tritonparse 0.2.4.dev20251007071533__tar.gz → 0.2.4.dev20251008071501__tar.gz

{tritonparse-0.2.4.dev20251007071533 → tritonparse-0.2.4.dev20251008071501}/.gitignore

@@ -70,4 +70,5 @@ env.bak/
 venv.bak/
 *.mdc
 repro_output/
+.github/copilot-instructions.md
 # end

{tritonparse-0.2.4.dev20251007071533 → tritonparse-0.2.4.dev20251008071501}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tritonparse
-Version: 0.2.4.dev20251007071533
+Version: 0.2.4.dev20251008071501
 Summary: TritonParse: A Compiler Tracer, Visualizer, and mini-Reproducer Generator for Triton Kernels
 Author-email: Yueming Hao <yhao@meta.com>
 License-Expression: BSD-3-Clause
@@ -27,13 +27,22 @@ Dynamic: license-file
 
 ## ✨ Key Features
 
-- **🚀 Launch Difference Analysis** - Automatically detect and visualize variations in kernel launch parameters, helping you pinpoint performance bottlenecks and debug launch configurations.
-- **🔍 Interactive Visualization** - Explore Triton kernels with detailed metadata and stack traces
-- **📊 Multi-format IR Support** - View TTGIR, TTIR, LLIR, PTX, and AMDGCN in one place
-- **🔄 Side-by-side Comparison** - Compare IR stages with synchronized highlighting
-- **📝 Structured Logging** - Capture detailed compilation and launch events with source mapping
-- **🌐 Ready-to-use Interface** - No installation required, works in your browser
-- **🔒 Privacy-first** - All processing happens locally in your browser, no data uploaded
+### 🔍 Visualization & Analysis
+- **🚀 Launch Difference Analysis** - Detect and visualize kernel launch parameter variations
+- **📊 IR Code View** - Side-by-side IR viewing with synchronized highlighting and line mapping
+- **🔄 File Diff View** - Compare kernels across different trace files side-by-side
+- **📝 Multi-format IR Support** - View TTGIR, TTIR, LLIR, PTX, and AMDGCN
+- **🎯 Interactive Code Views** - Click-to-highlight corresponding lines across IR stages
+
+### 📊 Structured Logging & Analysis
+- **📝 Compilation & Launch Tracing** - Capture detailed events with source mapping
+- **🔍 Stack Trace Integration** - Full Python stack traces for debugging
+- **📈 Metadata Extraction** - Comprehensive kernel statistics
+
+### 🛠️ Developer Tools
+- **🔧 Reproducer Generation** - Generate standalone Python scripts to reproduce kernels
+- **🌐 Browser-based Interface** - No installation required, works in your browser
+- **🔒 Privacy-first** - All processing happens locally, no data uploaded
 
 ## 🚀 Quick Start
 
@@ -41,22 +50,22 @@ Dynamic: license-file
 
 ```python
 import tritonparse.structured_logging
+import tritonparse.utils
 
-# Initialize logging with launch tracing enabled
+# Initialize logging
 tritonparse.structured_logging.init("./logs/", enable_trace_launch=True)
 
 # Your Triton/PyTorch code here
 # ... your kernels ...
 
 # Parse and generate trace files
-import tritonparse.utils
-tritonparse.utils.unified_parse("./logs/")
+tritonparse.utils.unified_parse("./logs/", out="./parsed_output")
 ```
-The example terminal output is:
-```bash
-tritonparse log file list: /tmp/tmp1gan7zky/log_file_list.json
-INFO:tritonparse:Copying parsed logs from /tmp/tmp1gan7zky to /scratch/findhao/tritonparse/tests/parsed_output
 
+<details>
+<summary>📝 Example output (click to expand)</summary>
+
+```bash
 ================================================================================
 📁 TRITONPARSE PARSING RESULTS
 ================================================================================
@@ -64,13 +73,13 @@ INFO:tritonparse:Copying parsed logs from /tmp/tmp1gan7zky to /scratch/findhao/t
 📊 Total files generated: 2
 
