Commits


Jing Fang authored and GitHub committed 11bf3097360
add transform part of the dq matmul tool chain (#21374) ### Description This is a partial change from [fajin/qdqmatmulnbitstoolchain](https://github.com/microsoft/onnxruntime/pull/21180). The original PR is blocked by Web CI failures. MatMulNBits is a heavily optimized matmul operation. Currently a MatMul can be converted to MatMulNBits to speed up the model inference. However, MatMulNBits is an ORT only op. To make the graph compatible with ONNX ops and utilize MatMulNBits at the same time, we introduce Q/DQ support for MatMulNBits. To convert MatMul ops in a model to MatMulNBits: 1. use matmul_4bits_quantizer.py to convert MatMul to DQ + MatMul using QDQ mode. 2. In ORT session, DQ + MatMul is fused to MatMulNBits #### Note MatMulNBits assume B weight is uint4. When no zp is provided, zp defaults to 8, which is different from DQ. DQ defaults zp to 0 when no zp provided. And DQ supports int4. Therefore some conversions are introduced during DQ + MatMul --> MatMulNBits step. #### Perf Using QDQ format will increase the model initialization time and memory consumption. With current implement, model init time increased from ~4s to ~9s, and memory consumption increased from ~2.8GB to ~4.8GB. The memory increase is due to 1. in optimizer, after transpose the B weight, a in-memory tensor proto is created using protobuf's arena. 2. in finalize step, when saving initializer and prepacking, ORT arena is used to create buffers for initializers. The memory allocated by arenas cannot be fully deallocated. If disable ORT arena memory allocation, the memory consumptions of both QDQ format and original format are ~2.2GB. The time increase is mainly due to multiple memory copy, but can be further optimized. ### Motivation and Context Please see description for details.