kernel-lens 0.1.0__tar.gz

kernel_lens-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Duc Justin
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

kernel_lens-0.1.0/PKG-INFO

@@ -0,0 +1,47 @@
+Metadata-Version: 2.4
+Name: kernel-lens
+Version: 0.1.0
+Summary: A production-grade, multi-backend compiler for Triton kernels (ONNX & TensorRT)
+Author: Compiler Engineer
+License: MIT License
+        
+        Copyright (c) 2026 Duc Justin
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+        
+Project-URL: Source Code, https://github.com/yourusername/kernel-lens
+Project-URL: Bug Tracker, https://github.com/yourusername/kernel-lens/issues
+Classifier: Programming Language :: Python :: 3
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=2.1.0
+Requires-Dist: triton>=2.1.0
+Requires-Dist: onnx>=1.15.0
+Provides-Extra: ort
+Requires-Dist: onnxruntime-gpu>=1.17.0; extra == "ort"
+Provides-Extra: trt
+Requires-Dist: tensorrt>=8.6.0; extra == "trt"
+Provides-Extra: all
+Requires-Dist: onnxruntime-gpu>=1.17.0; extra == "all"
+Requires-Dist: tensorrt>=8.6.0; extra == "all"
+Dynamic: license-file

kernel_lens-0.1.0/kernel_lens/__init__.py

@@ -0,0 +1,11 @@
+# kernel_lens/__init__.py
+
+from .compiler.core import compile
+from .utils.deployment import extract_libs
+from .compiler.core import load
+
+__all__ = [
+    "compile",
+    "load",
+    "extract_libs",
+]