 📄 Generated files:
---------------------------------------------------
    1. 📝 dedicated_log_triton_trace_findhao__mapped.ndjson.gz (7.2KB)
    2. 📝 log_file_list.json (181B)
 ================================================================================
 ✅ Parsing completed successfully!
 ================================================================================
 ```
+</details>
 
 ### 2. Visualize Results
 
@@ -106,18 +115,13 @@ pip install triton
 
 | 📖 Guide | Description |
 |----------|-------------|
-| **[🏠 Wiki Home](https://github.com/meta-pytorch/tritonparse/wiki)** | Complete documentation and navigation |
-| **[📦 Installation Guide](https://github.com/meta-pytorch/tritonparse/wiki/01.-Installation)** | Detailed setup for all scenarios |
-| **[📋 Usage Guide](https://github.com/meta-pytorch/tritonparse/wiki/02.-Usage-Guide)** | Complete workflow and examples |
-| **[🌐 Web Interface Guide](https://github.com/meta-pytorch/tritonparse/wiki/03.-Web-Interface-Guide)** | Master the visualization interface |
-| **[🔧 Developer Guide](https://github.com/meta-pytorch/tritonparse/wiki/04.-Developer-Guide)** | Contributing and development setup |
-| **[❓ FAQ](https://github.com/meta-pytorch/tritonparse/wiki/06.-FAQ)** | Frequently asked questions |
-
-## 🛠️ Tech Stack
-
-- **Frontend**: React 19, TypeScript, Vite, Tailwind CSS, Monaco Editor
-- **Backend**: Python with Triton integration, structured logging
-- **Deployment**: GitHub Pages, automatic deployment
+| **[🏠 Wiki Home](https://github.com/meta-pytorch/tritonparse/wiki)** | Complete documentation and quick navigation |
+| **[📦 Installation](https://github.com/meta-pytorch/tritonparse/wiki/01.-Installation)** | Setup guide for all scenarios |
+| **[📋 Usage Guide](https://github.com/meta-pytorch/tritonparse/wiki/02.-Usage-Guide)** | Complete workflow, examples, and reproducer |
+| **[🌐 Web Interface](https://github.com/meta-pytorch/tritonparse/wiki/03.-Web-Interface-Guide)** | Master the visualization interface |
+| **[🔧 Developer Guide](https://github.com/meta-pytorch/tritonparse/wiki/04.-Developer-Guide)** | Contributing and architecture overview |
+| **[📝 Code Formatting](https://github.com/meta-pytorch/tritonparse/wiki/05.-Code-Formatting)** | Formatting standards and tools |
+| **[❓ FAQ](https://github.com/meta-pytorch/tritonparse/wiki/06.-FAQ)** | Quick answers and troubleshooting |
 
 ## 📊 Understanding Triton Compilation
 
@@ -130,9 +134,10 @@ Each stage can be inspected and compared to understand optimization transformati
 ## 🤝 Contributing
 
 We welcome contributions! Please see our **[Developer Guide](https://github.com/meta-pytorch/tritonparse/wiki/04.-Developer-Guide)** for:
-- Development setup
-- Code formatting standards
-- Pull request process
+- Development setup and prerequisites
+- Code formatting standards (**[Formatting Guide](https://github.com/meta-pytorch/tritonparse/wiki/05.-Code-Formatting)**)
+- Pull request and code review process
+- Testing guidelines
 - Architecture overview
 
 ## 📞 Support & Community

{tritonparse-0.2.4.dev20251007071533 → tritonparse-0.2.4.dev20251008071501}/README.md

@@ -9,13 +9,22 @@
 
 ## ✨ Key Features
 
-- **🚀 Launch Difference Analysis** - Automatically detect and visualize variations in kernel launch parameters, helping you pinpoint performance bottlenecks and debug launch configurations.
-- **🔍 Interactive Visualization** - Explore Triton kernels with detailed metadata and stack traces
-- **📊 Multi-format IR Support** - View TTGIR, TTIR, LLIR, PTX, and AMDGCN in one place
-- **🔄 Side-by-side Comparison** - Compare IR stages with synchronized highlighting
-- **📝 Structured Logging** - Capture detailed compilation and launch events with source mapping
-- **🌐 Ready-to-use Interface** - No installation required, works in your browser
-- **🔒 Privacy-first** - All processing happens locally in your browser, no data uploaded
+### 🔍 Visualization & Analysis
+- **🚀 Launch Difference Analysis** - Detect and visualize kernel launch parameter variations
+- **📊 IR Code View** - Side-by-side IR viewing with synchronized highlighting and line mapping
+- **🔄 File Diff View** - Compare kernels across different trace files side-by-side
+- **📝 Multi-format IR Support** - View TTGIR, TTIR, LLIR, PTX, and AMDGCN
+- **🎯 Interactive Code Views** - Click-to-highlight corresponding lines across IR stages
+
+### 📊 Structured Logging & Analysis
+- **📝 Compilation & Launch Tracing** - Capture detailed events with source mapping
+- **🔍 Stack Trace Integration** - Full Python stack traces for debugging
+- **📈 Metadata Extraction** - Comprehensive kernel statistics
+
+### 🛠️ Developer Tools
+- **🔧 Reproducer Generation** - Generate standalone Python scripts to reproduce kernels
+- **🌐 Browser-based Interface** - No installation required, works in your browser
+- **🔒 Privacy-first** - All processing happens locally, no data uploaded
 
 ## 🚀 Quick Start
 
@@ -23,22 +32,22 @@
 
