vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/build.yml
#	.gitignore
#	CMakeLists.txt
#	Makefile
#	README.md
#	ci/run.sh
#	flake.lock
#	flake.nix
#	ggml-cuda.cu
#	ggml-cuda.h
#	scripts/get-wikitext-2.sh
#	tests/CMakeLists.txt
This commit is contained in:
Concedo 2024-01-21 14:32:15 +08:00
commit 1cb8a5e955
79 changed files with 6273 additions and 2982 deletions
Download patch file Download diff file Expand all files Collapse all files

614
llama.cpp
View file

@ -193,8 +193,10 @@ enum llm_arch {
LLM_ARCH_BLOOM,
LLM_ARCH_STABLELM,
LLM_ARCH_QWEN,
LLM_ARCH_QWEN2,
LLM_ARCH_PHI2,
LLM_ARCH_PLAMO,
LLM_ARCH_CODESHELL,
LLM_ARCH_UNKNOWN,
};
@ -212,8 +214,10 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
{ LLM_ARCH_BLOOM, "bloom" },
{ LLM_ARCH_STABLELM, "stablelm" },
{ LLM_ARCH_QWEN, "qwen" },
{ LLM_ARCH_QWEN2, "qwen2" },
{ LLM_ARCH_PHI2, "phi2" },
{ LLM_ARCH_PLAMO, "plamo" },
{ LLM_ARCH_CODESHELL, "codeshell" },
};
enum llm_kv {
@ -567,6 +571,23 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_QWEN2,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_PHI2,
{
@ -601,6 +622,26 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_CODESHELL,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_UNKNOWN,
@ -1398,6 +1439,9 @@ struct llama_cparams {
bool mul_mat_q;
bool offload_kqv;
ggml_backend_sched_eval_callback cb_eval;
void * cb_eval_user_data;
};
struct llama_layer {
@ -1606,7 +1650,7 @@ struct llama_model {
std::unique_ptr<llama_mmap> mapping;
// objects representing data potentially being locked in memory
llama_mlock mlock_buf;
std::vector<std::unique_ptr<llama_mlock>> mlock_bufs;
llama_mlock mlock_mmap;
// for quantize-stats only
@ -2855,6 +2899,17 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_QWEN2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 24: model.type = e_model::MODEL_1B; break;
case 32: model.type = e_model::MODEL_7B; break;
case 40: model.type = e_model::MODEL_13B; break;
case 80: model.type = e_model::MODEL_70B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_PHI2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@ -2885,6 +2940,14 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_CODESHELL:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
switch (hparams.n_layer) {
case 42: model.type = e_model::MODEL_SMALL; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
default: (void)0;
}
@ -3465,7 +3528,12 @@ static bool llm_load_tensors(
{
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd});
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});
if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_OUTPUT, "weight").c_str()) >= 0) {
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});
} else {
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
ml.n_created--; // artificial tensor
}
}
for (int i = 0; i < n_layer; ++i) {
@ -3696,6 +3764,41 @@ static bool llm_load_tensors(
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff/2});
}
} break;
case LLM_ARCH_QWEN2:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
// output
{
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});
}
for (int i = 0; i < n_layer; ++i) {
ggml_context * ctx_layer = ctx_for_layer(i);
ggml_context * ctx_split = ctx_for_layer_split(i);
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
// optional bias tensors
layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd});
layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa});
layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa});
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
}
} break;
case LLM_ARCH_PHI2:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@ -3806,6 +3909,42 @@ static bool llm_load_tensors(
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff});
}
} break;
case LLM_ARCH_CODESHELL:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
// output
{
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd});
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});
}
for (int i = 0; i < n_layer; ++i) {
ggml_context * ctx_layer = ctx_for_layer(i);
ggml_context * ctx_split = ctx_for_layer_split(i);
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd});
layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa});
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd});
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd});
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd});
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff});
