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

Ag ggml graph caching #229

Merged
merged 3 commits into from
Jul 11, 2024
Merged

Ag ggml graph caching #229

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
bca068f
ggml: avoid rebuild of GGML graph for each token (#7456)
agray3 Jun 4, 2024
b7956a8
fix seg fault
agray3 Jul 8, 2024
a34900a
restrict to nsplit=2
agray3 Jul 10, 2024
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
5 changes: 5 additions & 0 deletions ggml/include/ggml-backend.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,6 +230,11 @@ extern "C" {
GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);

// Utility to query whether cached GGML graph is in use
GGML_API bool ggml_use_cached_graph(ggml_backend_sched_t sched);

// Set whether or not to use GGML graph caching
GGML_API void ggml_set_cached_graph(ggml_backend_sched_t sched, bool set_value);

#ifdef __cplusplus
}
Expand Down
33 changes: 32 additions & 1 deletion ggml/src/ggml-backend.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1036,6 +1036,13 @@ struct ggml_backend_sched_split {
struct ggml_cgraph graph;
};

// Object to facilitate GML graph caching
struct ggml_cached_graph {
bool is_active;
ggml_backend_t input_backend;
struct ggml_tensor * input_cpy[GGML_SCHED_MAX_SPLIT_INPUTS];
};

struct ggml_backend_sched {
bool is_reset; // true if the scheduler has been reset since the last graph split
bool is_alloc;
Expand Down Expand Up @@ -1087,6 +1094,8 @@ struct ggml_backend_sched {
__attribute__((aligned(GGML_MEM_ALIGN)))
#endif
char context_buffer[GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];

struct ggml_cached_graph cached_graph;
};

#define hash_id(tensor) ggml_hash_find_or_insert(sched->hash_set, tensor)
Expand Down Expand Up @@ -1753,6 +1762,14 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
struct ggml_tensor * input = split->inputs[j];
struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id][sched->cur_copy];

if (!sched->cached_graph.is_active) {
sched->cached_graph.input_backend = input_backend;
sched->cached_graph.input_cpy[j] = input_cpy;
}
else {
input_backend = sched->cached_graph.input_backend;
input_cpy = sched->cached_graph.input_cpy[j];
}
if (input->flags & GGML_TENSOR_FLAG_INPUT) {
// inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
Expand Down Expand Up @@ -1872,6 +1889,8 @@ ggml_backend_sched_t ggml_backend_sched_new(

ggml_backend_sched_reset(sched);

sched->cached_graph.is_active = false;

return sched;
}

Expand Down Expand Up @@ -1947,6 +1966,9 @@ enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, st
}

enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {

if(!sched->cached_graph.is_active)
{
if (!sched->is_reset && !sched->is_alloc) {
ggml_backend_sched_reset(sched);
}
Expand All @@ -1956,7 +1978,7 @@ enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sch
return GGML_STATUS_ALLOC_FAILED;
}
}

}
return ggml_backend_sched_compute_splits(sched);
}

Expand Down Expand Up @@ -2223,3 +2245,12 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t

return true;
}

bool ggml_use_cached_graph(ggml_backend_sched_t sched) {
return sched->cached_graph.is_active;
}

void ggml_set_cached_graph(ggml_backend_sched_t sched, bool set_value) {
sched->cached_graph.is_active = set_value;
}

130 changes: 123 additions & 7 deletions src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2712,6 +2712,17 @@ struct llama_model {
}
};

// Object used to allow caching of GGML graph between tokens where possible.
struct ggml_cached_graph {
bool is_active = false;
ggml_cgraph * gf;
size_t n;
ggml_backend_t backend_res;
ggml_backend_t backend_embd;
struct ggml_tensor * res;
struct ggml_tensor * embd;
};

struct llama_context {
llama_context(const llama_model & model) : model(model), t_start_us(model.t_start_us), t_load_us(model.t_load_us) {}
~llama_context() {
Expand Down Expand Up @@ -2813,6 +2824,8 @@ struct llama_context {

// control vectors
struct llama_control_vector cvec;

struct ggml_cached_graph cached_graph;
};

static size_t llama_get_device_count(const llama_model & model) {
Expand Down Expand Up @@ -14524,12 +14537,44 @@ static int llama_decode_internal(
ggml_backend_sched_reset(lctx.sched);
ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data);

ggml_cgraph * gf = llama_build_graph(lctx, u_batch, false);

ggml_cgraph * gf;
// the output is always the last tensor in the graph
struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 2];
struct ggml_tensor * res;
struct ggml_tensor * embd;

bool n_has_changed_since_last_token = false;
if(lctx.cached_graph.n != kv_self.n) n_has_changed_since_last_token = true;
lctx.cached_graph.n = kv_self.n;

// Re-build graph only if graph caching is not possible
if(!ggml_use_cached_graph(lctx.sched) || n_has_changed_since_last_token) {

gf = llama_build_graph(lctx, u_batch, false);