 ```python
 import tritonparse.structured_logging
+import tritonparse.utils
 
-# Initialize logging with launch tracing enabled
+# Initialize logging
 tritonparse.structured_logging.init("./logs/", enable_trace_launch=True)
 
 # Your Triton/PyTorch code here
 # ... your kernels ...
 
 # Parse and generate trace files
-import tritonparse.utils
-tritonparse.utils.unified_parse("./logs/")
+tritonparse.utils.unified_parse("./logs/", out="./parsed_output")
 ```
-The example terminal output is:
-```bash
-tritonparse log file list: /tmp/tmp1gan7zky/log_file_list.json
-INFO:tritonparse:Copying parsed logs from /tmp/tmp1gan7zky to /scratch/findhao/tritonparse/tests/parsed_output
 
+<details>
+<summary>📝 Example output (click to expand)</summary>
+
+```bash
 ================================================================================
 📁 TRITONPARSE PARSING RESULTS
 ================================================================================
@@ -46,13 +55,13 @@ INFO:tritonparse:Copying parsed logs from /tmp/tmp1gan7zky to /scratch/findhao/t
 📊 Total files generated: 2
 
 📄 Generated files:
---------------------------------------------------
    1. 📝 dedicated_log_triton_trace_findhao__mapped.ndjson.gz (7.2KB)
    2. 📝 log_file_list.json (181B)
 ================================================================================
 ✅ Parsing completed successfully!
 ================================================================================
 ```
+</details>
 
 ### 2. Visualize Results
 
@@ -88,18 +97,13 @@ pip install triton
 
 | 📖 Guide | Description |
 |----------|-------------|
-| **[🏠 Wiki Home](https://github.com/meta-pytorch/tritonparse/wiki)** | Complete documentation and navigation |
-| **[📦 Installation Guide](https://github.com/meta-pytorch/tritonparse/wiki/01.-Installation)** | Detailed setup for all scenarios |
-| **[📋 Usage Guide](https://github.com/meta-pytorch/tritonparse/wiki/02.-Usage-Guide)** | Complete workflow and examples |
-| **[🌐 Web Interface Guide](https://github.com/meta-pytorch/tritonparse/wiki/03.-Web-Interface-Guide)** | Master the visualization interface |
-| **[🔧 Developer Guide](https://github.com/meta-pytorch/tritonparse/wiki/04.-Developer-Guide)** | Contributing and development setup |
-| **[❓ FAQ](https://github.com/meta-pytorch/tritonparse/wiki/06.-FAQ)** | Frequently asked questions |
-
-## 🛠️ Tech Stack
-
-- **Frontend**: React 19, TypeScript, Vite, Tailwind CSS, Monaco Editor
-- **Backend**: Python with Triton integration, structured logging
-- **Deployment**: GitHub Pages, automatic deployment
+| **[🏠 Wiki Home](https://github.com/meta-pytorch/tritonparse/wiki)** | Complete documentation and quick navigation |
+| **[📦 Installation](https://github.com/meta-pytorch/tritonparse/wiki/01.-Installation)** | Setup guide for all scenarios |
+| **[📋 Usage Guide](https://github.com/meta-pytorch/tritonparse/wiki/02.-Usage-Guide)** | Complete workflow, examples, and reproducer |
+| **[🌐 Web Interface](https://github.com/meta-pytorch/tritonparse/wiki/03.-Web-Interface-Guide)** | Master the visualization interface |
+| **[🔧 Developer Guide](https://github.com/meta-pytorch/tritonparse/wiki/04.-Developer-Guide)** | Contributing and architecture overview |
+| **[📝 Code Formatting](https://github.com/meta-pytorch/tritonparse/wiki/05.-Code-Formatting)** | Formatting standards and tools |
+| **[❓ FAQ](https://github.com/meta-pytorch/tritonparse/wiki/06.-FAQ)** | Quick answers and troubleshooting |
 
