mirrors/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-05-11 10:26:53 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

llama: update to commit 2016f07b (#10352)

Browse source
This commit is contained in:
Jeffrey Morgan 2025-04-25 09:26:02 +09:00 • committed by GitHub
parent 11dde41824
commit e9e5f61c45
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
46 changed files with 1967 additions and 1753 deletions
Download patch file Download diff file Expand all files Collapse all files

3
llama/llama.cpp/examples/llava/clip-impl.h vendored
View file

@ -50,7 +50,6 @@
// tensor name constants
//
#define TN_TOKEN_EMBD "%s.token_embd.weight"
#define TN_POS_EMBD "%s.position_embd.weight"
#define TN_CLASS_EMBD "v.class_embd"
#define TN_PATCH_EMBD "v.patch_embd.weight" // not rename tensor with ".0" postfix for backwrad compat
@ -66,8 +65,6 @@
#define TN_LN_2 "%s.blk.%d.ln2.%s"
#define TN_LN_PRE "%s.pre_ln.%s"
#define TN_LN_POST "%s.post_ln.%s"
#define TN_TEXT_PROJ "text_projection.weight"
#define TN_VIS_PROJ "visual_projection.weight"
#define TN_LLAVA_PROJ "mm.%d.%s"
#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s"
#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"

922
llama/llama.cpp/examples/llava/clip.cpp vendored
View file

File diff suppressed because it is too large Load diff

23
llama/llama.cpp/src/llama-arch.cpp vendored
View file

@ -145,6 +145,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
{ LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION, "%s.attention.block_skip_connection" },
{ LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers" },
{ LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
{ LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
@ -1142,6 +1144,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_ATTN_Q_B, "blk.%d.attn_q_b" },
{ LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" },
{ LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" },
{ LLM_TENSOR_ATTN_K_B, "blk.%d.attn_k_b" },
{ LLM_TENSOR_ATTN_V_B, "blk.%d.attn_v_b" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
@ -1636,23 +1640,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_KV_A_MQA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_DEC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_DEC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_QKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_FFN_DOWN_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_FFN_GATE_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_FFN_UP_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_Q_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_KV_A_MQA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_K_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_V_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_DEC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_DEC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_DEC_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},

4
llama/llama.cpp/src/llama-arch.h vendored
View file

@ -149,6 +149,8 @@ enum llm_kv {
LLM_KV_ATTENTION_SCALE,
LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
LLM_KV_ATTENTION_KEY_LENGTH_MLA,
LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_DIMENSION_SECTIONS,
@ -311,6 +313,8 @@ enum llm_tensor {
LLM_TENSOR_ATTN_Q_B,
LLM_TENSOR_ATTN_KV_A_MQA,
LLM_TENSOR_ATTN_KV_B,
LLM_TENSOR_ATTN_K_B,
LLM_TENSOR_ATTN_V_B,
LLM_TENSOR_ATTN_Q_A_NORM,
LLM_TENSOR_ATTN_KV_A_NORM,
LLM_TENSOR_ATTN_SUB_NORM,

6
llama/llama.cpp/src/llama-context.cpp vendored
View file

@ -10,6 +10,7 @@
#include <cstring>
#include <stdexcept>
#include <cinttypes>
#include <cmath>
//
// llama_context
@ -473,7 +474,6 @@ ggml_tensor * llama_context::build_rope_shift(
const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
const auto & yarn_ext_factor = cparams.yarn_ext_factor;
const auto & yarn_attn_factor = cparams.yarn_attn_factor;
const auto & yarn_beta_fast = cparams.yarn_beta_fast;
const auto & yarn_beta_slow = cparams.yarn_beta_slow;
@ -482,6 +482,10 @@ ggml_tensor * llama_context::build_rope_shift(
const auto & n_rot = hparams.n_rot;
const auto & rope_type = hparams.rope_type;
// See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly.
// See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor;
ggml_tensor * tmp;
if (ggml_is_quantized(cur->type)) {

