 720 |
+
* GPUBuffer and the depth kernels share GPU memory: zero copy, zero readback
|
| 721 |
+
* except the 12 sampled tokens once per frame.
|
| 722 |
+
* ====================================================================== */
|
| 723 |
+
const D_IN=1024, D_MODEL=768, D_FF=3072, VOCAB=12294, NUM_CODEBOOKS=12,
|
| 724 |
+
CODEBOOK_SIZE=1024, NUM_RESERVED=6, NUM_LAYERS=2, NUM_HEADS=6, UPH=128;
|
| 725 |
+
const HD=NUM_HEADS*UPH, QKV=3*HD, SOFT_CAP=30.0, EPS=1e-6, R_SOFTPLUS_0=1.442695041;
|
| 726 |
+
const D_BLK=32; // q4 block size
|
| 727 |
+
const RED=32, ROWS=8, T_WG=ROWS*RED; // bandwidth tiling
|
| 728 |
+
const nWGT=N=>Math.ceil(N/ROWS);
|
| 729 |
+
let USE_SUBGROUPS=false;
|
| 730 |
+
|
| 731 |
+
const CONSTS=`
|
| 732 |
+
const D_IN : u32 = ${D_IN}u;
|
| 733 |
+
const D_MODEL : u32 = ${D_MODEL}u;
|
| 734 |
+
const D_FF : u32 = ${D_FF}u;
|
| 735 |
+
const VOCAB : u32 = ${VOCAB}u;
|
| 736 |
+
const NUM_HEADS : u32 = ${NUM_HEADS}u;
|
| 737 |
+
const UPH : u32 = ${UPH}u;
|
| 738 |
+
const HD : u32 = ${HD}u;
|
| 739 |
+
const QKV : u32 = ${QKV}u;
|
| 740 |
+
const SOFT_CAP : f32 = ${SOFT_CAP};
|
| 741 |
+
const EPS : f32 = ${EPS};
|
| 742 |
+
const R_SOFTPLUS_0 : f32 = ${R_SOFTPLUS_0};
|
| 743 |
+
const NUM_RESERVED : u32 = ${NUM_RESERVED}u;
|
| 744 |
+
const CODEBOOK_SIZE : u32 = ${CODEBOOK_SIZE}u;
|
| 745 |
+
const RED : u32 = ${RED}u;
|
| 746 |
+
const ROWS : u32 = ${ROWS}u;
|
| 747 |
+
const BLK : u32 = ${D_BLK}u;
|
| 748 |
+
`;
|
| 749 |
+
// q4 HEADER: SMALL fp32 params @8 (wv) + PACKED P0/P1 @0/6 + SCALES @7 (deq).
|
| 750 |
+
function buildHeaders(P_SPLIT_WORDS){
|
| 751 |
+
const SMALLBUF=`
|
| 752 |
+
@group(0) @binding(8) var<storage, read> SMALL : array<f32>;
|
| 753 |
+
fn wv(i : u32) -> f32 { return SMALL[i]; }
|
| 754 |
+
fn softplus(x: f32) -> f32 { return log(1.0 + exp(-abs(x))) + max(x, 0.0); }
|
| 755 |
+
fn gelu_tanh(x: f32) -> f32 {
|
| 756 |
+
let c : f32 = 0.7978845608028654;
|
| 757 |
+
// clamp: Metal fast-math tanh overflows to 0/NaN for |arg|>~44.
|
| 758 |
+
// |t|>=12 already rounds tanh(t) to +/-1 in f32, so this is exact.
|
| 759 |
+
let t = clamp(c * (x + 0.044715 * x * x * x), -12.0, 12.0);
|
| 760 |
+
return 0.5 * x * (1.0 + tanh(t));
|
| 761 |
+
}`;
|
| 762 |
+
const QBUF=`
|
| 763 |
+
@group(0) @binding(0) var<storage, read> P0 : array<u32>;
|
| 764 |
+
@group(0) @binding(6) var<storage, read> P1 : array<u32>;
|
| 765 |
+
@group(0) @binding(7) var<storage, read> SC : array<f32>;
|
| 766 |
+
const P_SPLIT_WORDS : u32 = ${P_SPLIT_WORDS}u;
|
| 767 |
+
fn qword(wi : u32) -> u32 {
|
| 768 |
+
if (wi < P_SPLIT_WORDS) { return P0[wi]; }
|
| 769 |
+
return P1[wi - P_SPLIT_WORDS];
|
| 770 |
+
}
|
| 771 |
+
fn deq(qOff : u32, sOff : u32, K : u32, col : u32, k : u32) -> f32 {
|
| 772 |
+
let word = qword(qOff + col * (K >> 3u) + (k >> 3u));
|
| 773 |
+
let nib = (word >> ((k & 7u) * 4u)) & 0xfu;
|
| 774 |
+
let s = SC[sOff + col * (K >> 5u) + (k >> 5u)];
|
| 775 |
+
return (f32(nib) - 8.0) * s;
|
| 776 |
+
}`;
|
| 777 |
+
return CONSTS+QBUF+SMALLBUF;
|
| 778 |
+
}
|
| 779 |
+
// q4 bandwidth matmul (replaces gridMatmulT/gridMatmul2). uniform {qOff,sOff,colBase,normOff,biasOff}.
|
| 780 |
+
function qMatmul(HEADER, opts){
|
| 781 |
+
const K=opts.K, sg=USE_SUBGROUPS;
|
| 782 |
+
const inDecl=`@group(0) @binding(${opts.inBinding}) var<storage, read> vin : array<f32>;`;
|
| 783 |
+
const outDecl=`@group(0) @binding(${opts.outBinding}) var<storage, read_write> vout : array<f32>;`;
|
| 784 |
+
const uniformDecl=`struct U { qOff:u32, sOff:u32, colBase:u32, normOff:u32, biasOff:u32, };
|
| 785 |
+
@group(0) @binding(5) var<uniform> u : U;`;
|
| 786 |
+
const normReduce=opts.rmsnorm?`
|
| 787 |
+
var lss : f32 = 0.0;
|
| 788 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) { let xv = vin[i]; lss = lss + xv * xv; }
|
| 789 |
+
red[tid] = lss; workgroupBarrier();
|
| 790 |
+
for (var s : u32 = ${T_WG/2}u; s > 0u; s = s >> 1u) { if (tid < s) { red[tid] = red[tid] + red[tid + s]; } workgroupBarrier(); }
|
| 791 |
+
let rms = 1.0 / sqrt(red[0] / f32(${K}u) + EPS); workgroupBarrier();
|
| 792 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) { inS[i] = (vin[i] * rms) * wv(u.normOff + i); }
|
| 793 |
+
workgroupBarrier();`:`
|
| 794 |
+
for (var i : u32 = tid; i < ${K}u; i = i + ${T_WG}u) { inS[i] = vin[i]; }
|
| 795 |
+
workgroupBarrier();`;
|
| 796 |
+
let epi="";
|
| 797 |
+
if(opts.epilogue==="geluBias") epi=`acc = gelu_tanh(acc + wv(u.biasOff + col));`;
|
| 798 |
+
else if(opts.epilogue==="softcap"){ const b=opts.hasBias?`acc = acc + wv(u.biasOff + col);`:``; epi=`${b}
|
| 799 |
+
acc = SOFT_CAP * tanh(clamp(acc / SOFT_CAP, -12.0, 12.0));`; }
|
| 800 |
+
else epi=opts.hasBias?`acc = acc + wv(u.biasOff + col);`:``;
|
| 801 |
+
const reduceWrite=sg?`
|
| 802 |
+
let total = subgroupAdd(acc);
|
| 803 |
+
if (lane == 0u) { var accf : f32 = total; ${epi.replace(/\bacc\b/g,"accf")} vout[outIdx] = accf; }`:`
|
| 804 |
+
part[tid] = acc; workgroupBarrier();
|
| 805 |
+
for (var s : u32 = ${RED/2}u; s > 0u; s = s >> 1u) { if (lane < s) { part[tid] = part[tid] + part[tid + s]; } workgroupBarrier(); }
|
| 806 |
+
if (lane == 0u) { var accf : f32 = part[row * RED]; ${epi.replace(/\bacc\b/g,"accf")} vout[outIdx] = accf; }`;
|
| 807 |
+
const enableSg=sg?"enable subgroups;\n":"";
|
| 808 |
+
return enableSg+HEADER+`
|
| 809 |
+
${inDecl}
|
| 810 |
+
${outDecl}
|
| 811 |
+
${uniformDecl}
|
| 812 |
+
var<workgroup> inS : array<f32, ${K}>;
|
| 813 |
+
${opts.rmsnorm?`var<workgroup> red : array<f32, ${T_WG}>;`:``}
|
| 814 |
+
${sg?"":`var<workgroup> part : array<f32, ${T_WG}>;`}
|
| 815 |
+
@compute @workgroup_size(${T_WG})
|
| 816 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>, @builtin(workgroup_id) wid : vec3<u32>) {
|
| 817 |
+
let tid = lid.x; let row = tid / RED; let lane = tid % RED;
|
| 818 |
+
${normReduce}
|
| 819 |
+
let outIdx = wid.x * ROWS + row;
|
| 820 |
+
let col = u.colBase + outIdx;
|
| 821 |
+
var acc : f32 = 0.0;
|
| 822 |
+
for (var k : u32 = lane; k < ${K}u; k = k + RED) { acc = acc + inS[k] * deq(u.qOff, u.sOff, ${K}u, col, k); }
|
| 823 |
+
${reduceWrite}
|
| 824 |
+
}`;
|
| 825 |
+
}
|
| 826 |
+
// q4 combined QKV (folds attn-pre RMSNorm). uniform {normOff,qQ,qS,kQ,kS,vQ,vS,slot}.
|
| 827 |
+
function buildQKV(HEADER){
|
| 828 |
+
const sg=USE_SUBGROUPS, enableSg=sg?"enable subgroups;\n":"";
|
| 829 |
+
const reduceWrite=sg?`
|
| 830 |
+
let sq = subgroupAdd(qa); let sk = subgroupAdd(ka); let sv = subgroupAdd(va);
|
| 831 |
+
if (lane == 0u) { qbuf[oi] = sq; let cbase = u.slot * HD; kcache[cbase + oi] = sk; vcache[cbase + oi] = sv; }`:`
|
| 832 |
+
pq[tid] = qa; pk[tid] = ka; pv[tid] = va; workgroupBarrier();
|
| 833 |
+
for (var s : u32 = ${RED/2}u; s > 0u; s = s >> 1u) {
|
| 834 |
+
if (lane < s) { pq[tid] = pq[tid] + pq[tid + s]; pk[tid] = pk[tid] + pk[tid + s]; pv[tid] = pv[tid] + pv[tid + s]; }
|
| 835 |
+
workgroupBarrier(); }
|
| 836 |
+
if (lane == 0u) { qbuf[oi] = pq[row * RED]; let cbase = u.slot * HD; kcache[cbase + oi] = pk[row * RED]; vcache[cbase + oi] = pv[row * RED]; }`;
|
| 837 |
+
return enableSg+HEADER+`
|
| 838 |
+
@group(0) @binding(1) var<storage, read> vin : array<f32>;
|
| 839 |
+
@group(0) @binding(2) var<storage, read_write> qbuf : array<f32>;
|
| 840 |
+
@group(0) @binding(3) var<storage, read_write> kcache : array<f32>;
|
| 841 |
+
@group(0) @binding(4) var<storage, read_write> vcache : array<f32>;
|
| 842 |
+
struct U { normOff:u32, qQ:u32, qS:u32, kQ:u32, kS:u32, vQ:u32, vS:u32, slot:u32, };
|
| 843 |
+
@group(0) @binding(5) var<uniform> u : U;
|
| 844 |
+
var<workgroup> inS : array<f32, ${D_MODEL}>;
|
| 845 |
+
var<workgroup> red : array<f32, ${T_WG}>;
|
| 846 |
+
${sg?"":`var<workgroup> pq : array<f32, ${T_WG}>;
|
| 847 |
+
var<workgroup> pk : array<f32, ${T_WG}>;
|
| 848 |
+
var<workgroup> pv : array<f32, ${T_WG}>;`}
|
| 849 |
+
@compute @workgroup_size(${T_WG})
|
| 850 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>, @builtin(workgroup_id) wid : vec3<u32>) {
|
| 851 |
+
let tid = lid.x; let row = tid / RED; let lane = tid % RED;
|
| 852 |
+
var lss : f32 = 0.0;
|
| 853 |
+
for (var i : u32 = tid; i < D_MODEL; i = i + ${T_WG}u) { let xv = vin[i]; lss = lss + xv * xv; }
|
| 854 |
+
red[tid] = lss; workgroupBarrier();
|
| 855 |
+
for (var s : u32 = ${T_WG/2}u; s > 0u; s = s >> 1u) { if (tid < s) { red[tid] = red[tid] + red[tid + s]; } workgroupBarrier(); }
|
| 856 |
+
let rms = 1.0 / sqrt(red[0] / f32(D_MODEL) + EPS); workgroupBarrier();
|
| 857 |
+
for (var i : u32 = tid; i < D_MODEL; i = i + ${T_WG}u) { inS[i] = (vin[i] * rms) * wv(u.normOff + i); }
|
| 858 |
+
workgroupBarrier();
|
| 859 |
+
let oi = wid.x * ROWS + row;
|
| 860 |
+
var qa : f32 = 0.0; var ka : f32 = 0.0; var va : f32 = 0.0;
|
| 861 |
+
for (var k : u32 = lane; k < D_MODEL; k = k + RED) {
|
| 862 |
+
let hv = inS[k];
|
| 863 |
+
qa = qa + hv * deq(u.qQ, u.qS, D_MODEL, oi, k);
|
| 864 |
+
ka = ka + hv * deq(u.kQ, u.kS, D_MODEL, oi, k);
|
| 865 |
+
va = va + hv * deq(u.vQ, u.vS, D_MODEL, oi, k);
|
| 866 |
+
}
|
| 867 |
+
${reduceWrite}
|
| 868 |
+
}`;
|
| 869 |
+
}
|
| 870 |
+
function buildAttnCore(HEADER){ return HEADER+`
|
| 871 |
+
@group(0) @binding(1) var<storage, read> qbuf : array<f32>;
|
| 872 |
+
@group(0) @binding(2) var<storage, read> kcache : array<f32>;
|
| 873 |
+
@group(0) @binding(3) var<storage, read> vcache : array<f32>;
|
| 874 |
+
@group(0) @binding(4) var<storage, read_write> ctxbuf : array<f32>;
|
| 875 |
+
struct U { per_dim_scale:u32, cacheLen:u32, };
|
| 876 |
+
@group(0) @binding(5) var<uniform> u : U;
|
| 877 |
+
var<workgroup> logitsW : array<f32, ${NUM_HEADS*NUM_CODEBOOKS}>;
|
| 878 |
+
var<workgroup> red : array<f32, 2>;
|
| 879 |
+
@compute @workgroup_size(256)
|
| 880 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>) {
|
| 881 |
+
let t = lid.x; let T = u.cacheLen;
|
| 882 |
+
let qscale = R_SOFTPLUS_0 * (1.0 / sqrt(f32(UPH)));
|
| 883 |
+
for (var h : u32 = 0u; h < NUM_HEADS; h = h + 1u) {
|
| 884 |
+
for (var k : u32 = t; k < T; k = k + 256u) {
|
| 885 |
+
var acc : f32 = 0.0; let kbase = k * HD + h * UPH; let qbase = h * UPH;
|
| 886 |
+
for (var d : u32 = 0u; d < UPH; d = d + 1u) {
|
| 887 |
+
let sv = qscale * softplus(wv(u.per_dim_scale + d));
|
| 888 |
+
acc = acc + (qbuf[qbase + d] * sv) * kcache[kbase + d];
|
| 889 |
+
}
|
| 890 |
+
logitsW[h * T + k] = acc;
|
| 891 |
+
}
|
| 892 |
+
}
|
| 893 |
+
workgroupBarrier();
|
| 894 |
+
for (var h : u32 = 0u; h < NUM_HEADS; h = h + 1u) {
|
| 895 |
+
if (t == 0u) { var m : f32 = logitsW[h * T + 0u]; for (var k : u32 = 1u; k < T; k = k + 1u) { m = max(m, logitsW[h * T + k]); } red[0] = m; }
|
| 896 |
+
workgroupBarrier(); let m = red[0];
|
| 897 |
+
if (t == 0u) { var s : f32 = 0.0; for (var k : u32 = 0u; k < T; k = k + 1u) { let e = exp(logitsW[h * T + k] - m); logitsW[h * T + k] = e; s = s + e; } red[1] = s; }
|
| 898 |
+
workgroupBarrier(); let denom = red[1];
|
| 899 |
+
for (var d : u32 = t; d < UPH; d = d + 256u) {
|
| 900 |
+
var acc : f32 = 0.0;
|
| 901 |
+
for (var k : u32 = 0u; k < T; k = k + 1u) { let w = logitsW[h * T + k] / denom; acc = acc + w * vcache[k * HD + h * UPH + d]; }
|
| 902 |
+
ctxbuf[h * UPH + d] = acc;
|
| 903 |
+
}
|
| 904 |
+
workgroupBarrier();
|
| 905 |
+
}
|