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
@ -3842,8 +3981,10 @@ static bool llm_load_tensors(
else {
buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
if (buf != nullptr && use_mlock && ggml_backend_buffer_is_host(buf)) {
model.mlock_buf.init (ggml_backend_buffer_get_base(buf));
model.mlock_buf.grow_to(ggml_backend_buffer_get_size(buf));
model.mlock_bufs.emplace_back(new llama_mlock);
auto & mlock_buf = model.mlock_bufs.back();
mlock_buf->init (ggml_backend_buffer_get_base(buf));
mlock_buf->grow_to(ggml_backend_buffer_get_size(buf));
}
}
if (buf == nullptr) {
@ -4204,6 +4345,7 @@ static struct ggml_tensor * llm_build_kqv(
const llama_model & model,
const llama_hparams & hparams,
const llama_kv_cache & kv,
struct ggml_cgraph * graph,
struct ggml_tensor * wo,
struct ggml_tensor * wo_b,
struct ggml_tensor * q_cur,
@ -4282,6 +4424,8 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens);
cb(cur, "kqv_merged_cont", il);
ggml_build_forward_expand(graph, cur);
cur = ggml_mul_mat(ctx, wo, cur);
if (wo_b) {
cb(cur, "kqv_wo", il);
@ -4294,6 +4438,44 @@ static struct ggml_tensor * llm_build_kqv(
return cur;
}
static struct ggml_tensor * llm_build_kv(
struct ggml_context * ctx,
const llama_model & model,
const llama_hparams & hparams,
const llama_kv_cache & kv,
struct ggml_cgraph * graph,
struct ggml_tensor * wo,
struct ggml_tensor * wo_b,
struct ggml_tensor * k_cur,
struct ggml_tensor * v_cur,
struct ggml_tensor * q_cur,
struct ggml_tensor * kq_mask,
int64_t n_ctx,
int32_t n_tokens,
int32_t kv_head,
int32_t n_kv,
float max_alibi_bias,
float kq_scale,
const llm_build_cb & cb,
int il) {
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(graph, k_cur);
ggml_build_forward_expand(graph, v_cur);
ggml_build_forward_expand(graph, q_cur);
llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il);
struct ggml_tensor * cur;
cur = llm_build_kqv(ctx, model, hparams, kv, graph,
wo, wo_b,
q_cur, kq_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, kq_scale, cb, il);
cb(cur, "kqv_out", il);
return cur;
}
struct llm_build_context {
const llama_model & model;
const llama_hparams & hparams;
@ -4451,12 +4633,6 @@ struct llm_build_context {
cb(Vcur, "Vcur", il);
}
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(gf, Qcur);
ggml_build_forward_expand(gf, Kcur);
ggml_build_forward_expand(gf, Vcur);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
@ -4471,11 +4647,9 @@ struct llm_build_context {
);
cb(Kcur, "Kcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -4652,14 +4826,13 @@ struct llm_build_context {
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
// apply ALiBi for 13B model
const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f;
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -4781,11 +4954,9 @@ struct llm_build_context {
);
cb(Kcur, "Kcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -4882,11 +5053,9 @@ struct llm_build_context {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -5089,12 +5258,9 @@ struct llm_build_context {
);
cb(Vcur, "Vcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
// TODO: not tested, could be broken
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Q, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Q, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -5181,11 +5347,9 @@ struct llm_build_context {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
cb(Qcur, "Qcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -5279,11 +5443,9 @@ struct llm_build_context {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -5374,11 +5536,9 @@ struct llm_build_context {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -5486,11 +5646,9 @@ struct llm_build_context {
);
cb(Kcur, "Kcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -5603,11 +5761,9 @@ struct llm_build_context {
);
cb(Kcur, "Kcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -5652,6 +5808,126 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_qwen2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
cb(inpL, "inp_embd", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
cb(inp_pos, "inp_pos", -1);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
cb(KQ_mask, "KQ_mask", -1);
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
// norm
cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(gf, Qcur);
ggml_build_forward_expand(gf, Kcur);
ggml_build_forward_expand(gf, Vcur);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// feed-forward network
