fix: hold stacked outputs in a separate map

.github/workflows/pypi-gpu.yml

@@ -26,7 +26,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v4
         with:
-          python-version: 3.7
+          python-version: 3.12
           architecture: x64
 
       - name: Set pyproject.toml version to match github tag

.github/workflows/rust.yml

@@ -184,7 +184,7 @@ jobs:
 
   wasm32-tests:
     runs-on: ubuntu-latest
-    # needs: [build,  library-tests, docs]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
@@ -207,7 +207,7 @@ jobs:
 
   tutorial:
     runs-on: ubuntu-latest
-    needs: [build, library-tests, docs]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
@@ -224,7 +224,7 @@ jobs:
 
   mock-proving-tests:
     runs-on: non-gpu
-    # needs: [build,  library-tests, docs]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
@@ -281,7 +281,7 @@ jobs:
 
   prove-and-verify-evm-tests:
     runs-on: non-gpu
-    needs: [build, library-tests]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
@@ -354,7 +354,7 @@ jobs:
 
   prove-and-verify-tests:
     runs-on: non-gpu
-    needs: [build, library-tests]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
@@ -460,7 +460,7 @@ jobs:
 
   prove-and-verify-mock-aggr-tests:
     runs-on: self-hosted
-    needs: [build, library-tests]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
@@ -495,7 +495,7 @@ jobs:
 
   prove-and-verify-aggr-tests:
     runs-on: large-self-hosted
-    needs: [build, library-tests]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
@@ -512,7 +512,7 @@ jobs:
 
   prove-and-verify-aggr-evm-tests:
     runs-on: large-self-hosted
-    needs: [build, library-tests]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
@@ -568,7 +568,7 @@ jobs:
       - name: Install solc
         run: (hash svm 2>/dev/null || cargo install svm-rs) && svm install 0.8.20 && solc --version
       - name: Setup Virtual Env and Install python dependencies
-        run: python -m venv .env; source .env/bin/activate; pip install -r requirements.txt;
+        run: python -m venv .env --clear; source .env/bin/activate; pip install -r requirements.txt;
       - name: Install Anvil
         run: cargo install --git https://github.com/foundry-rs/foundry --rev c2233ec9fe61e0920c61c6d779bc707252852037 --profile local --locked anvil --force
       - name: Build python ezkl
@@ -578,7 +578,7 @@ jobs:
 
   accuracy-measurement-tests:
     runs-on: ubuntu-latest-32-cores
-    # needs: [build, library-tests, docs]
+    needs: [build, library-tests, docs, python-tests, python-integration-tests]
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v4
@@ -594,7 +594,7 @@ jobs:
           crate: cargo-nextest
           locked: true
       - name: Setup Virtual Env and Install python dependencies
-        run: python -m venv .env; source .env/bin/activate; pip install -r requirements.txt;
+        run: python -m venv .env --clear; source .env/bin/activate; pip install -r requirements.txt;
       - name: Build python ezkl
         run: source .env/bin/activate; unset CONDA_PREFIX; maturin develop --features python-bindings --release
       - name: Div rebase
@@ -628,12 +628,14 @@ jobs:
         run: (hash svm 2>/dev/null || cargo install svm-rs) && svm install 0.8.20 && solc --version
       - name: Install Anvil
         run: cargo install --git https://github.com/foundry-rs/foundry --rev c2233ec9fe61e0920c61c6d779bc707252852037 --profile local --locked anvil --force
+      - name: Install pip
+        run: python -m ensurepip --upgrade
       - name: Setup Virtual Env and Install python dependencies
-        run: python -m venv .env; source .env/bin/activate; pip install -r requirements.txt;
+        run: python -m venv .env --clear; source .env/bin/activate; pip install -r requirements.txt; python -m ensurepip --upgrade
       - name: Build python ezkl
         run: source .env/bin/activate; unset CONDA_PREFIX; maturin develop --features python-bindings --release
       - name: Tictactoe tutorials
-        run: source .env/bin/activate; cargo nextest run py_tests::tests::tictactoe_
+        run: source .env/bin/activate; cargo nextest run py_tests::tests::tictactoe_ --test-threads 1
       # - name: authenticate-kaggle-cli
       #   shell: bash
       #   env:

examples/onnx/lstm_large/input.json

@@ -0,0 +1,13 @@
+{
+  "input_data": [
+    [
+      0.8894134163856506,
+      0.8894201517105103
+    ]
+  ],
+  "output_data": [
+    [
+      0.8436377
+    ]
+  ]
+}

