root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-13 12:10:18 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

ggml-opt: fix data corruption (ggml/1022)

Browse Source
This commit is contained in:
Johannes Gäßler 2024-11-20 14:56:04 +01:00 committed by Georgi Gerganov
parent 9abe9eeae9
commit 02e4eaf22f
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
5 changed files with 129 additions and 118 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-backend.cpp
View File

@ -252,6 +252,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
} }
void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_ASSERT(tensor);
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
if (size == 0) { if (size == 0) {
@ -266,6 +267,7 @@ void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, siz
} }
void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_ASSERT(tensor);
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
if (size == 0) { if (size == 0) {

3
ggml/src/ggml-impl.h
View File

@ -295,6 +295,9 @@ struct ggml_cgraph {
enum ggml_cgraph_eval_order order; enum ggml_cgraph_eval_order order;
}; };
// returns a slice of cgraph with nodes [i0, i1)
// the slice does not have leafs or gradients
// if you need the gradients, get them from the original graph
struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1); struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
// Memory allocation // Memory allocation

147
ggml/src/ggml-opt.cpp
View File

@ -14,51 +14,51 @@
#include <vector> #include <vector>
struct ggml_opt_dataset { struct ggml_opt_dataset {
struct ggml_context * ctx; struct ggml_context * ctx = nullptr;
ggml_backend_buffer_t buf; ggml_backend_buffer_t buf = nullptr;
struct ggml_tensor * data; struct ggml_tensor * data = nullptr;
struct ggml_tensor * labels; struct ggml_tensor * labels = nullptr;
int64_t ndata; int64_t ndata = -1;
int64_t ndata_shard; int64_t ndata_shard = -1;
size_t nbs_data; size_t nbs_data = -1;
size_t nbs_labels; size_t nbs_labels = -1;
std::vector<int64_t> permutation; std::vector<int64_t> permutation;
}; };
struct ggml_opt_context { struct ggml_opt_context {
ggml_backend_sched_t backend_sched; ggml_backend_sched_t backend_sched = nullptr;
ggml_cgraph * allocated_graph; ggml_cgraph * allocated_graph = nullptr;
ggml_cgraph * allocated_graph_copy; ggml_cgraph * allocated_graph_copy = nullptr;
struct ggml_context * ctx_static; struct ggml_context * ctx_static = nullptr;
struct ggml_context * ctx_static_cpu; struct ggml_context * ctx_static_cpu = nullptr;
struct ggml_context * ctx_compute; struct ggml_context * ctx_compute = nullptr;
struct ggml_context * ctx_copy; struct ggml_context * ctx_copy = nullptr;
ggml_backend_buffer_t buf_static; ggml_backend_buffer_t buf_static = nullptr;
ggml_backend_buffer_t buf_static_cpu; ggml_backend_buffer_t buf_static_cpu = nullptr;
std::mt19937 rng; std::mt19937 rng;
struct ggml_tensor * inputs; struct ggml_tensor * inputs = nullptr;
struct ggml_tensor * outputs; struct ggml_tensor * outputs = nullptr;
struct ggml_tensor * labels; struct ggml_tensor * labels = nullptr;
struct ggml_tensor * loss; struct ggml_tensor * loss = nullptr;
struct ggml_tensor * pred; struct ggml_tensor * pred = nullptr;
struct ggml_tensor * ncorrect; struct ggml_tensor * ncorrect = nullptr;
struct ggml_cgraph * gf; struct ggml_cgraph * gf = nullptr;
struct ggml_cgraph * gb_grad; struct ggml_cgraph * gb_grad = nullptr;
struct ggml_cgraph * gb_opt; struct ggml_cgraph * gb_opt = nullptr;
int64_t iter; int64_t iter = 1;
int32_t opt_period; int32_t opt_period = 1;
int32_t opt_i; int32_t opt_i = 0;
bool loss_per_datapoint; bool loss_per_datapoint = false;
ggml_opt_get_optimizer_params get_opt_pars; ggml_opt_get_optimizer_params get_opt_pars = nullptr;
void * get_opt_pars_ud; void * get_opt_pars_ud = nullptr;
struct ggml_tensor * adamw_params; struct ggml_tensor * adamw_params = nullptr;
}; };
struct ggml_opt_result { struct ggml_opt_result {
@ -67,8 +67,8 @@ struct ggml_opt_result {
std::vector<int32_t> pred; std::vector<int32_t> pred;
int64_t ncorrect = 0; int64_t ncorrect = 0;
bool loss_per_datapoint = false; int64_t opt_period = -1;
int64_t opt_period = -1; bool loss_per_datapoint = false;
}; };
// ====== Dataset ====== // ====== Dataset ======
@ -188,11 +188,11 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
} }
struct ggml_opt_params ggml_opt_default_params( struct ggml_opt_params ggml_opt_default_params(
ggml_backend_sched_t backend_sched, ggml_backend_sched_t backend_sched,
struct ggml_context * ctx_compute, struct ggml_context * ctx_compute,
struct ggml_tensor * inputs, struct ggml_tensor * inputs,
struct ggml_tensor * outputs, struct ggml_tensor * outputs,
enum ggml_opt_loss_type loss_type) { enum ggml_opt_loss_type loss_type) {
return { return {
/*backend_sched =*/ backend_sched, /*backend_sched =*/ backend_sched,
/*ctx_compute =*/ ctx_compute, /*ctx_compute =*/ ctx_compute,
@ -237,25 +237,33 @@ static ggml_tensor * map_tensor(std::map<ggml_tensor *, ggml_tensor *> & tensor_
return new_tensor; return new_tensor;
} }
