vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	CMakeLists.txt
#	Makefile
#	README.md
#	flake.lock
#	llama.cpp
This commit is contained in:
Concedo 2024-02-07 22:21:32 +08:00
commit ec2dbd99a3
21 changed files with 2614 additions and 1863 deletions
Download patch file Download diff file Expand all files Collapse all files

267
llama.cpp
View file

@ -229,6 +229,7 @@ enum llm_arch {
LLM_ARCH_CODESHELL,
LLM_ARCH_ORION,
LLM_ARCH_INTERNLM2,
LLM_ARCH_MINICPM,
LLM_ARCH_UNKNOWN,
};
@ -252,6 +253,7 @@ static std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_CODESHELL, "codeshell" },
{ LLM_ARCH_ORION, "orion" },
{ LLM_ARCH_INTERNLM2, "internlm2" },
{ LLM_ARCH_MINICPM, "minicpm" },
};
enum llm_kv {
@ -714,6 +716,29 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_MINICPM,
{
{ 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_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ 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_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" },
{ LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" },
{ LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
},
},
{
LLM_ARCH_UNKNOWN,
{
@ -1358,7 +1383,7 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) {
#elif defined(GGML_USE_CUBLAS)
buft = ggml_backend_cuda_buffer_type(gpu);
#elif defined(GGML_USE_VULKAN)
buft = ggml_backend_vk_buffer_type();
buft = ggml_backend_vk_buffer_type(gpu);
#elif defined(GGML_USE_SYCL)
buft = ggml_backend_sycl_buffer_type(gpu);
#elif defined(GGML_USE_CLBLAST)
@ -1395,6 +1420,33 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(int fallback_g
GGML_UNUSED(tensor_split);
}
static size_t llama_get_device_count() {
#if defined(GGML_USE_CUBLAS)
return ggml_backend_cuda_get_device_count();
#elif defined(GGML_USE_VULKAN)
return ggml_backend_vk_get_device_count();
#else
return 1;
#endif
}
static size_t llama_get_device_memory(int device) {
#if defined(GGML_USE_CUBLAS)
size_t total;
size_t free;
ggml_backend_cuda_get_device_memory(device, &total, &free);
return free;
#elif defined(GGML_USE_VULKAN)
size_t total;
size_t free;
ggml_backend_vk_get_device_memory(device, &total, &free);
return free;
#else
return 1;
GGML_UNUSED(device);
#endif
}
//
// globals
//
@ -1418,6 +1470,7 @@ enum e_model {
MODEL_UNKNOWN,
MODEL_0_5B,
MODEL_1B,
MODEL_2B,
MODEL_3B,
MODEL_4B,
MODEL_7B,
@ -1769,6 +1822,10 @@ struct llama_context {
ggml_backend_free(backend);
}
#ifdef GGML_USE_VULKAN
ggml_vk_free_cpu_assist();
#endif
ggml_backend_buffer_free(buf_input);
ggml_free(ctx_input);
}
@ -2794,6 +2851,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
static const char * llama_model_type_name(e_model type) {
switch (type) {
case MODEL_1B: return "1B";
case MODEL_2B: return "2B";
case MODEL_3B: return "3B";
case MODEL_7B: return "7B";
case MODEL_8B: return "8B";
@ -2933,6 +2991,13 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_MINICPM:
{
switch (hparams.n_layer) {
case 40: model.type = e_model::MODEL_2B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_FALCON:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@ -3474,22 +3539,18 @@ static bool llm_load_tensors(
model.buft_layer[i] = llama_default_buffer_type_cpu(true);
}
#ifdef GGML_USE_CUBLAS
if (split_mode == LLAMA_SPLIT_LAYER) {
// calculate the split points
int device_count = ggml_backend_cuda_get_device_count();
int device_count = llama_get_device_count();
bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + device_count, [](float x) { return x == 0.0f; });
float splits[GGML_CUDA_MAX_DEVICES];
std::vector<float> splits(device_count);
if (all_zero) {
// default split, by free memory
for (int i = 0; i < device_count; ++i) {
size_t total;
size_t free;
ggml_backend_cuda_get_device_memory(i, &total, &free);
splits[i] = free;
splits[i] = llama_get_device_memory(i);
}
} else {
std::copy(tensor_split, tensor_split + device_count, splits);
std::copy(tensor_split, tensor_split + device_count, splits.begin());
}
// sum and normalize the splits to get the split points
@ -3505,19 +3566,17 @@ static bool llm_load_tensors(
// assign the repeating layers to the devices according to the splits
int act_gpu_layers = std::min(n_gpu_layers, (int)n_layer + 1);
for (int64_t i = i_gpu_start; i < n_layer; ++i) {
int layer_gpu = std::upper_bound(splits, splits + device_count, float(i - i_gpu_start)/act_gpu_layers) - splits;
int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(i - i_gpu_start)/act_gpu_layers) - splits.begin();
model.buft_layer[i] = llama_default_buffer_type_offload(layer_gpu);
}
// assign the output layer
if (n_gpu_layers > n_layer) {