// Set whether GGML graph caching is in use within GGML module, based on
// whether caching was activated here during the previous token
ggml_set_cached_graph(lctx.sched,lctx.cached_graph.is_active);

// Disable future graph caching in presence of env var,
// if there are multiple devices, if batch size is greater than 1,
// or if nsplits is not 2.
// TO DO enable graph caching for these cases
bool disable_cached_ggml_graph = (getenv("GGML_DISABLE_GRAPH_CACHING") != nullptr)
|| (llama_get_device_count(model) > 1)
|| (ggml_backend_sched_get_n_splits(lctx.sched) != 2);
for (int i = 0 ; i < gf->n_nodes; i++) {
if (gf->nodes[i]->op == GGML_OP_ADD && gf->nodes[i]->src[1] && gf->nodes[i]->src[1]->ne[1] > 1) {
disable_cached_ggml_graph = true;
break;
}
}

// Set whether graph caching should be used for future tokens
lctx.cached_graph.is_active=!disable_cached_ggml_graph;

// the output is always the last tensor in the graph
res = gf->nodes[gf->n_nodes - 1];
embd = gf->nodes[gf->n_nodes - 2];
if (lctx.n_outputs == 0) {
// no output
res = nullptr;
Expand All @@ -14545,10 +14590,71 @@ static int llama_decode_internal(
embd = nullptr; // do not extract embeddings when not needed
GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
}
lctx.cached_graph.res = res;
lctx.cached_graph.embd = embd;
// LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);

ggml_backend_sched_alloc_graph(lctx.sched, gf);

}
else {
gf = lctx.cached_graph.gf;
res = lctx.cached_graph.res;
embd = lctx.cached_graph.embd;
}
lctx.cached_graph.gf = gf;

if(ggml_use_cached_graph(lctx.sched)) {

// If using flash attention, find mask node so it can be skipped when updating
// KV cache paramaters in cached graph nodes below
void * flash_attn_mask_node = nullptr;
if(cparams.flash_attn) {
for (int i = 0; i < gf->n_nodes; i++) {
ggml_tensor * node = gf->nodes[i];
if (node->op == GGML_OP_FLASH_ATTN_EXT) {
flash_attn_mask_node = node->src[3];
break;
}
}
}

// Temporarily store KV cache parameters that will need updated in cached graph.
const struct llama_hparams & hparams = model.hparams;
const int64_t n_layer = hparams.n_layer;
const int64_t kv_head = kv_self.head;
std::vector<void *> kv_cache_ptrs;
for (int il = 0; il < n_layer; ++il) {
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
ggml_tensor * tmp_tensor = kv_self.k_l[il];
size_t tmp_offset = (ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa))*kv_head;
kv_cache_ptrs.push_back(static_cast<char*>(tmp_tensor->data) + tmp_offset);
tmp_tensor = kv_self.v_l[il];
if (cparams.flash_attn) {
tmp_offset = (kv_head)*ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa);
} else {
tmp_offset = (kv_head)*ggml_element_size(kv_self.v_l[il]);
}
kv_cache_ptrs.push_back(static_cast<char*>(tmp_tensor->data) + tmp_offset);
}

// Update KV cache parameters in cached graph.
int copy_op_count = 0;
if(gf != nullptr && gf->nodes != nullptr){
for (int i = 0; i < gf->n_nodes; i++) {
ggml_tensor * node = gf->nodes[i];
if (node->op == GGML_OP_CPY) {
if (node != flash_attn_mask_node) {
node->src[1]->data = kv_cache_ptrs[copy_op_count];
copy_op_count++;
}
}
}
}

}

llama_set_inputs(lctx, u_batch);

llama_graph_compute(lctx, gf, n_threads);
Expand All @@ -14571,11 +14677,15 @@ static int llama_decode_internal(
// extract logits
if (res) {
ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(lctx.sched, res);
GGML_ASSERT(backend_res != nullptr);
GGML_ASSERT(lctx.logits != nullptr);

float * logits_out = lctx.logits + n_outputs_prev*n_vocab;
const int32_t n_outputs_new = lctx.n_outputs;
if(!ggml_use_cached_graph(lctx.sched))
lctx.cached_graph.backend_res = backend_res;
else
backend_res = lctx.cached_graph.backend_res;

GGML_ASSERT(backend_res != nullptr);
GGML_ASSERT(lctx.logits != nullptr);

if (n_outputs_new) {
GGML_ASSERT( n_outputs_prev + n_outputs_new <= n_outputs);
Expand All @@ -14587,6 +14697,12 @@ static int llama_decode_internal(
// extract embeddings
if (embd) {
ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd);


if(!ggml_use_cached_graph(lctx.sched))
lctx.cached_graph.backend_embd = backend_embd;
else
backend_embd = lctx.cached_graph.backend_embd;
GGML_ASSERT(backend_embd != nullptr);

switch (cparams.pooling_type) {
Expand Down
Loading