static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * graph) { static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) {
std::map<ggml_tensor *, ggml_tensor *> tensor_map; std::map<ggml_tensor *, ggml_tensor *> tensor_map;
ggml_cgraph * new_graph = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); ggml_cgraph * dst = ggml_new_graph_custom(ctx, src->size, /*grads =*/ true);
for (int i = 0; i < graph->n_leafs; i++) { for (int i = 0; i < src->n_leafs; i++) {
ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->leafs[i])); ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->leafs[i]));
} }
for (int i = 0; i < graph->n_nodes; i++) { GGML_ASSERT(dst->n_leafs == src->n_leafs);
ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->nodes[i])); for (int i = 0; i < src->n_nodes; i++) {
ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->nodes[i]));
} }
for (int i = 0; i < graph->n_nodes; ++i) { GGML_ASSERT(dst->n_nodes == src->n_nodes);
const size_t igrad_src = ggml_hash_find(&graph->visited_hash_set, graph->nodes[i]); for (int i = 0; i < src->n_nodes; ++i) {
const size_t igrad_dst = ggml_hash_find(&new_graph->visited_hash_set, new_graph->nodes[i]); const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
graph->grads[igrad_dst] = new_graph->grads[igrad_src]; const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
graph->grad_accs[igrad_dst] = new_graph->grad_accs[igrad_src];
GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
dst->grads[igrad_dst] = src->grads[igrad_src];
dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
} }
return new_graph; return dst;
} }
static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) { static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
@ -284,18 +292,13 @@ static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph
ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
ggml_opt_context_t result = new struct ggml_opt_context; ggml_opt_context_t result = new struct ggml_opt_context;
result->backend_sched = params.backend_sched; result->backend_sched = params.backend_sched;
result->allocated_graph = nullptr; result->ctx_compute = params.ctx_compute;
result->allocated_graph_copy = nullptr; result->inputs = params.inputs;
result->ctx_compute = params.ctx_compute; result->outputs = params.outputs;
result->ctx_copy = nullptr; result->opt_period = params.opt_period;
result->inputs = params.inputs; result->get_opt_pars = params.get_opt_pars;
result->outputs = params.outputs; result->get_opt_pars_ud = params.get_opt_pars_ud;
result->iter = 1;
result->opt_period = params.opt_period;
result->opt_i = 0;
result->get_opt_pars = params.get_opt_pars;
result->get_opt_pars_ud = params.get_opt_pars_ud;
GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically"); GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
GGML_ASSERT(result->opt_period >= 1); GGML_ASSERT(result->opt_period >= 1);
@ -348,7 +351,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
switch (params.loss_type) { switch (params.loss_type) {
case GGML_OPT_LOSS_TYPE_MEAN: { case GGML_OPT_LOSS_TYPE_MEAN: {
result->labels = nullptr;
result->loss = ggml_sum(result->ctx_static, result->outputs); result->loss = ggml_sum(result->ctx_static, result->outputs);
ggml_set_name(result->loss, "loss_sum"); ggml_set_name(result->loss, "loss_sum");
const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs)); const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
@ -358,7 +360,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
break; break;
} }
case GGML_OPT_LOSS_TYPE_SUM: { case GGML_OPT_LOSS_TYPE_SUM: {
result->labels = nullptr;
result->loss = ggml_sum(result->ctx_static, result->outputs); result->loss = ggml_sum(result->ctx_static, result->outputs);
ggml_set_name(result->loss, "loss_sum"); ggml_set_name(result->loss, "loss_sum");
result->loss_per_datapoint = false; result->loss_per_datapoint = false;
@ -413,14 +414,7 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
} }
if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) { if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
result->gb_grad = nullptr;
result->gb_opt = nullptr;
result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0)); result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
result->buf_static_cpu = nullptr;
ggml_opt_alloc_graph(result, result->gf);
return result; return result;
} }
@ -429,14 +423,8 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate); ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) { if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
result->gb_opt = nullptr;
result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0)); result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
result->buf_static_cpu = nullptr;
ggml_opt_alloc_graph(result, result->gb_grad);
ggml_graph_reset(result->gb_grad); ggml_graph_reset(result->gb_grad);
return result; return result;
} }
@ -466,7 +454,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type()); result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
ggml_opt_alloc_graph(result, result->gb_opt);
ggml_graph_reset(result->gb_opt); ggml_graph_reset(result->gb_opt);
return result; return result;

94
ggml/src/ggml.c
View File

@ -5019,8 +5019,10 @@ static void ggml_hash_map_free(struct hash_map * map) {
} }
// utility functions to change gradients // utility functions to change gradients
// if a is in acc_table, modify gradients in-place and mark result as gradient accumulator // isrc is the index of tensor in cgraph->visited_has_set.keys
// else if a is in zero_table, replace a // the corresponding gradient (accumulators) are also at position isrc
// if tensor has a gradient accumulator, modify that accumulator in-place
// else if there is no gradient for tensor, set the corresponding value
// else, just add/subtract/etc. the gradients // else, just add/subtract/etc. the gradients
static void ggml_add_or_set( static void ggml_add_or_set(
@ -5028,11 +5030,14 @@ static void ggml_add_or_set(
struct ggml_cgraph * cgraph, struct ggml_cgraph * cgraph,
size_t isrc, size_t isrc,
struct ggml_tensor * tensor) { struct ggml_tensor * tensor) {
struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
GGML_ASSERT(src);
if (cgraph->grads[isrc]) { if (cgraph->grads[isrc]) {
cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]); cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, /*inplace =*/ cgraph->grad_accs[isrc]);
} else { } else {
cgraph->grads[isrc] = tensor; cgraph->grads[isrc] = tensor;
} }
ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
ggml_build_forward_expand(cgraph, cgraph->grads[isrc]); ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
} }
@ -5040,18 +5045,20 @@ static void ggml_acc_or_set(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_cgraph * cgraph, struct ggml_cgraph * cgraph,
size_t isrc, size_t isrc,
struct ggml_tensor * src,
struct ggml_tensor * tensor, struct ggml_tensor * tensor,
const size_t nb1, const size_t nb1,
const size_t nb2, const size_t nb2,
const size_t nb3, const size_t nb3,
const size_t offset) { const size_t offset) {
struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
GGML_ASSERT(src);
if (cgraph->grads[isrc]) { if (cgraph->grads[isrc]) {
cgraph->grads[isrc] = ggml_acc_impl(ctx, cgraph->grads[isrc], tensor, nb1, nb2, nb3, offset, cgraph->grad_accs[isrc]); cgraph->grads[isrc] = ggml_acc_impl(ctx, cgraph->grads[isrc], tensor, nb1, nb2, nb3, offset, cgraph->grad_accs[isrc]);
} else { } else {
struct ggml_tensor * a_zero = ggml_scale(ctx, src, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN struct ggml_tensor * a_zero = ggml_scale(ctx, src, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN
cgraph->grads[isrc] = ggml_acc_impl(ctx, a_zero, tensor, nb1, nb2, nb3, offset, false); cgraph->grads[isrc] = ggml_acc_impl(ctx, a_zero, tensor, nb1, nb2, nb3, offset, false);
} }
ggml_format_name(cgraph->grads[isrc], "grad for %s", cgraph->visited_hash_set.keys[isrc]->name);
ggml_build_forward_expand(cgraph, cgraph->grads[isrc]); ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
} }
@ -5059,13 +5066,15 @@ static void ggml_add1_or_set(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_cgraph * cgraph, struct ggml_cgraph * cgraph,
size_t isrc, size_t isrc,
struct ggml_tensor * src,
struct ggml_tensor * tensor) { struct ggml_tensor * tensor) {
struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
GGML_ASSERT(src);
if (cgraph->grads[isrc]) { if (cgraph->grads[isrc]) {
cgraph->grads[isrc] = ggml_add1_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]); cgraph->grads[isrc] = ggml_add1_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
} else { } else {
cgraph->grads[isrc] = ggml_repeat(ctx, tensor, src); cgraph->grads[isrc] = ggml_repeat(ctx, tensor, src);
} }
ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
ggml_build_forward_expand(cgraph, cgraph->grads[isrc]); ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
} }
@ -5074,11 +5083,14 @@ static void ggml_sub_or_set(
struct ggml_cgraph * cgraph, struct ggml_cgraph * cgraph,
size_t isrc, size_t isrc,
struct ggml_tensor * tensor) { struct ggml_tensor * tensor) {
struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
GGML_ASSERT(src);
if (cgraph->grads[isrc]) { if (cgraph->grads[isrc]) {