int layer_gpu = std::upper_bound(splits, splits + device_count, float(act_gpu_layers - 1)/act_gpu_layers) - splits;
int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(act_gpu_layers - 1)/act_gpu_layers) - splits.begin();
model.buft_output = llama_default_buffer_type_offload(layer_gpu);
} else {
model.buft_output = llama_default_buffer_type_cpu(true);
}
} else
#endif
{
} else {
ggml_backend_buffer_type_t split_buft;
if (split_mode == LLAMA_SPLIT_ROW) {
split_buft = llama_default_buffer_type_split(main_gpu, tensor_split);
@ -3596,13 +3655,16 @@ static bool llm_load_tensors(
switch (model.arch) {
case LLM_ARCH_LLAMA:
case LLM_ARCH_REFACT:
case LLM_ARCH_MINICPM:
{
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});
if (model.arch != LLM_ARCH_MINICPM){
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) {
@ -6853,6 +6915,153 @@ struct llm_build_context {
return gf;
}
// ref: https://arxiv.org/abs/2203.03466
// https://github.com/ggerganov/llama.cpp/issues/5276#issuecomment-1925774738
// based on the original build_llama() function
struct ggml_cgraph * build_minicpm() {
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);
const int64_t n_embd = hparams.n_embd;
//TODO: if the model varies, these parameters need to be read from the model
const int64_t n_embd_base = 256;
const float scale_embd = 12.0f;
const float scale_depth = 1.4f;
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
cb(inpL, "inp_embd", -1);
// scale the input embeddings
inpL = ggml_scale(ctx0, inpL, scale_embd);
cb(inpL, "inp_scaled", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
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_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
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, lctx.inp_K_shift, LLM_ROPE, 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);
if (model.layers[il].bq) {
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);
if (model.layers[il].bk) {
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);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
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,
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, 0, 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);
}
// scale_res - scale the hidden states for residual connection
const float scale_res = scale_depth/sqrtf(float(n_layer));
cur = ggml_scale(ctx0, cur, scale_res);
cb(cur, "hidden_scaled", -1);
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);
}
// scale the hidden states for residual connection
cur = ggml_scale(ctx0, cur, scale_res);
cb(cur, "hidden_scaled_ffn", -1);
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 scaling
const float scale_lmhead = float(n_embd_base)/float(n_embd);
cur = ggml_scale(ctx0, cur, scale_lmhead);
cb(cur, "lmhead_scaling", -1);
// lm_head
cur = ggml_mul_mat(ctx0, model.tok_embd, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
};
static struct ggml_cgraph * llama_build_graph(
@ -7015,6 +7224,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_internlm2();
} break;
case LLM_ARCH_MINICPM:
{
result = llm.build_minicpm();
} break;
default:
GGML_ASSERT(false);
}
@ -9778,8 +9991,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
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_IQ3_XXS && qs.model.hparams.n_gqa() >= 4) {
new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : !qs.has_imatrix ? GGML_TYPE_Q3_K : GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
@ -9818,9 +10031,9 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
}
//else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
// if (i_layer < n_layer/8) new_type = GGML_TYPE_Q5_K;
//}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && !qs.has_imatrix) {
new_type = i_layer < n_layer/8 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = i_layer < n_layer/16 ? GGML_TYPE_Q5_K
: arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
@ -10816,13 +11029,15 @@ struct llama_context * llama_new_context_with_model(
}
#elif defined(GGML_USE_VULKAN)
if (model->n_gpu_layers > 0) {
ggml_backend_t backend = ggml_backend_vk_init();
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__);
llama_free(ctx);
return nullptr;
for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
ggml_backend_t backend = ggml_backend_vk_init(device);
if (backend == nullptr) {
LLAMA_LOG_ERROR("%s: failed to initialize Vulkan%d backend\n", __func__, device);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
}
ctx->backends.push_back(backend);
}
#elif defined(GGML_USE_SYCL)
if (model->n_gpu_layers > 0) {