main function
void
main()
Implementation
void main() {
CudaEngine.initialize(debug: false);
Random random = Random();
int numHeads = 8;
double learningRate = 0.01;
int epochsPerTest = 10; // Lowered to 10 so you don't wait hours on the massive sizes
print('====================================================================================================');
print(' GPU MULTI-HEAD ATTENTION HEAVYWEIGHT STRESS TEST ');
print('====================================================================================================');
print('Heads: $numHeads | LR: $learningRate | Epochs/Test: $epochsPerTest');
print('----------------------------------------------------------------------------------------------------');
print(' SeqLen | dModel | Compile(ms) | Avg Time(ms) | Throughput (GB/s) | Compute (TFLOPs) | Final Loss');
print('----------------------------------------------------------------------------------------------------');
List<int> sequences = <int>[128, 512, 1024];
List<int> dimensions = <int>[256, 512, 1024,2048, 4096, 8192];
for (int s = 0; s < sequences.length; s = s + 1) {
int seqLength = sequences[s];
for (int d = 0; d < dimensions.length; d = d + 1) {
int dModel = dimensions[d];
// 1. Generate Input Matrix
List<List<double>> hInput = <List<double>>[];
for (int i = 0; i < seqLength; i = i + 1) {
List<double> row = <double>[];
for (int j = 0; j < dModel; j = j + 1) {
row.add((random.nextDouble() * 2.0) - 1.0);
}
hInput.add(row);
}
// 2. Generate Target Matrix
List<List<double>> hTarget = <List<double>>[];
for (int i = 0; i < seqLength; i = i + 1) {
List<double> row = <double>[];
for (int j = 0; j < dModel; j = j + 1) {
row.add(0.5);
}
hTarget.add(row);
}
GPUTensor<Matrix> input = GPUTensor<Matrix>(hInput);
GPUTensor<Matrix> target = GPUTensor<Matrix>(hTarget);
// 3. Build Layer
MultiHeadAttentionTL mhaLayer = MultiHeadAttentionTL(dModel, numHeads);
mhaLayer.build(input);
Stopwatch compileSw = Stopwatch();
compileSw.start();
// ===================================================================
// FORWARD TAPE
// ===================================================================
CommandBuffer fTape = CommandBuffer();
List<GPUTensor> intermediates = <GPUTensor>[];
GPUTensor<Matrix> output = mhaLayer.forward(input, fTape, intermediates) as GPUTensor<Matrix>;
GPUTensor<Scalar> loss = GPUTensor<Scalar>(0.0);
fTape.putInt(OP_MSE_LOSS_FORWARD);
fTape.putString(output.id);
fTape.putString(target.id);
fTape.putString(loss.id);
loss.creator = GPUNode(
<GPUTensor>[output, target],
(CommandBuffer bTape) {
bTape.putInt(OP_MSE_LOSS_BACKWARD);
bTape.putString('${output.id}_grad');
bTape.putString(output.id);
bTape.putString(target.id);
bTape.putString('${loss.id}_grad');
},
opName: 'mse_loss_manual',
);
Uint8List forwardBytes = fTape.bytes();
// ===================================================================
// BACKWARD & OPTIMIZER TAPES
// ===================================================================
CommandBuffer bTape = CommandBuffer();
SGDGPU optimizer = SGDGPU(mhaLayer.parameters, learningRate);
optimizer.zeroGrad(bTape);
for (int i = 0; i < intermediates.length; i = i + 1) {
bTape.putInt(OP_ZERO_GRAD);
bTape.putString('${intermediates[i].id}_grad');
}
bTape.putInt(OP_ZERO_GRAD);
bTape.putString('${input.id}_grad');
bTape.putInt(OP_ZERO_GRAD);
bTape.putString('${output.id}_grad');
bTape.putInt(OP_FILL);
bTape.putString('${loss.id}_grad');
bTape.putFloat(1.0);
loss.backward(bTape);
Uint8List backwardBytes = bTape.bytes();
CommandBuffer oTape = CommandBuffer();
optimizer.step(oTape);
Uint8List optimizeBytes = oTape.bytes();
compileSw.stop();
// ===================================================================
// EXECUTION LOOP & METRIC CALCULATIONS
// ===================================================================
// Math for FLOPs (Forward = ~8*S*D^2 + 4*S^2*D. Backward = ~2x Forward)
double seqD = seqLength.toDouble();
double modD = dModel.toDouble();
double headsD = numHeads.toDouble();
double flopsFwd = (8.0 * seqD * modD * modD) + (4.0 * seqD * seqD * modD);
double totalFlopsStep = 3.0 * flopsFwd; // Fwd + Bwd
// Math for Memory Traffic (Bytes Read/Written to VRAM)
// Estimates base matrices + O(N^2) attention matrices. Bwd is roughly 2x Fwd.
double bytesFwd = 4.0 * ((3.0 * modD * modD) + (8.0 * seqD * modD) + (2.0 * headsD * seqD * seqD));
double totalBytesStep = 3.0 * bytesFwd;
Stopwatch runSw = Stopwatch();
// Warmup
CudaEngine.run(forwardBytes);
runSw.start();
for (int epoch = 1; epoch <= epochsPerTest; epoch = epoch + 1) {
CudaEngine.run(forwardBytes);
CudaEngine.run(backwardBytes);
CudaEngine.run(optimizeBytes);
}
runSw.stop();
loss.toCpu();
double currentLoss = loss.value;
double avgEpochSec = (runSw.elapsedMicroseconds / 1000000.0) / epochsPerTest;
double avgEpochMs = avgEpochSec * 1000.0;
double tflops = (totalFlopsStep / avgEpochSec) / 1000000000000.0;
double gbps = (totalBytesStep / avgEpochSec) / 1000000000.0;
String sSeq = seqLength.toString().padRight(6);
String sDim = dModel.toString().padRight(6);
String sComp = compileSw.elapsedMilliseconds.toString().padRight(11);
String sAvg = avgEpochMs.toStringAsFixed(1).padRight(12);
String sGbps = gbps.toStringAsFixed(2).padRight(17);
String sTflops = tflops.toStringAsFixed(4).padRight(16);
String sLoss = currentLoss.toStringAsFixed(5);
print(' $sSeq | $sDim | $sComp | $sAvg | $sGbps | $sTflops | $sLoss');
// Free Memory
mhaLayer.free();
input.free();
target.free();
loss.free();
output.free();
for (int i = 0; i < intermediates.length; i = i + 1) {
intermediates[i].free();
}
}
}
print('----------------------------------------------------------------------------------------------------');
print('Benchmark Complete.');
}