README.md 11.6 KB
Newer Older
1
2
<img src=./images/logo-row.svg />

3
<div align="center">
4

5
# Tile Language
6

7
</div>
8

9
Tile Language (**tile-lang**) is a concise domain-specific language designed to streamline the development of high-performance GPU/CPU kernels (e.g., GEMM, Dequant GEMM, FlashAttention, LinearAttention). By employing a Pythonic syntax with an underlying compiler infrastructure on top of [TVM](https://tvm.apache.org/), tile-lang allows developers to focus on productivity without sacrificing the low-level optimizations necessary for state-of-the-art performance.
10

11
<img src=./images/MatmulExample.png />
12

Lei Wang's avatar
Lei Wang committed
13
## Latest News
14
- 14/04/2025 🚀: Added high-performance FlashMLA implementation for AMD MI300X, achieving performance parity with hand-optimized assembly kernels of Aiter! See [example_mla_amd](./examples/deepseek_mla/amd/README.md) for details.
15
- 03/03/2025 🚀: Added high-performance MLA Decoding support using only 80 lines of Python code, achieving performance on par with FlashMLA on H100 (see [example_mla_decode.py](./examples/deepseek_mla/example_mla_decode.py))! We also provide [documentation](./examples/deepseek_mla/README.md) explaining how TileLang achieves this.
16
- 02/15/2025 ✨: Added WebGPU Codegen support, see [Pull Request #86](https://github.com/tile-ai/tilelang/pull/86)!
Lei Wang's avatar
Lei Wang committed
17
- 02/12/2025 ✨: Excited to announce the release of [v0.1.0](https://github.com/tile-ai/tilelang/releases/tag/v0.1.0)!
18
- 02/10/2025 🚀: Added debug tools for TileLang—`T.print` for printing variables/buffers ([docs](https://tilelang.com/tutorials/debug_tools_for_tilelang.html)) and a memory layout plotter ([examples/plot_layout](./examples/plot_layout)).
Lei Wang's avatar
Lei Wang committed
19
20
- 01/20/2025 ✨: We are excited to announce that tile-lang, a dsl for high performance AI workloads, is now open source and available to the public!

21
## Tested Devices
22
Although tile-lang aims to be portable across a range of Devices, it has been specifically tested and validated on the following devices: for NVIDIA GPUs, this includes the H100 (with Auto TMA/WGMMA support), A100, V100, RTX 4090, RTX 3090, and RTX A6000; for AMD GPUs, it includes the MI250 (with Auto MatrixCore support) and the MI300X (with Async Copy support).
23
24
25
26
27
28
29
30

## OP Implementation Examples
**tile-lang** provides the building blocks to implement a wide variety of operators. Some examples include:

- [Matrix Multiplication](./examples/gemm/)
- [Dequantization GEMM](./examples/dequantize_gemm/)
- [Flash Attention](./examples/flash_attention/)
- [Flash Linear Attention](./examples/linear_attention/)
31
- [Flash MLA Decoding](./examples/deepseek_mla/)
32
- [Native Sparse Attention](./examples/native_sparse_attention/)
33

34
35
Within the `examples` directory, you will also find additional complex kernels—such as convolutions, forward/backward passes for FlashAttention, more operators will continuously be added.

36
37
38

## Benchmark Summary

39
TileLang achieves exceptional performance across a variety of computational patterns. Comprehensive benchmark scripts and settings are available at [tilelang-benchmark](https://github.com/tile-ai/tilelang-benchmark). Below are selected results showcasing its capabilities:
40

41
42
43
44
45
46
47
48
49
50
51
- MLA Decoding Performance on H100

  <div style="display: flex; gap: 10px; justify-content: center;">
    <div style="flex: 1;">
      <img src="./examples/deepseek_mla/figures/bs64_float16.png" alt="mla decode performance bs64 on H100" width="100%" />
    </div>
    <div style="flex: 1;">
      <img src="./examples/deepseek_mla/figures/bs128_float16.png" alt="mla decode performance bs128 on H100" width="100%" />
    </div>
  </div>
  
52
- Flash Attention Performance on H100
53

54
  <div align="center">    <img src="./images/mha_performance_h100.png" alt="operator performance on H100" width=80% />
55
56
  </div>

57
- Matmul Performance on GPUs (RTX 4090, A100, H100, MI300X)
58
59

  <div>
60
    <img src="./images/op_benchmark_consistent_gemm_fp16.png" alt="gemm fp16 performance on Gpus" />
61
62
  </div>