kernel_lens-0.1.0/kernel_lens/backends/builder.py

@@ -0,0 +1,129 @@
+import os
+import subprocess
+import urllib.request
+import tarfile
+import onnxruntime
+
+def build_ort_plugin(ort_plugins_dir: str, cache_dir: str):
+    """
+    Natively compiles the generated C++ files into a Shared Library (.so)
+    replacing the need for an external bash script.
+    """
+    # 1. Get exact ORT version from the current Python environment
+    ort_version = onnxruntime.__version__.split('+')[0]
+    # print(f"     [Builder] Detected ONNX Runtime v{ort_version}")
+    
+    ort_release_dir = os.path.join(cache_dir, f"onnxruntime-linux-x64-gpu-{ort_version}")
+    ort_tgz = f"{ort_release_dir}.tgz"
+    
+    # 2. Download exact matching C++ Developer Release if missing
+    if not os.path.exists(ort_release_dir):
+        url = f"https://github.com/microsoft/onnxruntime/releases/download/v{ort_version}/onnxruntime-linux-x64-gpu-{ort_version}.tgz"
+        # print(f"     [Builder] Downloading ORT C++ headers from {url}...")
+        urllib.request.urlretrieve(url, ort_tgz)
+        with tarfile.open(ort_tgz, "r:gz") as tar:
+            tar.extractall(path=cache_dir)
+            
+    ort_inc = os.path.join(ort_release_dir, "include")
+    ort_lib = os.path.join(ort_release_dir, "lib")
+    
+    # # 3. Dynamically find CUDA paths via nvcc
+    # print("     [Builder] Querying system for CUDA configuration...")
+    try:
+        nvcc_path = subprocess.check_output(["which", "nvcc"]).decode().strip()
+        cuda_home = os.path.dirname(os.path.dirname(nvcc_path))
+    except Exception:
+        # Fallback to standard Linux path
+        cuda_home = "/usr/local/cuda"
+        
+    cuda_inc = os.path.join(cuda_home, "include")
+    cuda_lib = os.path.join(cuda_home, "lib64")
+    # print(f"     [Builder] Detected CUDA at {cuda_home}")
+    
+    # 4. Compile the .cu files into object files
+    cu_files = [f for f in os.listdir(ort_plugins_dir) if f.endswith(".cu")]
+    obj_files = []
+    
+    for cu_file in cu_files:
+        cu_path = os.path.join(ort_plugins_dir, cu_file)
+        obj_path = os.path.join(ort_plugins_dir, cu_file.replace(".cu", ".o"))
+        obj_files.append(obj_path)
+        
+        cmd = [
+            "nvcc", "-c", cu_path, "-o", obj_path, "-O3", "-Xcompiler", "-fPIC",
+            f"-I{ort_inc}", f"-I{cuda_inc}"
+        ]
+        subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+        
+    # 5. Compile the registrar (register_ops.cpp)
+    reg_cpp = os.path.join(ort_plugins_dir, "register_ops.cpp")
+    reg_obj = os.path.join(ort_plugins_dir, "register_ops.o")
+    obj_files.append(reg_obj)
+    
+    cmd = [
+        "g++", "-c", reg_cpp, "-o", reg_obj, "-O3", "-fPIC",
+        f"-I{ort_inc}", f"-I{cuda_inc}"
+    ]
+    subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    
+    # 6. Link everything into the final .so library (with RPATH baked in)
+    so_path = os.path.join(ort_plugins_dir, "libtriton_ort_plugins.so")
+    abs_ort_lib = os.path.abspath(ort_lib)
+    
+    cmd = [
+        "g++", "-shared", "-o", so_path
+    ] + obj_files + [
+        f"-L{ort_lib}", "-lonnxruntime",
+        f"-L{cuda_lib}", "-lcuda", "-lcudart",
+        f"-Wl,-rpath,{abs_ort_lib}"
+    ]
+    subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    
+    # print(f"     [Builder] Compilation successful! Plugin saved to {so_path}")
+
+def build_trt_plugin(trt_plugins_dir: str, cache_dir: str):
+    """
+    Natively compiles the generated C++ files into a TensorRT Shared Library (.so).
+    """
+    # print("     [Builder] Querying system for CUDA configuration...")
+    try:
+        nvcc_path = subprocess.check_output(["which", "nvcc"]).decode().strip()
+        cuda_home = os.path.dirname(os.path.dirname(nvcc_path))
+    except Exception:
+        cuda_home = "/usr/local/cuda"
+        
+    cuda_inc = os.path.join(cuda_home, "include")
+    cuda_lib = os.path.join(cuda_home, "lib64")
+    
+    # In TRT 8.6+, the library is often split into nvinfer and nvinfer_plugin
+    # We will assume standard system paths for TRT (/usr/lib/x86_64-linux-gnu or LD_LIBRARY_PATH)
+    
+    cu_files = [f for f in os.listdir(trt_plugins_dir) if f.endswith(".cu")]
+    obj_files = []
+    
+    for cu_file in cu_files:
+        cu_path = os.path.join(trt_plugins_dir, cu_file)
+        obj_path = os.path.join(trt_plugins_dir, cu_file.replace(".cu", ".o"))
+        obj_files.append(obj_path)
+        
+        cmd = [
+            "nvcc", "-c", cu_path, "-o", obj_path, "-O3", "-Xcompiler", "-fPIC",
+            f"-I{cuda_inc}", "-Wno-deprecated-gpu-targets"
+        ]
+        subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+        
+    so_path = os.path.join(trt_plugins_dir, "libtriton_trt_plugins.so")
+    
+    cmd = [
+        "g++", "-shared", "-o", so_path
+    ] + obj_files + [
+        f"-L{cuda_lib}", "-lcuda", "-lcudart", "-lnvinfer"
+    ]
+    
+    try:
+        subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    except subprocess.CalledProcessError as e:
+        print("\n[ERROR] TensorRT compilation failed. Ensure TensorRT is installed and in your LD_LIBRARY_PATH.")
+        raise e
+        
+    # print(f"     [Builder] TRT Compilation successful! Plugin saved to {so_path}")