cgraph->grads[isrc] = ggml_sub_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]); cgraph->grads[isrc] = ggml_sub_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
} else { } else {
cgraph->grads[isrc] = ggml_neg(ctx, tensor); cgraph->grads[isrc] = ggml_neg(ctx, tensor);
} }
ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
ggml_build_forward_expand(cgraph, cgraph->grads[isrc]); ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
} }
@ -5095,12 +5107,12 @@ static void ggml_compute_backward(
struct ggml_tensor * src1 = tensor->src[1]; struct ggml_tensor * src1 = tensor->src[1];
struct ggml_tensor * src2 = tensor->src[2]; struct ggml_tensor * src2 = tensor->src[2];
struct ggml_hash_set * hash_set = &cgraph->visited_hash_set; struct ggml_hash_set * hash_set = &cgraph->visited_hash_set;
const size_t isrc0 = ggml_hash_find(hash_set, src0); const size_t isrc0 = src0 ? ggml_hash_find(hash_set, src0) : (size_t) -1;
const size_t isrc1 = ggml_hash_find(hash_set, src1); const size_t isrc1 = src1 ? ggml_hash_find(hash_set, src1) : (size_t) -1;
const size_t isrc2 = ggml_hash_find(hash_set, src2); const size_t isrc2 = src2 ? ggml_hash_find(hash_set, src2) : (size_t) -1;
const bool src0_needs_grads = isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0]; const bool src0_needs_grads = src0 && isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0];
const bool src1_needs_grads = isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1]; const bool src1_needs_grads = src1 && isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1];
const bool src2_needs_grads = isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2]; const bool src2_needs_grads = src2 && isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2];
switch (tensor->op) { switch (tensor->op) {
case GGML_OP_DUP: { case GGML_OP_DUP: {
@ -5200,7 +5212,7 @@ static void ggml_compute_backward(
} break; } break;
case GGML_OP_SUM: { case GGML_OP_SUM: {
if (src0_needs_grads) { if (src0_needs_grads) {
ggml_add1_or_set(ctx, cgraph, isrc0, src0, grad); ggml_add1_or_set(ctx, cgraph, isrc0, grad);
} }
} break; } break;
case GGML_OP_SUM_ROWS: { case GGML_OP_SUM_ROWS: {
@ -5210,7 +5222,7 @@ static void ggml_compute_backward(
} break; } break;
case GGML_OP_MEAN: { case GGML_OP_MEAN: {
if (src0_needs_grads) { if (src0_needs_grads) {
ggml_add1_or_set(ctx, cgraph, isrc0, src0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false)); ggml_add1_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false));
} }
} break; } break;
case GGML_OP_REPEAT: { case GGML_OP_REPEAT: {
@ -5363,7 +5375,7 @@ static void ggml_compute_backward(
nb3 = (nb3 / n0) * ng; nb3 = (nb3 / n0) * ng;
} }
ggml_acc_or_set(ctx, cgraph, isrc0, src0, grad, nb1, nb2, nb3, offset); ggml_acc_or_set(ctx, cgraph, isrc0, grad, nb1, nb2, nb3, offset);
} }
} break; } break;
case GGML_OP_PERMUTE: { case GGML_OP_PERMUTE: {
@ -5597,10 +5609,9 @@ void ggml_build_backward_expand(
const int n_nodes_f = cgraph->n_nodes; const int n_nodes_f = cgraph->n_nodes;
const size_t hash_size = ggml_hash_size(2*cgraph->size); memset(cgraph->grads, 0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *));
memset(cgraph->grads, 0, hash_size*sizeof(struct ggml_tensor *)); memset(cgraph->grad_accs, 0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *));
memset(cgraph->grad_accs, 0, hash_size*sizeof(struct ggml_tensor *)); bool * grads_needed = calloc(cgraph->visited_hash_set.size, sizeof(bool));
bool * grads_needed = calloc(hash_size, sizeof(bool));
{ {
bool any_params = false; bool any_params = false;
@ -5621,7 +5632,7 @@ void ggml_build_backward_expand(
continue; continue;
} }
bool node_needs_grad = node->flags & GGML_TENSOR_FLAG_PARAM; bool node_needs_grad = (node->flags & GGML_TENSOR_FLAG_PARAM) || (node->flags & GGML_TENSOR_FLAG_LOSS);
bool ignore_src[GGML_MAX_SRC] = {false}; bool ignore_src[GGML_MAX_SRC] = {false};
switch (node->op) { switch (node->op) {
// gradients in node->src[0] for one reason or another have no effect on output gradients // gradients in node->src[0] for one reason or another have no effect on output gradients
