English | 简体中文
FlagPrism is a multi-backend debugging and performance-analysis toolkit for Triton programs, designed to provide a consistent observability workflow across NVIDIA GPUs and diverse AI accelerators. Built for the FlagTree ecosystem, it observes Triton kernels at both compile time and runtime, connects source-level context with Triton IR operations and device events, and turns the collected data into reports that help developers understand correctness, memory behavior, and performance across heterogeneous accelerator backends.
FlagPrism contains two user-facing components:
- Debugger captures values, numerical summaries,
memory-address summaries, complete tensor data, and statement/operation
metadata from selected regions inside
@triton.jitkernels. See the Debugger README for configuration, collection levels, report formats, examples, and current limitations. - Profiler records execution context, timelines, operation counts, estimated bytes, hardware metrics, and vendor-profiler data. It aggregates these records into tree, timeline, Hatchet, metadata, and vendor-specific outputs. See the Profiler README for APIs, modes, command-line tools, visualization, and backend-specific usage.
The table below tracks the FlagPrism enablement roadmap. It describes Debugger and Profiler integration, not the availability of the corresponding FlagTree compiler backend. We are rapidly expanding FlagPrism to support more devices and accelerator backends.
| NVIDIA | Huawei Ascend | T-Head | Hygon | Moore Threads | MetaX | ILUVATAR | Enflame |
|---|---|---|---|---|---|---|---|
| In progress Target: Sep 2026 |
✅ | — | — | Merging | Not started Target: Oct 2026 |
✅ | In progress Target: Sep 2026 |
FlagTree is a unified, multi-backend
compiler that enables Triton programs to run across diverse AI accelerators.
FlagPrism complements FlagTree with observability tools that help developers
debug kernel correctness, analyze runtime behavior, identify performance
bottlenecks, and optimize Triton workloads. FlagPrism is integrated and
maintained as the third_party/FlagPrism submodule of FlagTree.
FlagPrism is designed around a unified observability experience across heterogeneous accelerators. It separates what developers want to observe from how each backend collects that information, allowing Debugger and Profiler to offer consistent concepts and workflows across devices. The design emphasizes low intrusion into user programs, reuse between debugging and profiling, predictable results, and incremental support for new accelerator backends.
The Debugger instruments selected regions inside Triton kernels and combines static source and IR metadata with runtime value and memory information. It helps developers inspect numerical results, memory access, and kernel data flow through statement-level reports, operation-level reports, and tensor artifacts. See the Debugger README for details.
The Profiler combines program context, kernel events, instrumentation data, and vendor performance metrics to analyze how Triton workloads execute on a device. It correlates and aggregates the collected information into call trees, timelines, and portable performance reports. See the Profiler README for details.
FlagPrism is developed and built together with FlagTree. Download FlagTree, initialize the FlagPrism submodule, and build the complete project from the FlagTree repository root:
git clone https://github.com/flagos-ai/FlagTree.git
cd FlagTree
git submodule update --init --recursive third_party/FlagPrism
python3 -m pip install . --no-build-isolationFlagPrism/
├── Debugger/ # Compiler passes, runtime, decoder, and Python API
├── Profiler/ # Profiler dialect, runtime, adapters, and Python API
├── cmake/FlagPrism.cmake # Unified CMake integration policy
├── python/flagprism_build.py # FlagTree wheel and package integration
└── docs/ # Architecture and backend adaptation documents
FlagPrism is licensed under the MIT License, using the same license as FlagTree.