[WIP] Feature/Deft Kernel

.github/workflows/mypy.yml

@@ -27,7 +27,7 @@ jobs:
     - name: Install dependencies
       run: |
         curl -sSL https://install.python-poetry.org | python3
-        poetry install --all-extras
+        poetry install --all-extras -vvv
     - name: Type-checking package with mypy
       run: |
         # Run this mypy instance against our main package.

.github/workflows/pytest.yml

@@ -27,7 +27,7 @@ jobs:
     - name: Install dependencies
       run: |
         curl -sSL https://install.python-poetry.org | python3
-        poetry install --all-extras
+        poetry install --all-extras -vvv
     - name: Test with pytest
       env:
         OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

.gitignore

@@ -0,0 +1,189 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+docs/source/getting_started/examples/*.rst
+!**/*.template.rst
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+.idea/
+
+# VSCode
+.vscode/
+
+# DS Store
+.DS_Store
+
+# Results
+*.csv
+
+# Python pickle files
+*.pkl
+
+# Sphinx documentation
+_build/
+
+# vim swap files
+*.swo
+*.swp
+
+# hip files generated by PyTorch
+*.hip
+*_hip*
+hip_compat.h
+
+# Benchmark dataset
+*.json

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "3rdparty/flashinfer"]
+	path = 3rdparty/flashinfer
+	url = https://github.com/flashinfer-ai/flashinfer.git

CMakeLists.txt

@@ -0,0 +1,52 @@
+cmake_minimum_required(VERSION 3.23.1)
+project(deft
+		VERSION 2024
+		DESCRIPTION "An IO-aware fast attention kernel for efficient tree-structured interactions with LLMs"
+		LANGUAGES CUDA CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CUDA_STANDARD 17)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+set(FLASHINFER_DIR ${PROJECT_SOURCE_DIR}/3rdparty/flashinfer)
+set(FLASHINFER_INCLUDE_DIR ${FLASHINFER_DIR}/include)
+
+add_subdirectory(${FLASHINFER_DIR})
+
+set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -g -lineinfo -Xptxas -v -lineinfo")
+find_package(Python3 REQUIRED)
+find_package(CUDAToolkit REQUIRED)
+if(NOT Python3_FOUND)
+  message(FATAL_ERROR "Python3 not found.")
+endif()
+if(NOT CUDAToolkit_FOUND)
+  message(FATAL_ERROR "CUDA not found.")
+endif()
+message(STATUS "Python3 found at ${Python3_EXECUTABLE}")
+message(STATUS "CUDA version is ${CUDAToolkit_VERSION}")
+set(DEFT_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/csrc)
+include_directories(${DEFT_SOURCE_DIR})
+include_directories(${CUDAToolkit_INCLUDE_DIRS})
+include_directories(${FLASHINFER_INCLUDE_DIR})
+
+link_directories(${CUDAToolkit_LIBRARY_DIR})
+
+set(DEFT_ENABLE_TESTS CACHE BOOL "Enable tests for DEFT" ON)
+
+if(DEFT_ENABLE_TESTS)
+  enable_testing()
+  set(DEFT_TEST_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/tests/csrc)
+  file(GLOB_RECURSE TEST_DUMMY_SRC ${DEFT_TEST_SOURCE_DIR}/test_dummy.cu)
+  add_executable(test_dummy ${TEST_DUMMY_SRC})
+  target_link_libraries(test_dummy gtest_main gtest)
+  target_link_libraries(test_dummy ${CUDAToolkit_LIBRARIES})
+  add_test(NAME test_dummy COMMAND test_dummy)
+
+  file(GLOB_RECURSE TEST_DEFT_SRC ${DEFT_TEST_SOURCE_DIR}/test_deft_attention.cu)
+  add_executable(test_deft_attention ${TEST_DEFT_SRC})
+  target_link_libraries(test_deft_attention gtest_main gtest)
+  target_link_libraries(test_deft_attention ${CUDAToolkit_LIBRARIES})
+  target_include_directories(test_deft_attention PRIVATE ${FLASHINFER_DIR}/src)
+  add_dependencies(test_deft_attention dispatch_inc)
+  target_link_libraries(test_deft_attention decode_kernels)
+  add_test(NAME test_deft_attention COMMAND test_deft_attention)
+endif()

README.md

@@ -20,7 +20,7 @@ We propose DeFT, an IO-aware attention algorithm for efficient tree-structured i
 
 - [2024/05] We update the second version of DeFT paper with a better algorithm for general tree-structured LLM inference: [DeFT: Decoding with Flash Tree-attention for Efficient Tree-structured LLM Inference](https://arxiv.org/abs/2404.00242)!
 - [2024/03] [DeFT: Flash Tree-Attention With IO-Awareness for Efficient Tree-Search-Based LLM Inference](https://openreview.net/pdf?id=HqfLHoX8bR) has been accepted as Oral presentation in [ICLR'24 AGI Workshop](https://iclr.cc/virtual/2024/23126)!
-
+****
 