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il);
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, NULL,
model.layers[il].ffn_gate, NULL,
model.layers[il].ffn_down, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
cb(cur, "ffn_out", il);
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = llm_build_norm(ctx0, cur, hparams,
model.output_norm, NULL,
LLM_NORM_RMS, cb, -1);
cb(cur, "result_norm", -1);
// lm_head
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_phi2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
@ -5733,11 +6009,9 @@ struct llm_build_context {
);
cb(Kcur, "Kcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f, cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f, cb, il);
cb(cur, "kqv_out", il);
}
@ -5838,11 +6112,9 @@ struct llm_build_context {
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
struct ggml_tensor * sa_out = cur;
@ -5939,11 +6211,118 @@ struct llm_build_context {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, model, hparams, kv_self,
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
// add the input
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
cb(ffn_inp, "ffn_inp", il);
// FF
{
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm,
model.layers[il].ffn_norm_b,
LLM_NORM, cb, il);
cb(cur, "ffn_norm", il);
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b,
NULL, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b,
NULL,
LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
cb(cur, "ffn_out", il);
}
inpL = ggml_add(ctx0, cur, ffn_inp);
cb(inpL, "l_out", il);
}
cur = llm_build_norm(ctx0, inpL, hparams,
model.output_norm,
model.output_norm_b,
LLM_NORM, cb, -1);
cb(cur, "result_norm", -1);
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_codeshell() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
cb(inpL, "inp_embd", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
cb(inp_pos, "inp_pos", -1);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
cb(KQ_mask, "KQ_mask", -1);
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm,
model.layers[il].attn_norm_b,
LLM_NORM, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
cb(cur, "bqkv", il);
struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
cb(tmpq, "tmpq", il);
cb(tmpk, "tmpk", il);
cb(Vcur, "Vcur", il);
struct ggml_tensor * Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
struct ggml_tensor * Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -6011,6 +6390,14 @@ static struct ggml_cgraph * llama_build_graph(
ggml_set_name(cur, name);
}
if (!lctx.cparams.offload_kqv) {
if (strcmp(name, "kqv_merged_cont") == 0) {
// all nodes between the KV store and the attention output are run on the CPU
ggml_backend_sched_set_node_backend(lctx.sched, cur, lctx.backend_cpu);
}
}
//
// allocate input tensors and set input data
//
@ -6167,6 +6554,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_qwen();
} break;
case LLM_ARCH_QWEN2:
{
result = llm.build_qwen2();
} break;
case LLM_ARCH_PHI2:
{
result = llm.build_phi2();
@ -6179,6 +6570,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_gpt2();
} break;
case LLM_ARCH_CODESHELL:
{
result = llm.build_codeshell();
} break;
default:
GGML_ASSERT(false);
}
@ -6284,6 +6679,7 @@ static int llama_decode_internal(
//printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
ggml_backend_sched_reset(lctx.sched);
ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data);
ggml_cgraph * gf = llama_build_graph(lctx, batch);
@ -8235,39 +8631,59 @@ static void llama_log_softmax(float * array, size_t size) {
}
}
void llama_sample_apply_guidance(
struct llama_context * ctx,
float * logits,
float * logits_guidance,
float scale) {
GGML_ASSERT(ctx);
const auto t_start_sample_us = ggml_time_us();
const auto n_vocab = llama_n_vocab(llama_get_model(ctx));
llama_log_softmax(logits, n_vocab);
llama_log_softmax(logits_guidance, n_vocab);
for (int i = 0; i < n_vocab; ++i) {
auto & l = logits[i];
const auto & g = logits_guidance[i];
l = scale * (l - g) + g;
}
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
}
void llama_sample_classifier_free_guidance(
struct llama_context * ctx,
llama_token_data_array * candidates,
struct llama_context * guidance_ctx,
float scale) {
int64_t t_start_sample_us = ggml_time_us();
GGML_ASSERT(ctx);
int64_t t_start_sample_us;
auto n_vocab = llama_n_vocab(llama_get_model(ctx));
t_start_sample_us = ggml_time_us();
const size_t n_vocab = llama_n_vocab(llama_get_model(ctx));
GGML_ASSERT(n_vocab == (int)candidates->size);
GGML_ASSERT(n_vocab == candidates->size);
GGML_ASSERT(!candidates->sorted);
std::vector<float> logits_base;
logits_base.reserve(candidates->size);
for (size_t i = 0; i < candidates->size; ++i) {
logits_base.push_back(candidates->data[i].logit);
}
llama_log_softmax(logits_base.data(), candidates->size);
float* logits_guidance = llama_get_logits(guidance_ctx);
llama_log_softmax(logits_guidance, n_vocab);
for (int i = 0; i < n_vocab; ++i) {
float logit_guidance = logits_guidance[i];
float logit_base = logits_base[i];
candidates->data[i].logit = scale * (logit_base - logit_guidance) + logit_guidance;
std::vector<float> logits_base(n_vocab);
for (size_t i = 0; i < n_vocab; ++i) {
logits_base[i] = candidates->data[i].logit;
}
if (ctx) {
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
float * logits_guidance = llama_get_logits(guidance_ctx);
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
llama_sample_apply_guidance(ctx, logits_base.data(), logits_guidance, scale);
t_start_sample_us = ggml_time_us();
for (size_t i = 0; i < n_vocab; ++i) {
candidates->data[i].logit = logits_base[i];
}
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
}
llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int32_t m, float * mu) {
@ -8691,6 +9107,8 @@ struct quantize_state_internal {
int n_k_quantized = 0;
int n_fallback = 0;
bool has_imatrix = false;
quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params)
: model(model)
, params(params)
@ -8792,7 +9210,12 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
}
else if (name == "token_embd.weight") new_type = GGML_TYPE_Q2_K;
} else if (name.find("attn_v.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) {
new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
@ -8863,6 +9286,13 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
new_type = GGML_TYPE_Q5_K;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0)
&& qs.has_imatrix && i_layer < n_layer/8) {
// Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0.
// We only do it when an imatrix is provided because a) we want to make sure that one can always get the
// same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
}
++qs.i_feed_forward_w2;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (arch != LLM_ARCH_FALCON) {
@ -8986,6 +9416,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
imatrix_data = static_cast<const std::unordered_map<std::string, std::vector<float>>*>(params->imatrix);
if (imatrix_data) {
LLAMA_LOG_INFO("================================ Have weights data with %d entries\n",int(imatrix_data->size()));
qs.has_imatrix = true;
}
}
@ -9045,8 +9476,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// placeholder for the meta data
::zeros(fout, meta_size);
std::set<ggml_type> used_iq2;
for (int i = 0; i < ml.n_tensors; ++i) {
struct ggml_tensor * tensor = ml.get_tensor_meta(i);
@ -9099,11 +9528,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
} else {
const size_t nelements = ggml_nelements(tensor);
if ((new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_XS) && used_iq2.find(new_type) == used_iq2.end()) {
ggml_init_iq2_quantization(new_type);
used_iq2.insert(new_type);
}
const float * imatrix = nullptr;
if (imatrix_data) {
auto it = imatrix_data->find(tensor->name);
@ -9229,10 +9653,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
fout.close();
for (auto type : used_iq2) {
ggml_deinit_iq2_quantization(type);
}
gguf_free(ctx_out);
LLAMA_LOG_INFO("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
@ -9578,6 +9998,8 @@ struct llama_context_params llama_context_default_params() {
/*.yarn_beta_fast =*/ 32.0f,
/*.yarn_beta_slow =*/ 1.0f,
/*.yarn_orig_ctx =*/ 0,
/*.cb_eval =*/ nullptr,
/*.cb_eval_user_data =*/ nullptr,
/*.type_k =*/ GGML_TYPE_F16,
/*.type_v =*/ GGML_TYPE_F16,
/*.mul_mat_q =*/ true,
@ -9638,6 +10060,7 @@ void llama_backend_free(void) {
#ifdef GGML_USE_MPI
ggml_mpi_backend_free();
#endif
ggml_quantize_free();
}
int64_t llama_time_us(void) {
@ -9718,6 +10141,9 @@ struct llama_context * llama_new_context_with_model(
hparams.n_yarn_orig_ctx != 0 ? hparams.n_yarn_orig_ctx :
hparams.n_ctx_train;
cparams.cb_eval = params.cb_eval;
cparams.cb_eval_user_data = params.cb_eval_user_data;
auto rope_scaling_type = params.rope_scaling_type;
if (rope_scaling_type == LLAMA_ROPE_SCALING_UNSPECIFIED) {
rope_scaling_type = hparams.rope_scaling_type_train;