Quant backward#290

Merged

LoserCheems merged 7 commits into

mainfrom

May 12, 2026

Commits on May 12, 2026

Update input assertions for quantized tensors in forward and backward functions
LoserCheems
committed
Update scale tensor data types to match query tensor in attention functions
LoserCheems
committed
Fix amax tensor data type to match input tensor dtype in quantize_fp8 function
LoserCheems
committed
Add query, key, and value scale handling in backward functions
LoserCheems
committed
Add query, key, and value scaling in backward functions
LoserCheems
committed
Add query, key, and value scaling in backward functions
LoserCheems
committed
Refactor attention functions to save query, key, and value scales in context
LoserCheems
committed