| 906 |
+
}`;
|
| 907 |
+
}
|
| 908 |
+
function buildResid(HEADER){ return HEADER+`
|
| 909 |
+
@group(0) @binding(1) var<storage, read_write> x : array<f32>;
|
| 910 |
+
@group(0) @binding(2) var<storage, read> y : array<f32>;
|
| 911 |
+
struct U { post:u32, };
|
| 912 |
+
@group(0) @binding(5) var<uniform> u : U;
|
| 913 |
+
var<workgroup> red : array<f32, 256>;
|
| 914 |
+
@compute @workgroup_size(256)
|
| 915 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>) {
|
| 916 |
+
let t = lid.x; var ss : f32 = 0.0;
|
| 917 |
+
for (var i : u32 = t; i < D_MODEL; i = i + 256u) { let v = y[i]; ss = ss + v*v; }
|
| 918 |
+
red[t] = ss; workgroupBarrier();
|
| 919 |
+
for (var s : u32 = 128u; s > 0u; s = s >> 1u) { if (t < s) { red[t] = red[t] + red[t + s]; } workgroupBarrier(); }
|
| 920 |
+
let rms = 1.0 / sqrt(red[0] / f32(D_MODEL) + EPS); workgroupBarrier();
|
| 921 |
+
for (var i : u32 = t; i < D_MODEL; i = i + 256u) { x[i] = x[i] + (y[i] * rms) * wv(u.post + i); }
|
| 922 |
+
}`;
|
| 923 |
+
}
|
| 924 |
+
function buildFinalLN(HEADER){ return HEADER+`
|
| 925 |
+
@group(0) @binding(1) var<storage, read> x : array<f32>;
|
| 926 |
+
@group(0) @binding(2) var<storage, read_write> xn : array<f32>;
|
| 927 |
+
struct U { final_scale:u32, final_bias:u32, };
|
| 928 |
+
@group(0) @binding(5) var<uniform> u : U;
|
| 929 |
+
var<workgroup> red : array<f32, 256>;
|
| 930 |
+
@compute @workgroup_size(256)
|
| 931 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>) {
|
| 932 |
+
let t = lid.x; var sm : f32 = 0.0;
|
| 933 |
+
for (var i : u32 = t; i < D_MODEL; i = i + 256u) { sm = sm + x[i]; }
|
| 934 |
+
red[t] = sm; workgroupBarrier();
|
| 935 |
+
for (var s : u32 = 128u; s > 0u; s = s >> 1u) { if (t < s) { red[t] = red[t] + red[t + s]; } workgroupBarrier(); }
|
| 936 |
+
let mean = red[0] / f32(D_MODEL); workgroupBarrier();
|
| 937 |
+
var vs : f32 = 0.0;
|
| 938 |
+
for (var i : u32 = t; i < D_MODEL; i = i + 256u) { let dd = x[i] - mean; vs = vs + dd*dd; }
|
| 939 |
+
red[t] = vs; workgroupBarrier();
|
| 940 |
+
for (var s : u32 = 128u; s > 0u; s = s >> 1u) { if (t < s) { red[t] = red[t] + red[t + s]; } workgroupBarrier(); }
|
| 941 |
+
let rstd = 1.0 / sqrt(red[0] / f32(D_MODEL) + EPS); workgroupBarrier();
|
| 942 |
+
for (var i : u32 = t; i < D_MODEL; i = i + 256u) { xn[i] = ((x[i] - mean) * rstd) * wv(u.final_scale + i) + wv(u.final_bias + i); }
|
| 943 |
+
}`;
|
| 944 |
+
}
|
| 945 |
+
// sampler: gumbel over the 1024-wide active slice. noise is COMPACT [Q*1024].
|
| 946 |
+
function buildSample(){ return CONSTS+`
|
| 947 |
+
@group(0) @binding(1) var<storage, read> logits1024 : array<f32>;
|
| 948 |
+
@group(0) @binding(2) var<storage, read> noise : array<f32>;
|
| 949 |
+
@group(0) @binding(3) var<storage, read_write> outTok : array<u32>;
|
| 950 |
+
struct U { lo:u32, temperature:f32, slot:u32, noiseBase:u32, };
|
| 951 |
+
@group(0) @binding(5) var<uniform> u : U;
|
| 952 |
+
var<workgroup> bestScore : array<f32, 256>;
|
| 953 |
+
var<workgroup> bestIdx : array<u32, 256>;
|
| 954 |
+
@compute @workgroup_size(256)
|
| 955 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>) {
|
| 956 |
+
let t = lid.x; let lo = u.lo;
|
| 957 |
+
var bScore : f32 = -3.0e38; var bIdx : u32 = 0xffffffffu;
|
| 958 |
+
for (var j : u32 = t; j < CODEBOOK_SIZE; j = j + 256u) {
|
| 959 |
+
var score : f32 = logits1024[j];
|
| 960 |
+
let gidx = lo + j;
|
| 961 |
+
if (u.temperature > 0.0) {
|
| 962 |
+
var un = noise[u.noiseBase + j];
|
| 963 |
+
un = clamp(un, 1.0e-10, 1.0 - 1.0e-7);
|
| 964 |
+
let g = -log(-log(un));
|
| 965 |
+
score = score + g * u.temperature;
|
| 966 |
+
}
|
| 967 |
+
if (score > bScore || (score == bScore && gidx < bIdx)) { bScore = score; bIdx = gidx; }
|
| 968 |
+
}
|
| 969 |
+
bestScore[t] = bScore; bestIdx[t] = bIdx; workgroupBarrier();
|
| 970 |
+
for (var s : u32 = 128u; s > 0u; s = s >> 1u) {
|
| 971 |
+
if (t < s) { let a = bestScore[t]; let ai = bestIdx[t]; let b = bestScore[t + s]; let bi = bestIdx[t + s];
|
| 972 |
+
if (b > a || (b == a && bi < ai)) { bestScore[t] = b; bestIdx[t] = bi; } }
|
| 973 |
+
workgroupBarrier();
|
| 974 |
+
}
|
| 975 |
+
if (t == 0u) { outTok[u.slot] = bestIdx[0]; }
|
| 976 |
+
}`;
|
| 977 |
+
}
|
| 978 |
+
function buildEmbed(HEADER){ return HEADER+`
|
| 979 |
+
@group(0) @binding(1) var<storage, read> tokBuf : array<u32>;
|
| 980 |
+
@group(0) @binding(2) var<storage, read_write> xout : array<f32>;
|
| 981 |
+
struct EU { embeddingOff : u32, embed_scale : f32, slot : u32 };
|
| 982 |
+
@group(0) @binding(3) var<uniform> u : EU;
|
| 983 |
+
@compute @workgroup_size(256)
|
| 984 |
+
fn main(@builtin(local_invocation_id) lid : vec3<u32>) {
|
| 985 |
+
let t = lid.x; let tok = tokBuf[u.slot];
|
| 986 |
+
let base = u.embeddingOff + tok * D_IN;
|
| 987 |
+
for (var i : u32 = t; i < D_IN; i = i + 256u) { xout[i] = wv(base + i) * u.embed_scale; }
|
| 988 |
+
}`;
|
| 989 |
+
}
|
| 990 |
+
|
| 991 |
+
// Build the whole depth engine on ORT's device. Returns { runDepth }.