Lei Wang's avatar
Lei Wang committed
63
64
65
66
67
68
- Dequantize Matmul Performance on A100

  <div>
    <img src="./images/op_benchmark_a100_wq_gemv.png" alt="dequantize gemv performance on A100" />
  </div>

69
70
71
72
73
74
75
76
77
78
79
80
## Installation
### Method 1: Install with Pip

The quickest way to get started is to install the latest release from PyPI:

```bash
pip install tilelang
```

Alternatively, you can install directly from the GitHub repository:

```bash
81
pip install git+https://github.com/tile-ai/tilelang
82
83
84
85
86
```

Or install locally:

```bash
87
88
89
90
# install required system dependencies
sudo apt-get update
sudo apt-get install -y python3-setuptools gcc libtinfo-dev zlib1g-dev build-essential cmake libedit-dev libxml2-dev

91
pip install -e . -v # remove -e option if you don't want to install in editable mode, -v for verbose output
92
93
94
95
```

### Method 2: Build from Source
We currently provide three ways to install **tile-lang** from source:
96
 - [Install from Source (using your own TVM installation)](./docs/get_started/Installation.md#method-1-install-from-source-using-your-own-tvm-installation)
97
98
 - [Install from Source (using the bundled TVM submodule)](./docs/get_started/Installation.md#method-2-install-from-source-using-the-bundled-tvm-submodule)
 - [Install Using the Provided Script](./docs/get_started/Installation.md#method-3-install-using-the-provided-script)
99

100
101
102
103
104
105
106
107
108
109
### Method 3: Install with Nightly Version

For users who want access to the latest features and improvements before official releases, we provide nightly builds of **tile-lang**.

```bash
pip install tilelang -f https://tile-ai.github.io/whl/nightly/cu121/
# or pip install tilelang --find-links https://tile-ai.github.io/whl/nightly/cu121/
```

> **Note:** Nightly builds contain the most recent code changes but may be less stable than official releases. They're ideal for testing new features or if you need a specific bugfix that hasn't been released yet.
110
111
112

## Quick Start

113
In this section, you'll learn how to write and execute a straightforward GEMM (matrix multiplication) kernel using tile-lang, followed by techniques for layout optimizations, pipelining, and L2-cache–friendly swizzling.
114

Lei Wang's avatar
Lei Wang committed
115
### GEMM Example with Annotations (Layout, L2 Cache Swizzling, and Pipelining, etc.)
116
117
118
119

