Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

CUDA acceleration when using LoRAs #1970

Merged
merged 1 commit into from
Jun 28, 2023
Merged

CUDA acceleration when using LoRAs #1970

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 0 additions & 7 deletions examples/common.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -414,13 +414,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
exit(1);
}

#ifdef GGML_USE_CUBLAS
if (!params.lora_adapter.empty() && params.n_gpu_layers > 0) {
fprintf(stderr, "%s: error: the simultaneous use of LoRAs and GPU acceleration is not supported", __func__);
exit(1);
}
#endif // GGML_USE_CUBLAS

if (escape_prompt) {
process_escapes(params.prompt);
}
Expand Down
53 changes: 42 additions & 11 deletions ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,6 +194,15 @@ static __global__ void add_f32(const float * x, const float * y, float * dst, co
dst[i] = x[i] + y[i];
}

static __global__ void add_f16_f32_f16(const half * x, const float * y, half * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;

if (i >= k) {
return;
}
dst[i] = __hadd(x[i], __float2half(y[i]));
}

static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;

Expand Down Expand Up @@ -1209,6 +1218,11 @@ static void add_f32_cuda(const float * x, const float * y, float * dst, const in
add_f32<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, k);
}

static void add_f16_f32_f16_cuda(const half * x, const float * y, half * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE;
add_f16_f32_f16<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, k);
}

static void mul_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) {
const int num_blocks = (kx + CUDA_MUL_BLOCK_SIZE - 1) / CUDA_MUL_BLOCK_SIZE;
mul_f32<<<num_blocks, CUDA_MUL_BLOCK_SIZE, 0, stream>>>(x, y, dst, kx, ky);
Expand Down Expand Up @@ -1675,15 +1689,21 @@ inline void ggml_cuda_op_add(
float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
cudaStream_t & cudaStream_main){

GGML_ASSERT(src0_ddf_i != nullptr);
GGML_ASSERT(src0_ddq_i != nullptr || src0_ddf_i != nullptr);
GGML_ASSERT(src1_ddf_i != nullptr);
GGML_ASSERT(dst_ddf_i != nullptr);

const int64_t ne0 = src0->ne[0];
const int64_t i01_diff = i01_high - i01_low;

// compute
add_f32_cuda(src0_ddf_i, src1_ddf_i, dst_ddf_i, ne0*i01_diff, cudaStream_main);
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
add_f32_cuda(src0_ddf_i, src1_ddf_i, dst_ddf_i, ne0*i01_diff, cudaStream_main);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
add_f16_f32_f16_cuda((half *) src0_ddq_i, src1_ddf_i, (half *) dst_ddf_i, ne0*i01_diff, cudaStream_main);
} else {
GGML_ASSERT(false);
}
CUDA_CHECK(cudaGetLastError());

(void) src1;
Expand Down Expand Up @@ -2281,8 +2301,14 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
}

void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
ggml_cuda_op(src0, src1, dst, ggml_cuda_op_add, true, true);
// ggml_cuda_add permits f16 dst even though this could in theory cause problems with the pointer arithmetic in ggml_cuda_op.
// Due to flatten_rows == true this does in practice not make a difference however.
// Better solution would be nice but right now that would require disproportionate changes.
GGML_ASSERT(
(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) &&
src1->type == GGML_TYPE_F32 &&
(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16));
ggml_cuda_op(src0, src1, dst, ggml_cuda_op_add, false, true);
}

void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
Expand Down Expand Up @@ -2535,7 +2561,7 @@ void ggml_cuda_free_data(struct ggml_tensor * tensor) {
delete extra;
}

void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch) {
void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bool force_inplace) {
if (scratch && g_scratch_size == 0) {
return;
}
Expand All @@ -2544,22 +2570,23 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch) {
if (tensor->src0 != nullptr && tensor->src0->backend == GGML_BACKEND_CPU) {
const ggml_op src0_op = tensor->src0->op;
if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW) {
ggml_cuda_assign_buffers_impl(tensor->src0, scratch);
ggml_cuda_assign_buffers_impl(tensor->src0, scratch, force_inplace);
}
}
if (tensor->op == GGML_OP_CPY && tensor->src1->backend == GGML_BACKEND_CPU) {
ggml_cuda_assign_buffers_impl(tensor->src1, scratch);
ggml_cuda_assign_buffers_impl(tensor->src1, scratch, force_inplace);
}

tensor->backend = GGML_BACKEND_GPU;
struct ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu;

const bool inplace = (tensor->src0 != nullptr && tensor->src0->data == tensor->data) ||
tensor->op == GGML_OP_VIEW;
tensor->op == GGML_OP_VIEW ||
force_inplace;
const size_t size = ggml_nbytes(tensor);