 ## 📊 Understanding Triton Compilation
 
@@ -112,9 +116,10 @@ Each stage can be inspected and compared to understand optimization transformati
 ## 🤝 Contributing
 
 We welcome contributions! Please see our **[Developer Guide](https://github.com/meta-pytorch/tritonparse/wiki/04.-Developer-Guide)** for:
-- Development setup
-- Code formatting standards
-- Pull request process
+- Development setup and prerequisites
+- Code formatting standards (**[Formatting Guide](https://github.com/meta-pytorch/tritonparse/wiki/05.-Code-Formatting)**)
+- Pull request and code review process
+- Testing guidelines
 - Architecture overview
 
 ## 📞 Support & Community

{tritonparse-0.2.4.dev20251007071533 → tritonparse-0.2.4.dev20251008071501}/tritonparse/context_manager.py

@@ -17,6 +17,8 @@ class TritonParseManager:
         self,
         enable_trace_launch=False,
         split_inductor_compilations=True,
+        enable_tensor_blob_storage=False,
+        tensor_storage_quota=None,
         **parse_kwargs,
     ):
         """
@@ -25,17 +27,28 @@ class TritonParseManager:
         Args:
             enable_trace_launch: Whether to enable trace launch
             split_inductor_compilations: Whether to split inductor compilations in the output
+            enable_tensor_blob_storage: Whether to enable tensor blob storage
+            tensor_storage_quota: Storage quota in bytes for tensor blobs (default: 100GB)
             **parse_kwargs: Additional keyword arguments to pass to unified_parse
         """
         self.enable_trace_launch = enable_trace_launch
         self.split_inductor_compilations = split_inductor_compilations
+        self.enable_tensor_blob_storage = enable_tensor_blob_storage
+        self.tensor_storage_quota = tensor_storage_quota
         self.parse_kwargs = parse_kwargs
         self.dir_path = None
         self.output_link = None
 
     def __enter__(self):
         self.dir_path = createUniqueTempDirectory()
-        init(self.dir_path, enable_trace_launch=self.enable_trace_launch)
+        init_kwargs = {
+            "enable_trace_launch": self.enable_trace_launch,
+            "enable_tensor_blob_storage": self.enable_tensor_blob_storage,
+        }
+        if self.tensor_storage_quota is not None:
+            init_kwargs["tensor_storage_quota"] = self.tensor_storage_quota
+
+        init(self.dir_path, **init_kwargs)
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):

