CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : make starcoder graph build more consistent with others

Browse Source
This commit is contained in:
Georgi Gerganov
2023-09-15 17:57:10 +03:00
parent f82328ab65
commit 92a4f86879
1 changed files with 61 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

98
llama.cpp
View File

@@ -3446,7 +3446,9 @@ static struct ggml_cgraph * llm_build_starcoder(
const int64_t n_layer = hparams.n_layer; const int64_t n_layer = hparams.n_layer;
const int64_t n_ctx = hparams.n_ctx; const int64_t n_ctx = hparams.n_ctx;
const int64_t n_head = hparams.n_head; const int64_t n_head = hparams.n_head;
const int64_t n_head_kv = hparams.n_head_kv;
const int64_t n_embd_head = hparams.n_embd_head(); const int64_t n_embd_head = hparams.n_embd_head();
const int64_t n_embd_gqa = hparams.n_embd_gqa();
GGML_ASSERT(n_embd_head == hparams.n_rot); GGML_ASSERT(n_embd_head == hparams.n_rot);
@@ -3508,28 +3510,44 @@ static struct ggml_cgraph * llm_build_starcoder(
position = ggml_get_rows(ctx0, model.pos_embeddings, inp_positions); position = ggml_get_rows(ctx0, model.pos_embeddings, inp_positions);
} }
struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
ggml_allocr_alloc(lctx.alloc, KQ_scale);
if (!ggml_allocr_is_measure(lctx.alloc)) {
ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
}
ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
inpL = ggml_add(ctx0, token, position); inpL = ggml_add(ctx0, token, position);
ggml_set_name(inpL, "inpL");
for (int il = 0; il < n_layer; ++il) { for (int il = 0; il < n_layer; ++il) {
{ {
// Norm // Norm
cur = ggml_norm(ctx0, inpL, norm_eps); cur = ggml_norm(ctx0, inpL, norm_eps);
cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b); cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b);
} }
{ {
// Self Attention // Self Attention
cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv); cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv);
struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd); struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd); struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd); struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
// store key and value to memory struct ggml_tensor * Qcur = tmpq;
if (N >= 1) { struct ggml_tensor * Kcur = tmpk;
struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
struct ggml_tensor * v = ggml_view_1d(ctx0, kv_self.v, N*n_embd, (ggml_element_size(kv_self.v)*n_embd)*(il*n_ctx + n_past)); {
struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, N));
ggml_set_name(Vcur, "Vcur");
struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + n_past));
ggml_set_name(k, "k");
struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd_gqa,
( n_ctx)*ggml_element_size(kv_self.v),
(il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + n_past*ggml_element_size(kv_self.v));
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
@@ -3541,56 +3559,62 @@ static struct ggml_cgraph * llm_build_starcoder(
Qcur, Qcur,
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)), ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
0, 2, 1, 3); 0, 2, 1, 3);
ggml_set_name(Q, "Q");
struct ggml_tensor * K = struct ggml_tensor * K =
ggml_permute(ctx0, ggml_view_3d(ctx0, kv_self.k,
ggml_reshape_3d(ctx0, n_embd_head, n_past + N, n_head_kv,
ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd), ggml_element_size(kv_self.k)*n_embd_gqa,
n_embd/n_head, n_head, n_past + N), ggml_element_size(kv_self.k)*n_embd_head,
0, 2, 1, 3); //TODO: need to be tiled ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
ggml_set_name(K, "K");
// K * Q // K * Q
// [n_past + N, N, 12]
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
ggml_set_name(KQ, "KQ");
// KQ_scaled = KQ / sqrt(n_embd/n_head) // KQ_scaled = KQ / sqrt(n_embd_head)
// [n_past + N, N, 12] // KQ_scaled shape [n_past + N, N, n_head, 1]
struct ggml_tensor * KQ_scaled = struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
ggml_scale_inplace(ctx0, ggml_set_name(KQ_scaled, "KQ_scaled");
KQ,
ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
);
// KQ_masked = mask_past(KQ_scaled) // KQ_masked = mask_past(KQ_scaled)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
ggml_set_name(KQ_masked, "KQ_masked");
// KQ = soft_max(KQ_masked) // KQ = soft_max(KQ_masked)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
ggml_set_name(KQ_soft_max, "KQ_soft_max");
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous() // split cached V into n_head heads
// [n_past + N, 64, 12] struct ggml_tensor * V =
struct ggml_tensor * V_trans = ggml_view_3d(ctx0, kv_self.v,
ggml_cpy(ctx0, n_past + N, n_embd_head, n_head_kv,
ggml_permute(ctx0, ggml_element_size(kv_self.v)*n_ctx,
ggml_reshape_3d(ctx0, ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.v)*n_embd), ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
n_embd/n_head, n_head, n_past + N), ggml_set_name(V, "V");
1, 2, 0, 3),
ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head));
// KQV = transpose(V) * KQ_soft_max #if 1
// [64, N, 12] struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max); ggml_set_name(KQV, "KQV");
#else
// make V contiguous in memory to speed up the matmul, however we waste time on the copy
// on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation
// is there a better way?
struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd_head, n_head));
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max);
#endif
// KQV_merged = KQV.permute(0, 2, 1, 3) // KQV_merged = KQV.permute(0, 2, 1, 3)
// [64, 12, N]
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
ggml_set_name(KQV_merged, "KQV_merged");
// cur = KQV_merged.contiguous().view(n_embd, N)
cur = ggml_cpy(ctx0, cur = ggml_cpy(ctx0,
KQV_merged, KQV_merged,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
ggml_set_name(cur, "KQV_merged_contiguous");
} }
// Projection // Projection