CUDA_CHECK(cudaSetDevice(g_main_device));
if (inplace && tensor->src0->backend == GGML_BACKEND_GPU) {
if (inplace && (tensor->src0->backend == GGML_BACKEND_GPU || tensor->src0->backend == GGML_BACKEND_GPU_SPLIT)) {
struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src0->extra;
char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
size_t offset = 0;
Expand Down Expand Up @@ -2598,11 +2625,15 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch) {
}

void ggml_cuda_assign_buffers(struct ggml_tensor * tensor) {
ggml_cuda_assign_buffers_impl(tensor, true);
ggml_cuda_assign_buffers_impl(tensor, true, false);
}

void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor) {
ggml_cuda_assign_buffers_impl(tensor, false);
ggml_cuda_assign_buffers_impl(tensor, false, false);
}

void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor) {
ggml_cuda_assign_buffers_impl(tensor, false, true);
}

void ggml_cuda_set_main_device(int main_device) {
Expand Down
1 change: 1 addition & 0 deletions ggml-cuda.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
void ggml_cuda_free_data(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
void ggml_cuda_set_main_device(int main_device);
void ggml_cuda_set_scratch_size(size_t scratch_size);
void ggml_cuda_free_scratch(void);
Expand Down
36 changes: 35 additions & 1 deletion llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2907,14 +2907,15 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
return false;
}
}
ggml_tensor* lora_tensor;
ggml_tensor * lora_tensor;
if (n_dims == 2) {
lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]);
}
else {
fprintf(stderr, "%s: unsupported tensor dimension %d\n", __func__, n_dims);
return 1;
}
ggml_set_name(lora_tensor, "lora_tensor");

// load tensor data
size_t offset = fin.tellg();
Expand All @@ -2930,6 +2931,21 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
lora_tensors.find(base_name + ".loraB") != lora_tensors.end()) {

ggml_tensor * dest_t = model_tensors[base_name];

offload_func_t offload_func = llama_nop;
offload_func_t offload_func_force_inplace = llama_nop;

#ifdef GGML_USE_CUBLAS
if (dest_t->backend == GGML_BACKEND_GPU || dest_t->backend == GGML_BACKEND_GPU_SPLIT) {
if (dest_t->type != GGML_TYPE_F16) {
throw std::runtime_error(format(
"%s: error: the simultaneous use of LoRAs and GPU acceleration is only supported for f16 models", __func__));
}
offload_func = ggml_cuda_assign_buffers;
offload_func_force_inplace = ggml_cuda_assign_buffers_force_inplace;
}
#endif // GGML_USE_CUBLAS

ggml_tensor * base_t;
if (model_loader) {
// load from base model
Expand Down Expand Up @@ -2957,7 +2973,12 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
}

ggml_tensor * loraA = lora_tensors[base_name + ".loraA"];
GGML_ASSERT(loraA->type == GGML_TYPE_F32);
ggml_set_name(loraA, "loraA");

ggml_tensor * loraB = lora_tensors[base_name + ".loraB"];
GGML_ASSERT(loraB->type == GGML_TYPE_F32);
ggml_set_name(loraB, "loraB");

if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
fprintf(stderr, "%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
Expand All @@ -2967,19 +2988,32 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *

// w = w + BA*s
ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
offload_func(BA);
ggml_set_name(BA, "BA");

if (scaling != 1.0f) {
ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling);
ggml_set_name(scale_tensor, "scale_tensor");

BA = ggml_scale_inplace(lora_ctx, BA, scale_tensor);
offload_func(BA);
ggml_set_name(BA, "BA_scaled");
}

ggml_tensor * r;
if (base_t == dest_t) {
r = ggml_add_inplace(lora_ctx, dest_t, BA);
offload_func_force_inplace(r);
ggml_set_name(r, "r_add_inplace");
}
else {
r = ggml_add(lora_ctx, base_t, BA);
offload_func(r);
ggml_set_name(r, "r_add");

r = ggml_cpy(lora_ctx, r, dest_t);
offload_func(r);
ggml_set_name(r, "r_cpy");
}

struct ggml_cgraph gf = ggml_build_forward(r);
Expand Down