cmriat · junjzhang · May 22, 2026 · May 22, 2026 · May 22, 2026 · May 22, 2026
diff --git a/.gitignore b/.gitignore
@@ -22,4 +22,6 @@ logs
 
 trace_processor
 
-docs/_build/
+docs/_build/
+
+.claude/
diff --git a/README.md b/README.md
@@ -33,9 +33,9 @@ Four properties define the surface:
   `TensorBusClient` and hand it tensors. No model wrappers, no
   restructuring of your distributed init, no framework to adopt.
 
-> **Supported placements:** `Shard` and `Replicate`. `Partial` is rejected
-> with `NotImplementedError` — redistribute it to `Replicate` or `Shard` on
-> the source mesh before handing the DTensor to Etha.
+> **Placements:** source supports `Shard` / `Replicate` / `Partial`
+> (collapsed to `Replicate` via a source-side all-reduce before send).
+> Target `Partial` is rejected.
 
 ## Architecture
 

diff --git a/bench/README.md b/bench/README.md
@@ -93,42 +93,65 @@ profile_plot(snapshot)
 ---
 
 # Test results
-Test on 16 GPUs with 2 nodes.
 
-source mesh placement: [Replicate(), Replicate(), Shard(0), Shard(1)]
-target mesh placement: [Replicate(), _StridedShard(1, split_factor=2), Replicate(), Shard(1)]
+Tested on 16 GPUs across 2 nodes. Each mesh combination is run under three
+source placements; target placement is fixed to
+`[Replicate, Replicate, StridedShard(0, split=k), Shard(0)]`
+where `k = target_mesh[3]`.
 
-## Mesh Configuration 1
-* 1, 1, 4, 2 -> 4, 2, 1, 1
-![Mesh 1](./results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1.png)
+| Config | Source placement |
+|---|---|
+| `no_partial`   | `[Replicate, Shard(0), Replicate, Shard(1)]` |
+| `replicate_dp` | `[Replicate, Replicate, Replicate, Shard(1)]` |
+| `partial_dp`   | `[Replicate, Partial("sum"), Replicate, Shard(1)]` |
 
-## Mesh Configuration 2
-* 8, 1, 1, 1 -> 1, 2, 2, 2
-![Mesh 2](./results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2.png)
+## Mesh 1 — (1, 1, 4, 2) → (4, 2, 1, 1)
 
-## Mesh Configuration 3
-* 1, 2, 4, 1 -> 8, 1, 1, 1
-![Mesh 3](./results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1.png)
+| no_partial | replicate_dp | partial_dp |
+|---|---|---|
+| ![](./results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_no_partial.png) | ![](./results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_replicate_dp.png) | ![](./results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_partial_dp.png) |
 
-## Mesh Configuration 4
-* 2, 2, 2, 1 -> 1, 1, 4, 2
-![Mesh 4](./results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2.png)
+## Mesh 2 — (8, 1, 1, 1) → (1, 2, 2, 2)
 
-## Mesh Configuration 5
-* 2, 4, 1, 1 -> 1, 2, 4, 1
-![Mesh 5](./results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1.png)
+| no_partial | replicate_dp | partial_dp |
+|---|---|---|
+| ![](./results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_no_partial.png) | ![](./results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_replicate_dp.png) | ![](./results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_partial_dp.png) |
 
-## Mesh Configuration 6
-* 4, 1, 1, 2 -> 2, 1, 1, 4
-![Mesh 6](./results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4.png)
+## Mesh 3 — (1, 2, 4, 1) → (8, 1, 1, 1)
 
-## Mesh Configuration 7
-* 4, 1, 2, 1 -> 1, 1, 1, 8
-![Mesh 7](./results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8.png)
+| no_partial | replicate_dp | partial_dp |
+|---|---|---|
+| ![](./results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_no_partial.png) | ![](./results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_replicate_dp.png) | ![](./results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_partial_dp.png) |
 
