Add CUDA backend support for NVIDIA GPUs with automatic detection

[Copilot]

Add CUDA Backend Support ✅

This PR implements CUDA backend support for Iris, enabling the framework to run seamlessly on both AMD GPUs (via HIP) and NVIDIA GPUs (via CUDA) with transparent backend auto-detection.

Summary

Iris now supports both AMD GPUs (HIP backend) and NVIDIA GPUs (CUDA backend) with automatic backend detection based on available GPU libraries. All backend logic is consolidated into a single iris/hip.py file with conditional branching.

Changes Made

1. Unified Backend Module (iris/hip.py):

   • Auto-detects backend at module load time by trying to load libamdhip64.so or libcudart.so
   • Sets _is_amd_backend flag for internal use
   • All functions branch internally based on backend type
   • Dynamic IPC handle size (64 bytes for HIP, 128 bytes for CUDA)
   • Exports get_ipc_handle_size() for use by iris.py

2. Dynamic Sizing (iris/iris.py):

   • Added get_ipc_handle_size() import
   • Uses dynamic IPC handle size instead of hardcoded 64
   • Ensures proper IPC handle allocation for each backend

3. Key Features:

   • Automatic backend detection based on available GPU libraries
   • Minimal code changes - original hip.py structure preserved with conditional branching
   • Consolidated implementation - all backend logic in single file
   • Zero configuration required - works out of the box
   • Falls back to HIP by default for backward compatibility

4. Code Quality:

   • Minimal changes to existing codebase
   • No separate backend files - everything consolidated in iris/hip.py
   • Clean conditional logic throughout
   • No build-time configuration or environment variables

Usage

# Just install normally - backend auto-detected
pip install git+https://github.com/ROCm/iris.git

# Works automatically on both AMD and NVIDIA GPUs
import iris
ctx = iris.iris(heap_size=1 << 30)

Implementation Details

Backend Detection (in iris/hip.py):

• Tries to load libamdhip64.so → sets _is_amd_backend = True
• Falls back to libcudart.so → sets _is_amd_backend = False
• Defaults to HIP for backward compatibility if both fail

Conditional Functions:

• hip_try() - branches to hipGetErrorString or cudaGetErrorString
• hipIpcMemHandle_t - 64 bytes for HIP, 128 bytes for CUDA
• open_ipc_handle() - calls hipIpcOpenMemHandle or cudaIpcOpenMemHandle
• All device/memory functions branch based on _is_amd_backend

Backend-Specific Behavior:

• get_rocm_version() - returns (-1, -1) for CUDA
• get_arch_string() - returns GCN arch for HIP, compute capability for CUDA (e.g., "sm_90")
• get_num_xcc() - returns actual count for AMD, 1 for NVIDIA

Code Statistics

• Modified files: iris/hip.py, iris/iris.py
• Total additions: ~40 lines (conditional logic)
• Files changed: 2
• Commits: 17

Backend Detection Logic

1. Try to load libamdhip64.so → use HIP backend
2. Try to load libcudart.so → use CUDA backend
3. Default to HIP for backward compatibility

Supported Hardware

AMD GPUs (HIP Backend):

• MI300X, MI350X, MI355X
• Requirements: ROCm 6.3.1+, PyTorch with ROCm

NVIDIA GPUs (CUDA Backend):

• All CUDA-capable GPUs with compute capability 7.0+
• Requirements: CUDA 11.0+, PyTorch with CUDA

Notes

• Minimal changes per review feedback - consolidated into single file with conditional branching
• Pure auto-detection with no configuration needed
• Original hip.py structure preserved - only added conditional logic
• Zero breaking changes - fully backward compatible
• All changes follow existing Iris code style and patterns

References

• Based on concepts from astroC86/iris cuda-port branch
• Related to PR CUDA Port #118
• Resolves issue Add CUDA backend #199

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Add CUDA backend</issue_title>
<issue_description>#### Summary
Allow Iris to be installed with --config-settings backend=nvidia so that iris.hip.* imports transparently redirect to CUDA equivalents (iris.cuda.*). This enables NVIDIA backend support without changing source code.

Goal

At install or runtime, users should be able to do:

pip install git+https://github.com/ROCm/iris.git --config-settings backend=nvidia

and have:

from iris.hip.runtime import hip_malloc

automatically resolve to:

from iris.cuda.runtime import hip_malloc

Cherry pick commits from https://github.com/astroC86/iris/tree/astroC86/cuda-port and #118</issue_description>

Comments on the Issue (you are @copilot in this section)

Fixes #199

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.