Integrating Custom Backends via JSON Platform Configs

[can-gaa-hou]

Summary

This change extends KernelAgent’s multi-platform support by allowing users to integrate a custom backend (or override existing platform settings) via an external JSON file. The JSON is loaded at import time and merged into the in-memory platform registry, enabling new --target-platform <name> values (e.g., npu) without modifying repository code.

Motivation

KernelAgent already supports CUDA and Intel XPU via a built-in PlatformConfig registry. However, adding a new backend (e.g., an NPU platform) typically requires:

• Modifying the in-repo platform registry,
• Updating prompts and device checks,
• Shipping and maintaining these changes for each new backend.

This PR introduces a lightweight escape hatch: JSON-defined platform configs that can be set per-environment. It lowers the friction for downstream teams with custom backends (rocm, tpu, npu, etc.) experimenting.

Usage

Define a custom platform (example: NPU)

1. Create a JSON file:

{
  "npu": {
    "name": "npu",
    "device_string": "npu",
    "guidance_block": "...",
    "kernel_guidance": "...",
    "cuda_hacks_to_strip": []
  }
}

2. Export the env var (must be set before importing KernelAgent modules):

export KERNELAGENT_PLATFORM_JSON=/abs/path/to/platforms.json

3. Select the platform via existing CLI flags (examples shown in README):

python -m Fuser.auto_agent --problem /abs/path/to/problem.py --target-platform npu

Programmatic usage

from triton_kernel_agent import TritonKernelAgent

agent = TritonKernelAgent(
    num_workers=4,
    max_rounds=8,
    model_name="gpt-5",
    target_platform="npu",  # can now pass a string
)

Testing

This PR does not add new unit tests since it requires external JSON files, but the following verification has been tested:

• Tested on CUDA (default behavior unchanged)
• Tested on a custom backend (NPU)
• CLI --target-platform flag works for auto_agent, pipeline, and cli
• UI platform dropdown works in all three UIs