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

ggml : change ggml_scale to take a float instead of tensor (#4573)

Browse source 
* ggml : change ggml_scale to take a float instead of tensor

* ggml : fix CPU implementation

* tests : fix test-grad0

ggml-ci
This commit is contained in:
Georgi Gerganov 2023-12-21 23:20:49 +02:00 • committed by GitHub
parent 769a7bc85e
commit afefa319f1
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
12 changed files with 82 additions and 205 deletions
Download patch file Download diff file Expand all files Collapse all files

15
examples/baby-llama/baby-llama.cpp
View file

@ -575,10 +575,7 @@ static struct ggml_tensor * forward(
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// KQ_scaled shape [n_past + N, N, n_head, 1]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
// KQ_masked = mask_past(KQ_scaled)
// KQ_masked shape [n_past + N, N, n_head, 1]
@ -844,10 +841,7 @@ static struct ggml_tensor * forward_batch(
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// KQ_scaled shape [n_past + N, N, n_head, n_batch]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
// KQ_masked = mask_past(KQ_scaled)
@ -1131,10 +1125,7 @@ static struct ggml_tensor * forward_lora(
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// KQ_scaled shape [n_past + N, N, n_head, 1]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
// KQ_masked = mask_past(KQ_scaled)
// KQ_masked shape [n_past + N, N, n_head, 1]

2
examples/export-lora/export-lora.cpp
View file

@ -309,7 +309,7 @@ static struct ggml_cgraph * build_graph_lora(
) {
struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
if (scaling != 1.0f) {
ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
ab = ggml_scale(ctx, ab, scaling);
}
struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);

42
examples/finetune/finetune.cpp
View file

@ -269,7 +269,7 @@ static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_h
float rope_freq_scale = 1.0f;
GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
if (rope_freq_scale != 1.0f) {
hparams->rope_freq_scale = 1.0f / rope_freq_scale;
}
@ -612,6 +612,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
const int n_rot = hparams.n_embd_head();
const int n_embd_head = hparams.n_embd_head();
const int n_embd_gqa = hparams.n_embd_gqa();
const float rms_norm_eps = hparams.f_norm_rms_eps;
const float rope_freq_base = hparams.rope_freq_base;
const float rope_freq_scale = hparams.rope_freq_scale;
@ -680,10 +681,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
checkpoints.push_back(t01);
}
struct ggml_tensor * kv_scale = NULL;
if (!enable_flash_attn) {
kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
}
const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
for (int il = 0; il < n_layer; ++il) {
struct my_llama_layer & layer = model->layers[il];
@ -781,32 +779,32 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
// make sure some tensors are not reallocated by inserting new temporary nodes depending on them
int n_leafs_before = gb->n_leafs;
int n_nodes_before = gb->n_nodes;
struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
// output tensors
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
// input gradient
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
ggml_allocr_alloc(alloc, t36->grad);
// KQ_pos
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
// make sure base model tensors data cannot be used in viewable operations
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, 1.0f));
for (int il = 0; il < n_layer; ++il) {
struct my_llama_layer & layer = model->layers[il];
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
}
// allocating checkpoints in one block to reduce memory fragmentation

8
examples/llava/clip.cpp
View file

@ -330,12 +330,6 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
ggml_repeat(ctx0, model.pre_ln_b, embeddings));
}
struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
ggml_allocr_alloc(ctx->alloc, KQ_scale);
if (!ggml_allocr_is_measure(ctx->alloc)) {
ggml_set_f32(KQ_scale, 1.0f / sqrt((float)d_head));
}
// loop over layers
for (int il = 0; il < n_layer - 1; il++) {
struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
@ -356,7 +350,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
struct ggml_tensor * Q =
ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].q_b, cur), ggml_mul_mat(ctx0, model.layers[il].q_w, cur));
Q = ggml_scale_inplace(ctx0, Q, KQ_scale);
Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);

14
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -369,10 +369,7 @@ static struct ggml_tensor * llama_build_train_graphs(
checkpoints.push_back(t00);
checkpoints.push_back(t01);
struct ggml_tensor * kv_scale = NULL;
if (!enable_flash_attn) {
kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
}
const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
for (int il = 0; il < n_layer; ++il) {
struct my_llama_layer & layer = model->layers[il];
@ -444,14 +441,13 @@ static struct ggml_tensor * llama_build_train_graphs(
// make sure some tensors are not reallocated by inserting new temporary nodes depending on them
int n_leafs_before = gb->n_leafs;
int n_nodes_before = gb->n_nodes;
struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
// output tensors
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
// input gradient
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
// KQ_pos
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
ggml_allocr_alloc(alloc, t36->grad);