mlx-cluster 0.0.3__tar.gz → 0.0.5__tar.gz

{mlx_cluster-0.0.3/mlx_cluster.egg-info → mlx_cluster-0.0.5}/PKG-INFO

@@ -1,7 +1,7 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.4
 Name: mlx_cluster
-Version: 0.0.3
-Summary: C++ and Metal extensions for MLX CTC Loss
+Version: 0.0.5
+Summary: C++ extension for generating random graphs
 Author-email: Vinay Pandya <vinayharshadpandya27@gmail.com>
 Project-URL: Homepage, https://github.com/vinayhpandya/mlx_cluster
 Project-URL: Issues, https://github.com/vinayhpandya/mlx_cluster/Issues
@@ -9,15 +9,22 @@ Classifier: Development Status :: 3 - Alpha
 Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: C++
 Classifier: License :: OSI Approved :: MIT License
-Classifier: Operating System :: OS Independent
+Classifier: Operating System :: MacOS
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Provides-Extra: dev
 Provides-Extra: test
+Requires-Dist: mlx-graphs>=0.0.8; extra == "test"
+Requires-Dist: torch>=2.2.0; extra == "test"
+Requires-Dist: mlx>=0.26.0; extra == "test"
 Requires-Dist: pytest==7.4.4; extra == "test"
-Requires-Dist: torch; extra == "test"
-Requires-Dist: mlx_graphs; extra == "test"
+Requires-Dist: scipy>=1.13.0; extra == "test"
+Requires-Dist: requests==2.31.0; extra == "test"
+Requires-Dist: fsspec[http]==2024.2.0; extra == "test"
+Requires-Dist: tqdm==4.66.1; extra == "test"
+Dynamic: license-file
+Dynamic: requires-python
 
 # mlx_cluster
 
@@ -50,24 +57,37 @@ for testing purposes you need to have `mlx-graphs`  and `torch_geometric` instal
 
 
 ```
-from mlx_graphs.utils.sorting import sort_edge_index
-from mlx_graphs.loaders import Dataloader
-from mlx_graphs_extension import random_walk
+# Can also use mlx for generating starting indices
+import torch
+from torch.utils.data import DataLoader
+
+loader = DataLoader(range(2708), batch_size=2000)
+start_indices = next(iter(loader))
 
 
+from mlx_graphs.datasets import PlanetoidDataset
+from mlx_graphs.utils.sorting import sort_edge_index
+from torch.utils.data import DataLoader
+from mlx_cluster import random_walk
+
 cora_dataset = PlanetoidDataset(name="cora", base_dir="~")
-start = mx.arange(0, 1000)
-start_time = time.time()
+# For some reason int_64t and int_32t are not compatible
 edge_index = cora_dataset.graphs[0].edge_index.astype(mx.int64)
-num_nodes = cora_dataset.graphs[0].num_nodes
+
+# Convert edge index into a CSR matrix
 sorted_edge_index = sort_edge_index(edge_index=edge_index)
 row_mlx = sorted_edge_index[0][0]
 col_mlx = sorted_edge_index[0][1]
-unique_vals, counts_mlx = np.unique(np.array(row_mlx, copy=False), return_counts=True)
+_, counts_mlx = np.unique(np.array(row_mlx, copy=False), return_counts=True)
 cum_sum_mlx = counts_mlx.cumsum()
-rand = mx.random.uniform(shape=[start.shape[0], 100])
 row_ptr_mlx = mx.concatenate([mx.array([0]), mx.array(cum_sum_mlx)])
-random_walk(row_ptr_mlx, col_mlx, start, rand, 1000, stream = mx.gpu)
+start_indices = mx.array(start_indices.numpy())
+
+rand_data = mx.random.uniform(shape=[start_indices.shape[0], 5])
+
+node_sequence = random_walk(
+    row_ptr_mlx, col_mlx, start_indices, rand_data, 5, stream=mx.cpu
+)
 ```
 
 ## TODO

{mlx_cluster-0.0.3 → mlx_cluster-0.0.5}/README.md

@@ -29,24 +29,37 @@ for testing purposes you need to have `mlx-graphs`  and `torch_geometric` instal
 
 
 ```
-from mlx_graphs.utils.sorting import sort_edge_index
-from mlx_graphs.loaders import Dataloader
-from mlx_graphs_extension import random_walk
+# Can also use mlx for generating starting indices
+import torch
+from torch.utils.data import DataLoader
+
+loader = DataLoader(range(2708), batch_size=2000)
+start_indices = next(iter(loader))
 
 
+from mlx_graphs.datasets import PlanetoidDataset
+from mlx_graphs.utils.sorting import sort_edge_index
+from torch.utils.data import DataLoader
+from mlx_cluster import random_walk
+
 cora_dataset = PlanetoidDataset(name="cora", base_dir="~")
-start = mx.arange(0, 1000)
-start_time = time.time()
+# For some reason int_64t and int_32t are not compatible
 edge_index = cora_dataset.graphs[0].edge_index.astype(mx.int64)
-num_nodes = cora_dataset.graphs[0].num_nodes
+
+# Convert edge index into a CSR matrix
 sorted_edge_index = sort_edge_index(edge_index=edge_index)
 row_mlx = sorted_edge_index[0][0]
 col_mlx = sorted_edge_index[0][1]
-unique_vals, counts_mlx = np.unique(np.array(row_mlx, copy=False), return_counts=True)
+_, counts_mlx = np.unique(np.array(row_mlx, copy=False), return_counts=True)
 cum_sum_mlx = counts_mlx.cumsum()
-rand = mx.random.uniform(shape=[start.shape[0], 100])
 row_ptr_mlx = mx.concatenate([mx.array([0]), mx.array(cum_sum_mlx)])
-random_walk(row_ptr_mlx, col_mlx, start, rand, 1000, stream = mx.gpu)
+start_indices = mx.array(start_indices.numpy())
+
+rand_data = mx.random.uniform(shape=[start_indices.shape[0], 5])
+
+node_sequence = random_walk(
+    row_ptr_mlx, col_mlx, start_indices, rand_data, 5, stream=mx.cpu
+)
 ```
 