29
llama/llama.cpp/src/llama-graph.cpp vendored
View file

@ -1194,6 +1194,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
ggml_tensor * v,
ggml_tensor * kq_b,
ggml_tensor * kq_mask,
ggml_tensor * v_mla,
bool v_trans,
float kq_scale) const {
//const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
@ -1205,8 +1206,6 @@ ggml_tensor * llm_graph_context::build_attn_mha(
//const auto & n_embd_head_k = hparams.n_embd_head_k;
//const auto & n_embd_head_v = hparams.n_embd_head_v;
const auto n_embd_head_v = v_trans ? v->ne[1] : v->ne[0];
const auto n_tokens = q->ne[1];
const auto n_head = q->ne[2];
const auto n_kv = k->ne[1];
@ -1235,7 +1234,12 @@ ggml_tensor * llm_graph_context::build_attn_mha(
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
cur = ggml_reshape_2d(ctx0, cur, n_embd_head_v*n_head, n_tokens);
if (v_mla) {
cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
cur = ggml_mul_mat(ctx0, v_mla, cur);
}
cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
} else {
ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
@ -1273,9 +1277,14 @@ ggml_tensor * llm_graph_context::build_attn_mha(
ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
// for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA
if (v_mla) {
kqv = ggml_mul_mat(ctx0, v_mla, kqv);
}
cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
if (!cparams.offload_kqv) {
// all nodes between the KV store and the attention output are run on the CPU
@ -1310,6 +1319,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * v_mla,
float kq_scale,
int il) const {
GGML_UNUSED(n_tokens);
@ -1331,7 +1341,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
//cb(k, "v", il);
ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale);
ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
cb(cur, "kqv_out", il);
@ -1385,6 +1395,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * v_mla,
float kq_scale,
int il) const {
// these nodes are added to the graph together so that they are not reordered
@ -1470,7 +1481,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v,
0);
ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_trans, kq_scale);
ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, v_trans, kq_scale);
cb(cur, "kqv_out", il);
if (wo) {
@ -1529,6 +1540,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * v_mla,
float kq_scale,
int il) const {
// these nodes are added to the graph together so that they are not reordered
@ -1548,7 +1560,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
//cb(k, "v", il);
ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale);
ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
cb(cur, "kqv_out", il);
@ -1717,4 +1729,3 @@ void llm_graph_context::build_pooling(
ggml_build_forward_expand(gf, cur);
}

10
llama/llama.cpp/src/llama-graph.h vendored
View file

@ -517,11 +517,12 @@ struct llm_graph_context {
ggml_tensor * build_attn_mha(
ggml_cgraph * gf,
ggml_tensor * q, // [n_embd_head_q, n_tokens, n_head_q]
ggml_tensor * k, // [n_embd_head_k, n_tokens, n_head_k]
ggml_tensor * v, // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
ggml_tensor * q, // [n_embd_head_q, n_tokens, n_head_q]
ggml_tensor * k, // [n_embd_head_k, n_tokens, n_head_k]
ggml_tensor * v, // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
ggml_tensor * kq_b,
ggml_tensor * kq_mask,
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
bool v_trans,
float kq_scale) const;
@ -536,6 +537,7 @@ struct llm_graph_context {
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
@ -550,6 +552,7 @@ struct llm_graph_context {
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
@ -564,6 +567,7 @@ struct llm_graph_context {
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;

4
llama/llama.cpp/src/llama-hparams.h vendored
View file

@ -46,6 +46,10 @@ struct llama_hparams {
uint32_t n_rel_attn_bkts = 0;
uint32_t n_vocab = 0;
// note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
uint32_t n_embd_head_k_mla = 0;
uint32_t n_embd_head_v_mla = 0;
// for WavTokenizer
struct llama_hparams_posnet posnet;
struct llama_hparams_convnext convnext;

