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

mpt : removed ne01 + n_past == ne00 assertion from alibi (cuda/f32) and rope_shift from build_mpt

Browse source
This commit is contained in:
Jan Ploski 2023-10-03 21:53:31 +02:00
parent 470801292d
commit 1364bcd712
3 changed files with 5 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml-cuda.cu
View file

@ -6295,12 +6295,12 @@ inline void ggml_cuda_op_alibi(
const int64_t ne02 = src0->ne[2]; const int64_t ne02 = src0->ne[2];
const int64_t nrows = ggml_nrows(src0); const int64_t nrows = ggml_nrows(src0);
const int n_past = ((int32_t *) dst->op_params)[0]; //const int n_past = ((int32_t *) dst->op_params)[0];
const int n_head = ((int32_t *) dst->op_params)[1]; const int n_head = ((int32_t *) dst->op_params)[1];
float max_bias; float max_bias;
memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float)); memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
GGML_ASSERT(ne01 + n_past == ne00); //GGML_ASSERT(ne01 + n_past == ne00);
GGML_ASSERT(n_head == ne02); GGML_ASSERT(n_head == ne02);
const int n_heads_log2_floor = 1 << (int) floor(log2(n_head)); const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));

4
ggml.c
View file

@ -12889,7 +12889,7 @@ static void ggml_compute_forward_alibi_f32(
return; return;
} }
const int n_past = ((int32_t *) dst->op_params)[0]; //const int n_past = ((int32_t *) dst->op_params)[0];
const int n_head = ((int32_t *) dst->op_params)[1]; const int n_head = ((int32_t *) dst->op_params)[1];
float max_bias; float max_bias;
memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float)); memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
@ -12910,7 +12910,7 @@ static void ggml_compute_forward_alibi_f32(
//const int nb3 = src0->nb[3]; //const int nb3 = src0->nb[3];
GGML_ASSERT(nb0 == sizeof(float)); GGML_ASSERT(nb0 == sizeof(float));
GGML_ASSERT(ne1 + n_past == ne0); //GGML_ASSERT(ne1 + n_past == ne0);
GGML_ASSERT(n_head == ne2); GGML_ASSERT(n_head == ne2);
// add alibi to src0 (KQ_scaled) // add alibi to src0 (KQ_scaled)

32
llama.cpp
View file

@ -4076,8 +4076,6 @@ static struct ggml_cgraph * llm_build_mpt(
const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n;
const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift;
//printf("kv_head = %d, n_kv = %d, n_tokens = %d, n_ctx = %d, is_measure = %d, has_shift = %d\n", //printf("kv_head = %d, n_kv = %d, n_tokens = %d, n_ctx = %d, is_measure = %d, has_shift = %d\n",
// kv_head, n_kv, n_tokens, n_ctx, ggml_allocr_is_measure(lctx.alloc), kv_self.has_shift); // kv_head, n_kv, n_tokens, n_ctx, ggml_allocr_is_measure(lctx.alloc), kv_self.has_shift);
@ -4176,34 +4174,6 @@ static struct ggml_cgraph * llm_build_mpt(
} }
} }
// shift the entire K-cache if needed
// TODO: Do we need to handle it? (MPT uses alibi instead of rope)
/* if (do_rope_shift) {
struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx);
offload_func_kq(K_shift);
ggml_set_name(K_shift, "K_shift");
ggml_allocr_alloc(lctx.alloc, K_shift);
if (!ggml_allocr_is_measure(lctx.alloc)) {
int * data = (int *) K_shift->data;
for (int i = 0; i < n_ctx; ++i) {
data[i] = kv_self.cells[i].delta;
}
}
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * tmp =
ggml_rope_custom_inplace(ctx0,
ggml_view_3d(ctx0, kv_self.k,
n_embd_head, n_head_kv, n_ctx,
ggml_element_size(kv_self.k)*n_embd_head,
ggml_element_size(kv_self.k)*n_embd_gqa,
ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il),
K_shift, n_embd_head, 2, 0, freq_base, freq_scale);
offload_func_kq(tmp);
ggml_build_forward_expand(gf, tmp);
}
}*/
for (int il = 0; il < n_layer; ++il) { for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * attn_norm; struct ggml_tensor * attn_norm;
@ -4306,7 +4276,7 @@ static struct ggml_cgraph * llm_build_mpt(
// TODO: replace with ggml_add() // TODO: replace with ggml_add()
struct ggml_tensor * KQ_scaled_alibi = struct ggml_tensor * KQ_scaled_alibi =
ggml_alibi(ctx0, KQ_scaled, std::max(kv_head, n_kv - n_tokens), n_head, max_alibi_bias); ggml_alibi(ctx0, KQ_scaled, 0, n_head, max_alibi_bias);
offload_func_kq(KQ_scaled_alibi); offload_func_kq(KQ_scaled_alibi);
ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi");