Open Source LLM Software Stack — OpenAI Triton
Motivation
Matrix multiplications are a key building block of most modern high-performance computing systems. They are notoriously hard to optimize, hence their implementation is generally done by hardware vendors themselves as part of so-called “kernel libraries” (e.g., cuBLAS). Unfortunately, these libraries are often proprietary and cannot be easily customized to accommodate the needs of modern deep learning workloads (e.g., fused activation functions).³
Unfortunately, these libraries only support a restricted set of tensor operations, leaving the implementation of novel primitives to experts.
Triton
Triton is an open-source intermediate language and compiler for specifying and compiling tile ( statically shaped multi-dimensional sub-arrays ) programs into efficient GPU code.
Triton exposes intra-instance parallelism via operations on blocks — small arrays whose dimensions are powers of two — rather than a Single Instruction, Multiple Thread (SIMT)7 execution model. In doing so, Triton effectively abstracts away all the issues related to concurrency within CUDA thread blocks (e.g., memory coalescing, shared memory synchronization/conflicts, tensor core scheduling).
The execution of CUDA code on GPUs is supported by an SPMD programming model in which each kernel is associated with an identifable thread-block in a so-called launch grid. The Triton programming model is similar, but each kernel is single-threaded — though automatically parallelized — and associated with a set of global ranges that varies from instance to instance (see Figure 3). This approach leads to simpler kernels in which CUDA-like concurrency primitives (shared memory synchronization, inter-thread
communication, etc.) are inexistent.
import torch
import triton
import triton.language as tl
# `triton.jit`'ed functions can be auto-tuned by using the `triton.autotune` decorator, which consumes:
# - A list of `triton.Config` objects that define different configurations of
# meta-parameters (e.g., `BLOCK_SIZE_M`) and compilation options (e.g., `num_warps`) to try
# - An auto-tuning *key* whose change in values will trigger evaluation of all the
# provided configs
@triton.autotune(
configs=[
triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 256, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 8}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 256, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=5, num_warps=2),
triton.Config({'BLOCK_SIZE_M': 32, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=5, num_warps=2),
],
key=['M', 'N', 'K'],
)
@triton.jit
def matmul_kernel(
# Pointers to matrices
a_ptr, b_ptr, c_ptr,
# Matrix dimensions
M, N, K,
# The stride variables represent how much to increase the ptr by when moving by 1
# element in a particular dimension. E.g. `stride_am` is how much to increase `a_ptr`
# by to get the element one row down (A has M rows).
stride_am, stride_ak,
stride_bk, stride_bn,
stride_cm, stride_cn,
# Meta-parameters
BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr,
ACTIVATION: tl.constexpr,
):
"""Kernel for computing the matmul C = A x B.
A has shape (M, K), B has shape (K, N) and C has shape (M, N)
"""
# -----------------------------------------------------------
# Map program ids `pid` to the block of C it should compute.
# This is done in a grouped ordering to promote L2 data reuse.
# See above `L2 Cache Optimizations` section for details.
pid = tl.program_id(axis=0)
num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
num_pid_in_group = GROUP_SIZE_M * num_pid_n
group_id = pid // num_pid_in_group
first_pid_m = group_id * GROUP_SIZE_M
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
pid_m = first_pid_m + (pid % group_size_m)
pid_n = (pid % num_pid_in_group) // group_size_m
# ----------------------------------------------------------
# Create pointers for the first blocks of A and B.
# We will advance this pointer as we move in the K direction
# and accumulate
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
# See above `Pointer Arithmetics` section for details
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
offs_k = tl.arange(0, BLOCK_SIZE_K)