 ## Abstract
 Given the increasing demand for tree-structured interactions with LLMs, we introduce DeFT (Decoding with Flash Tree-Attention), an IO-aware tree attention algorithm tailored for tree-structured inference. Unlike traditional sequence-based decoding, tree-structured decoding better accommodates modern task requirements, including self-consistency, few-shot prompting, multi-step reasoning, and multi-model/head coordination. However, existing sequence-based inference systems are ill-suited for tree-structured decoding, resulting in redundancy in computation, memory footprints, and memory access, thereby undermining inference efficiency. To address this challenge, DeFT maintains memory-efficient attention calculation with low memory footprints through two key stages: (1) QKV Preparation: We propose a KV-Guided Grouping Strategy with Tree Split to intelligently group QKV, optimizing GPU resource utilization while minimizing memory reads/writes for KV cache between GPU global memory and on-chip shared memory; (2)Attention Calculation: We compute partial attention of each QKV group in a fused kernel and employ a Tree-topology-aware Global Reduction strategy to obtain final attention. By reducing 73-99% KV cache IO and nearly 100% IO for partial results during attention calculation (e.g., Softmax), DeFT achieves up to 2.52/3.82x speedup in the end-to-end/attention latency across three practical tree-based workloads: namely, few-shot prompting, multi-step reasoning, and speculative decoding, over state-of-the-art attention algorithms.
@@ -42,10 +42,23 @@ poetry install
 CUDA_VISIBLE_DEVICES=0 python examples/
 ```
 
+
+### Run Tests
+
+<!-- We profile DeFT kernel performance with [nvbench](https://github.com/NVIDIA/nvbench) and you can compile and run the benchmarks with the following commands: -->
+
+```bash
+git submodule update --init --recursive # clone flashinfer
+cmake -B build
+cmake --build build
+cd build
+ctest
+```
+
 ## FAQ
 
 1. **What is the difference between two versions of DeFT papers in arXiv?**
-   
+
     DeFT-v1
 
 
@@ -72,4 +85,4 @@ If you find DeFT useful or relevant to your project and research, please kindly
   journal={arXiv preprint arXiv:2404.00242},
   year={2024}
 }
-```
+```

build.py

@@ -0,0 +1,29 @@
+import logging
+from typing import Any, Dict
+
+from torch.utils.cpp_extension import CUDAExtension
+
+logger = logging.getLogger(__name__)
+
+
+ext_modules = []
+ext_modules.append(
+    CUDAExtension(
+        name='deft._kernels',
+        sources=['csrc/deft_api.cpp', 'csrc/deft/ops.cu'],
+        include_dirs=[
+            'csrc',
+        ],
+    )
+)
+
+
+def build(setup_kwargs: Dict[str, Any]) -> None:
+    setup_kwargs.update(
+        {
+            # 'ext_modules': ext_modules,
+            # 'cmdclass': {
+            #     'build_ext': BuildExtension.with_options(use_ninja=False),
+            # },
+        }
+    )

config.cmake

@@ -0,0 +1,41 @@
+# Whether to compile fp8 kernels or not.
+set(FLASHINFER_ENABLE_FP8 OFF)
+# Whether to compile bf16 kernels or not.
+set(FLASHINFER_ENABLE_BF16 OFF)
+# Whether to compile tvm bindings or not.
+set(FLASHINFER_TVM_BINDING OFF)
+# Whether to compile prefill kernel tests/benchmarks or not.
+set(FLASHINFER_PREFILL ON)
+# Whether to compile decode kernel tests/benchmarks or not.
+set(FLASHINFER_DECODE ON)
+# Whether to compile page kernel tests/benchmarks or not.
+set(FLASHINFER_PAGE ON)
+# Whether to compile cascade kernel tests/benchmarks or not.
+set(FLASHINFER_CASCADE ON)
+# Whether to compile sampling kernel tests/benchmarks or not.
+set(FLASHINFER_SAMPLING OFF)
+# Whether to compile normalization kernel tests/benchmarks or not.
+set(FLASHINFER_NORMALIZATION OFF)
+# Whether to compile fastdiv tests
+set(FLASHINFER_FASTDIV_TEST ON)
+# Whether to compile distributed tests
+set(FLASHINFER_DISTRIBUTED OFF)
+# The following configurations can impact the binary
+# size of the generated library
+set(FLASHINFER_GEN_LOGITS_POST_HOOKS 0)
+set(FLASHINFER_GEN_HEAD_DIMS 128)
+set(FLASHINFER_GEN_KV_LAYOUTS 0)
+set(FLASHINFER_GEN_POS_ENCODING_MODES 0)
+set(FLASHINFER_GEN_ALLOW_FP16_QK_REDUCTIONS "false")
+set(FLASHINFER_GEN_MASK_MODES 0)
+
+# Set target cuda architectures for tests/benchmarks, defaults to native.
+# "native" is a special value for CMAKE_CUDA_ARCHITECTURES which means use the architectures of the host's GPU.
+# it's new in CMake 3.24, if you are using an older of CMake or you want to use a different value, you can
+# set its value here. Supported CUDA architctures include 80;86;89;90
+# NOTE(Zihao): using "native" might be slow because whenever compile a cuda file with `-arch=native`, nvcc will spawn
+# a `__nvcc_device_query` process to get the architecture of the host's GPU, which could stall the compilation process.
+# So it's recommended to set it to a specific value if you know the architecture of the target GPU.
+# Example:
+# set(FLASHINFER_CUDA_ARCHITECTURES 80)
+set(FLASHINFER_CUDA_ARCHITECTURES 80)