|
| 992 |
+
async function buildDepth(device, layout, binBytes){
|
| 993 |
+
const EMBED_SCALE=layout.embed_scale;
|
| 994 |
+
const q4=n=>layout.q4[n];
|
| 995 |
+
const smallOff=n=>layout.small[n].offset;
|
| 996 |
+
const layerOffsets=li=>{ const p="l"+li+"_"; return {
|
| 997 |
+
attn_pre:smallOff(p+"attn_pre"), attn_post:smallOff(p+"attn_post"),
|
| 998 |
+
per_dim_scale:smallOff(p+"per_dim_scale"),
|
| 999 |
+
ffn_pre:smallOff(p+"ffn_pre"), ffn_post:smallOff(p+"ffn_post"),
|
| 1000 |
+
bi:smallOff(p+"bi"), bo:smallOff(p+"bo") }; };
|
| 1001 |
+
|
| 1002 |
+
const makeStorage=(floats,extra=0)=>device.createBuffer({ size:Math.max(4,floats*4),
|
| 1003 |
+
usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST|GPUBufferUsage.COPY_SRC|extra });
|
| 1004 |
+
|
| 1005 |
+
// section byte ranges within depth_q4_ort.bin
|
| 1006 |
+
const PACKED_BYTE_OFF=layout.packed_byte_off, SCALE_BYTE_OFF=layout.scale_byte_off, SMALL_BYTE_OFF=layout.small_byte_off;
|
| 1007 |
+
const PACKED_WORDS=layout.packed_words, SCALE_FLOATS=layout.scale_floats, SMALL_FLOATS=layout.small_floats;
|
| 1008 |
+
const P_SPLIT_WORDS=Math.ceil(PACKED_WORDS/2);
|
| 1009 |
+
const P0_BYTES=P_SPLIT_WORDS*4, P1_BYTES=PACKED_WORDS*4-P0_BYTES;
|
| 1010 |
+
const p0Buf=device.createBuffer({ size:Math.max(4,P0_BYTES), usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 1011 |
+
const p1Buf=device.createBuffer({ size:Math.max(4,P1_BYTES), usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 1012 |
+
device.queue.writeBuffer(p0Buf,0,binBytes,PACKED_BYTE_OFF,P0_BYTES);
|
| 1013 |
+
device.queue.writeBuffer(p1Buf,0,binBytes,PACKED_BYTE_OFF+P0_BYTES,P1_BYTES);
|
| 1014 |
+
const scBuf=device.createBuffer({ size:Math.max(4,SCALE_FLOATS*4), usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 1015 |
+
device.queue.writeBuffer(scBuf,0,binBytes,SCALE_BYTE_OFF,SCALE_FLOATS*4);
|
| 1016 |
+
const smBuf=device.createBuffer({ size:Math.max(4,SMALL_FLOATS*4), usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST });
|
| 1017 |
+
device.queue.writeBuffer(smBuf,0,binBytes,SMALL_BYTE_OFF,SMALL_FLOATS*4);
|
| 1018 |
+
console.log(`depth q4: PACKED ${(PACKED_WORDS*4/1048576).toFixed(1)}MB (P0 ${(P0_BYTES/1048576).toFixed(1)} / P1 ${(P1_BYTES/1048576).toFixed(1)}), SCALES ${(SCALE_FLOATS*4/1048576).toFixed(1)}MB, SMALL ${(SMALL_FLOATS*4/1048576).toFixed(1)}MB`);
|
| 1019 |
+
|
| 1020 |
+
const xinBuf=makeStorage(D_IN), x768Buf=makeStorage(D_MODEL), qBuf=makeStorage(HD), ctxBuf=makeStorage(HD),
|
| 1021 |
+
attnOutBuf=makeStorage(D_MODEL), ffBuf=makeStorage(D_FF), ffnOutBuf=makeStorage(D_MODEL),
|
| 1022 |
+
xnBuf=makeStorage(D_MODEL), logits1024Buf=makeStorage(CODEBOOK_SIZE);
|
| 1023 |
+
const kBufs=[makeStorage(NUM_CODEBOOKS*D_MODEL),makeStorage(NUM_CODEBOOKS*D_MODEL)];
|
| 1024 |
+
const vBufs=[makeStorage(NUM_CODEBOOKS*D_MODEL),makeStorage(NUM_CODEBOOKS*D_MODEL)];
|
| 1025 |
+
const tokBuf=device.createBuffer({ size:NUM_CODEBOOKS*4, usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST|GPUBufferUsage.COPY_SRC });
|
| 1026 |
+
const tokReadBuf=device.createBuffer({ size:NUM_CODEBOOKS*4, usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST });
|
| 1027 |
+
const noiseBuf=makeStorage(NUM_CODEBOOKS*CODEBOOK_SIZE);
|
| 1028 |
+
|
| 1029 |
+
const makeUniform=bytes=>device.createBuffer({ size:bytes, usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST });
|
| 1030 |
+
const pipeline=code=>{ const m=device.createShaderModule({code}); return device.createComputePipeline({ layout:"auto", compute:{module:m,entryPoint:"main"} }); };
|
| 1031 |
+
|
| 1032 |
+
const HEADER=buildHeaders(P_SPLIT_WORDS);
|
| 1033 |
+
const pAdapter=pipeline(qMatmul(HEADER,{K:D_IN,inBinding:1,outBinding:2,rmsnorm:false,epilogue:"none",hasBias:false}));
|
| 1034 |
+
const pQKV=pipeline(buildQKV(HEADER));
|
| 1035 |
+
const pAttnCore=pipeline(buildAttnCore(HEADER));
|
| 1036 |
+
const pWo=pipeline(qMatmul(HEADER,{K:HD,inBinding:1,outBinding:2,rmsnorm:false,epilogue:"none",hasBias:false}));
|
| 1037 |
+
const pResid=pipeline(buildResid(HEADER));
|
| 1038 |
+
const pWi=pipeline(qMatmul(HEADER,{K:D_MODEL,inBinding:1,outBinding:2,rmsnorm:true,epilogue:"geluBias",hasBias:true}));
|
| 1039 |
+
const pWoFfn=pipeline(qMatmul(HEADER,{K:D_FF,inBinding:1,outBinding:2,rmsnorm:false,epilogue:"none",hasBias:true}));
|
| 1040 |
+
const pFinalLN=pipeline(buildFinalLN(HEADER));
|
| 1041 |
+
const pToLogits=pipeline(qMatmul(HEADER,{K:D_MODEL,inBinding:1,outBinding:2,rmsnorm:false,epilogue:"softcap",hasBias:true}));
|
| 1042 |
+
const pSample=pipeline(buildSample());
|
| 1043 |
+
const pEmbed=pipeline(buildEmbed(HEADER));
|
| 1044 |
+
|
| 1045 |
+
// qMatmul U { qOff,sOff,colBase,normOff,biasOff } = 20 -> pad 32.