2
llama/llama.cpp/src/llama-kv-cache.cpp vendored
View file

@ -27,7 +27,7 @@ bool llama_kv_cache_unified::init(
recurrent = llama_model_is_recurrent(&model);
v_trans = !recurrent && !cparams.flash_attn;
can_shift = !recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
can_shift = !recurrent;
LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n",
__func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, can_shift);

334
llama/llama.cpp/src/llama-model.cpp vendored
View file

@ -1170,6 +1170,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
}
ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla, false);
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale);
@ -3281,8 +3283,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
{
const bool is_lite = (hparams.n_layer == 27);
const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
// note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
const int64_t n_embd_head_qk_rope = hparams.n_rot;
const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
const int64_t q_lora_rank = hparams.n_lora_q;
const int64_t kv_lora_rank = hparams.n_lora_kv;
@ -3308,14 +3316,22 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
if (!is_lite) {
layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}, 0);
layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
} else {
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, 0);
}
layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0);
layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_head * ( n_embd_head_v), n_embd}, 0);
layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
// note: only old legacy GGUF files will have the unsplit wkv_b tensor in
if (is_mla) {
layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
} else {
layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, 0);
}
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@ -4394,6 +4410,8 @@ void llama_model::print_info() const {
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv);
LLAMA_LOG_INFO("%s: n_embd_head_k_mla = %d\n", __func__, hparams.n_embd_head_k_mla);
LLAMA_LOG_INFO("%s: n_embd_head_v_mla = %d\n", __func__, hparams.n_embd_head_v_mla);
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
@ -4600,7 +4618,7 @@ struct llm_build_llama : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, kq_scale, il);
Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}
@ -4903,14 +4921,14 @@ struct llm_build_mllama: public llm_graph_context {
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
if (il == n_layer - 1) {
// skip computing output for unused tokens
@ -5053,7 +5071,7 @@ struct llm_build_deci : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, kq_scale, il);
Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
}
if (il == n_layer - 1) {
@ -5195,7 +5213,7 @@ struct llm_build_baichuan : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -5310,7 +5328,7 @@ struct llm_build_xverse : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -5435,7 +5453,7 @@ struct llm_build_falcon : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -5565,7 +5583,7 @@ struct llm_build_grok : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f, il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
}
if (il == n_layer - 1) {
@ -5716,7 +5734,7 @@ struct llm_build_dbrx : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -5830,7 +5848,7 @@ struct llm_build_starcoder : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -5929,7 +5947,7 @@ struct llm_build_refact : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -6083,7 +6101,7 @@ struct llm_build_bert : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
cb(cur, "kqv_out", il);
if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) {
@ -6200,7 +6218,7 @@ struct llm_build_bloom : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -6341,7 +6359,7 @@ struct llm_build_mpt : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -6487,7 +6505,7 @@ struct llm_build_stablelm : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -6610,7 +6628,7 @@ struct llm_build_qwen : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -6730,7 +6748,7 @@ struct llm_build_qwen2 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -6851,7 +6869,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -6978,7 +6996,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -7131,7 +7149,7 @@ struct llm_build_qwen3 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -7252,7 +7270,7 @@ struct llm_build_qwen3moe : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -7392,7 +7410,7 @@ struct llm_build_phi2 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f, il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
}
if (il == n_layer - 1) {
@ -7521,7 +7539,7 @@ struct llm_build_phi3 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f, il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
}
if (il == n_layer - 1) {
@ -7656,7 +7674,7 @@ struct llm_build_plamo : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
ggml_tensor * sa_out = cur;
@ -7763,7 +7781,7 @@ struct llm_build_gpt2 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -7879,7 +7897,7 @@ struct llm_build_codeshell : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -8008,7 +8026,7 @@ struct llm_build_orion : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -8135,7 +8153,7 @@ struct llm_build_internlm2 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -8332,7 +8350,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
q_states, k_states, v_states, nullptr, kq_scale, il);
q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
