Upload custom kernels
Browse files
build/torch-universal/triton_llama_mlp/__init__.py
ADDED
|
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from . import layers
|
| 2 |
+
__all__ = ["layers"]
|
build/torch-universal/triton_llama_mlp/layers.py
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from .mlp import TritonLlamaMLP
|
| 2 |
+
|
| 3 |
+
__all__ = ["TritonLlamaMLP"]
|
build/torch-universal/triton_llama_mlp/mlp.py
ADDED
|
@@ -0,0 +1,210 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
import triton
|
| 4 |
+
import triton.language as tl
|
| 5 |
+
from typing import Callable, Optional
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
@triton.jit
|
| 9 |
+
def matmul_kernel(
|
| 10 |
+
# Pointers to matrices
|
| 11 |
+
a_ptr, b_ptr, c_ptr,
|
| 12 |
+
# Matrix dimensions
|
| 13 |
+
M, N, K,
|
| 14 |
+
# The stride variables represent how much to increase the ptr by when moving by 1
|
| 15 |
+
# element in a particular dimension. E.g. `stride_am` is how much to increase `a_ptr`
|
| 16 |
+
# by to get the element one row down (A has M rows).
|
| 17 |
+
stride_am, stride_ak, #
|
| 18 |
+
stride_bk, stride_bn, #
|
| 19 |
+
stride_cm, stride_cn,
|
| 20 |
+
# Meta-parameters
|
| 21 |
+
BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr, #
|
| 22 |
+
GROUP_SIZE_M: tl.constexpr, #
|
| 23 |
+
ACTIVATION: tl.constexpr = None #
|
| 24 |
+
):
|
| 25 |
+
"""Kernel for computing the matmul C = A x B.
|
| 26 |
+
A has shape (M, K), B has shape (K, N) and C has shape (M, N)
|
| 27 |
+
"""
|
| 28 |
+
# -----------------------------------------------------------
|
| 29 |
+
# Map program ids `pid` to the block of C it should compute.
|
| 30 |
+
# This is done in a grouped ordering to promote L2 data reuse.
|
| 31 |
+
# See above `L2 Cache Optimizations` section for details.
|
| 32 |
+
pid = tl.program_id(axis=0)
|
| 33 |
+
num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
|
| 34 |
+
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
| 35 |
+
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
| 36 |
+
group_id = pid // num_pid_in_group
|
| 37 |
+
first_pid_m = group_id * GROUP_SIZE_M
|
| 38 |
+
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
| 39 |
+
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
| 40 |
+
pid_n = (pid % num_pid_in_group) // group_size_m
|
| 41 |
+
|
| 42 |
+
# ----------------------------------------------------------
|
| 43 |
+
# Create pointers for the first blocks of A and B.
|
| 44 |
+
# We will advance this pointer as we move in the K direction
|
| 45 |
+
# and accumulate
|
| 46 |
+
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
| 47 |
+
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
| 48 |
+
# See above `Pointer Arithmetic` section for details
|
| 49 |
+
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
|
| 50 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
| 51 |
+
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
| 52 |
+
a_ptrs = a_ptr + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
|
| 53 |
+
b_ptrs = b_ptr + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
|
| 54 |
+
|
| 55 |
+
# -----------------------------------------------------------
|
| 56 |
+
# Iterate to compute a block of the C matrix.
|
| 57 |
+
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
| 58 |
+
# of fp32 values for higher accuracy.
|
| 59 |
+
# `accumulator` will be converted back to fp16 after the loop.
|
| 60 |
+
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
| 61 |
+
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
| 62 |
+
# Load the next block of A and B, generate a mask by checking the K dimension.
|
| 63 |
+
# If it is out of bounds, set it to 0.
|
| 64 |
+
a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
|
| 65 |
+
b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
|
| 66 |
+
# We accumulate along the K dimension.
|
| 67 |
+
accumulator += tl.dot(a, b)
|
| 68 |
+
# Advance the ptrs to the next K block.
|
| 69 |
+
a_ptrs += BLOCK_SIZE_K * stride_ak
|
| 70 |
+
b_ptrs += BLOCK_SIZE_K * stride_bk
|
| 71 |
+
# You can fuse arbitrary activation functions here
|
| 72 |
+
# while the accumulator is still in FP32!
|
| 73 |
+
|
| 74 |
+
c = accumulator.to(tl.float32)
|
| 75 |
+
|
| 76 |
+
# -----------------------------------------------------------
|
| 77 |
+
# Write back the block of the output matrix C with masks.
|
| 78 |
+
offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
| 79 |
+
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
| 80 |
+
c_ptrs = c_ptr + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
|
| 81 |
+
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
|
| 82 |
+
tl.store(c_ptrs, c, mask=c_mask)
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
# We can fuse `leaky_relu` by providing it as an `ACTIVATION` meta-parameter in `matmul_kernel`.
|
| 86 |
+
@triton.jit
|
| 87 |
+
def silu(x):
|
| 88 |
+
return x * tl.sigmoid(x)
|
| 89 |
+
|
| 90 |
+
def matmul(a, b, activation=""):
|
| 91 |
+
# Check constraints.
|
| 92 |
+
assert a.shape[1] == b.shape[0], "Incompatible dimensions"
|
| 93 |
+
assert a.is_contiguous(), "Matrix A must be contiguous"
|
| 94 |
+
M, K = a.shape
|
| 95 |
+
K, N = b.shape
|
| 96 |
+
BLOCK_SIZE_M = 32
|
| 97 |
+
BLOCK_SIZE_N = 32
|
| 98 |
+
BLOCK_SIZE_K = 32
|
| 99 |
+
GROUP_SIZE_M = 8
|
| 100 |
+
# Allocates output.
|
| 101 |
+
c = torch.empty((M, N), device=a.device, dtype=torch.float)
|
| 102 |
+
# 1D launch kernel where each block gets its own program.
|
| 103 |
+
grid = lambda META: (triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(N, META['BLOCK_SIZE_N']), )
|
| 104 |
+
matmul_kernel[grid](
|
| 105 |
+
a, b, c, #
|
| 106 |
+
M, N, K, #
|
| 107 |
+
a.stride(0), a.stride(1), #
|
| 108 |
+
b.stride(0), b.stride(1), #
|
| 109 |
+
c.stride(0), c.stride(1), #
|
| 110 |
+
BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K, #
|
| 111 |
+
GROUP_SIZE_M, #
|
| 112 |
+
ACTIVATION=activation #
|
| 113 |
+
)
|
| 114 |
+
return c
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
class TritonLlamaMLP(nn.Module):
|
| 118 |
+
"""LlamaMLP implementation using Triton kernels for matrix multiplication"""
|
| 119 |
+
gate_proj: nn.Linear
|
| 120 |
+
up_proj: nn.Linear
|
| 121 |
+
down_proj: nn.Linear
|
| 122 |
+
act_fn: Callable
|
| 123 |
+
def forward(self, x):
|
| 124 |
+
# Replace nn.Linear with matmul using triton kernel
|
| 125 |
+
# Save original shape for reshaping back later
|
| 126 |
+
original_shape = x.shape
|
| 127 |
+
# Reshape input to 2D for matmul: (*, hidden_size) -> (batch_size*seq_len, hidden_size)
|
| 128 |
+
x_2d = x.reshape(-1, x.size(-1))
|
| 129 |
+
|
| 130 |
+
# Gate projection
|
| 131 |
+
gate_output = matmul(x_2d, self.gate_proj.weight.t())
|
| 132 |
+
if self.gate_proj.bias is not None:
|
| 133 |
+
gate_output += self.gate_proj.bias
|
| 134 |
+
|
| 135 |
+
# Up projection
|
| 136 |
+
up_output = matmul(x_2d, self.up_proj.weight.t())
|
| 137 |
+
if self.up_proj.bias is not None:
|
| 138 |
+
up_output += self.up_proj.bias
|
| 139 |
+
|
| 140 |
+
# Apply activation function and element-wise multiplication
|
| 141 |
+
intermediate_output = self.act_fn(gate_output) * up_output
|
| 142 |
+
|
| 143 |
+
# Final projection
|
| 144 |
+
down_output = matmul(intermediate_output, self.down_proj.weight.t())
|
| 145 |
+
if self.down_proj.bias is not None:
|
| 146 |
+
down_output += self.down_proj.bias
|
| 147 |
+
|
| 148 |
+
# Reshape back to original dimensions: (batch_size*seq_len, hidden_size) -> (*, hidden_size)
|
| 149 |
+
return down_output.reshape(original_shape)
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
# def test_triton_llama_mlp():
|
| 153 |
+
# """Test that TritonLlamaMLP produces the same output as LlamaMLP from transformers."""
|
| 154 |
+
# import torch
|
| 155 |
+
# import torch.nn.functional as F
|
| 156 |
+
|
| 157 |
+
# # Skip test if CUDA is not available
|
| 158 |
+
# if not torch.cuda.is_available():
|
| 159 |
+
# print("CUDA not available, skipping test")
|
| 160 |
+
# return True
|
| 161 |
+
|
| 162 |
+
# # Define test parameters
|
| 163 |
+
# batch_size = 2
|
| 164 |
+
# seq_len = 4
|
| 165 |
+
# hidden_size = 128
|
| 166 |
+
# intermediate_size = 256
|
| 167 |
+
|
| 168 |
+
# # Create input tensor
|
| 169 |
+
# x = torch.randn(batch_size, seq_len, hidden_size, device="cuda", dtype=torch.float)
|
| 170 |
+
|
| 171 |
+
# # Create a standard PyTorch implementation for comparison
|
| 172 |
+
# class StandardLlamaMLP(nn.Module):
|
| 173 |
+
# def __init__(self, hidden_size, intermediate_size):
|
| 174 |
+
# super().__init__()
|
| 175 |
+
# self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
| 176 |
+
# self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
| 177 |
+
# self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
|
| 178 |
+
# self.act_fn = F.silu
|
| 179 |
+
|
| 180 |
+
# def forward(self, x):
|
| 181 |
+
# return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
| 182 |
+
|
| 183 |
+
# # Initialize models
|
| 184 |
+
# standard_mlp = StandardLlamaMLP(hidden_size, intermediate_size).to("cuda").to(torch.float)
|
| 185 |
+
|
| 186 |
+
# # Create our Triton implementation
|
| 187 |
+
# triton_mlp = TritonLlamaMLP()
|
| 188 |
+
# triton_mlp.gate_proj = standard_mlp.gate_proj
|
| 189 |
+
# triton_mlp.up_proj = standard_mlp.up_proj
|
| 190 |
+
# triton_mlp.down_proj = standard_mlp.down_proj
|
| 191 |
+
# triton_mlp.act_fn = standard_mlp.act_fn
|
| 192 |
+
|
| 193 |
+
# # Run both implementations
|
| 194 |
+
# with torch.no_grad():
|
| 195 |
+
# standard_output = standard_mlp(x)
|
| 196 |
+
# triton_output = triton_mlp(x)
|
| 197 |
+
|
| 198 |
+
# # Compare outputs
|
| 199 |
+
# max_diff = torch.max(torch.abs(standard_output - triton_output))
|
| 200 |
+
# print(f"Maximum difference between standard and Triton implementation: {max_diff}")
|
| 201 |
+
|
| 202 |
+
# # Check if outputs are close enough (allowing for some floating point differences)
|
| 203 |
+
# is_close = torch.allclose(standard_output, triton_output, rtol=1e-2, atol=1e-2)
|
| 204 |
+
# print(f"Outputs match within tolerance: {is_close}")
|
| 205 |
+
|
| 206 |
+
# return is_close
|
| 207 |
+
|
| 208 |
+
# if __name__ == "__main__":
|
| 209 |
+
# test_triton_llama_mlp()
|
| 210 |
+
|
torch-ext/triton_llama_mlp/mlp.py
CHANGED
|
@@ -71,7 +71,7 @@ def matmul_kernel(
|
|
| 71 |