|
| 1046 |
+
const uAdapter=makeUniform(32), uWo=[makeUniform(32),makeUniform(32)], uWi=[makeUniform(32),makeUniform(32)], uWoFfn=[makeUniform(32),makeUniform(32)];
|
| 1047 |
+
const uToLogits=Array.from({length:NUM_CODEBOOKS},()=>makeUniform(32));
|
| 1048 |
+
const uQKV=Array.from({length:NUM_CODEBOOKS},()=>[makeUniform(32),makeUniform(32)]);
|
| 1049 |
+
const uAttnCore=Array.from({length:NUM_CODEBOOKS},()=>[makeUniform(16),makeUniform(16)]);
|
| 1050 |
+
const uResidAttn=[makeUniform(16),makeUniform(16)], uResidFfn=[makeUniform(16),makeUniform(16)], uFinalLN=makeUniform(16);
|
| 1051 |
+
const uSample=Array.from({length:NUM_CODEBOOKS},()=>makeUniform(16));
|
| 1052 |
+
const uEmbed=Array.from({length:NUM_CODEBOOKS},()=>makeUniform(16));
|
| 1053 |
+
|
| 1054 |
+
// fill a 5-u32 qMatmul uniform { qOff, sOff, colBase, normOff, biasOff }
|
| 1055 |
+
const fillMM=(buf,qOff,sOff,colBase,normOff,biasOff)=>{ const a=new Uint32Array(8);
|
| 1056 |
+
a[0]=qOff; a[1]=sOff; a[2]=colBase; a[3]=normOff; a[4]=biasOff; device.queue.writeBuffer(buf,0,a); };
|
| 1057 |
+
|
| 1058 |
+
{ const a=q4("adapter"); fillMM(uAdapter, a.packed_off, a.scale_off, 0, 0, 0); }
|
| 1059 |
+
for(let li=0;li<NUM_LAYERS;li++){ const Lr=layerOffsets(li), p="l"+li+"_";
|
| 1060 |
+
const Wq=q4(p+"Wq"), Wk=q4(p+"Wk"), Wv=q4(p+"Wv");
|
| 1061 |
+
for(let q=0;q<NUM_CODEBOOKS;q++){
|
| 1062 |
+
const a=new Uint32Array(8); a[0]=Lr.attn_pre;
|
| 1063 |
+
a[1]=Wq.packed_off; a[2]=Wq.scale_off; a[3]=Wk.packed_off; a[4]=Wk.scale_off;
|
| 1064 |
+
a[5]=Wv.packed_off; a[6]=Wv.scale_off; a[7]=q; // slot
|
| 1065 |
+
device.queue.writeBuffer(uQKV[q][li],0,a);
|
| 1066 |
+
const c=new Uint32Array(4); c[0]=Lr.per_dim_scale; c[1]=q+1; device.queue.writeBuffer(uAttnCore[q][li],0,c); }
|
| 1067 |
+
{ const a=new Uint32Array(4); a[0]=Lr.attn_post; device.queue.writeBuffer(uResidAttn[li],0,a); }
|
| 1068 |
+
{ const Wo=q4(p+"Wo"); fillMM(uWo[li], Wo.packed_off, Wo.scale_off, 0, 0, 0); }
|
| 1069 |
+
{ const Wi=q4(p+"Wi"); fillMM(uWi[li], Wi.packed_off, Wi.scale_off, 0, Lr.ffn_pre, Lr.bi); }
|
| 1070 |
+
{ const Wofn=q4(p+"Wo_ffn"); fillMM(uWoFfn[li], Wofn.packed_off, Wofn.scale_off, 0, 0, Lr.bo); }
|
| 1071 |
+
{ const a=new Uint32Array(4); a[0]=Lr.ffn_post; device.queue.writeBuffer(uResidFfn[li],0,a); }
|
| 1072 |
+
}
|
| 1073 |
+
{ const a=new Uint32Array(4); a[0]=smallOff("final_scale"); a[1]=smallOff("final_bias"); device.queue.writeBuffer(uFinalLN,0,a); }
|
| 1074 |
+
for(let q=0;q<NUM_CODEBOOKS;q++){ const e=new ArrayBuffer(16);
|
| 1075 |
+
new Uint32Array(e,0,1)[0]=smallOff("embedding"); new Float32Array(e,4,1)[0]=EMBED_SCALE; new Uint32Array(e,8,1)[0]=q;
|
| 1076 |
+
device.queue.writeBuffer(uEmbed[q],0,e); }
|
| 1077 |
+
{ const tl=q4("to_logits_w");
|
| 1078 |
+
for(let q=0;q<NUM_CODEBOOKS;q++){ const lo=NUM_RESERVED+q*CODEBOOK_SIZE;
|
| 1079 |
+
fillMM(uToLogits[q], tl.packed_off, tl.scale_off, lo, 0, smallOff("to_logits_b")); } }
|
| 1080 |
+
|
| 1081 |
+
// q4 matmul bind group: P0@0, P1@6, SC@7, SMALL@8, in@1, out@2, uniform@5.
|
| 1082 |
+
// usesSmall=false (adapter/Wo: no rmsnorm/bias) -> wv() pruned -> don't bind SMALL@8.
|
| 1083 |
+
function bgMM(p,inBuf,outBuf,uBuf,usesSmall=true){
|
| 1084 |
+
const entries=[{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},
|
| 1085 |
+
{binding:1,resource:{buffer:inBuf}},{binding:2,resource:{buffer:outBuf}},{binding:5,resource:{buffer:uBuf}}];
|
| 1086 |
+
if(usesSmall) entries.push({binding:8,resource:{buffer:smBuf}});
|
| 1087 |
+
return device.createBindGroup({ layout:p.getBindGroupLayout(0), entries });
|
| 1088 |
+
}
|
| 1089 |
+
const bgAdapterXin=bgMM(pAdapter,xinBuf,x768Buf,uAdapter,false);
|
| 1090 |
+
const mkAdapterBG=inBuf=>bgMM(pAdapter,inBuf,x768Buf,uAdapter,false);
|
| 1091 |
+
const bgWo=[0,1].map(li=>bgMM(pWo,ctxBuf,attnOutBuf,uWo[li],false));
|
| 1092 |
+
const bgQKV=Array.from({length:NUM_CODEBOOKS},(_,q)=>[0,1].map(li=>device.createBindGroup({
|
| 1093 |
+
layout:pQKV.getBindGroupLayout(0), entries:[
|
| 1094 |
+
{binding:0,resource:{buffer:p0Buf}},{binding:6,resource:{buffer:p1Buf}},{binding:7,resource:{buffer:scBuf}},{binding:8,resource:{buffer:smBuf}},
|
| 1095 |
+
{binding:1,resource:{buffer:x768Buf}},{binding:2,resource:{buffer:qBuf}},
|
| 1096 |
+
{binding:3,resource:{buffer:kBufs[li]}},{binding:4,resource:{buffer:vBufs[li]}},{binding:5,resource:{buffer:uQKV[q][li]}}] })));
|
| 1097 |
+
const bgAttnCore=Array.from({length:NUM_CODEBOOKS},(_,q)=>[0,1].map(li=>device.createBindGroup({
|
| 1098 |
+
layout:pAttnCore.getBindGroupLayout(0), entries:[
|
| 1099 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 1100 |
+
{binding:1,resource:{buffer:qBuf}},{binding:2,resource:{buffer:kBufs[li]}},
|
| 1101 |
+
{binding:3,resource:{buffer:vBufs[li]}},{binding:4,resource:{buffer:ctxBuf}},{binding:5,resource:{buffer:uAttnCore[q][li]}}] })));
|
| 1102 |
+
const bgResidAttn=[0,1].map(li=>device.createBindGroup({ layout:pResid.getBindGroupLayout(0), entries:[
|
| 1103 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 1104 |
+
{binding:1,resource:{buffer:x768Buf}},{binding:2,resource:{buffer:attnOutBuf}},{binding:5,resource:{buffer:uResidAttn[li]}}] }));
|
| 1105 |
+
const bgWi=[0,1].map(li=>bgMM(pWi,x768Buf,ffBuf,uWi[li]));
|
| 1106 |
+
const bgWoFfn=[0,1].map(li=>bgMM(pWoFfn,ffBuf,ffnOutBuf,uWoFfn[li]));
|
| 1107 |
+
const bgResidFfn=[0,1].map(li=>device.createBindGroup({ layout:pResid.getBindGroupLayout(0), entries:[
|
| 1108 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 1109 |
+
{binding:1,resource:{buffer:x768Buf}},{binding:2,resource:{buffer:ffnOutBuf}},{binding:5,resource:{buffer:uResidFfn[li]}}] }));
|
| 1110 |
+
const bgFinalLN=device.createBindGroup({ layout:pFinalLN.getBindGroupLayout(0), entries:[
|
| 1111 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 1112 |
+
{binding:1,resource:{buffer:x768Buf}},{binding:2,resource:{buffer:xnBuf}},{binding:5,resource:{buffer:uFinalLN}}] });
|
| 1113 |
+
const bgToLogits=Array.from({length:NUM_CODEBOOKS},(_,q)=>bgMM(pToLogits,xnBuf,logits1024Buf,uToLogits[q]));
|
| 1114 |
+
const bgSample=Array.from({length:NUM_CODEBOOKS},(_,q)=>device.createBindGroup({ layout:pSample.getBindGroupLayout(0), entries:[
|
| 1115 |
+
{binding:1,resource:{buffer:logits1024Buf}},{binding:2,resource:{buffer:noiseBuf}},
|
| 1116 |
+
{binding:3,resource:{buffer:tokBuf}},{binding:5,resource:{buffer:uSample[q]}}] }));
|
| 1117 |
+
const bgEmbed=Array.from({length:NUM_CODEBOOKS},(_,q)=>device.createBindGroup({ layout:pEmbed.getBindGroupLayout(0), entries:[
|
| 1118 |
+
{binding:8,resource:{buffer:smBuf}},
|
| 1119 |
+
{binding:1,resource:{buffer:tokBuf}},{binding:2,resource:{buffer:xinBuf}},{binding:3,resource:{buffer:uEmbed[q]}}] }));
|
| 1120 |
+
|
| 1121 |
+
const WG_ADAPTER=nWGT(D_MODEL), WG_QKV=nWGT(HD), WG_WO=nWGT(D_MODEL),
|
| 1122 |
+
WG_WI=nWGT(D_FF), WG_WOFFN=nWGT(D_MODEL), WG_TOLOGITS=nWGT(CODEBOOK_SIZE);
|
| 1123 |
+
|
| 1124 |
+
const adapterBGCache=new WeakMap();
|
| 1125 |
+
|
| 1126 |
+
// runDepth: step-0 input is the temporal-out GPUBuffer (GPU-resident, bound
|
| 1127 |
+
// directly — needs only STORAGE usage). noiseFlat COMPACT [Q*1024]. Returns
|
| 1128 |
+
// { codes:[12 relative], toks:[12 global] }. Single tiny readback.
|
| 1129 |
+
async function runDepth(temporalOutBuf, noiseFlat, temperature){
|
| 1130 |
+
device.queue.writeBuffer(noiseBuf,0,noiseFlat);
|
| 1131 |
+
for(let q=0;q<NUM_CODEBOOKS;q++){ const lo=NUM_RESERVED+q*CODEBOOK_SIZE; const buf=new ArrayBuffer(16);
|
| 1132 |
+
new Uint32Array(buf,0,1)[0]=lo; new Float32Array(buf,4,1)[0]=temperature;
|
| 1133 |
+
new Uint32Array(buf,8,1)[0]=q; new Uint32Array(buf,12,1)[0]=q*CODEBOOK_SIZE;
|
| 1134 |
+
device.queue.writeBuffer(uSample[q],0,buf); }
|
| 1135 |
+
|
| 1136 |
+
let bgAdapter0=adapterBGCache.get(temporalOutBuf);
|
| 1137 |
+
if(!bgAdapter0){ bgAdapter0=mkAdapterBG(temporalOutBuf); adapterBGCache.set(temporalOutBuf,bgAdapter0); }
|
| 1138 |
+
|
| 1139 |
+
for(let q=0;q<NUM_CODEBOOKS;q++){
|
| 1140 |
+
const enc=device.createCommandEncoder(); const pass=enc.beginComputePass();
|
| 1141 |
+
pass.setPipeline(pAdapter);
|
| 1142 |
+
pass.setBindGroup(0, q===0?bgAdapter0:bgAdapterXin);
|
| 1143 |
+
pass.dispatchWorkgroups(WG_ADAPTER);
|
| 1144 |
+
for(let li=0;li<NUM_LAYERS;li++){
|
| 1145 |
+
pass.setPipeline(pQKV); pass.setBindGroup(0,bgQKV[q][li]); pass.dispatchWorkgroups(WG_QKV);
|
| 1146 |
+
pass.setPipeline(pAttnCore); pass.setBindGroup(0,bgAttnCore[q][li]); pass.dispatchWorkgroups(1);
|
| 1147 |
+
pass.setPipeline(pWo); pass.setBindGroup(0,bgWo[li]); pass.dispatchWorkgroups(WG_WO);
|
| 1148 |
+
pass.setPipeline(pResid); pass.setBindGroup(0,bgResidAttn[li]); pass.dispatchWorkgroups(1);
|
| 1149 |
+
pass.setPipeline(pWi); pass.setBindGroup(0,bgWi[li]); pass.dispatchWorkgroups(WG_WI);
|
| 1150 |
+
pass.setPipeline(pWoFfn); pass.setBindGroup(0,bgWoFfn[li]); pass.dispatchWorkgroups(WG_WOFFN);
|
| 1151 |
+
pass.setPipeline(pResid); pass.setBindGroup(0,bgResidFfn[li]); pass.dispatchWorkgroups(1);
|
| 1152 |
+
}
|
| 1153 |
+
pass.setPipeline(pFinalLN); pass.setBindGroup(0,bgFinalLN); pass.dispatchWorkgroups(1);
|
| 1154 |
+
pass.setPipeline(pToLogits); pass.setBindGroup(0,bgToLogits[q]); pass.dispatchWorkgroups(WG_TOLOGITS);
|
| 1155 |
+
pass.setPipeline(pSample); pass.setBindGroup(0,bgSample[q]); pass.dispatchWorkgroups(1);
|
| 1156 |
+
if(q<NUM_CODEBOOKS-1){ pass.setPipeline(pEmbed); pass.setBindGroup(0,bgEmbed[q]); pass.dispatchWorkgroups(1); }
|
| 1157 |
+
pass.end(); device.queue.submit([enc.finish()]);
|
| 1158 |
+
}
|
| 1159 |
+
const enc2=device.createCommandEncoder();
|
| 1160 |
+
enc2.copyBufferToBuffer(tokBuf,0,tokReadBuf,0,NUM_CODEBOOKS*4);
|
| 1161 |
+
device.queue.submit([enc2.finish()]);
|
| 1162 |
+
await tokReadBuf.mapAsync(GPUMapMode.READ);
|
| 1163 |
+
const toksU=new Uint32Array(tokReadBuf.getMappedRange().slice(0)); tokReadBuf.unmap();
|
| 1164 |
+
const toks=[], codes=[];
|
| 1165 |
+
for(let q=0;q<NUM_CODEBOOKS;q++){ const tok=Number(toksU[q]); toks.push(tok);
|
| 1166 |
+
codes.push(((tok-NUM_RESERVED)%CODEBOOK_SIZE+CODEBOOK_SIZE)%CODEBOOK_SIZE); }
|
| 1167 |
+
return {codes, toks};
|
| 1168 |
+
}
|
| 1169 |
+
return { runDepth };
|
| 1170 |
+
}
|
| 1171 |
+
|
| 1172 |
+
// ── ORT tensor helpers ──
|
| 1173 |
+
const NR=()=>cfg.num_reserved_tokens, CB=()=>cfg.codebook_size, Q=()=>cfg.num_codebooks, D=()=>cfg.model_dims;
|
| 1174 |
+
const i64=(arr,dims)=>new ort.Tensor("int64",BigInt64Array.from(arr.map(x=>BigInt(x))),dims);
|
| 1175 |
+
const f32=(arr,dims)=>new ort.Tensor("float32",arr instanceof Float32Array?arr:Float32Array.from(arr),dims);
|
| 1176 |
+
|
| 1177 |
+
// ── conditioning -> source (ORT encoder). Returns the source as a Float32Array. ──
|
| 1178 |
+
async function encodeStyle(tokens){
|
| 1179 |
+
const off=cfg.conditioning.cond_offset;
|
| 1180 |
+
const vals=[...tokens, ...Array(128).fill(-1), -1, 20,10,8].map(v=>v+off);
|
| 1181 |
+
const {source}=await S.encoder.run({cond:i64(vals,[1,1,vals.length])});
|
| 1182 |
+
const data = source.data ?? await source.getData(); // CPU-location encoder output
|
| 1183 |
+
const srcF32 = data instanceof Float32Array ? Float32Array.from(data) : Float32Array.from(data);
|
| 1184 |
+
try{ source.dispose && source.dispose(); }catch(e){}
|
| 1185 |
+
return srcF32; // [256]
|
| 1186 |
+
}
|
| 1187 |
+
|
| 1188 |
+
// ── per-frame state ──
|
| 1189 |
+
// prevToks: the 12 global tokens from the previous frame's depth (next temporal `prev`).
|
| 1190 |
+
let prevToks=null, sourceF32=null, history=[], pendingStyle=null;
|
| 1191 |
+
function dispose(t){ try{ t && t.dispose && t.dispose(); }catch(e){} }
|
| 1192 |
+
function reset(){
|
| 1193 |
+
// temporal `prev` for the very first frame: 12 zeros (BOS), as in the ORT path.
|
| 1194 |
+
prevToks=Array(Q()).fill(0);
|
| 1195 |
+
history=[];
|
| 1196 |
+
if(TEMPORAL) TEMPORAL.resetKV(); // reset self/cross KV growth to length 0
|
| 1197 |
+
if(DSJ){ if(DSTATE) DSJ.keys.forEach((k,i)=>dispose(DSTATE["s_"+i]));
|
| 1198 |
+
DSTATE={}; DSJ.keys.forEach((k,i)=>{ const sh=DSJ.shapes[k]; DSTATE["s_"+i]=f32(new Float32Array(sh.reduce((a,b)=>a*b,1)),sh); }); }
|
| 1199 |
+
}
|
| 1200 |
+
|
| 1201 |
+
// ── per-frame generation: custom WGSL temporal → custom WGSL depth ──
|
| 1202 |
+
let DEC_MS=0, DEPTH_MS=0, TEMP_MS=0;
|
| 1203 |
+
async function genFrame(temp){
|
| 1204 |
+
// 1) custom WGSL temporal step. prevToks[12] global ids -> out in xBuf[1024]
|
| 1205 |
+
// (GPU-resident). Self/cross KV slot-appended on-device. Zero readback.
|
| 1206 |
+
const t0=performance.now();
|
| 1207 |
+
const outBuf=TEMPORAL.runTemporal(prevToks);
|
| 1208 |
+
TEMP_MS=performance.now()-t0;
|
| 1209 |
+
// 2) custom WGSL depth reads the temporal-out GPUBuffer directly (zero copy).
|
| 1210 |
+
const noise=Float32Array.from({length:Q()*CODEBOOK_SIZE},()=>Math.random());
|
| 1211 |
+
const t1=performance.now();
|
| 1212 |
+
const {codes,toks}=await DEPTH.runDepth(outBuf, noise, temp);
|
| 1213 |
+
DEPTH_MS=performance.now()-t1;
|
| 1214 |
+
prevToks=toks; // next temporal `prev` = the 12 global tokens
|
| 1215 |
+
if(window.__LOG){ try{ const _d=await TEMPORAL.debugReadOut(); window.__FN=(window.__FN||0)+1;
|
| 1216 |
+
console.log('f'+window.__FN, 'tmpAbs='+_d.absmax.toFixed(2), 'mean='+_d.mean.toFixed(3), 'std='+_d.std.toFixed(2), 'f4=['+_d.first4.map(x=>x.toFixed(2)).join(',')+']', (_d.hasNaN?'NaN!':''), 'tlen='+TEMPORAL.kvLen(), 'codes=['+codes.join(',')+']'); }catch(e){console.warn('log',e);} }
|
| 1217 |
+
return codes; // 12 relative codes for decode
|
| 1218 |
+
}
|
| 1219 |
+
|
| 1220 |
+
// codes [F,12] -> embeddings [1,F,256] (gather+sum over 12 levels)
|
| 1221 |
+
function codesToEmb(frames){ const F=frames.length, dd=256, out=new Float32Array(F*dd);
|
| 1222 |
+
for(let t=0;t<F;t++) for(let q=0;q<12;q++){ const c=frames[t][q], o=(q*1024+c)*dd;
|
| 1223 |
+
for(let i=0;i<dd;i++) out[t*dd+i]+=QUANT[o+i]; } return f32(out,[1,F,dd]); }
|
| 1224 |
+
// stateful streaming decode (GPU-resident conv-state caches; decodes only NEW frames)
|
| 1225 |
+
async function streamDecode(frames){
|
| 1226 |
+
const emb=codesToEmb(frames);
|
| 1227 |
+
const feed={embeddings:emb}; DSJ.keys.forEach((k,i)=>feed["s_"+i]=DSTATE["s_"+i]);
|
| 1228 |
+
const r=await S.decoder.run(feed);
|
| 1229 |
+
DSJ.keys.forEach((k,i)=>{ dispose(DSTATE["s_"+i]); DSTATE["s_"+i]=r["ns_"+i]; });
|
| 1230 |
+
dispose(emb);
|
| 1231 |
+
return r.waveform;
|
| 1232 |
+
}
|
| 1233 |
+
|
| 1234 |
+
// ── load ──
|
| 1235 |
+
$("load").onclick=async()=>{
|
| 1236 |
+
$("load").disabled=true; $("prog").style.display="block"; setBar(0);
|
| 1237 |
+
$("stat").textContent=`loading ${sz} · webgpu · custom WGSL temporal + WGSL depth…`;
|
| 1238 |
+
const base=REPO;
|
| 1239 |
+
try{
|
| 1240 |
+
cfg=await fetchJSON(base+"/config.json");
|
| 1241 |
+
presets=await fetchJSON(base+"/presets.json");
|
| 1242 |
+
DSJ=await fetchJSON(DEC_STATE_URL);
|
| 1243 |
+
$("prompt").innerHTML=Object.keys(presets).map(p=>`<option>${p}</option>`).join("");
|
| 1244 |
+
const ep=["webgpu"];
|
| 1245 |
+
HIFLUSH={extra:{"ep.webgpuexecutionprovider.maxPendingDispatches":"100000"}};
|
| 1246 |
+
// ORT graphs: encoder (-> source) + fp32 streaming decoder. Temporal + depth are WGSL.
|
| 1247 |
+
const g={encoder:"encoder",decoder:"spectrostream_decoder"};
|
| 1248 |
+
const extMap=cfg.external_data||{};
|
| 1249 |
+
const items=[];
|
| 1250 |
+
for(const k in g){
|
| 1251 |
+
const fn = k==="decoder" ? "spectrostream_stream.onnx" : `${g[k]}.q4.onnx`;
|
| 1252 |
+
items.push({id:"g_"+k, url: k==="decoder" ? DEC_URL : `${base}/onnx/${fn}`});
|
| 1253 |
+
const df=extMap[fn]; if(df) items.push({id:"d_"+k, path:df, url:`${base}/onnx/${df}`}); }
|
| 1254 |
+
items.push({id:"emb", url:base+"/onnx/token_embedding_fp16.bin"});
|
| 1255 |
+
items.push({id:"quant", url:base+"/onnx/quantizer_fp16.bin"});
|
| 1256 |
+
// WGSL temporal weights (q4: packed nibbles + scales + small fp32 params)
|
| 1257 |
+
items.push({id:"temp_layout", url:TEMP_BASE+"temporal_q4_ort_layout.json", json:true});
|
| 1258 |
+
items.push({id:"temp_bin", url:TEMP_BASE+"temporal_q4_ort.bin"});
|
| 1259 |
+
// WGSL depth weights (q4)
|
| 1260 |
+
items.push({id:"depth_layout", url:DEPTH_BASE+"depth_q4_ort_layout.json", json:true});
|
| 1261 |
+
items.push({id:"depth_bin", url:DEPTH_BASE+"depth_q4_ort.bin"});
|
| 1262 |
+
const N=items.length, bufs={};
|
| 1263 |
+
for(let i=0;i<N;i++){ const it=items[i], name=it.url.split("/").pop();
|
| 1264 |
+
if(it.json){ bufs[it.id]=await fetchJSON(it.url); setBar((i+1)/N); continue; }
|
| 1265 |
+
bufs[it.id]=await fetchBuf(it.url, f=>{ setBar((i+f)/N); $("stat").textContent=`downloading ${name} — ${(f*100).toFixed(0)}% (${i+1}/${N})`; }); }
|
| 1266 |
+
setBar(1); $("stat").textContent="compiling ORT graphs…";
|
| 1267 |
+
for(const k in g){ const opts={executionProviders:ep}, d=items.find(x=>x.id==="d_"+k);
|
| 1268 |
+
if(d) opts.externalData=[{path:d.path, data:bufs["d_"+k]}];
|
| 1269 |
+
if(k==="decoder"){ opts.preferredOutputLocation={}; DSJ.keys.forEach((_,i)=>opts.preferredOutputLocation["ns_"+i]="gpu-buffer"); }
|
| 1270 |
+
S[k]=await ort.InferenceSession.create(bufs["g_"+k], opts); }
|
| 1271 |
+
EMB=f16decode(bufs.emb); QUANT=f16decode(bufs.quant);
|
| 1272 |
+
|
| 1273 |
+
// build BOTH custom WGSL engines on ORT's SAME WebGPU device (zero-copy handoff)
|
| 1274 |
+
const ortDevice = ort.env.webgpu.device;
|
| 1275 |
+
if(!ortDevice) throw new Error("ort.env.webgpu.device is null — WebGPU EP not initialized");
|
| 1276 |
+
// capture any GPU validation/out-of-memory errors that would otherwise be
|
| 1277 |
+
// silently swallowed (pipelines/bind groups invalid on ORT's shared device
|
| 1278 |
+
// -> garbage out). Surfaced via window.__DBG for headless verification.
|
| 1279 |
+
const __dbgErrs=[];
|
| 1280 |
+
try{ ortDevice.addEventListener("uncapturederror",e=>{ __dbgErrs.push(String(e.error&&e.error.message||e.error||e)); }); }catch(e){}
|
| 1281 |
+
window.__DBG = { ready:false, errors:__dbgErrs };
|
| 1282 |
+
$("stat").textContent="building WGSL temporal on ORT device…";
|
| 1283 |
+
TEMPORAL = await buildTemporal(ortDevice, bufs.temp_layout, bufs.temp_bin);
|
| 1284 |
+
$("stat").textContent="building WGSL depth on ORT device…";
|
| 1285 |
+
DEPTH = await buildDepth(ortDevice, bufs.depth_layout, bufs.depth_bin);
|
| 1286 |
+
// record layout/blob fingerprints so a hosted-vs-local byte mismatch is visible.