tritonparse-0.2.4.dev20251008071501/tritonparse/reproducer/templates/example.py

@@ -0,0 +1,387 @@
+"""
+This file is automatically generated by TritonParse reproducer.
+It contains a smallest testing example for a Triton kernel.
+"""
+
+import gzip
+import hashlib
+import importlib
+import io
+import json
+import logging
+import sys
+from functools import lru_cache
+from pathlib import Path
+from typing import Union
+
+import torch
+
+# {{KERNEL_SYSPATH_PLACEHOLDER}}
+
+# {{KERNEL_IMPORT_PLACEHOLDER}}
+
+TRITON_KERNELS_CUSTOM_TYPES = (
+    importlib.util.find_spec("triton_kernels") is not None
+    and importlib.util.find_spec("triton_kernels.tensor") is not None
+)
+
+
+@lru_cache(maxsize=1)
+def _get_triton_tensor_types():
+    """
+    Import and cache Triton custom tensor types.
+
+    Returns:
+        tuple: (Tensor, Storage, StridedLayout) classes from triton_kernels.tensor.
+
+    Raises:
+        ImportError: If the optional module 'triton_kernels.tensor' is not available.
+    """
+    mod = importlib.import_module("triton_kernels.tensor")
+    return (
+        mod.Tensor,
+        mod.Storage,
+        mod.StridedLayout,
+    )
+
+
+def load_tensor(tensor_file_path: Union[str, Path], device: str = None) -> torch.Tensor:
+    """
+    Load a tensor from its file path and verify its integrity using the hash in the filename.
+
+    Args:
+        tensor_file_path (str | Path): Direct path to the tensor file. Supports both:
+                               - .bin.gz: gzip-compressed tensor (hash is of uncompressed data)
+                               - .bin: uncompressed tensor (for backward compatibility)
+        device (str, optional): Device to load the tensor to (e.g., 'cuda:0', 'cpu').
+                               If None, keeps the tensor on its original device.
+
+    Returns:
+        torch.Tensor: The loaded tensor (moved to the specified device if provided)
+
+    Raises:
+        FileNotFoundError: If the tensor file doesn't exist
+        RuntimeError: If the tensor cannot be loaded
+        ValueError: If the computed hash doesn't match the filename hash
+    """
+    blob_path = Path(tensor_file_path)
+
+    if not blob_path.exists():
+        raise FileNotFoundError(f"Tensor blob not found: {blob_path}")
+
+    # Detect compression by file extension
+    is_compressed = blob_path.name.endswith(".bin.gz")
+
+    # Read file contents (decompress if needed)
+    try:
+        with open(blob_path, "rb") as f:
+            file_obj = gzip.GzipFile(fileobj=f, mode="rb") if is_compressed else f
+            file_contents = file_obj.read()
+    except (OSError, gzip.BadGzipFile) as e:
+        if is_compressed:
+            raise RuntimeError(f"Failed to decompress gzip file {blob_path}: {str(e)}")
+        else:
+            raise RuntimeError(f"Failed to read file {blob_path}: {str(e)}")
+
+    # Extract expected hash from filename
+    # abc123.bin.gz -> abc123 or abc123.bin -> abc123
+    expected_hash = blob_path.name.removesuffix(".bin.gz" if is_compressed else ".bin")
+
+    # Compute hash of uncompressed data
+    computed_hash = hashlib.blake2b(file_contents).hexdigest()
+
+    # Verify hash matches filename
+    if computed_hash != expected_hash:
+        raise ValueError(
+            f"Hash verification failed: expected '{expected_hash}' but computed '{computed_hash}'"
+        )
+
+    try:
+        # Load the tensor from memory buffer
+        tensor = torch.load(io.BytesIO(file_contents), map_location=device)
+        return tensor
+    except Exception as e:
+        raise RuntimeError(f"Failed to load tensor from {blob_path}: {str(e)}")
+
+
+def create_args_from_json_file(json_path):
+    with open(json_path, "r") as f:
+        data = json.load(f)
+    return create_args_from_json(data)
+
+
+def create_args_from_json(data):
+    """
+    Parse a reproducer JSON and build kernel grid and argument dictionary.
+
+    Args:
+        json_path (str): Path to the JSON file describing the kernel launch.
+
+    Returns:
+        tuple[list, dict]: Grid specification list and map of argument name to value.
+    """
+    # Handle data format validation and extraction
+    if isinstance(data, list):
+        if len(data) != 1:
+            print(
+                f"Error: Expected single element list, got list with {len(data)} elements"
+            )
+            sys.exit(1)
+        data = data[0]
+    elif not isinstance(data, dict):
+        print(f"Error: Expected list or dict, got {type(data)}")
+        sys.exit(1)
+
+    grid = data.get("grid", [])
+    args_dict = {}
+    extracted_args = data.get("extracted_args", {})
+
+    for arg_name, arg_info in extracted_args.items():
+        args_dict[arg_name] = _create_arg_from_info(arg_info)
+
+    return grid, args_dict
+
+