@ -5638,7 +5649,7 @@ void ggml_build_backward_expand(
} break; } break;
// gradients in node->src[1] for one reason or another have no effect on output gradients // gradients in node->src[1] for one reason or another have no effect on output gradients
case GGML_OP_CPY: // gradients in CPY target are irrelevant case GGML_OP_CPY: // gradients in CPY target are irrelevant
case GGML_OP_GET_ROWS: // row indices not differentiable case GGML_OP_GET_ROWS: // row indices not differentiable
case GGML_OP_GET_ROWS_BACK: // same as for GET_ROWS case GGML_OP_GET_ROWS_BACK: // same as for GET_ROWS
case GGML_OP_ROPE: // positions not differentiable case GGML_OP_ROPE: // positions not differentiable
@ -5665,9 +5676,12 @@ void ggml_build_backward_expand(
node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE); node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node); const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
GGML_ASSERT(igrad != GGML_HASHSET_FULL);
GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad));
if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) { if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
cgraph->grads[igrad] = ggml_dup_tensor(ctx_static, node); cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node);
cgraph->grad_accs[igrad] = cgraph->grads[igrad]; cgraph->grads[igrad] = cgraph->grad_accs[igrad];
ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name);
} }
grads_needed[igrad] = true; grads_needed[igrad] = true;
} }
@ -5761,15 +5775,15 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1) { struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1) {
struct ggml_cgraph cgraph = { struct ggml_cgraph cgraph = {
/*.size =*/ 0, /*.size =*/ 0,
/*.n_nodes =*/ i1 - i0, /*.n_nodes =*/ i1 - i0,
/*.n_leafs =*/ 0, /*.n_leafs =*/ 0,
/*.nodes =*/ cgraph0->nodes + i0, /*.nodes =*/ cgraph0->nodes + i0,
/*.grads =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL, /*.grads =*/ NULL, // gradients would need visited_hash_set
/*.grad_accs =*/ cgraph0->grad_accs ? cgraph0->grad_accs + i0 : NULL, /*.grad_accs =*/ NULL,
/*.leafs =*/ NULL, /*.leafs =*/ NULL,
/*.hash_table =*/ { 0, NULL, NULL }, /*.visited_hash_set =*/ { 0, NULL, NULL },
/*.order =*/ cgraph0->order, /*.order =*/ cgraph0->order,
}; };
return cgraph; return cgraph;
@ -5799,12 +5813,22 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
} }
} }
if (dst->grads) {
memset(dst->grads, 0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *));
memset(dst->grad_accs, 0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *));
}
if (src->grads) { if (src->grads) {
GGML_ASSERT(dst->grads != NULL); GGML_ASSERT(dst->grads != NULL);
GGML_ASSERT(dst->grad_accs != NULL); GGML_ASSERT(dst->grad_accs != NULL);
for (int i = 0; i < src->n_nodes; ++i) { for (int i = 0; i < src->n_nodes; ++i) {
const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]); const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]); const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
dst->grads[igrad_dst] = src->grads[igrad_src]; dst->grads[igrad_dst] = src->grads[igrad_src];
dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src]; dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
} }
@ -5839,12 +5863,8 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
if (node->op == GGML_OP_OPT_STEP_ADAMW) { if (node->op == GGML_OP_OPT_STEP_ADAMW) {
// clear momenta // clear momenta
if (node->src[2]->data) { ggml_set_zero(node->src[2]);
ggml_set_zero(node->src[2]); ggml_set_zero(node->src[3]);
}
if (node->src[3]->data) {
ggml_set_zero(node->src[3]);
}
} }
// initial gradients of loss should be 1, 0 otherwise // initial gradients of loss should be 1, 0 otherwise

1
tests/test-backend-ops.cpp
View File

@ -819,7 +819,6 @@ struct test_case {
} }
} }
// TODO: refactor so that this check is only needed once
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (!ggml_backend_supports_op(backend, t)) { if (!ggml_backend_supports_op(backend, t)) {
printf("not supported [%s] ", ggml_backend_name(backend)); printf("not supported [%s] ", ggml_backend_name(backend));