kernel_lens-0.1.0/kernel_lens/backends/ort_gen.py

@@ -0,0 +1,266 @@
+import os
+import textwrap
+import re
+
+class ORTGenerator:
+    def __init__(self, manifests, ops_package="triton_custom"):
+        self.manifests = manifests
+        self.ops_package = ops_package
+
+    def _generate_kernel_h(self, manifest) -> str:
+        op_name = f"{manifest.kernel_name}Op"
+        kernel_name = f"{manifest.kernel_name}Kernel"
+        
+        inputs_to_node = manifest.arguments
+        outputs_from_node = [a for a in manifest.arguments if a.kind == 'output']
+        
+        input_types_cpp = []
+        mem_types_cpp = []
+        
+        for a in inputs_to_node:
+            if a.kind == 'scalar':
+                # --- NEW: Keep Scalars on the CPU so we can safely read them! ---
+                mem_types_cpp.append("OrtMemTypeCPUInput")
+                if 'float' in a.dtype.lower():
+                    input_types_cpp.append("ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE")
+                else:
+                    input_types_cpp.append("ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64")
+            else:
+                # --- Keep Tensors on the GPU ---
+                mem_types_cpp.append("OrtMemTypeDefault")
+                input_types_cpp.append("ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT")
+                
+        input_types_str = ",\n            ".join(input_types_cpp)
+        mem_types_str = ",\n            ".join(mem_types_cpp)
+        
+        tpl = f'''#pragma once
+// Unlock the Custom Op Initialization API
+#define ORT_API_MANUAL_INIT
+#include <onnxruntime_cxx_api.h>
+#include <cuda.h>
+
+namespace custom {{
+
+struct {kernel_name} {{
+    void Compute(OrtKernelContext* context);
+}};
+
+struct {op_name} : Ort::CustomOpBase<{op_name}, {kernel_name}> {{
+    {op_name}() = default;
+
+    void* CreateKernel(const OrtApi& api, const OrtKernelInfo* info) const {{
+        return new {kernel_name}();
+    }}
+
+    const char* GetName() const {{ return "{manifest.kernel_name}"; }}
+    const char* GetExecutionProviderType() const {{ return "CUDAExecutionProvider"; }}
+
+    size_t GetInputTypeCount() const {{ return {len(inputs_to_node)}; }}
+    ONNXTensorElementDataType GetInputType(size_t index) const {{
+        static const ONNXTensorElementDataType types[] = {{
+            {input_types_str}
+        }};
+        return types[index];
+    }}
+
+    size_t GetOutputTypeCount() const {{ return {max(1, len(outputs_from_node))}; }}
+    ONNXTensorElementDataType GetOutputType(size_t index) const {{
+        return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT;
+    }}
+
+    // --- THE MAGIC FIX: Dynamic Memory Placement ---
+    OrtMemType GetInputMemoryType(size_t index) const {{
+        static const OrtMemType types[] = {{
+            {mem_types_str}
+        }};
+        return types[index];
+    }}
+    
+    OrtCustomOpInputOutputCharacteristic GetInputCharacteristic(size_t index) const {{
+        return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED;
+    }}
+}};
+
+}} // namespace custom
+'''
+        return textwrap.dedent(tpl).strip()
+
+    def _generate_kernel_cu(self, manifest) -> str:
+        kernel_name = f"{manifest.kernel_name}Kernel"
+        
+        arg_setup_lines = []
+        arg_setup_lines.append("Ort::KernelContext ctx(context);")
+        arg_setup_lines.append("auto ref_in = ctx.GetInput(0);")
+        arg_setup_lines.append("auto info = ref_in.GetTensorTypeAndShapeInfo();")
+        arg_setup_lines.append("std::vector<int64_t> dim_values = info.GetShape();")
+        
+        import re