a_ptrs = a_ptr + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
b_ptrs = b_ptr + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
# -----------------------------------------------------------
# Iterate to compute a block of the C matrix.
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
# of fp32 values for higher accuracy.
# `accumulator` will be converted back to fp16 after the loop.
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
# Load the next block of A and B, generate a mask by checking the K dimension.
# If it is out of bounds, set it to 0.
a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
# We accumulate along the K dimension.
accumulator += tl.dot(a, b)
# Advance the ptrs to the next K block.
a_ptrs += BLOCK_SIZE_K * stride_ak
b_ptrs += BLOCK_SIZE_K * stride_bk
# You can fuse arbitrary activation functions here
# while the accumulator is still in FP32!
if ACTIVATION == "leaky_relu":
accumulator = leaky_relu(accumulator)
c = accumulator.to(tl.float16)
# -----------------------------------------------------------
# Write back the block of the output matrix C with masks.
offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
c_ptrs = c_ptr + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
tl.store(c_ptrs, c, mask=c_mask)
# We can fuse `leaky_relu` by providing it as an `ACTIVATION` meta-parameter in `_matmul`.
@triton.jit
def leaky_relu(x):
x = x + 1
return tl.where(x >= 0, x, 0.01 * x)
a_ptrs, b_ptrs are input matrices tiles. tl.load(a_ptrs, .. ) loads a_ptrs tile; tl.dot(a, b) is a matrix product of two tiles. a_ptrs pointer to the next block is advanced in BLOCK_SIZE_K increments.
def matmul(a, b, activation=""):
# Check constraints.
assert a.shape[1] == b.shape[0], "Incompatible dimensions"
assert a.is_contiguous(), "Matrix A must be contiguous"
assert b.is_contiguous(), "Matrix B must be contiguous"
M, K = a.shape
K, N = b.shape
# Allocates output.
c = torch.empty((M, N), device=a.device, dtype=a.dtype)
# 1D launch kernel where each block gets its own program.
grid = lambda META: (
triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(N, META['BLOCK_SIZE_N']),
)
matmul_kernel[grid](
a, b, c,
M, N, K,
a.stride(0), a.stride(1),
b.stride(0), b.stride(1),
c.stride(0), c.stride(1),
ACTIVATION=activation
)
return c
This differs from PyTorch’s⁴ internal CUDA code, whose use of temporary memory makes it more general but significantly slower (below). The bottom line here is not that Triton is inherently better, but that it simplifies the development of specialized kernels that can be much faster than those found in general-purpose libraries.
The @triton.jit decorator works by walking the Abstract Syntax Tree (AST) of the provided Python function so as to generate Triton-IR on-the-fly using a common SSA construction algorithm.8 The resulting IR code is then simplified, optimized and automatically parallelized by our compiler backend, before being converted into high-quality LLVM-IR—and eventually PTX—for execution on recent NVIDIA GPUs.
We have found that the use of blocked program representations via Triton-IR allows our compiler to automatically perform a wide variety of important program optimizations. For example, data can be automatically stashed to shared memory by looking at the operands of computationally intensive block-level operations (e.g., tl.dot)—and allocated/synchronized using standard liveness analysis techniques.
On the other hand, Triton programs can be efficiently and automatically parallelized both (1) across SMs by executing different kernel instances concurrently, and (2) within SMs by analyzing the iteration space of each block-level operation and partitioning it adequately across different SIMD units, as shown below.
Triton performs significantly worse than cuBLAS on transformers, except for large matrices.
¹ Ran on NVIDIA T4 GPU ( Colab ). We were not able to confirm / reproduce better tutorial results reported in Triton tutorial.
³ Out of all the Domain Specific Languages and JIT-compilers available, Triton is perhaps most similar to Numba: kernels are defined as decorated Python functions, and launched concurrently with different program_id’s on a grid of so-called instances.
Numba reads the Python bytecode for a decorated function and combines this with information about the types of the input arguments to the function. It analyzes and optimizes your code, and finally uses the LLVM compiler library to generate a machine code version of your function, tailored to your CPU capabilities. This compiled version is then used every time your function is called.
³ Another potential benefit of Triton is portability — same code will run unmodified on AMD GPUs.
⁴ PyTorch @torch.compile ( makes PyTorch code run faster by JIT-compiling PyTorch code into optimized kernels ) ,uses Triton.
