vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml/examples: add backend support for numerical optimization (ggml/949)

Browse source 
* CUDA eval works

* stochastic gradient descent op

* Adam except decay

* CUDA CROSS_ENTROPY_LOSS_BACK

* CUDA mnist-fc training works

* backend CLI arg

* refactor gguf load

* remove sched from opt_step_adam

* implement l1 regularization (weight decay)

* extra call to add optimizer

* initialize gradients with ggml_graph_reset

* gradient accumulation

* increment iter per eval instead of epoch

* adjust backend interfaces

* fix ggml_graph_reset without backend

* fix ggml graph export/import

* fixup

* rename

* revert ggml_opt changes

* more general CUDA repeat_back

* update documentation, fix CNN

* validation split

* add clarifying comment

* optimize PyTorch training

* adjust buffer size, thread count

* fix 0.0f validation split

* Update examples/mnist/mnist-common.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* fix gradient accumulation

* tensor flag for accumulators -> tensor hash set

* Update include/ggml.h

Co-authored-by: slaren <slarengh@gmail.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: slaren <slarengh@gmail.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: slaren <slarengh@gmail.com>

* fix test prints

* Update src/ggml-backend.c

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* better CUDA support for noncontiguous out_prod

* add comment

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Johannes Gäßler 2024-09-20 19:04:44 +03:00 • committed by Georgi Gerganov
parent a6809c6a2e
commit 424c5d00a9
24 changed files with 883 additions and 129 deletions
Download patch file Download diff file Expand all files Collapse all files

60
ggml/src/ggml-cuda/cross-entropy-loss.cu
View file

@ -71,6 +71,32 @@ static __global__ void cross_entropy_loss_f32(const float * logits, const float
dst[blockIdx.x] = loss;
}
static __global__ void cross_entropy_loss_back_f32(const float * logits, const float * labels, const float * loss, float * dst, const int nclasses) {
extern __shared__ float tmp[];
float maxval = -INFINITY;
for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
const float val = logits[blockIdx.x*nclasses + i];
maxval = fmaxf(maxval, val);
tmp[i] = val;
}
maxval = warp_reduce_max(maxval);
float sum = 0.0f;
for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
const float val = expf(tmp[i] - maxval);
sum += val;
tmp[i] = val;
}
sum = warp_reduce_sum(sum);
const float sm_scale = 1.0f/sum;
const float d_by_nrows = *loss/gridDim.x;
for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
dst[blockIdx.x*nclasses + i] = (tmp[i]*sm_scale - labels[blockIdx.x*nclasses + i])*d_by_nrows;
}
}
void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
@ -104,3 +130,37 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
// Combine results from individual blocks:
sum_f32_cuda(pool, dst_tmp.ptr, dst_d, blocks_num.x, stream);
}
void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * opt0 = dst->src[2];
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(opt0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(src1));
GGML_ASSERT(ggml_is_contiguous(opt0));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(ggml_are_same_shape(src0, src1));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
const int64_t ne00 = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
const float * src0_d = (const float *) src0->data;
const float * src1_d = (const float *) src1->data;
const float * opt0_d = (const float *) opt0->data;
float * dst_d = (float *) dst->data;
cudaStream_t stream = ctx.stream();
const dim3 blocks_dim(WARP_SIZE, 1, 1);
const dim3 blocks_num(nrows, 1, 1);
const int shmem = ne00*sizeof(float);
cross_entropy_loss_back_f32<<<blocks_num, blocks_dim, shmem, stream>>>(src0_d, src1_d, opt0_d, dst_d, ne00);
}