+        
+        # --- ROBUST GRID EVALUATION ---
+        grid_strs = []
+        if hasattr(manifest, '_sym_grid_asts') and manifest._sym_grid_asts:
+            for g in manifest._sym_grid_asts:
+                # Always extract pure SymPy string FIRST
+                expr = str(g.node.expr) if hasattr(g, 'node') else str(g)
+                expr = re.sub(r'([a-zA-Z0-9_]+)\*\*([a-zA-Z0-9_]+)', r'std::pow(\1, \2)', expr)
+                expr = expr.replace("//", "/")
+                expr = re.sub(r'floor\((.*?)\)', r'(\1)', expr)
+                
+                # Replace SymPy symbols with ORT C++
+                # Use regex to avoid replacing s0 inside s01
+                expr = re.sub(r'\bs0\b', "(int64_t)dim_values[0]", expr)
+                expr = re.sub(r'\bs1\b', "(int64_t)(dim_values.size() > 1 ? dim_values[1] : 1)", expr)
+                expr = re.sub(r'\bs2\b', "(int64_t)(dim_values.size() > 2 ? dim_values[2] : 1)", expr)
+                grid_strs.append(expr)
+                
+        while len(grid_strs) < 3:
+            grid_strs.append("1")
+            
+        grid_eval_lines = [
+            f"unsigned int grid_x = std::max(1u, (unsigned int)({grid_strs[0]}));",
+            f"unsigned int grid_y = std::max(1u, (unsigned int)({grid_strs[1]}));",
+            f"unsigned int grid_z = std::max(1u, (unsigned int)({grid_strs[2]}));"
+        ]
+
+        arg_setup_lines.append("// Stable memory addresses for CUDA kernel launch")
+        for slot_idx, arg in enumerate(manifest.arguments):
+            if arg.kind == 'input':
+                arg_setup_lines.append(f"const void* arg_{slot_idx} = nullptr;")
+            elif arg.kind == 'scalar':
+                if 'float' in arg.dtype.lower():
+                    arg_setup_lines.append(f"float arg_{slot_idx} = 0.0f;")
+                else:
+                    arg_setup_lines.append(f"int32_t arg_{slot_idx} = 0;")
+            elif arg.kind == 'output':
+                arg_setup_lines.append(f"void* arg_{slot_idx} = nullptr;")
+                
+        padding_idx = len(manifest.arguments)
+        arg_setup_lines.append(f"int64_t pad_{padding_idx} = 0;")
+
+        arg_setup_lines.append("std::vector<void*> kp;")
+
+        input_idx = 0
+        output_idx = 0
+        
+        for slot_idx, arg in enumerate(manifest.arguments):
+            if arg.kind == 'input':
+                arg_setup_lines.append(f"auto input_{slot_idx} = ctx.GetInput({input_idx});")
+                arg_setup_lines.append(f"arg_{slot_idx} = (const void*)input_{slot_idx}.GetTensorData<float>();")
+                arg_setup_lines.append(f"kp.push_back(&arg_{slot_idx});")
+                input_idx += 1
+            elif arg.kind == 'scalar':
+                arg_setup_lines.append(f"auto input_{slot_idx} = ctx.GetInput({input_idx});")
+                # NOW THIS IS 100% SAFE because ONNX stored it in CPU memory!
+                if 'float' in arg.dtype.lower():
+                    arg_setup_lines.append(f"arg_{slot_idx} = (float)(*input_{slot_idx}.GetTensorData<double>());")
+                else:
+                    arg_setup_lines.append(f"arg_{slot_idx} = (int32_t)(*input_{slot_idx}.GetTensorData<int64_t>());")
+                
+                arg_name = getattr(arg, 'name', '')
+                if 'stride' in arg_name.lower():
+                    arg_setup_lines.append(f"if (arg_{slot_idx} != 1) kp.push_back(&arg_{slot_idx});")
+                else:
+                    arg_setup_lines.append(f"kp.push_back(&arg_{slot_idx});")
+                input_idx += 1
+            elif arg.kind == 'output':