# You can fuse arbitrary activation functions here
|
| 72 |
# while the accumulator is still in FP32!
|
| 73 |
|
| 74 |
-
c = accumulator.to(tl.
|
| 75 |
|
| 76 |
# -----------------------------------------------------------
|
| 77 |
# Write back the block of the output matrix C with masks.
|
|
@@ -98,7 +98,7 @@ def matmul(a, b, activation=""):
|
|
| 98 |
BLOCK_SIZE_K = 32
|
| 99 |
GROUP_SIZE_M = 8
|
| 100 |
# Allocates output.
|
| 101 |
-
c = torch.empty((M, N), device=a.device, dtype=torch.
|
| 102 |
# 1D launch kernel where each block gets its own program.
|
| 103 |
grid = lambda META: (triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(N, META['BLOCK_SIZE_N']), )
|
| 104 |
matmul_kernel[grid](
|
|
@@ -166,7 +166,7 @@ class TritonLlamaMLP(nn.Module):
|
|
| 166 |
# intermediate_size = 256
|
| 167 |
|
| 168 |
# # Create input tensor
|
| 169 |
-
# x = torch.randn(batch_size, seq_len, hidden_size, device="cuda", dtype=torch.
|
| 170 |
|
| 171 |
# # Create a standard PyTorch implementation for comparison
|
| 172 |
# class StandardLlamaMLP(nn.Module):
|
|
@@ -181,7 +181,7 @@ class TritonLlamaMLP(nn.Module):
|
|
| 181 |
# return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
| 182 |
|
| 183 |
# # Initialize models
|
| 184 |
-
# standard_mlp = StandardLlamaMLP(hidden_size, intermediate_size).to("cuda").to(torch.
|
| 185 |
|
| 186 |
# # Create our Triton implementation
|
| 187 |
# triton_mlp = TritonLlamaMLP()
|
|
|
|
| 71 |
# You can fuse arbitrary activation functions here
|
| 72 |
# while the accumulator is still in FP32!
|
| 73 |
|
| 74 |
+
c = accumulator.to(tl.float32)
|
| 75 |
|
| 76 |
# -----------------------------------------------------------
|
| 77 |
# Write back the block of the output matrix C with masks.
|
|
|
|
| 98 |
BLOCK_SIZE_K = 32
|
| 99 |
GROUP_SIZE_M = 8
|
| 100 |
# Allocates output.
|
| 101 |
+
c = torch.empty((M, N), device=a.device, dtype=torch.float)
|
| 102 |
# 1D launch kernel where each block gets its own program.
|
| 103 |
grid = lambda META: (triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(N, META['BLOCK_SIZE_N']), )
|
| 104 |
matmul_kernel[grid](
|
|
|
|
| 166 |
# intermediate_size = 256
|
| 167 |
|
| 168 |
# # Create input tensor
|
| 169 |
+
# x = torch.randn(batch_size, seq_len, hidden_size, device="cuda", dtype=torch.float)
|
| 170 |
|
| 171 |
# # Create a standard PyTorch implementation for comparison
|
| 172 |
# class StandardLlamaMLP(nn.Module):
|
|
|
|
| 181 |
# return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
| 182 |
|
| 183 |
# # Initialize models
|
| 184 |
+
# standard_mlp = StandardLlamaMLP(hidden_size, intermediate_size).to("cuda").to(torch.float)
|
| 185 |
|
| 186 |
# # Create our Triton implementation
|
| 187 |
# triton_mlp = TritonLlamaMLP()
|