src/graph/model.rs

@@ -803,13 +803,18 @@ impl Model {
                     let input_state_idx = input_state_idx(&input_mappings);
 
                     let mut output_mappings = vec![];
-                    for mapping in b.output_mapping.iter() {
+                    for (i, mapping) in b.output_mapping.iter().enumerate() {
                         let mut mappings = vec![];
                         if let Some(outlet) = mapping.last_value_slot {
                             mappings.push(OutputMapping::Single {
                                 outlet,
                                 is_state: mapping.state,
                             });
+                        } else if mapping.state {
+                            mappings.push(OutputMapping::Single {
+                                outlet: i,
+                                is_state: mapping.state,
+                            });
                         }
                         if let Some(last) = mapping.scan {
                             mappings.push(OutputMapping::Stacked {
@@ -818,6 +823,7 @@ impl Model {
                                 is_state: false,
                             });
                         }
+
                         output_mappings.push(mappings);
                     }
 
@@ -1264,8 +1270,8 @@ impl Model {
                     let num_iter = number_of_iterations(&input_mappings, input_dims.collect());
 
                     debug!(
-                        "{} iteration(s) in a subgraph with inputs {:?} and sources {:?}",
-                        num_iter, inputs, model.graph.inputs
+                        "{} iteration(s) in a subgraph with inputs {:?}, sources {:?}, and outputs {:?}",
+                        num_iter, inputs, model.graph.inputs, model.graph.outputs
                     );
 
                     let mut full_results: Vec<ValTensor<Fp>> = vec![];
@@ -1297,6 +1303,7 @@ impl Model {
                         let res = model.layout_nodes(config, region, &mut subgraph_results)?;
 
                         let mut outlets = BTreeMap::new();
+                        let mut stacked_outlets = BTreeMap::new();
 
                         for (mappings, outlet_res) in output_mappings.iter().zip(res) {
                             for mapping in mappings {
@@ -1309,25 +1316,42 @@ impl Model {
                                             let stacked_res = full_results[*outlet]
                                                 .clone()
                                                 .concat_axis(outlet_res.clone(), axis)?;
-
-                                            outlets.insert(outlet, stacked_res);
-                                        } else {
-                                            outlets.insert(outlet, outlet_res.clone());
+                                            stacked_outlets.insert(outlet, stacked_res);
                                         }
+                                        outlets.insert(outlet, outlet_res.clone());
                                     }
                                 }
                             }
                         }
 
-                        full_results = outlets.into_values().collect_vec();
+                        // now extend with stacked elements
+                        let mut pre_stacked_outlets = outlets.clone();
+                        pre_stacked_outlets.extend(stacked_outlets);
+
+                        let outlets = outlets.into_values().collect_vec();
+
+                        full_results = pre_stacked_outlets.into_values().collect_vec();
 
                         let output_states = output_state_idx(output_mappings);
                         let input_states = input_state_idx(&input_mappings);
 
-                        assert_eq!(input_states.len(), output_states.len());
+                        assert_eq!(
+                            input_states.len(),
+                            output_states.len(),
+                            "input and output states must be the same length, got {:?} and {:?}",
+                            input_mappings,
+                            output_mappings
+                        );
 
                         for (input_idx, output_idx) in input_states.iter().zip(output_states) {
-                            values[*input_idx] = full_results[output_idx].clone();
+                            assert_eq!(
+                                values[*input_idx].dims(),
+                                outlets[output_idx].dims(),
+                                "input and output dims must be the same, got {:?} and {:?}",
+                                values[*input_idx].dims(),
+                                outlets[output_idx].dims()
+                            );
+                            values[*input_idx] = outlets[output_idx].clone();
                         }
                     }
 

tests/integration_tests.rs

@@ -200,7 +200,7 @@ mod native_tests {
         "1l_tiny_div",
     ];
 
-    const TESTS: [&str; 91] = [
+    const TESTS: [&str; 92] = [
         "1l_mlp", //0
         "1l_slice",
         "1l_concat",
@@ -296,6 +296,7 @@ mod native_tests {
         "reducel1",
         "reducel2", // 89
         "1l_lppool",
+        "lstm_large", // 91
     ];
 
     const WASM_TESTS: [&str; 46] = [
@@ -534,7 +535,7 @@ mod native_tests {
             }
         });
 
-            seq!(N in 0..=90 {
+            seq!(N in 0..=91 {
 
             #(#[test_case(TESTS[N])])*
             #[ignore]