catlass-epilogue-composition

Attach a BlockEpilogue to the GEMM using MatmulEpilogue with a TileElemWise component such as TileElemWiseAdd. For ReLU or multi-op chains, use the EVG TreeVisitor path with VisitorCompute<Relu> instead of composing torch ops in Python.

Use BasicMatmul for pure GEMM, MatmulEpilogue for simple elementwise fusion done on AIC, BasicMatmulTlaVisitor for complex fusion graphs executed on AIV, and BasicMatmulTlaUbVisitor when you want accumulation to stay in UB and reduce GM round trips.

EVG GM (BasicMatmulTlaVisitor) routes MMAD results through a workspace with AIV performing tree fusion, while EVG UB (BasicMatmulTlaUbVisitor) keeps accumulation in UB by switching EpilogueVisitor<false> to <true>, reducing global memory traffic.

No. Writing F.relu(matmul(...)) in kernel.py deviates from CATLASS optimization goals and may trigger static check failures. The fusion logic must live in the .asc kernel, a custom epilogue header, or catlass_torch.cpp, with ModelNew calling the matching torch.ops.catlass op.

You can add a custom Tile Epilogue header following existing examples and attach it to BlockEpilogue in the .asc file. However, if the AR task does not list that header in editable_files, you are limited to combining existing Tile components.