 ## TODO

mlx_cluster-0.0.5/bindings.cpp

@@ -0,0 +1,81 @@
+#include <nanobind/nanobind.h>
+#include <nanobind/stl/variant.h>
+#include <random_walks/RandomWalk.h>
+#include <random_walks/BiasedRandomWalk.h>
+
+namespace nb = nanobind;
+using namespace nb::literals;
+using namespace mlx::core;
+
+NB_MODULE(_ext, m){
+
+    m.def(
+      "random_walk",
+      [](const mx::array& rowptr,
+        const mx::array& col,
+        const mx::array& start,
+        const mx::array& rand,
+        int  walk_length,
+        nb::object stream = nb::none()) {
+
+          // call the real C++ implementation
+          auto outs = mlx_random_walk::random_walk(
+              rowptr, col, start, rand, walk_length,
+              stream.is_none() ? mx::StreamOrDevice{}
+                              : nb::cast<mx::StreamOrDevice>(stream));
+
+          // vector -> tuple (move avoids a copy)
+          return nb::make_tuple(std::move(outs[0]), std::move(outs[1]));
+      },
+      "rowptr"_a, "col"_a, "start"_a, "rand"_a, "walk_length"_a,
+      nb::kw_only(), "stream"_a = nb::none(),
+      R"(
+          Uniform random walks.
+
+          Returns:
+              (nodes, edges) tuple of arrays
+      )",
+      nb::rv_policy::move);
+
+      m.def(
+        "rejection_sampling",
+        [](const mx::array& rowptr,
+          const mx::array& col,
+          const mx::array& start,
+          int walk_length,
+          float p,
+          float q,
+          nb::object stream = nb::none()
+        ){
+          auto outs = mlx_biased_random_walk::rejection_sampling(
+            rowptr, col, start, walk_length, p, q,
+            stream.is_none() ? mx::StreamOrDevice{}
+                              : nb::cast<mx::StreamOrDevice>(stream));
+          return nb::make_tuple(std::move(outs[0]), std::move(outs[1]));
+        },
+        "rowptr"_a,
+        "col"_a,
+        "start"_a,
+        "walk_length"_a,
+        "p"_a,
+        "q"_a,
+        nb::kw_only(), "stream"_a = nb::none(),
+      R"(
+        Sample nodes from the graph by sampling neighbors based
+        on probablity p and q
+
+        Args:
+            rowptr (array): rowptr of graph in csr format.
+            col (array): edges in csr format.
+            start (array): starting node of graph from which 
+                            biased sampling will be performed.
+            walk_length (int) : walk length of random graph
+            p : Likelihood of immediately revisiting a node in the walk.
+            q : Control parameter to interpolate between
+                breadth-first strategy and depth-first strategy
+
+        Returns:
+            (nodes, edges) tuple of arrays
+      )",
+      nb::rv_policy::move);
+}

{mlx_cluster-0.0.3 → mlx_cluster-0.0.5/mlx_cluster.egg-info}/PKG-INFO

@@ -1,7 +1,7 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.4
 Name: mlx_cluster
-Version: 0.0.3
-Summary: C++ and Metal extensions for MLX CTC Loss
+Version: 0.0.5
+Summary: C++ extension for generating random graphs
 Author-email: Vinay Pandya <vinayharshadpandya27@gmail.com>
 Project-URL: Homepage, https://github.com/vinayhpandya/mlx_cluster
 Project-URL: Issues, https://github.com/vinayhpandya/mlx_cluster/Issues
@@ -9,15 +9,22 @@ Classifier: Development Status :: 3 - Alpha
 Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: C++
 Classifier: License :: OSI Approved :: MIT License
-Classifier: Operating System :: OS Independent
+Classifier: Operating System :: MacOS
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Provides-Extra: dev
 Provides-Extra: test
+Requires-Dist: mlx-graphs>=0.0.8; extra == "test"
+Requires-Dist: torch>=2.2.0; extra == "test"
+Requires-Dist: mlx>=0.26.0; extra == "test"
 Requires-Dist: pytest==7.4.4; extra == "test"
-Requires-Dist: torch; extra == "test"
-Requires-Dist: mlx_graphs; extra == "test"
+Requires-Dist: scipy>=1.13.0; extra == "test"
+Requires-Dist: requests==2.31.0; extra == "test"
+Requires-Dist: fsspec[http]==2024.2.0; extra == "test"
+Requires-Dist: tqdm==4.66.1; extra == "test"
+Dynamic: license-file
+Dynamic: requires-python
 
 # mlx_cluster
 
@@ -50,24 +57,37 @@ for testing purposes you need to have `mlx-graphs`  and `torch_geometric` instal
 
 
 ```
-from mlx_graphs.utils.sorting import sort_edge_index
-from mlx_graphs.loaders import Dataloader
-from mlx_graphs_extension import random_walk
+# Can also use mlx for generating starting indices
+import torch
+from torch.utils.data import DataLoader
+
+loader = DataLoader(range(2708), batch_size=2000)
+start_indices = next(iter(loader))
 
 
+from mlx_graphs.datasets import PlanetoidDataset
+from mlx_graphs.utils.sorting import sort_edge_index
+from torch.utils.data import DataLoader
+from mlx_cluster import random_walk
+
 cora_dataset = PlanetoidDataset(name="cora", base_dir="~")
-start = mx.arange(0, 1000)
-start_time = time.time()
+# For some reason int_64t and int_32t are not compatible
 edge_index = cora_dataset.graphs[0].edge_index.astype(mx.int64)
-num_nodes = cora_dataset.graphs[0].num_nodes
+
+# Convert edge index into a CSR matrix
 sorted_edge_index = sort_edge_index(edge_index=edge_index)
 row_mlx = sorted_edge_index[0][0]
 col_mlx = sorted_edge_index[0][1]
-unique_vals, counts_mlx = np.unique(np.array(row_mlx, copy=False), return_counts=True)
+_, counts_mlx = np.unique(np.array(row_mlx, copy=False), return_counts=True)
 cum_sum_mlx = counts_mlx.cumsum()
-rand = mx.random.uniform(shape=[start.shape[0], 100])
 row_ptr_mlx = mx.concatenate([mx.array([0]), mx.array(cum_sum_mlx)])
-random_walk(row_ptr_mlx, col_mlx, start, rand, 1000, stream = mx.gpu)
+start_indices = mx.array(start_indices.numpy())
+
+rand_data = mx.random.uniform(shape=[start_indices.shape[0], 5])
+
+node_sequence = random_walk(
+    row_ptr_mlx, col_mlx, start_indices, rand_data, 5, stream=mx.cpu
+)
 ```
 