+def _apply_stride_and_offset(tensor, shape, stride, storage_offset):
+    """
+    Apply custom stride and storage offset to a tensor if needed.
+
+    Args:
+        tensor: The base contiguous tensor
+        shape: The desired shape
+        stride: The desired stride (or None for contiguous)
+        storage_offset: The desired storage offset
+
+    Returns:
+        torch.Tensor: The strided tensor view or original tensor if contiguous
+    """
+    if stride is None:
+        return tensor
+
+    # Calculate expected contiguous stride
+    expected_contiguous_stride = []
+    s = 1
+    for dim_size in reversed(shape):
+        expected_contiguous_stride.insert(0, s)
+        s *= dim_size
+
+    # If stride matches contiguous stride and no storage offset, return as-is
+    if tuple(stride) == tuple(expected_contiguous_stride) and storage_offset == 0:
+        return tensor
+
+    # Calculate required storage size
+    if len(shape) > 0 and len(stride) > 0:
+        max_offset = storage_offset
+        for dim_stride, dim_size in zip(stride, shape):
+            if dim_size > 0:
+                max_offset += dim_stride * (dim_size - 1)
+        storage_size = max_offset + 1
+    else:
+        storage_size = storage_offset + 1
+
+    # Create larger storage tensor and create strided view
+    storage_tensor = torch.empty(storage_size, dtype=tensor.dtype, device=tensor.device)
+
+    # Create strided view
+    strided_view = storage_tensor.as_strided(
+        size=shape, stride=stride, storage_offset=storage_offset
+    )
+
+    # Copy data from the base tensor into the strided layout
+    strided_view.copy_(tensor.flatten()[: strided_view.numel()].view(shape))
+
+    return strided_view
+
+
+def _create_base_tensor(arg_info) -> torch.Tensor:
+    if arg_info.get("blob_path"):
+        return load_tensor(arg_info.get("blob_path"), arg_info.get("device"))
+
+    # Extract basic tensor properties
+    dtype_str = arg_info.get("dtype")
+    try:
+        torch_dtype = getattr(torch, dtype_str.split(".")[-1])
+    except AttributeError:
+        logging.error(f"Unsupported dtype: {dtype_str}. Defaulting to float32.")
+        torch_dtype = torch.float32
+
+    shape = arg_info.get("shape", [])
+    device = arg_info.get("device", "cpu")
+
+    # Extract statistical information if available
+    mean = arg_info.get("mean")
+    std = arg_info.get("std")
+    min_val = arg_info.get("min")
+    max_val = arg_info.get("max")
+    has_stats = (
+        mean is not None
+        and std is not None
+        and min_val is not None
+        and max_val is not None
+    )
+
+    if arg_info.get("tensor_capture_error", False):
+        logging.error(
+            f"Error: Tensor '{arg_info.get('name', '')}' had capture error. Generating random tensor instead."
+        )
+
+    # Use a dummy tensor to check properties of the dtype
+    tensor_props = torch.empty(0, dtype=torch_dtype)
+
+    # Case 1: Floating point types
+    if tensor_props.is_floating_point():
+        if has_stats:
+            # Generate tensor with statistical properties matching original data
+            if std == 0 or min_val == max_val:
+                # Constant tensor
+                return torch.full(shape, mean, dtype=torch_dtype, device=device)
+            # Generate normal distribution with mean and std, then clamp to [min, max]
+            tensor = torch.randn(shape, dtype=torch.float32, device=device) * std + mean
+            tensor = torch.clamp(tensor, min=min_val, max=max_val)
+            return tensor.to(torch_dtype)
+        else:
+            # Fallback to original random generation
+            if torch_dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
+                tmp = torch.rand(shape, dtype=torch.float32, device=device)
+                return tmp.to(torch_dtype)
+            else:
+                return torch.empty(shape, dtype=torch_dtype, device=device).random_()
+
+    # Case 2: Integer types
+    elif torch_dtype in [
+        torch.int8,
+        torch.int16,
+        torch.int32,
+        torch.int64,
+        torch.uint8,
+        torch.bool,
+    ]:
+        if has_stats and torch_dtype != torch.bool:
+            # Generate tensor with statistical properties, then round for integers
+            if std == 0 or min_val == max_val:
+                # Constant tensor
+                return torch.full(shape, int(mean), dtype=torch_dtype, device=device)
+            tensor = torch.randn(shape, dtype=torch.float32, device=device) * std + mean
+            tensor = torch.clamp(tensor, min=min_val, max=max_val)
+            return torch.round(tensor).to(torch_dtype)
+        else:
+            # Fallback to original random generation
+            return torch.empty(shape, dtype=torch_dtype, device=device).random_()
+
+    # Case 3: Complex numbers need special handling
+    elif tensor_props.is_complex():
+        # Complex types: fallback to original logic for now
+        # TODO: Could be improved to use statistical info if available
+        float_dtype = torch.float32 if torch_dtype == torch.complex64 else torch.float64
+        real_part = torch.rand(shape, dtype=float_dtype, device=device)
+        imag_part = torch.rand(shape, dtype=float_dtype, device=device)
+        return torch.complex(real_part, imag_part)
+
+    # Case 4: Handle other unsigned integers (like uint32) which fail with random_()
+    elif "uint" in str(torch_dtype):
+        if has_stats:
+            # Generate tensor with statistical properties for unsigned integers
+            if std == 0 or min_val == max_val:
+                return torch.full(shape, int(mean), dtype=torch_dtype, device=device)
+            tensor = torch.randn(shape, dtype=torch.float32, device=device) * std + mean
+            tensor = torch.clamp(tensor, min=min_val, max=max_val)
+            return torch.round(tensor).to(torch_dtype)
+        else:
+            # Fallback to original random generation
+            return torch.randint(0, 1000, shape, dtype=torch_dtype, device=device)
+
+    # Case 5: If we don't know how to handle the type, raise an error
+    else:
+        raise NotImplementedError(
+            f"Random data generation not implemented for dtype: {torch_dtype}"
+        )
+
+
+def _create_tensor(arg_info) -> torch.Tensor:
+    tensor = _create_base_tensor(arg_info)
+
+    # Apply stride and storage offset if needed
+    shape = arg_info.get("shape", [])
+    stride = arg_info.get("stride")
+    storage_offset = arg_info.get("storage_offset", 0)
+    return _apply_stride_and_offset(tensor, shape, stride, storage_offset)
+
+
+def _create_arg_from_info(arg_info):
+    """
+    Recursively construct a kernel argument from its JSON schema.
+
+    Args:
+        arg_info (dict): JSON object describing a single argument, including
+            fields like 'type', 'value', 'dtype', 'shape', 'device', etc.
+
+    Returns:
+        Any: The constructed Python object suitable for kernel invocation.
+
+    Raises:
+        RuntimeError: When required optional dependencies are missing.
+        NotImplementedError: When a dtype or type is not supported yet.
+    """
+    arg_type = arg_info.get("type")
+
+    if arg_type == "NoneType":
+        return None
+
+    if arg_type in ["int", "bool", "str", "float"]:
+        return arg_info.get("value")
+
+    elif arg_type == "tensor":
+        return _create_tensor(arg_info)
+
+    elif arg_type == "triton_kernels.tensor.Tensor":
+        if not TRITON_KERNELS_CUSTOM_TYPES:
+            raise RuntimeError(
+                "Optional dependency 'triton_kernels.tensor' is not installed; cannot construct Tensor."
+            )
+        Tensor, Storage, StridedLayout = _get_triton_tensor_types()
+        storage = _create_arg_from_info(arg_info.get("storage"))
+        dtype_str = arg_info.get("dtype")
+        torch_dtype = getattr(torch, dtype_str.split(".")[-1])
+        return Tensor(
+            storage=storage,
+            shape=arg_info.get("shape"),
+            shape_max=arg_info.get("shape_max"),
+            dtype=torch_dtype,
+        )
+
+    elif arg_type == "triton_kernels.tensor.Storage":
+        if not TRITON_KERNELS_CUSTOM_TYPES:
+            raise RuntimeError(
+                "Optional dependency 'triton_kernels.tensor' is not installed; cannot construct Storage."
+            )
+        Tensor, Storage, StridedLayout = _get_triton_tensor_types()
+        data = _create_arg_from_info(arg_info.get("data"))
+        layout = _create_arg_from_info(arg_info.get("layout"))
+        return Storage(data=data, layout=layout)
+
+    elif arg_type == "StridedLayout":
+        if not TRITON_KERNELS_CUSTOM_TYPES:
+            raise RuntimeError(
+                "Optional dependency 'triton_kernels.tensor' is not installed; cannot construct StridedLayout."
+            )
+        Tensor, Storage, StridedLayout = _get_triton_tensor_types()
+        return StridedLayout(shape=arg_info.get("initial_shape"))
+    else:
+        print(f"Warning: Unhandled argument type '{arg_type}'. Returning None.")
+        return None
+
+
+if __name__ == "__main__":
+    script_dir = Path(__file__).resolve().parent
+    json_file = script_dir / "{{JSON_FILE_NAME_PLACEHOLDER}}"
+    grid, args_dict = create_args_from_json_file(str(json_file))
+
+    print("Generated kernel arguments dictionary:")
+    for name, arg in args_dict.items():
+        print(f"  {name}: {arg}")
+    print(f"Grid: {grid}")
+
+    # {{KERNEL_INVOCATION_PLACEHOLDER}}
+
+    torch.cuda.synchronize()
+    print("Kernel execution finished.")