Installation

Installation#

FlagPrism is built as part of the FlagTree wheel via the third_party/FlagPrism submodule. It is not published as a standalone package.

Prerequisites#

  • A FlagTree source checkout with the FlagPrism submodule.

  • A supported backend toolchain:

    • Debugger dynamic collection: Ascend/CANN, Tianshu/CoreX, Moore Threads MUSA, or NVIDIA CUDA.

    • Profiler: NVIDIA CUDA, AMD ROCm, Ascend CANN, Moore Threads MUSA, or Tianshu CoreX.

  • Build dependencies required by FlagTree (LLVM/MLIR, Python build tooling).

Build with FlagTree#

From the FlagTree repository root:

git submodule update --init --recursive
python -m pip wheel . --no-build-isolation

This builds the FlagTree core, Debugger, and Profiler in one CMake graph and packages them into a single FlagTree wheel.

Build examples for each vendor:

# Ascend
FLAGTREE_BACKEND=ascend TRITON_BUILD_FLAGPRISM=ON MAX_JOBS=16 \
python -m pip install . --no-build-isolation

# Tianshu-only (excludes Ascend/CANN sources, uses the Tianshu/CoreX runtime path)
FLAGPRISM_BACKEND=tianshu TRITON_BUILD_FLAGPRISM=ON \
python -m pip install . --no-build-isolation

# Moore Threads
FLAGTREE_BACKEND=mthreads FLAGPRISM_BACKEND=mthreads \
TRITON_BUILD_FLAGPRISM=ON python -m pip install . --no-build-isolation

# NVIDIA-only (CUDA/CUPTI path, no ROCm headers or libraries required)
FLAGTREE_BACKEND=nvidia FLAGPRISM_BACKEND=nvidia \
TRITON_BUILD_FLAGPRISM=ON python -m pip install . --no-build-isolation

For development, an editable install is supported:

python3 -m pip install -e . --no-build-isolation --no-deps

Build switches#

Variable

Values

Effect

TRITON_BUILD_FLAGPRISM

ON (default), OFF

Builds the combined Debugger+Profiler suite, or produces a core-only wheel. This is the only component switch; the two tools cannot be enabled independently.

FLAGPRISM_BACKEND

e.g. tianshu, mthreads, nvidia

Selects a vendor backend; for example tianshu builds only Tianshu/CoreX adaptations and does not probe or link Ascend CANN.

FLAGTREE_BACKEND

e.g. ascend

Selects the FlagTree target backend used for the build.

Verifying the install#

import flagtree.debugger as debugger
import flagtree.profiler as profiler

print("debugger available:", debugger.is_available())
print("profiler imported:", profiler is not None)

A core-only wheel (TRITON_BUILD_FLAGPRISM=OFF) does not include flagtree.debugger. Use debugger.is_available() to check whether both the compiler and runtime native bindings are present.

Requirements#

  • Python 3.10+

  • PyTorch (with vendor extensions such as torch_npu for Ascend, where applicable)

  • A Triton/FlagTree-compatible device and driver for the selected backend