 ## TODO

{mlx_cluster-0.0.3 → mlx_cluster-0.0.5}/mlx_cluster.egg-info/SOURCES.txt

@@ -6,9 +6,6 @@ bindings.cpp
 pyproject.toml
 setup.py
 mlx_cluster/__init__.py
-mlx_cluster/_ext.cpython-311-darwin.so
-mlx_cluster/libmlx.dylib
-mlx_cluster/libmlx_cluster.dylib
 mlx_cluster/mlx_cluster.metallib
 mlx_cluster.egg-info/PKG-INFO
 mlx_cluster.egg-info/SOURCES.txt

mlx_cluster-0.0.5/mlx_cluster.egg-info/requires.txt

@@ -0,0 +1,12 @@
+
+[dev]
+
+[test]
+mlx-graphs>=0.0.8
+torch>=2.2.0
+mlx>=0.26.0
+pytest==7.4.4
+scipy>=1.13.0
+requests==2.31.0
+fsspec[http]==2024.2.0
+tqdm==4.66.1

{mlx_cluster-0.0.3 → mlx_cluster-0.0.5}/pyproject.toml

@@ -1,39 +1,48 @@
 [project]
 name = "mlx_cluster"
-version = "0.0.3"
+version = "0.0.5"
 authors = [
   { name = "Vinay Pandya", email = "vinayharshadpandya27@gmail.com" },
 ]
-description = "C++ and Metal extensions for MLX CTC Loss"
+description = "C++ extension for generating random graphs"
 readme = "README.md"
-requires-python = ">=3.8"
+requires-python = ">=3.10"
 classifiers = [
   "Development Status :: 3 - Alpha",
   "Programming Language :: Python :: 3",
   "Programming Language :: C++",
   "License :: OSI Approved :: MIT License",
-  "Operating System :: OS Independent",
+  "Operating System :: MacOS",
 ]
 
 [project.optional-dependencies]
 dev = []
 test = [
+  "mlx-graphs>=0.0.8",
+  "torch>=2.2.0",
+  "mlx>=0.26.0",
   "pytest==7.4.4",
-  "torch",
-  "mlx_graphs",
+  "scipy>=1.13.0",
+  "requests==2.31.0",
+  "fsspec[http]==2024.2.0",
+  "tqdm==4.66.1",
 ]
+
 [project.urls]
 Homepage = "https://github.com/vinayhpandya/mlx_cluster"
 Issues = "https://github.com/vinayhpandya/mlx_cluster/Issues"
 
+[tool.pytest.ini_options]
+addopts = "-ra"
+markers = [
+    "slow: marks tests that download data, compile kernels, or are otherwise time-consuming (deselect with -m 'not slow')",
+]
 
 [build-system]
 requires = [
+  "nanobind==2.4.0",
   "setuptools>=42",
   "cmake>=3.24",
-  "mlx>=0.17.0",
-  "nanobind@git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4",
+  "mlx>=0.26.0", 
 ]
-
-
 build-backend = "setuptools.build_meta"

{mlx_cluster-0.0.3 → mlx_cluster-0.0.5}/random_walks/BiasedRandomWalk.cpp

@@ -15,7 +15,8 @@
 #endif
 #include "random_walks/BiasedRandomWalk.h"
 
-namespace mlx::core {
+
+namespace mlx_biased_random_walk {
 
     bool inline is_neighbor(const int64_t *rowptr, const int64_t *col, int64_t v,
                         int64_t w) {
@@ -27,21 +28,20 @@ namespace mlx::core {
         return false;
     }
 
-    void BiasedRandomWalk::eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs) {
+    void BiasedRandomWalk::eval_cpu(const std::vector<mx::array>& inputs, std::vector<mx::array>& outputs) {
         auto& rowptr = inputs[0];
         auto& col = inputs[1];
         auto& start = inputs[2];
         auto& rand = inputs[3];
         int numel = start.size();
-
+        std::cout<<"Inside biased random walk"<<std::endl;
         // Initialize outputs
         assert(outputs.size() == 2);
         // Allocate memory for outputs if not already allocated
-        outputs[0].set_data(allocator::malloc_or_wait(numel*(walk_length_+1)*sizeof(int64_t)));
-        outputs[1].set_data(allocator::malloc_or_wait(numel*walk_length_*sizeof(int64_t)));
+        outputs[0].set_data(mx::allocator::malloc(numel*(walk_length_+1)*sizeof(int64_t)));
+        outputs[1].set_data(mx::allocator::malloc(numel*walk_length_*sizeof(int64_t)));
         auto& n_out = outputs[0];
         auto& e_out = outputs[1];
-
         auto* n_out_ptr = n_out.data<int64_t>();
         auto* e_out_ptr = e_out.data<int64_t>();
         auto* start_values = start.data<int64_t>();
@@ -53,7 +53,7 @@ namespace mlx::core {
         double prob_0 = 1. / p_ / max_prob;
         double prob_1 = 1. / max_prob;
         double prob_2 = 1. / q_ / max_prob;
-
+        
         for (int64_t n = 0; n < numel; n++) {
             int64_t t = start_values[n], v, x, e_cur, row_start, row_end;
             n_out_ptr[n * (walk_length_ + 1)] = t;
@@ -91,7 +91,6 @@ namespace mlx::core {
                     break;
                     }
                 }
-
                 n_out_ptr[n * (walk_length_ + 1) + (l + 1)] = x;
                 e_out_ptr[n * walk_length_ + l] = e_cur;
                 t = v;
@@ -101,9 +100,9 @@ namespace mlx::core {
    
     };
 
-    std::vector<array> BiasedRandomWalk::jvp(
-    const std::vector<array>& primals,
-    const std::vector<array>& tangents,
+    std::vector<mx::array> BiasedRandomWalk::jvp(
+    const std::vector<mx::array>& primals,
+    const std::vector<mx::array>& tangents,
     const std::vector<int>& argnums)
 {
     // Random walk is not differentiable, so we return zero tangents
@@ -121,8 +120,8 @@ namespace mlx::core {
 //         int numel = start.size();
         
