Skip to content

whisper : reduce ggml_context usage #2525

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

Merged
merged 3 commits into from
Oct 31, 2024
Merged

whisper : reduce ggml_context usage #2525

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
3689d49
whisper : reduce ggml_context usage
ggerganov Oct 30, 2024
987f314
ggml : allocate contexts on the heap (v2)
ggerganov Oct 31, 2024
552419f
ggml : aligned malloc -> malloc
ggerganov Oct 31, 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
2 changes: 1 addition & 1 deletion ggml/src/ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3853,7 +3853,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
}

if (ctx == NULL) {
GGML_PRINT_DEBUG("%s: no unused context found\n", __func__);
GGML_LOG_ERROR("%s: ran out of contexts (max = %d)\n", __func__, GGML_MAX_CONTEXTS);

ggml_critical_section_end();

Expand Down
28 changes: 15 additions & 13 deletions src/whisper.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -699,9 +699,9 @@ struct whisper_kv_cache {
struct ggml_tensor * k;
struct ggml_tensor * v;

struct ggml_context * ctx = nullptr;

ggml_backend_buffer_t buffer = nullptr;

std::vector<uint8_t> ctx_buf;
};

struct whisper_model {
Expand Down Expand Up @@ -941,9 +941,11 @@ static bool whisper_kv_cache_init(
const int64_t n_mem = n_text_layer*n_ctx;
const int64_t n_elements = n_text_state*n_mem;

cache.ctx_buf.resize(2*ggml_tensor_overhead());

struct ggml_init_params params = {
/*.mem_size =*/ 2*ggml_tensor_overhead(),
/*.mem_buffer =*/ nullptr,
/*.mem_size =*/ cache.ctx_buf.size(),
/*.mem_buffer =*/ cache.ctx_buf.data(),
/*.no_alloc =*/ true,
};

Expand All @@ -953,31 +955,31 @@ static bool whisper_kv_cache_init(
cache.cells.clear();
cache.cells.resize(n_ctx);

cache.ctx = ggml_init(params);
struct ggml_context * ctx = ggml_init(params);

if (!cache.ctx) {
if (!ctx) {
WHISPER_LOG_ERROR("%s: failed to allocate memory for the kv cache context\n", __func__);
return false;
}

cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
cache.k = ggml_new_tensor_1d(ctx, wtype, n_elements);
cache.v = ggml_new_tensor_1d(ctx, wtype, n_elements);

cache.buffer = ggml_backend_alloc_ctx_tensors(cache.ctx, backend);
cache.buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
if (!cache.buffer) {
WHISPER_LOG_ERROR("%s: failed to allocate memory for the kv cache\n", __func__);
return false;
}

ggml_backend_buffer_clear(cache.buffer, 0);

ggml_free(ctx);

return true;
}

static void whisper_kv_cache_free(struct whisper_kv_cache & cache) {
ggml_free(cache.ctx);
ggml_backend_buffer_free(cache.buffer);
cache.ctx = nullptr;
}

static bool whisper_kv_cache_find_slot(
Expand Down Expand Up @@ -2002,7 +2004,7 @@ static struct ggml_cgraph * whisper_build_graph_encoder(

auto & kv_pad = wstate.kv_pad;

WHISPER_ASSERT(!!kv_pad.ctx);
WHISPER_ASSERT(!!kv_pad.buffer);

const int n_ctx_pad = GGML_PAD(n_ctx, 256);

Expand Down Expand Up @@ -2416,7 +2418,7 @@ static struct ggml_cgraph * whisper_build_graph_decoder(

auto & kv_self = wstate.kv_self;

WHISPER_ASSERT(!!kv_self.ctx);
WHISPER_ASSERT(!!kv_self.buffer);

const int n_ctx = kv_self.size;
const int n_state = hparams.n_text_state;
Expand Down
Loading