+                arg_setup_lines.append(f"// Skip PyTorch's dummy input")
+                arg_setup_lines.append(f"auto dummy_in_{slot_idx} = ctx.GetInput({input_idx});")
+                input_idx += 1
+                
+                arg_setup_lines.append(f"auto output_{slot_idx} = ctx.GetOutput({output_idx}, dim_values.data(), dim_values.size());")
+                arg_setup_lines.append(f"arg_{slot_idx} = (void*)output_{slot_idx}.GetTensorMutableData<float>();")
+                arg_setup_lines.append(f"kp.push_back(&arg_{slot_idx});")
+                output_idx += 1
+                
+        arg_setup_lines.append(f"kp.push_back(&pad_{padding_idx});")
+
+        dynamic_args_cpp = "\n    ".join(arg_setup_lines)
+        grid_cpp = "\n    ".join(grid_eval_lines)
+
+        tpl = f'''#include "{manifest.kernel_name}Op.h"
+#include <stdexcept>
+#include <iostream>
+#include <vector>
+#include <cmath>
+#include <algorithm>
+
+namespace custom {{
+
+static const char* PTX_CODE = R"ptx(
+{manifest.ptx}
+)ptx";
+
+void {kernel_name}::Compute(OrtKernelContext* context) {{
+    {dynamic_args_cpp}
+
+    static CUmodule mModule = nullptr;
+    static CUfunction mKernel = nullptr;
+    
+    if (mModule == nullptr) {{
+        CUresult res = cuModuleLoadDataEx(&mModule, PTX_CODE, 0, nullptr, nullptr);
+        if (res != CUDA_SUCCESS) throw std::runtime_error("Failed to load PTX module");
+        res = cuModuleGetFunction(&mKernel, mModule, "{manifest.kernel_name}");
+        if (res != CUDA_SUCCESS) throw std::runtime_error("Failed to extract function");
+    }}
+
+    cudaStream_t stream = reinterpret_cast<cudaStream_t>(ctx.GetGPUComputeStream());
+    
+    {grid_cpp}
+    unsigned int block_x = 128; 
+
+    cuLaunchKernel(mKernel, grid_x, grid_y, grid_z, block_x, 1, 1, {manifest.shared_memory_bytes}, stream, kp.data(), nullptr);
+}}
+
+}} // namespace custom
+'''
+        return textwrap.dedent(tpl).strip()
+
+    def generate(self, output_dir: str):
+        if not os.path.exists(output_dir):
+            os.makedirs(output_dir)
+
+        for m in self.manifests:
+            base_path = os.path.join(output_dir, f"{m.kernel_name}Op")
+            with open(f"{base_path}.h", "w") as f:
+                f.write(self._generate_kernel_h(m))
+            with open(f"{base_path}.cu", "w") as f:
+                f.write(self._generate_kernel_cu(m))
+
+        registrar_code = '''#define ORT_API_MANUAL_INIT
+#include <onnxruntime_cxx_api.h>
+
+'''
+        for m in self.manifests:
+            registrar_code += f'#include "{m.kernel_name}Op.h"\n'
+
+        registrar_code += '''
+#ifndef ORT_EXPORT
+#ifdef _WIN32
+#define ORT_EXPORT __declspec(dllexport)
+#else
+#define ORT_EXPORT __attribute__((visibility("default")))
+#endif
+#endif
+
+extern "C" {
+    ORT_EXPORT OrtStatus* ORT_API_CALL RegisterCustomOps(OrtSessionOptions* options, const OrtApiBase* api_base) {
+        Ort::InitApi(api_base->GetApi(ORT_API_VERSION));
+        static Ort::CustomOpDomain custom_domain("triton_custom");
+'''
+        for m in self.manifests:
+            registrar_code += f'        static custom::{m.kernel_name}Op c_{m.kernel_name};\n'
+            registrar_code += f'        custom_domain.Add(&c_{m.kernel_name});\n'
+
+        registrar_code += '''
+        Ort::UnownedSessionOptions sess_options(options);
+        sess_options.Add(custom_domain);
+        return nullptr;
+    }
+}
+'''
+        with open(os.path.join(output_dir, "register_ops.cpp"), "w") as f:
+            f.write(registrar_code)
+
+def generate_ort_bindings(manifests, output_path: str):
+    gen = ORTGenerator(manifests)
+    gen.generate(output_path)