 //         assert(outputs.size() == 2);
-//         outputs[0].set_data(allocator::malloc_or_wait(numel * (walk_length_ + 1) * sizeof(int64_t)));
-//         outputs[1].set_data(allocator::malloc_or_wait(numel * walk_length_ * sizeof(int64_t)));
+//         outputs[0].set_data(allocator::malloc(numel * (walk_length_ + 1) * sizeof(int64_t)));
+//         outputs[1].set_data(allocator::malloc(numel * walk_length_ * sizeof(int64_t)));
 //         std::cout<<"after setting data"<<std::endl;
 //         auto& s = stream();
 //         auto& d = metal::device(s.device);
@@ -148,47 +147,42 @@ namespace mlx::core {
 //     }
 // #endif
 void BiasedRandomWalk::eval_gpu(
-       const std::vector<array>& inputs, std::vector<array>& outputs
+       const std::vector<mx::array>& inputs, std::vector<mx::array>& outputs
   )
 {
     throw std::runtime_error("Random walk has no GPU implementation.");
 }
-std::vector<array> BiasedRandomWalk::vjp(
-    const std::vector<array>& primals,
-    const std::vector<array>& cotangents,
+std::vector<mx::array> BiasedRandomWalk::vjp(
+    const std::vector<mx::array>& primals,
+    const std::vector<mx::array>& cotangents,
     const std::vector<int>& argnums,
-    const std::vector<array>& outputs)
+    const std::vector<mx::array>& outputs)
 {
     // Random walk is not differentiable, so we return zero gradients
     throw std::runtime_error("Random walk has no JVP implementation.");
 }
 
-std::pair<std::vector<array>, std::vector<int>> BiasedRandomWalk::vmap(
-    const std::vector<array>& inputs,
+std::pair<std::vector<mx::array>, std::vector<int>> BiasedRandomWalk::vmap(
+    const std::vector<mx::array>& inputs,
     const std::vector<int>& axes)
 {
     throw std::runtime_error("vmap not implemented for biasedRandomWalk");
 }
 
-bool BiasedRandomWalk::is_equivalent(const Primitive& other) const
-{
-    throw std::runtime_error("biased Random walk has no GPU implementation.");
-}
-
-std::vector<std::vector<int>> BiasedRandomWalk::output_shapes(const std::vector<array>& inputs)
+bool BiasedRandomWalk::is_equivalent(const mx::Primitive& other) const
 {
     throw std::runtime_error("biased Random walk has no GPU implementation.");
 }
 
-array rejection_sampling(const array& rowptr, const array& col, const array& start, int walk_length, const double p, 
-       const double q, StreamOrDevice s)
+std::vector<mx::array> rejection_sampling(const mx::array& rowptr, const mx::array& col, const mx::array& start, int walk_length, const double p, 
+       const double q, mx::StreamOrDevice s)
 {   
     int nodes = start.size();
     auto primitive = std::make_shared<BiasedRandomWalk>(to_stream(s), walk_length, p, q);
-    return array::make_arrays({{nodes,walk_length+1},{nodes, walk_length}},
+    return mx::array::make_arrays({{nodes,walk_length+1},{nodes, walk_length}},
      {rowptr.dtype(), rowptr.dtype()},
      primitive,
      {rowptr, col, start}
-    )[0];
+    );
 }
 }

mlx_cluster-0.0.5/random_walks/BiasedRandomWalk.h

@@ -0,0 +1,65 @@
+#pragma once
+
+#include <mlx/array.h>
+#include <mlx/ops.h>
+#include <mlx/primitives.h>
+
+namespace mx = mlx::core;
+namespace mlx_biased_random_walk{
+
+    class BiasedRandomWalk : public mx::Primitive {
+        public:
+            BiasedRandomWalk(mx::Stream stream, int walk_length, double p, double q)
+            : mx::Primitive(stream), walk_length_(walk_length), p_(p), q_(q) {}
+            void eval_cpu(const std::vector<mx::array>& inputs, std::vector<mx::array>& outputs)
+            override;
+            void eval_gpu(const std::vector<mx::array>& inputs, std::vector<mx::array>& outputs)
+            override;
+
+            /** The Jacobian-vector product. */
+            std::vector<mx::array> jvp(
+                const std::vector<mx::array>& primals,
+                const std::vector<mx::array>& tangents,
+                const std::vector<int>& argnums) override;
+          
+            /** The vector-Jacobian product. */
+            std::vector<mx::array> vjp(
+                const std::vector<mx::array>& primals,
+                const std::vector<mx::array>& cotangents,
+                const std::vector<int>& argnums,
+                const std::vector<mx::array>& outputs) override;
+          
+            /**
+             * The primitive must know how to vectorize itself across
+             * the given axes. The output is a pair containing the array
+             * representing the vectorized computation and the axis which
+             * corresponds to the output vectorized dimension.
+             */
+            std::pair<std::vector<mx::array>, std::vector<int>> vmap(
+                const std::vector<mx::array>& inputs,
+                const std::vector<int>& axes) override;
+
+            /** Print the primitive. */
+            virtual const char* name() const override {
+                return "biased random walk implementation";
+            }
+
+            /** Equivalence check **/
+            bool is_equivalent(const mx::Primitive& other) const override;
+        
+        private:
+            int walk_length_;
+            double p_;
+            double q_;
+
+    };
+
+    std::vector<mx::array> rejection_sampling(const mx::array& rowptr,
+     const mx::array& col,
+    const mx::array& start,
+       int walk_length,
+       const double p, 
+       const double q,
+        mx::StreamOrDevice s = {});
+
+};

{mlx_cluster-0.0.3 → mlx_cluster-0.0.5}/random_walks/RandomWalk.cpp

@@ -1,7 +1,7 @@
 #include <cassert>
 #include <iostream>
 #include <sstream>
-
+#include <dlfcn.h>
 #include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/utils.h"
@@ -14,19 +14,31 @@
 #endif
 #include "random_walks/RandomWalk.h"
 
-namespace mlx::core {
-    void RandomWalk::eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs) {
+
+namespace mlx_random_walk {
+    std::string current_binary_dir() {
+        static std::string binary_dir = []() {
+          Dl_info info;
+          if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+            throw std::runtime_error("Unable to get current binary dir.");
+          }
+          return std::filesystem::path(info.dli_fname).parent_path().string();
+        }();
+        return binary_dir;
+      }
+
+    void RandomWalk::eval_cpu(const std::vector<mx::array>& inputs, std::vector<mx::array>& outputs) {
     auto& rowptr = inputs[0];
     auto& col = inputs[1];
     auto& start = inputs[2];
     auto& rand = inputs[3];
     int numel = start.size();
-
+    std::cout<<"Its really inside cpu"<<std::endl;
     // Initialize outputs
     assert(outputs.size() == 2);
     // Allocate memory for outputs if not already allocated
-    outputs[0].set_data(allocator::malloc_or_wait(numel*(walk_length_+1)*sizeof(int64_t)));
-    outputs[1].set_data(allocator::malloc_or_wait(numel*walk_length_*sizeof(int64_t)));
+    outputs[0].set_data(mx::allocator::malloc(numel*(walk_length_+1)*sizeof(int64_t)));
+    outputs[1].set_data(mx::allocator::malloc(numel*walk_length_*sizeof(int64_t)));
     auto& n_out = outputs[0];
     auto& e_out = outputs[1];
 
@@ -37,7 +49,6 @@ namespace mlx::core {
     auto* col_values = col.data<int64_t>();
     auto* rand_values = rand.data<float>();
 
-    std::cout<<"After evaluating outputs"<<std::endl;
     for (int64_t n = 0; n < numel; n++) {
         int64_t n_cur = start_values[n];
         n_out_ptr[n * (walk_length_ + 1)] = n_cur;
@@ -61,9 +72,9 @@ namespace mlx::core {
    
     };
 
-    std::vector<array> RandomWalk::jvp(
-    const std::vector<array>& primals,
-    const std::vector<array>& tangents,
+    std::vector<mx::array> RandomWalk::jvp(
+    const std::vector<mx::array>& primals,
+    const std::vector<mx::array>& tangents,
     const std::vector<int>& argnums)
 {
     // Random walk is not differentiable, so we return zero tangents
@@ -71,8 +82,8 @@ namespace mlx::core {
 }
 #ifdef _METAL_
 void RandomWalk::eval_gpu(
-        const std::vector<array>& inputs,
-        std::vector<array>& outputs
+        const std::vector<mx::array>& inputs,
+        std::vector<mx::array>& outputs
     ){
        auto& rowptr = inputs[0];
         auto& col = inputs[1];
@@ -81,17 +92,16 @@ void RandomWalk::eval_gpu(
         int numel = start.size();
         
         assert(outputs.size() == 2);
-        outputs[0].set_data(allocator::malloc_or_wait(numel * (walk_length_ + 1) * sizeof(int64_t)));
-        outputs[1].set_data(allocator::malloc_or_wait(numel * walk_length_ * sizeof(int64_t)));
-        std::cout<<"after setting data"<<std::endl;
+        outputs[0].set_data(mx::allocator::malloc(numel * (walk_length_ + 1) * sizeof(int64_t)));
+        outputs[1].set_data(mx::allocator::malloc(numel * walk_length_ * sizeof(int64_t)));
         auto& s = stream();
-        auto& d = metal::device(s.device);
-
-        d.register_library("mlx_cluster");
-        auto kernel = d.get_kernel("random_walk", "mlx_cluster");
+        auto& d = mx::metal::device(s.device);
+        std::cout<<"Its really inside gpu"<<std::endl;
+        auto lib = d.get_library("mlx_cluster", current_binary_dir());
+        auto kernel = d.get_kernel("random_walk", lib);
 
         auto& compute_encoder = d.get_command_encoder(s.index);
-        compute_encoder->setComputePipelineState(kernel);
+        compute_encoder.set_compute_pipeline_state(kernel);
 
         compute_encoder.set_input_array(rowptr, 0);
         compute_encoder.set_input_array(col, 1);
@@ -99,50 +109,46 @@ void RandomWalk::eval_gpu(
         compute_encoder.set_input_array(rand, 3);
         compute_encoder.set_output_array(outputs[0], 4);
         compute_encoder.set_output_array(outputs[1], 5);
-        compute_encoder->setBytes(&walk_length_, sizeof(int32), 6);
+        compute_encoder.set_bytes(&walk_length_, sizeof(walk_length_), 6);
 
         MTL::Size grid_size = MTL::Size(numel, 1, 1);
         MTL::Size thread_group_size = MTL::Size(kernel->maxTotalThreadsPerThreadgroup(), 1, 1);
 
-        compute_encoder.dispatchThreads(grid_size, thread_group_size);
+        compute_encoder.dispatch_threads(grid_size, thread_group_size);
     }
 #endif
 
-std::vector<array> RandomWalk::vjp(
-    const std::vector<array>& primals,
-    const std::vector<array>& cotangents,
+std::vector<mx::array> RandomWalk::vjp(
+    const std::vector<mx::array>& primals,
+    const std::vector<mx::array>& cotangents,
     const std::vector<int>& argnums,
-    const std::vector<array>& outputs)
+    const std::vector<mx::array>& outputs)
 {
     // Random walk is not differentiable, so we return zero gradients
     throw std::runtime_error("Random walk has no GPU implementation.");
 }
 
-std::pair<std::vector<array>, std::vector<int>> RandomWalk::vmap(
-    const std::vector<array>& inputs,
+std::pair<std::vector<mx::array>, std::vector<int>> RandomWalk::vmap(
+    const std::vector<mx::array>& inputs,
     const std::vector<int>& axes)
 {
     throw std::runtime_error("vmap not implemented for RandomWalk");
 }
 
-bool RandomWalk::is_equivalent(const Primitive& other) const
-{
-    throw std::runtime_error("Random walk has no GPU implementation.");
-}
-
-std::vector<std::vector<int>> RandomWalk::output_shapes(const std::vector<array>& inputs)
+bool RandomWalk::is_equivalent(const mx::Primitive& other) const
 {
     throw std::runtime_error("Random walk has no GPU implementation.");
 }
 