}
if (il == n_layer - 1) {
@ -8462,7 +8480,7 @@ struct llm_build_gemma : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f, il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
}
if (il == n_layer - 1) {
@ -8584,7 +8602,7 @@ struct llm_build_gemma2 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f, il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
}
cur = build_norm(cur,
@ -8725,7 +8743,7 @@ struct llm_build_gemma3 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, hparams.f_attention_scale, il);
Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
}
cur = build_norm(cur,
@ -8865,7 +8883,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -9200,7 +9218,7 @@ struct llm_build_command_r : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -9335,7 +9353,7 @@ struct llm_build_cohere2 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -9466,7 +9484,7 @@ struct llm_build_olmo : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, nullptr,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -9586,7 +9604,7 @@ struct llm_build_olmo2 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
cur = build_norm(cur,
@ -9719,7 +9737,7 @@ struct llm_build_olmoe : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -9852,7 +9870,7 @@ struct llm_build_openelm : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -9966,7 +9984,7 @@ struct llm_build_gptneox : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -10116,7 +10134,7 @@ struct llm_build_arctic : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -10271,7 +10289,7 @@ struct llm_build_deepseek : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, kq_scale, il);
Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
}
if (il == n_layer - 1) {
@ -10361,15 +10379,22 @@ struct llm_build_deepseek2 : public llm_graph_context {
llm_build_deepseek2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
bool is_lite = (hparams.n_layer == 27);
const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
// note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
const int64_t n_embd_head_qk_rope = hparams.n_rot;
const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
const uint32_t kv_lora_rank = hparams.n_lora_kv;
// We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
// See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k));
const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
const uint32_t n_embd_head_qk_rope = hparams.n_rot;
const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
const uint32_t kv_lora_rank = hparams.n_lora_kv;
const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(n_embd_head_k));
const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
ggml_tensor * cur;
ggml_tensor * inpL;
@ -10395,16 +10420,14 @@ struct llm_build_deepseek2 : public llm_graph_context {
{
ggml_tensor * q = NULL;
if (!is_lite) {
// {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
cb(q, "q", il);
q = build_norm(q,
model.layers[il].attn_q_a_norm, NULL,
model.layers[il].attn_q_a_norm, nullptr,
LLM_NORM_RMS, il);
cb(q, "q", il);
// {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
cb(q, "q", il);
} else {
@ -10412,96 +10435,125 @@ struct llm_build_deepseek2 : public llm_graph_context {
cb(q, "q", il);
}
// split into {n_head * n_embd_head_qk_nope, n_tokens}
ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
ggml_row_size(q->type, hparams.n_embd_head_k),
ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
// split into {n_embd_head_qk_nope, n_head, n_tokens}
ggml_tensor * q_nope = ggml_view_3d(ctx0, q,
n_embd_head_qk_nope, n_head, n_tokens,
ggml_row_size(q->type, n_embd_head_k),
ggml_row_size(q->type, n_embd_head_k) * n_head,
0);
cb(q_nope, "q_nope", il);
// and {n_head * n_embd_head_qk_rope, n_tokens}
ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
ggml_row_size(q->type, hparams.n_embd_head_k),
ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
// and {n_embd_head_qk_rope, n_head, n_tokens}
ggml_tensor * q_pe = ggml_view_3d(ctx0, q,
n_embd_head_qk_rope, n_head, n_tokens,
ggml_row_size(q->type, n_embd_head_k),
ggml_row_size(q->type, n_embd_head_k) * n_head,
ggml_row_size(q->type, n_embd_head_qk_nope));
cb(q_pe, "q_pe", il);
// {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
cb(kv_pe_compresseed, "kv_pe_compresseed", il);
ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
cb(kv_cmpr_pe, "kv_cmpr_pe", il);
// split into {kv_lora_rank, n_tokens}
ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
kv_pe_compresseed->nb[1],
ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe,
kv_lora_rank, n_tokens,
ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
0);
cb(kv_compressed, "kv_compressed", il);
cb(kv_cmpr, "kv_cmpr", il);
// and {n_embd_head_qk_rope, n_tokens}
ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
kv_pe_compresseed->nb[1],
kv_pe_compresseed->nb[1],
ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
// and {n_embd_head_qk_rope, 1, n_tokens}
ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe,
n_embd_head_qk_rope, 1, n_tokens,
ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
cb(k_pe, "k_pe", il);
// TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont
kv_compressed = ggml_cont(ctx0, kv_compressed);
kv_compressed = build_norm(kv_compressed,
model.layers[il].attn_kv_a_norm, NULL,
LLM_NORM_RMS, il);
cb(kv_compressed, "kv_compressed", il);
// {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
cb(kv, "kv", il);
// split into {n_head * n_embd_head_qk_nope, n_tokens}
ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
0);
cb(k_nope, "k_nope", il);