Below is an example that demonstrates more advanced features: layout annotation, parallelized copy, and swizzle for improved L2 cache locality. This snippet shows how to adapt your kernel to maximize performance on complex hardware.

```python
120
import tilelang
121
122
123
124
125
126
127
128
129
import tilelang.language as T
# `make_mma_swizzle_layout` is a python defined layout function
# specifically designed for for MMA operations
# which ensures the consistency with the nvidia CUTLASS Library.
# to avoid bank conflicts and maximize the performance.
from tilelang.intrinsics import (
    make_mma_swizzle_layout as make_swizzle_layout,)

def matmul(M, N, K, block_M, block_N, block_K, dtype="float16", accum_dtype="float"):
130
    # add decorator @tilelang.jit if you want to return a torch function
131
132
    @T.prim_func
    def main(
133
134
135
        A: T.Tensor((M, K), dtype),
        B: T.Tensor((K, N), dtype),
        C: T.Tensor((M, N), dtype),
136
    ):
137
        # Initialize Kernel Context
138
139
140
141
142
        with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=128) as (bx, by):
            A_shared = T.alloc_shared((block_M, block_K), dtype)
            B_shared = T.alloc_shared((block_K, block_N), dtype)
            C_local  = T.alloc_fragment((block_M, block_N), accum_dtype)

Lei Wang's avatar
Lei Wang committed
143
144
            # Apply layout optimizations or define your own layout (Optional)
            # If not specified, we will deduce the layout automatically
145
146
147
148
            # T.annotate_layout({
            #     A_shared: make_swizzle_layout(A_shared),
            #     B_shared: make_swizzle_layout(B_shared),
            # })
149

Lei Wang's avatar
Lei Wang committed
150
            # Enable rasterization for better L2 cache locality (Optional)
151
            # T.use_swizzle(panel_size=10, enable=True)
152
153
154
155

            # Clear local accumulation
            T.clear(C_local)

156
            for ko in T.Pipelined(T.ceildiv(K, block_K), num_stages=3):
157
                # Copy tile of A
Lei Wang's avatar
Lei Wang committed
158
                # This is a sugar syntax for parallelized copy
159
                T.copy(A[by * block_M, ko * block_K], A_shared)
160
161

                # Demonstrate parallelized copy from global to shared for B
162
163
                for k, j in T.Parallel(block_K, block_N):
                    B_shared[k, j] = B[ko * block_K + k, bx * block_N + j]
164
165

                # Perform a tile-level GEMM on the shared buffers
Lei Wang's avatar
Lei Wang committed
166
                # Currently we dispatch to the cute/hip on Nvidia/AMD GPUs
167
168
169
170
171
172
                T.gemm(A_shared, B_shared, C_local)

            # Copy result back to global memory
            T.copy(C_local, C[by * block_M, bx * block_N])

    return main
173
174


175
# 1. Define the kernel (matmul) with the desired dimensions
176
177
178
179
180
181
func = matmul(1024, 1024, 1024, 128, 128, 32)

# 2. Compile the kernel into a torch function
# out_idx specifies the index of the output buffer in the argument list
# if out_idx is specified, the tensor will be created during runtime
# target currently can be "cuda" or "hip" or "cpu".
182
jit_kernel = tilelang.compile(func, out_idx=[2], target="cuda")
183
184
185
186
187
188
189
190
191

# 3. Test the kernel in Python with PyTorch data
import torch

# Create random input tensors on the GPU
a = torch.randn(1024, 1024, device="cuda", dtype=torch.float16)
b = torch.randn(1024, 1024, device="cuda", dtype=torch.float16)


192
# Run the kernel through the JIT-compiled function
193
194
195
196
197
198
199
200
201
202
203
204
205
c = jit_kernel(a, b)

# Reference multiplication using PyTorch
ref_c = a @ b

# Validate correctness
torch.testing.assert_close(c, ref_c, rtol=1e-2, atol=1e-2)
print("Kernel output matches PyTorch reference.")

# 4. Retrieve and inspect the generated CUDA source (optional)
cuda_source = jit_kernel.get_kernel_source()
print("Generated CUDA kernel:\n", cuda_source)

206
# 5.Pofile latency with the profiler
207
208
209
210
211
profiler = jit_kernel.get_profiler()

latency = profiler.do_bench()

print(f"Latency: {latency} ms")
212
213
```

214
215
216
217
### Dive Deep into TileLang Beyond GEMM

In addition to GEMM, we provide a variety of examples to showcase the versatility and power of TileLang, including:

218
- [Dequantize GEMM](./examples/dequantize_gemm/): Achieve high-performance dequantization by **fine-grained control over per-thread operations**, with many features now adopted as default behaviors in [BitBLAS](https://github.com/microsoft/BitBLAS), which utilizing magic layout transformation and intrins to accelerate dequantize gemm.
219
220
221
222
- [FlashAttention](./examples/flash_attention/): Enable cross-operator fusion with simple and intuitive syntax, and we also provide an example of auto tuning.
- [LinearAttention](./examples/linear_attention/): Examples include RetNet and Mamba implementations.
- [Convolution](./examples/convolution/): Implementations of Convolution with IM2Col.

Lei Wang's avatar
Lei Wang committed
223
224
225
226
## Upcoming Features

Check our [tilelang v0.2.0 release plan](https://github.com/tile-ai/tilelang/issues/79) for upcoming features.

227
228
---

FeiyangChen's avatar
FeiyangChen committed
229
TileLang has now been used in project [BitBLAS](https://github.com/microsoft/BitBLAS) and [AttentionEngine](https://github.com/microsoft/AttentionEngine).
230

231
232
233
234
235
236
## Join the Discussion

Welcome to join our Discord community for discussions, support, and collaboration!

[![Join our Discord](https://img.shields.io/badge/Discord-Join%20Us-blue?logo=discord&style=for-the-badge)](https://discord.gg/TUrHyJnKPG)

237
## Acknowledgements
238

239
We would like to express our gratitude to the [TVM](https://github.com/apache/tvm) community for their invaluable contributions. The initial version of this project was mainly developed by [LeiWang1999](https://github.com/LeiWang1999), [chengyupku](https://github.com/chengyupku) and [nox-410](https://github.com/nox-410) with supervision from Prof. [Zhi Yang](https://yangzhihome.github.io) at Peking University. Part of this work was carried out during an internship at Microsoft Research, where Dr. Lingxiao Ma, Dr. Yuqing Xia, Dr. Jilong Xue, and Dr. Fan Yang offered valuable advice and support. We deeply appreciate their mentorship and contributions.