-array random_walk(const array& rowptr, const array& col, const array& start, const array& rand, int walk_length, StreamOrDevice s)
+std::vector<mx::array> random_walk(const mx::array& rowptr, const mx::array& col, const mx::array& start, const mx::array& rand, int walk_length, mx::StreamOrDevice s)
 {   
+    std::cout<<"Inside random walk"<<std::endl;
     int nodes = start.size();
     auto primitive = std::make_shared<RandomWalk>(walk_length, to_stream(s));
-    return array::make_arrays({{nodes,walk_length+1},{nodes, walk_length}},
+    return mx::array::make_arrays({{nodes,walk_length+1},{nodes, walk_length}},
      {start.dtype(), start.dtype()},
      primitive,
      {rowptr, col, start, rand}
-    )[0];
+    );
 }
 }

mlx_cluster-0.0.5/random_walks/RandomWalk.h

@@ -0,0 +1,62 @@
+#pragma once
+
+#include <mlx/array.h>
+#include <mlx/ops.h>
+#include <mlx/primitives.h>
+
+namespace mx = mlx::core;
+namespace mlx_random_walk{
+
+    class RandomWalk : public mx::Primitive {
+        public:
+            explicit RandomWalk(int walk_length, mx::Stream stream):
+                mx::Primitive(stream), walk_length_(walk_length) {};
+            void eval_cpu(const std::vector<mx::array>& inputs, std::vector<mx::array>& outputs)
+            override;
+            void eval_gpu(const std::vector<mx::array>& inputs, std::vector<mx::array>& outputs)
+            override;
+
+            /** The Jacobian-vector product. */
+            std::vector<mx::array> jvp(
+                const std::vector<mx::array>& primals,
+                const std::vector<mx::array>& tangents,
+                const std::vector<int>& argnums) override;
+          
+            /** The vector-Jacobian product. */
+            std::vector<mx::array> vjp(
+                const std::vector<mx::array>& primals,
+                const std::vector<mx::array>& cotangents,
+                const std::vector<int>& argnums,
+                const std::vector<mx::array>& outputs) override;
+          
+            /**
+             * The primitive must know how to vectorize itself across
+             * the given axes. The output is a pair containing the array
+             * representing the vectorized computation and the axis which
+             * corresponds to the output vectorized dimension.
+             */
+            std::pair<std::vector<mx::array>, std::vector<int>> vmap(
+                const std::vector<mx::array>& inputs,
+                const std::vector<int>& axes) override;
+
+            /** Print the primitive. */
+            virtual const char* name() const override {
+                return "Random walk implementation";
+            }
+
+            /** Equivalence check **/
+            bool is_equivalent(const mx::Primitive& other) const override;
+        
+        private:
+            int walk_length_;
+
+    };
+
+    std::vector<mx::array> random_walk(const mx::array& rowptr,
+     const mx::array& col,
+    const mx::array& start,
+    const mx::array& rand,
+       int walk_length,
+        mx::StreamOrDevice s = {});
+
+};

{mlx_cluster-0.0.3 → mlx_cluster-0.0.5}/setup.py

@@ -4,20 +4,13 @@ from mlx import extension
 if __name__ == "__main__":
     setup(
         name="mlx_cluster",
-        version="0.0.3",
+        version="0.0.5",
         description="Sample C++ and Metal extensions for MLX primitives.",
         ext_modules=[extension.CMakeExtension("mlx_cluster._ext")],
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_cluster"],
         package_data={"mlx_cluster": ["*.so", "*.dylib", "*.metallib"]},
-        extras_require={
-            "dev": [],
-            "test": [
-                "mlx_graphs",
-                "torch",
-                "pytest",
-            ],
-        },
+        extras_require={"dev": []},
         zip_safe=False,
         python_requires=">=3.8",
     )

mlx_cluster-0.0.5/tests/test_random_walk.py

@@ -0,0 +1,72 @@
+import mlx.core as mx
+import numpy as np
+import time
+
+# Torch dataset
+import torch
+from torch.utils.data import DataLoader
+
+loader = DataLoader(range(2708), batch_size=2000)
+start_indices = next(iter(loader))
+
+
+from mlx_graphs.datasets import PlanetoidDataset
+from mlx_graphs.utils.sorting import sort_edge_index
+from torch.utils.data import DataLoader
+from mlx_cluster import random_walk
+import pytest
+import time
+
+@pytest.mark.slow  # give download/compile plenty of time on CI
+def test_random_walk(tmp_path):
+    """
+    Runs 1 000 random walks of length 10 on the Cora graph and checks:
+    1. output tensor shape == (num_start_nodes, walk_length + 1)
+    2. all returned node indices are valid ( < num_nodes )
+    """
+
+    # ---------- Dataset (downloaded to the temp dir) ----------
+    data_dir = tmp_path / "mlx_datasets"
+    cora = PlanetoidDataset(name="cora", base_dir=data_dir)
+
+    edge_index = cora.graphs[0].edge_index.astype(mx.int64)
+
+    # CSR conversion
+    sorted_edge_index = sort_edge_index(edge_index=edge_index)
+    row = sorted_edge_index[0][0]
+    col = sorted_edge_index[0][1]
+    _, counts = np.unique(np.array(row, copy=False), return_counts=True)
+    row_ptr = mx.concatenate([mx.array([0]), mx.array(counts.cumsum())])
+
+    # pick 1 000 random start nodes
+    num_starts = 1_000
+    rng = np.random.default_rng(42)
+    start_idx = mx.array(rng.integers(low=0, high=row.max().item() + 1,
+                                      size=num_starts, dtype=np.int64))
+
+    # random numbers for the kernel (shape [num_starts, walk_length])
+    walk_len = 10
+    rand_data = mx.random.uniform(shape=[num_starts, walk_len])
+
+    # ---------- Warm-up ----------
+    mx.eval(row_ptr, col, start_idx, rand_data)
+
+    # ---------- Run kernel ----------
+    t0 = time.time()
+    node_seq = random_walk(row_ptr, col, start_idx, rand_data,
+                           walk_len, stream=mx.cpu)
+    elapsed = time.time() - t0
+    print(f"Random-walk kernel took {elapsed:.3f} s")
+    print("Node sequence is ", node_seq[0])
+    # ---------- Assertions ----------
+    assert node_seq[0].shape == (num_starts, walk_len + 1)
+
+    # num_nodes = cora.graphs[0].num_nodes
+    # assert (node_seq < num_nodes).all().item(), \
+    #     "Random walk produced invalid node indices"
+    t0 = time.time()
+    node_seq_gpu = random_walk(row_ptr, col, start_idx, rand_data,
+                           walk_len, stream=mx.gpu)
+    elapsed = time.time() - t0
+    print(f"Random-walk kernel on gpu took {elapsed:.3f} s")
+    print("Node sequence is ", node_seq_gpu[0])