// and {n_head * n_embd_head_v, n_tokens}
ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
ggml_row_size(kv->type, (n_embd_head_qk_nope)));
cb(v_states, "v_states", il);
v_states = ggml_cont(ctx0, v_states);
cb(v_states, "v_states", il);
v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
0);
cb(v_states, "v_states", il);
q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
q_pe = ggml_rope_ext(
ctx0, q_pe, inp_pos, nullptr,
q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor_scaled, beta_fast, beta_slow
);
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(q_pe, "q_pe", il);
// shared RoPE key
k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
k_pe = ggml_rope_ext(
ctx0, k_pe, inp_pos, nullptr,
k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor_scaled, beta_fast, beta_slow
);
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(k_pe, "k_pe", il);
ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
cb(q_states, "q_states", il);
kv_cmpr = build_norm(kv_cmpr,
model.layers[il].attn_kv_a_norm, nullptr,
LLM_NORM_RMS, il);
cb(kv_cmpr, "kv_cmpr", il);
ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
cb(k_states, "k_states", il);
if (is_mla) {
// {n_embd_head_qk_nope, n_tokens, n_head}
q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
cb(q_nope, "q_nope_perm", il);
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
q_states, k_states, v_states, nullptr, kq_scale, il);
// {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
cb(q_nope_absorbed, "q_nope_absorbed", il);
// {kv_lora_rank, n_head, n_tokens}
q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
// {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
// note: rope must go first for in-place context shifting in build_rope_shift()
ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
cb(Qcur, "Qcur", il);
kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
cb(kv_cmpr, "kv_cmpr_reshape", il);
// {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
cb(Kcur, "Kcur", il);
// {kv_lora_rank, 1, n_tokens}
ggml_tensor * Vcur = kv_cmpr;
cb(Vcur, "Vcur", il);
// note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, model.layers[il].wv_b, kq_scale, il);
} else {
ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
cb(kv, "kv", il);
// split into {n_embd_head_qk_nope, n_head, n_tokens}
ggml_tensor * k_nope = ggml_view_3d(ctx0, kv,
n_embd_head_qk_nope, n_head, n_tokens,
ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
0);
cb(k_nope, "k_nope_view", il);
// and {n_embd_head_v, n_head, n_tokens}
ggml_tensor * Vcur = ggml_view_3d(ctx0, kv,
n_embd_head_v, n_head, n_tokens,
ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
ggml_row_size(kv->type, n_embd_head_qk_nope));
cb(Vcur, "Vcur_view", il);
Vcur = ggml_cont(ctx0, Vcur);
cb(Vcur, "Vcur_cont", il);
// note: rope must go first for in-place context shifting in build_rope_shift()
ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
cb(Kcur, "Kcur", il);
// note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
}
}
if (il == n_layer - 1) {
@ -10667,7 +10719,7 @@ struct llm_build_bitnet : public llm_graph_context {
cur = build_attn(inp_attn, gf,
NULL, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
cur = build_norm(cur,
model.layers[il].attn_sub_norm, NULL,
@ -10790,7 +10842,7 @@ struct llm_build_t5_enc : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo_enc, nullptr,
Qcur, Kcur, Vcur, kq_b, 1.0f, il);
Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
cb(cur, "kqv_out", il);
}
@ -10896,7 +10948,7 @@ struct llm_build_t5_dec : public llm_graph_context {
cur = build_attn(inp_attn_self, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, kq_b, 1.0f, il);
Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
cb(cur, "kqv_out", il);
}
@ -10928,7 +10980,7 @@ struct llm_build_t5_dec : public llm_graph_context {
cur = build_attn(inp_attn_cross, gf,
model.layers[il].wo_cross, nullptr,
Qcur, Kcur, Vcur, nullptr, 1.0f, il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
cb(cur, "kqv_out", il);
//ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
@ -11061,7 +11113,7 @@ struct llm_build_jais : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/float(n_embd_head), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il);
}
if (il == n_layer - 1) {
@ -11193,7 +11245,7 @@ struct llm_build_chatglm : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -11326,7 +11378,7 @@ struct llm_build_glm4 : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -11470,7 +11522,7 @@ struct llm_build_nemotron : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -11601,7 +11653,7 @@ struct llm_build_exaone : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {
@ -12503,7 +12555,7 @@ struct llm_build_chameleon : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, nullptr,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
if (hparams.swin_norm) {
cur = build_norm(cur,
@ -12683,14 +12735,14 @@ struct llm_build_solar : public llm_graph_context {
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, kq_scale, il);
Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}
@ -13018,7 +13070,7 @@ struct llm_build_plm : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
q_states, k_states, v_states, nullptr, kq_scale, il);
q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
}
if (il == n_layer - 1) {
@ -13141,7 +13193,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
cur = build_attn(inp_attn, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_rot)), il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
}
if (il == n_layer - 1) {