|
| 1287 |
+
const fp=(l,bin)=>({packed_byte_off:l.packed_byte_off,scale_byte_off:l.scale_byte_off,small_byte_off:l.small_byte_off,
|
| 1288 |
+
packed_words:l.packed_words,scale_floats:l.scale_floats,small_floats:l.small_floats,embed_scale:l.embed_scale,
|
| 1289 |
+
binBytes:bin.byteLength,nq4:Object.keys(l.q4||{}).length,nsmall:Object.keys(l.small||{}).length});
|
| 1290 |
+
window.__DBG = Object.assign(window.__DBG||{},{ tempLayout:fp(bufs.temp_layout,bufs.temp_bin), depthLayout:fp(bufs.depth_layout,bufs.depth_bin) });
|
| 1291 |
+
|
| 1292 |
+
// warm up: compile WebGPU shaders (temporal + depth) across the KV-growth range
|
| 1293 |
+
$("stat").textContent="warming up (compiling shaders)…"; setBar(0);
|
| 1294 |
+
try{ reset(); sourceF32=await encodeStyle(presets[$("prompt").value]); TEMPORAL.setSource(sourceF32);
|
| 1295 |
+
const W=(cfg.temporal&&cfg.temporal.max_past?cfg.temporal.max_past:T_KV_CAP)+4;
|
| 1296 |
+
for(let w=0;w<W;w++){ const wt=performance.now(); const c=await genFrame(1.1); history.push(c); const ms=performance.now()-wt;
|
| 1297 |
+
setBar((w+1)/W); $("stat").textContent=`warming up ${w+1}/${W} · ${ms.toFixed(0)} ms/frame (temporal ${TEMP_MS.toFixed(1)} · depth ${DEPTH_MS.toFixed(1)} ms)`;
|
| 1298 |
+
await new Promise(r=>setTimeout(r)); }
|
| 1299 |
+
// ---- debug hook: after warmup, capture temporal-out + first depth output ----
|
| 1300 |
+
try{
|
| 1301 |
+
reset(); TEMPORAL.setSource(sourceF32);
|
| 1302 |
+
TEMPORAL.runTemporal(prevToks); // one temporal step (BOS prev)
|
| 1303 |
+
const t=await TEMPORAL.debugReadOut(); // read temporal out[1024]
|
| 1304 |
+
const noise=Float32Array.from({length:Q()*CODEBOOK_SIZE},()=>Math.random());
|
| 1305 |
+
const {codes,toks}=await DEPTH.runDepth(TEMPORAL.outBuf, noise, 1.1);
|
| 1306 |
+
window.__DBG = Object.assign(window.__DBG||{}, {
|
| 1307 |
+
ready:true,
|
| 1308 |
+
tempOutAbsmax:t.absmax, tempOutHasNaN:t.hasNaN, tempOutFirst4:t.first4,
|
| 1309 |
+
depthCodes:codes, depthToks:toks, errors:__dbgErrs,
|
| 1310 |
+
});
|
| 1311 |
+
console.log("__DBG", window.__DBG);
|
| 1312 |
+
}catch(e){ window.__DBG = Object.assign(window.__DBG||{}, { ready:false, dbgError:String(e), errors:__dbgErrs }); console.warn("dbg:",e); }
|
| 1313 |
+
if(history.length>=10){ const w=await streamDecode(history.slice(-10)); dispose(w); }
|
| 1314 |
+
reset(); }catch(e){ console.warn("warmup:",e); window.__DBG=Object.assign(window.__DBG||{},{warmupError:String(e)}); }
|
| 1315 |
+
$("prog").style.display="none"; $("play").disabled=false;
|
| 1316 |
+
$("stat").innerHTML=`ready · <code>webgpu</code> · <code>WGSL temporal + WGSL depth</code> — warmed, press Play`;
|
| 1317 |
+
}catch(e){ $("prog").style.display="none"; $("stat").textContent="load failed: "+e.message; $("load").disabled=false; console.error(e); }
|
| 1318 |
+
};
|
| 1319 |
+
|
| 1320 |
+
// ── Web Audio streaming ──
|
| 1321 |
+
let actx=null,analyser=null,nextT=0,playing=false,freq=null;
|
| 1322 |
+
function playSamples(data,start,end){ const n=end-start; if(n<1)return;
|
| 1323 |
+
const buf=actx.createBuffer(2,n,48000), L=buf.getChannelData(0), R=buf.getChannelData(1);
|
| 1324 |
+
for(let i=0;i<n;i++){ L[i]=data[(start+i)*2]; R[i]=data[(start+i)*2+1]; }
|
| 1325 |
+
const s=actx.createBufferSource(); s.buffer=buf; s.connect(analyser);
|
| 1326 |
+
if(nextT<actx.currentTime+0.05)nextT=actx.currentTime+0.15; s.start(nextT); nextT+=n/48000; }
|
| 1327 |
+
function viz(){ const c=$("viz"),g=c.getContext("2d"),w=c.width,h=c.height; g.clearRect(0,0,w,h);
|
| 1328 |
+
if(analyser){ analyser.getByteFrequencyData(freq); const bw=w/48;
|
| 1329 |
+
for(let i=0;i<48;i++){ const v=freq[i*2]/255; g.fillStyle=`hsl(${70-v*40},85%,${45+v*15}%)`; g.fillRect(i*bw,h-v*h,bw-1,v*h);} }
|
| 1330 |
+
requestAnimationFrame(viz); } viz();
|
| 1331 |
+
|
| 1332 |
+
$("play").onclick=async()=>{
|
| 1333 |
+
actx=new AudioContext({sampleRate:48000}); await actx.resume(); nextT=actx.currentTime+0.2;
|
| 1334 |
+
analyser=actx.createAnalyser(); analyser.fftSize=128; analyser.connect(actx.destination); freq=new Uint8Array(analyser.frequencyBinCount);
|
| 1335 |
+
$("play").disabled=true; $("stop").disabled=false; playing=true;
|
| 1336 |
+
reset(); sourceF32=await encodeStyle(presets[$("prompt").value]); TEMPORAL.setSource(sourceF32);
|
| 1337 |
+
const CHUNK=10; let done=0, firstChunk=true; const t0=performance.now();
|
| 1338 |
+
// running averages for the breakdown line
|
| 1339 |
+
let aT=0,aD=0,aC=0,nT=0;
|
| 1340 |
+
while(playing){
|
| 1341 |
+
if(pendingStyle){ const old=sourceF32; sourceF32=await encodeStyle(presets[pendingStyle]); TEMPORAL.setSource(sourceF32); pendingStyle=null; }
|
| 1342 |
+
const nf=[];
|
| 1343 |
+
for(let i=0;i<CHUNK&&playing;i++){ const c=await genFrame(1.1); history.push(c); nf.push(c); done++;
|
| 1344 |
+
aT+=TEMP_MS; aD+=DEPTH_MS; nT++; }
|
| 1345 |
+
if(nf.length===CHUNK){
|
| 1346 |
+
const dt=performance.now();
|
| 1347 |
+
const wav=await streamDecode(nf);
|
| 1348 |
+
const data=wav.data ?? await wav.getData(); DEC_MS=performance.now()-dt;
|
| 1349 |
+
aC+=DEC_MS;
|
| 1350 |
+
const start=firstChunk?1920:0; firstChunk=false;
|
| 1351 |
+
playSamples(data, start, wav.dims[1]); dispose(wav);
|
| 1352 |
+
}
|
| 1353 |
+
const sec=done/cfg.frames_per_second, rt=sec/((performance.now()-t0)/1000);
|
| 1354 |
+
$("rt").textContent=`${rt.toFixed(2)}× realtime`;
|
| 1355 |
+
$("stat").textContent=`generating · ${sec.toFixed(1)}s · ${rt.toFixed(2)}× realtime`;
|
| 1356 |
+
// per-frame breakdown (decode is per-10-frame chunk -> /CHUNK to show per-frame)
|
| 1357 |
+
const chunks=Math.max(1,Math.round(nT/CHUNK));
|
| 1358 |
+
$("brk").textContent=`temporal ${(aT/Math.max(1,nT)).toFixed(1)} ms · depth ${(aD/Math.max(1,nT)).toFixed(1)} ms · decode ${(aC/Math.max(1,chunks)).toFixed(1)} ms/chunk`;
|
| 1359 |
+
}
|
| 1360 |
+
};
|
| 1361 |
+
$("stop").onclick=()=>{ playing=false; if(actx)actx.suspend(); $("play").disabled=false; $("stop").disabled=true; $("stat").textContent="stopped"; };
|
| 1362 |
+
$("prompt").onchange=()=>{ if(playing) pendingStyle=$("prompt").value; };
|
| 1363 |
+
</script>
|
| 1364 |
+
</body>
|
| 1365 |
+
</html>
|