AMDMIGraphX
AMDMIGraphX copied to clipboard
ROCmSoftwarePlatform
AMD's graph optimization engine.
### Problem Description This ticket is to implement gemm(i8)-dequantizelinear-softmax(fp16)-gemm(fp16) pattern to do a partial i8 attention kernel in rocmlir. Here is one of the examples test we currently have working...
This is a **tracking** issue and is supposed to synchronize both rocMLIR and MIGraphX teams regarding the SplitK GEMM scheme integration. The overall design is based on the [proposal](https://github.com/ROCm/AMDMIGraphX/discussions/2858). ##...
When compiling for gemm using splitk for MLIR, split the pointwise operators and generate code object using our regular pointwise code generation.
For the detailed perf report we can print the types as abbreviated types so we can attach to each tensor size: * double_type -> f64 * float_type -> f32 *...
When a model is quantized to FP16 LayerNorm is also quantized. This leads to an accuracy problem. Make the code changes needed to hold LayerNorm as always FP32 accumulation. Then...
We want to be able to insert `copy` operator for some concat cases where: * Almost all the inputs are a precompile_op * Only 1 copy is needed To do...