-## Mesh Configuration 8
-* 4, 2, 1, 1 -> 1, 1, 2, 4
-![Mesh 8](./results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4.png)
+## Mesh 4 — (2, 2, 2, 1) → (1, 1, 4, 2)
+
+| no_partial | replicate_dp | partial_dp |
+|---|---|---|
+| ![](./results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_no_partial.png) | ![](./results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_replicate_dp.png) | ![](./results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_partial_dp.png) |
+
+## Mesh 5 — (2, 4, 1, 1) → (1, 2, 4, 1)
+
+| no_partial | replicate_dp | partial_dp |
+|---|---|---|
+| ![](./results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_no_partial.png) | ![](./results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_replicate_dp.png) | ![](./results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_partial_dp.png) |
+
+## Mesh 6 — (4, 1, 1, 2) → (2, 1, 1, 4)
+
+| no_partial | replicate_dp | partial_dp |
+|---|---|---|
+| ![](./results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_no_partial.png) | ![](./results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_replicate_dp.png) | ![](./results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_partial_dp.png) |
+
+## Mesh 7 — (4, 1, 2, 1) → (1, 1, 1, 8)
+
+| no_partial | replicate_dp | partial_dp |
+|---|---|---|
+| ![](./results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_no_partial.png) | ![](./results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_replicate_dp.png) | ![](./results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_partial_dp.png) |
+
+## Mesh 8 — (4, 2, 1, 1) → (1, 1, 2, 4)
+
+| no_partial | replicate_dp | partial_dp |
+|---|---|---|
+| ![](./results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_no_partial.png) | ![](./results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_replicate_dp.png) | ![](./results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_partial_dp.png) |
 
 ---
 

diff --git a/bench/results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1.png b/bench/results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1.png
diff --git a/bench/results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_no_partial.png b/bench/results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_no_partial.png
diff --git a/bench/results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_partial_dp.png b/bench/results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_partial_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_replicate_dp.png b/bench/results/throughput_benchmark_mesh_01_1_1_4_2_4_2_1_1_replicate_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2.png b/bench/results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2.png
diff --git a/bench/results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_no_partial.png b/bench/results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_no_partial.png
diff --git a/bench/results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_partial_dp.png b/bench/results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_partial_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_replicate_dp.png b/bench/results/throughput_benchmark_mesh_02_8_1_1_1_1_2_2_2_replicate_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1.png b/bench/results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1.png
diff --git a/bench/results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_no_partial.png b/bench/results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_no_partial.png
diff --git a/bench/results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_partial_dp.png b/bench/results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_partial_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_replicate_dp.png b/bench/results/throughput_benchmark_mesh_03_1_2_4_1_8_1_1_1_replicate_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2.png b/bench/results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2.png
diff --git a/bench/results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_no_partial.png b/bench/results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_no_partial.png
diff --git a/bench/results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_partial_dp.png b/bench/results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_partial_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_replicate_dp.png b/bench/results/throughput_benchmark_mesh_04_2_2_2_1_1_1_4_2_replicate_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1.png b/bench/results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1.png
diff --git a/bench/results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_no_partial.png b/bench/results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_no_partial.png
diff --git a/bench/results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_partial_dp.png b/bench/results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_partial_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_replicate_dp.png b/bench/results/throughput_benchmark_mesh_05_2_4_1_1_1_2_4_1_replicate_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4.png b/bench/results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4.png
diff --git a/bench/results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_no_partial.png b/bench/results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_no_partial.png
diff --git a/bench/results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_partial_dp.png b/bench/results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_partial_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_replicate_dp.png b/bench/results/throughput_benchmark_mesh_06_4_1_1_2_2_1_1_4_replicate_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8.png b/bench/results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8.png
diff --git a/bench/results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_no_partial.png b/bench/results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_no_partial.png
diff --git a/bench/results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_partial_dp.png b/bench/results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_partial_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_replicate_dp.png b/bench/results/throughput_benchmark_mesh_07_4_1_2_1_1_1_1_8_replicate_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4.png b/bench/results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4.png
diff --git a/bench/results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_no_partial.png b/bench/results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_no_partial.png
diff --git a/bench/results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_partial_dp.png b/bench/results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_partial_dp.png
diff --git a/bench/results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_replicate_dp.png b/bench/results/throughput_benchmark_mesh_08_4_2_1_1_1_1_2_4_replicate_dp.png