2
llama/llama.cpp/src/llama-model.h vendored
View file

@ -174,6 +174,8 @@ struct llama_layer {
struct ggml_tensor * wq_b = nullptr;
struct ggml_tensor * wkv_a_mqa = nullptr;
struct ggml_tensor * wkv_b = nullptr;
struct ggml_tensor * wk_b = nullptr;
struct ggml_tensor * wv_b = nullptr;
struct ggml_tensor * wq_cross = nullptr;
struct ggml_tensor * wk_cross = nullptr;
struct ggml_tensor * wv_cross = nullptr;

5
llama/llama.cpp/src/llama-vocab.cpp vendored
View file

@ -1833,6 +1833,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (false
|| t.first == "<|fim_prefix|>" // Qwen
|| t.first == "<fim-prefix>"
|| t.first == "<fim_prefix>" // Granite
|| t.first == "<｜fim▁begin｜>" // DeepSeek
|| t.first == "<PRE>"
|| t.first == "▁<PRE>" // CodeLlama
@ -1851,6 +1852,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (false
|| t.first == "<|fim_suffix|>" // Qwen
|| t.first == "<fim-suffix>"
|| t.first == "<fim_suffix>" // Granite
|| t.first == "<｜fim▁hole｜>" // DeepSeek
|| t.first == "<SUF>"
|| t.first == "▁<SUF>" // CodeLlama
@ -1869,6 +1871,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (false
|| t.first == "<|fim_middle|>" // Qwen
|| t.first == "<fim-middle>"
|| t.first == "<fim_middle>" // Granite
|| t.first == "<｜fim▁end｜>" // DeepSeek
|| t.first == "<MID>"
|| t.first == "▁<MID>" // CodeLlama
@ -1887,6 +1890,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
if (false
|| t.first == "<|fim_pad|>" // Qwen
|| t.first == "<fim-pad>"
|| t.first == "<fim_pad>" // Granite
|| t.first == "<PAD>"
) {
special_fim_pad_id = t.second;
@ -1905,6 +1909,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<|repo_name|>"
|| t.first == "<fim-repo>"
|| t.first == "<REPO>"
|| t.first == "<reponame>" // Granite
) {
special_fim_rep_id = t.second;
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {