From db22649491c5dfe1f6bab2f016bb6b76d1e6255e Mon Sep 17 00:00:00 2001
From: "Schlimbach, Frank" <frank.schlimbach@intel.com>
Date: Mon, 23 Feb 2026 04:52:14 -0800
Subject: [PATCH 1/2] correcting/simplifying mlp-mpi readme

---
 examples/mlp-mpi/README.md | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/examples/mlp-mpi/README.md b/examples/mlp-mpi/README.md
index 5de6957..84668d8 100644
--- a/examples/mlp-mpi/README.md
+++ b/examples/mlp-mpi/README.md
@@ -2,19 +2,15 @@
 
 ## Prerequisites
 
-You need mpi4py in your python env. If you are not using OpenMPI (e.g. not MPICH-like like Intel MPI) you need to modify the first line in mlp_weight_stationary.mlir by replacing `"MPI:Implementation" = "MPICH"` with `"MPI:Implementation" = "OpenMPI"`.
+You need mpi4py in your python env. The default MPI implementation is MPICH.
+
+For OpenMPI, change `"MPI:Implementation" = "MPICH"` to `"MPI:Implementation" = "OpenMPI"` in the first line of mlp_weight_stationary.mlir.
 
 ## Running
 
 ```
-export MLIR_DIR=<path_to_mlir_build_dir>
 export MPI_DIR=<path_to_mpi_install>
-export LH_DIR=<path_to_lighthouse>
-PYTHONPATH=$LH_DIR:$MLIR_DIR/tools/mlir/python_packages/mlir_core \
-  mpirun -n <nRanks> \
-  python -u mlp-mpi.py \
-  --mpilib $MPI_DIR/libmpi.so \
-  --utils_dir $MLIR_DIR/lib \
-  -s 64 64 64
+uv sync --extra runtime_mpich
+uv run mpirun -n <nRanks> python -u mlp-mpi.py --mpilib $MPI_DIR/lib/libmpi.so
 ```
 Run with `--help` for more options.

From 75cea34d411e751b870e71ec6915182b83cf41ea Mon Sep 17 00:00:00 2001
From: "Schlimbach, Frank" <frank.schlimbach@intel.com>
Date: Mon, 23 Feb 2026 05:01:54 -0800
Subject: [PATCH 2/2] More readme

---
 examples/mlp-mpi/README.md | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/examples/mlp-mpi/README.md b/examples/mlp-mpi/README.md
index 84668d8..ef2f9d8 100644
--- a/examples/mlp-mpi/README.md
+++ b/examples/mlp-mpi/README.md
@@ -1,4 +1,10 @@
-# Computing a MLP sigmoid(A@B)@C on multiple ranks using MPI through MLIR
+# Computing an MLP sigmoid(A@B)@C on multiple ranks using MPI through MLIR
+
+This example shows how MLIR's sharding infrastructure can be used to distribute data and computation across multiple nodes with non-shared memory.
+
+Currently, only the lower part of the sharding pipeline is used: `shard-partition`, `convert-shard-to-mpi`, and lowering to LLVM. Therefore, the ingress MLIR is fully annotated.
+
+The example implements a "single MLP", following a 1D/2D weight-stationary partition strategy as described in figures 2a and 2b of https://arxiv.org/pdf/2211.05102.
 
 ## Prerequisites