mlx_cluster-0.0.5/tests/test_rejection_sampling.py

@@ -0,0 +1,62 @@
+import mlx.core as mx
+import numpy as np
+import time
+import pytest
+
+# Torch dataset
+import torch
+from torch.utils.data import DataLoader
+
+loader = DataLoader(range(2708), batch_size=2000)
+start_indices = next(iter(loader))
+# random_walks = torch.ops.torch_cluster.random_walk(
+#     row_ptr, col, start_indices, 5, 1.0, 3.0
+# )
+
+from mlx_graphs.datasets import PlanetoidDataset
+from mlx_graphs.utils.sorting import sort_edge_index
+from torch.utils.data import DataLoader
+from mlx_cluster import rejection_sampling
+
+@pytest.mark.slow  # give download/compile plenty of time on CI
+def test_random_walk(tmp_path):
+    """
+    Runs 1 000 random walks of length 10 on the Cora graph and checks:
+    1. output tensor shape == (num_start_nodes, walk_length + 1)
+    2. all returned node indices are valid ( < num_nodes )
+    """
+
+    # ---------- Dataset (downloaded to the temp dir) ----------
+    data_dir = tmp_path / "mlx_datasets"
+    cora = PlanetoidDataset(name="cora", base_dir=data_dir)
+
+    edge_index = cora.graphs[0].edge_index.astype(mx.int64)
+
+    # CSR conversion
+    sorted_edge_index = sort_edge_index(edge_index=edge_index)
+    row = sorted_edge_index[0][0]
+    col = sorted_edge_index[0][1]
+    _, counts = np.unique(np.array(row, copy=False), return_counts=True)
+    row_ptr = mx.concatenate([mx.array([0]), mx.array(counts.cumsum())])
+
+    # pick 1 000 random start nodes
+    num_starts = 1_000
+    rng = np.random.default_rng(42)
+    start_idx = mx.array(rng.integers(low=0, high=row.max().item() + 1,
+                                      size=num_starts, dtype=np.int64))
+
+    # random numbers for the kernel (shape [num_starts, walk_length])
+    walk_len = 10
+    rand_data = mx.random.uniform(shape=[num_starts, walk_len])
+
+    # ---------- Warm-up ----------
+    mx.eval(row_ptr, col, start_idx, rand_data)
+
+    # ---------- Run kernel ----------
+    t0 = time.time()
+    node_seq = rejection_sampling(row_ptr, col, start_idx, walk_len, 1.0, 3.0, stream=mx.cpu)
+    elapsed = time.time() - t0
+    print(f"Random-walk kernel took {elapsed:.3f} s")
+    print("Node sequence is ", node_seq)
+    # ---------- Assertions ----------
+    assert node_seq[0].shape == (num_starts, walk_len + 1)

mlx_cluster-0.0.3/bindings.cpp

@@ -1,65 +0,0 @@
-#include <nanobind/nanobind.h>
-#include <nanobind/stl/variant.h>
-#include <random_walks/RandomWalk.h>
-#include <random_walks/BiasedRandomWalk.h>
-
-namespace nb = nanobind;
-using namespace nb::literals;
-using namespace mlx::core;
-
-NB_MODULE(_ext, m){
-
-      m.def(
-        "random_walk",
-        &random_walk,
-        "rowptr"_a,
-        "col"_a,
-        "start"_a,
-        "rand"_a,
-        "walk_length"_a,
-        nb::kw_only(),
-      "stream"_a = nb::none(),
-      R"(
-        uniformly sample a graph
-
-
-        Args:
-            rowptr (array): rowptr of graph in csr format.
-            col (array): edges in csr format.
-            walk_length (int) : walk length of random graph
-
-        Returns:
-            array: consisting of nodes visited on random walk
-      )");
-
-      m.def(
-        "rejection_sampling",
-        &rejection_sampling,
-        "rowptr"_a,
-        "col"_a,
-        "start"_a,
-        "walk_length"_a,
-        "p"_a,
-        "q"_a,
-        nb::kw_only(),
-      "stream"_a = nb::none(),
-      R"(
-        Sample nodes from the graph by sampling neighbors based
-        on probablity p and q
-
-
-        Args:
-            rowptr (array): rowptr of graph in csr format.
-            col (array): edges in csr format.
-            start (array): starting node of graph from which 
-                            biased sampling will be performed.
-            walk_length (int) : walk length of random graph
-            p : Likelihood of immediately revisiting a node in the walk.
-            q : Control parameter to interpolate between
-                breadth-first strategy and depth-first strategy
-
-        Returns:
-            array: consisting of nodes visited on random walk
-      )");
-}
-

mlx_cluster-0.0.3/mlx_cluster.egg-info/requires.txt

@@ -1,7 +0,0 @@
-
-[dev]
-
-[test]
-pytest==7.4.4
-torch
-mlx_graphs

mlx_cluster-0.0.3/random_walks/BiasedRandomWalk.h

@@ -1,66 +0,0 @@
-#pragma once
-
-#include <mlx/array.h>
-#include <mlx/ops.h>
-#include <mlx/primitives.h>
-
-namespace mlx::core{
-
-    class BiasedRandomWalk : public Primitive {
-        public:
-            BiasedRandomWalk(Stream stream, int walk_length, double p, double q)
-            : Primitive(stream), walk_length_(walk_length), p_(p), q_(q) {}
-            void eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs)
-            override;
-            void eval_gpu(const std::vector<array>& inputs, std::vector<array>& outputs)
-            override;
-
-            /** The Jacobian-vector product. */
-            std::vector<array> jvp(
-                const std::vector<array>& primals,
-                const std::vector<array>& tangents,
-                const std::vector<int>& argnums) override;
-
-            /** The vector-Jacobian product. */
-            std::vector<array> vjp(
-                const std::vector<array>& primals,
-                const std::vector<array>& cotangents,
-                const std::vector<int>& argnums,
-                const std::vector<array>& outputs) override;
-
-            /**
-             * The primitive must know how to vectorize itself across
-             * the given axes. The output is a pair containing the array
-             * representing the vectorized computation and the axis which
-             * corresponds to the output vectorized dimension.
-             */
-            std::pair<std::vector<array>, std::vector<int>> vmap(
-                const std::vector<array>& inputs,
-                const std::vector<int>& axes) override;
-
-            /** Print the primitive. */
-            void print(std::ostream& os) override {
-                os << "biased random walk implementation";
-            }
-
-            /** Equivalence check **/
-            bool is_equivalent(const Primitive& other) const override;
-
-            std::vector<std::vector<int>> output_shapes(const std::vector<array>& inputs) override;
-        
-        private:
-            int walk_length_;
-            double p_;
-            double q_;
-
-    };
-
-    array rejection_sampling(const array& rowptr,
-     const array& col,
-    const array& start,
-       int walk_length,
-       const double p, 
-       const double q,
-        StreamOrDevice s = {});
-
-};

mlx_cluster-0.0.3/random_walks/RandomWalk.h

@@ -1,63 +0,0 @@
-#pragma once
-
-#include <mlx/array.h>
-#include <mlx/ops.h>
-#include <mlx/primitives.h>
-
-namespace mlx::core{
-
-    class RandomWalk : public Primitive {
-        public:
-            explicit RandomWalk(int walk_length, Stream stream):
-                Primitive(stream), walk_length_(walk_length) {};
-            void eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs)
-            override;
-            void eval_gpu(const std::vector<array>& inputs, std::vector<array>& outputs)
-            override;
-
-            /** The Jacobian-vector product. */
-            std::vector<array> jvp(
-                const std::vector<array>& primals,
-                const std::vector<array>& tangents,
-                const std::vector<int>& argnums) override;
-
-            /** The vector-Jacobian product. */
-            std::vector<array> vjp(
-                const std::vector<array>& primals,
-                const std::vector<array>& cotangents,
-                const std::vector<int>& argnums,
-                const std::vector<array>& outputs) override;
-
-            /**
-             * The primitive must know how to vectorize itself across
-             * the given axes. The output is a pair containing the array
-             * representing the vectorized computation and the axis which
-             * corresponds to the output vectorized dimension.
-             */
-            std::pair<std::vector<array>, std::vector<int>> vmap(
-                const std::vector<array>& inputs,
-                const std::vector<int>& axes) override;
-
-            /** Print the primitive. */
-            void print(std::ostream& os) override {
-                os << "Random walk implementation";
-            }
-
-            /** Equivalence check **/
-            bool is_equivalent(const Primitive& other) const override;
-
-            std::vector<std::vector<int>> output_shapes(const std::vector<array>& inputs) override;
-        
-        private:
-            int walk_length_;
-
-    };
-
-    array random_walk(const array& rowptr,
-     const array& col,
-    const array& start,
-    const array& rand,
-       int walk_length,
-        StreamOrDevice s = {});
-
-};

mlx_cluster-0.0.3/tests/test_random_walk.py

@@ -1,35 +0,0 @@
-import mlx.core as mx
-import numpy as np
-import time
-
-# Torch dataset
-from torch.utils.data import DataLoader
-
-loader = DataLoader(range(2708), batch_size=2000)
-start_indices = next(iter(loader))
-
-from mlx_graphs.datasets import PlanetoidDataset
-from mlx_graphs.utils.sorting import sort_edge_index
-from torch.utils.data import DataLoader
-from mlx_cluster import random_walk
-
-cora_dataset = PlanetoidDataset(name="cora", base_dir="~")
-# For some reason int_64t and int_32t are not compatible
-edge_index = cora_dataset.graphs[0].edge_index.astype(mx.int64)
-# Convert edge index into a CSR matrix
-sorted_edge_index = sort_edge_index(edge_index=edge_index)
-row_mlx = sorted_edge_index[0][0]
-col_mlx = sorted_edge_index[0][1]
-_, counts_mlx = np.unique(np.array(row_mlx, copy=False), return_counts=True)
-cum_sum_mlx = counts_mlx.cumsum()
-row_ptr_mlx = mx.concatenate([mx.array([0]), mx.array(cum_sum_mlx)])
-start_indices = mx.array(start_indices.numpy())
-
-rand_data = mx.random.uniform(shape=[start_indices.shape[0], 5])
-start_time = time.time()
-
-node_sequence = random_walk(
-    row_ptr_mlx, col_mlx, start_indices, rand_data, 5, stream=mx.cpu
-)
-print("Time taken to complete 1000 random walks : ", time.time() - start_time)
-print("MLX random walks are", node_sequence)

mlx_cluster-0.0.3/tests/test_rejection_sampling.py

@@ -1,34 +0,0 @@
-import mlx.core as mx
-import numpy as np
-import time
-
-# Torch dataloader
-from torch.utils.data import DataLoader
-
-
-loader = DataLoader(range(2708), batch_size=2000)
-start_indices = next(iter(loader))
-# random_walks = torch.ops.torch_cluster.random_walk(
-#     row_ptr, col, start_indices, 5, 1.0, 3.0
-# )
-
-from mlx_graphs.datasets import PlanetoidDataset
-from mlx_graphs.utils.sorting import sort_edge_index
-from mlx_cluster import rejection_sampling
-
-cora_dataset = PlanetoidDataset(name="cora", base_dir="~")
-edge_index = cora_dataset.graphs[0].edge_index.astype(mx.int64)
-sorted_edge_index = sort_edge_index(edge_index=edge_index)
-row_mlx = sorted_edge_index[0][0]
-col_mlx = sorted_edge_index[0][1]
-_, counts_mlx = np.unique(np.array(row_mlx, copy=False), return_counts=True)
-cum_sum_mlx = counts_mlx.cumsum()
-row_ptr_mlx = mx.concatenate([mx.array([0]), mx.array(cum_sum_mlx)])
-start_indices = mx.array(start_indices.numpy())
-rand_data = mx.random.uniform(shape=[start_indices.shape[0], 5])
-start_time = time.time()
-node_sequence = rejection_sampling(
-    row_ptr_mlx, col_mlx, start_indices, 5, 1.0, 3.0, stream=mx.cpu
-)
-print("Time taken to complete random walks : ", time.time() - start_time)
-print(node_sequence)