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:
#	Makefile
#	README.md
#	docs/token_generation_performance_tips.md
#	grammars/README.md
#	scripts/sync-ggml.sh
#	tests/CMakeLists.txt
#	tests/test-grad0.cpp
#	tests/test-opt.cpp
This commit is contained in:
Concedo 2023-11-15 16:59:53 +08:00
commit 35a97e14b2
49 changed files with 4972 additions and 3131 deletions
Download patch file Download diff file Expand all files Collapse all files

331
llama.cpp
View file

@ -92,6 +92,8 @@
#define LLAMA_ATTRIBUTE_FORMAT(...)
#endif
#define LLAMA_MAX_NODES 4096
//
// logging
//
@ -191,6 +193,7 @@ enum llm_arch {
LLM_ARCH_PERSIMMON,
LLM_ARCH_REFACT,
LLM_ARCH_BLOOM,
LLM_ARCH_STABLELM,
LLM_ARCH_UNKNOWN,
};
@ -206,6 +209,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
{ LLM_ARCH_PERSIMMON, "persimmon" },
{ LLM_ARCH_REFACT, "refact" },
{ LLM_ARCH_BLOOM, "bloom" },
{ LLM_ARCH_STABLELM, "stablelm" },
};
enum llm_kv {
@ -494,6 +498,25 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
},
},
{
LLM_ARCH_STABLELM,
{
{ 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_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,
{
@ -2225,6 +2248,16 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_STABLELM:
{
GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
switch (hparams.n_layer) {
case 32: model.type = e_model::MODEL_3B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
default: (void)0;
}
@ -2907,6 +2940,13 @@ static void llm_load_tensors(
ggml_backend_type backend_output;
if (n_gpu_layers > int(n_layer)) {
#ifdef GGML_USE_CUBLAS
if (n_gpu_layers > int(n_layer + 1)) {
LLAMA_LOG_ERROR("%s: CUDA backend missing Persimmon CUDA ops, can offload at most %ld layers. See: https://github.com/ggerganov/llama.cpp/issues/4038\n",
__func__, n_layer + 1);
throw std::runtime_error("Persimmon CUDA offload failed");
}
#endif
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying
// on Windows however this is detrimental unless everything is on the GPU
#ifndef _WIN32
@ -3108,6 +3148,81 @@ static void llm_load_tensors(
}
}
} break;
case LLM_ARCH_STABLELM:
{
model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
// output
{
ggml_backend_type backend_norm;
ggml_backend_type backend_output;
if (n_gpu_layers > int(n_layer)) {
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying
// on Windows however this is detrimental unless everything is on the GPU
#ifndef _WIN32
backend_norm = llama_backend_offload;
#else
backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
#endif // _WIN32
backend_output = llama_backend_offload_split;
} else {
backend_norm = GGML_BACKEND_CPU;
backend_output = GGML_BACKEND_CPU;
}
model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm);
model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm);
model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output);
if (backend_norm == GGML_BACKEND_GPU) {
vram_weights += ggml_nbytes(model.output_norm);
}
if (backend_output == GGML_BACKEND_GPU_SPLIT) {
vram_weights += ggml_nbytes(model.output);
}
}
const uint32_t n_ff = hparams.n_ff;
const int i_gpu_start = n_layer - n_gpu_layers;
model.layers.resize(n_layer);
for (uint32_t i = 0; i < n_layer; ++i) {
/*
llama_model_loader: - tensor 4: blk.0.attn_output.weight f16 [ 2560, 2560, 1, 1 ]
*/
const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT
const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend);
layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split);
layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split);
layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split);
layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);
layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend);
layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split);
layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split);
layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split);
if (backend == GGML_BACKEND_GPU) {
vram_weights +=
ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) +
ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) +
ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
}
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
@ -3641,7 +3756,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_llama() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
GGML_ASSERT(n_embd_head == hparams.n_rot);
@ -3753,7 +3868,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_baichuan() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -3873,7 +3988,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_falcon() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -3995,7 +4110,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_starcoder() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * pos;
@ -4094,7 +4209,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_persimmon() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_rot = n_embd_head / 2;
@ -4304,7 +4419,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_refact() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -4395,7 +4510,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_bloom() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -4489,7 +4604,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_mpt() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -4586,6 +4701,177 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_stablelm() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
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_scale
struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
cb(KQ_scale, "KQ_scale", -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, hparams.n_rot, 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,
model.layers[il].attn_norm_b,
LLM_NORM, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and RoPE them
struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(tmpq, "tmpq", il);
struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
cb(tmpk, "tmpk", il);
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
// RoPE the first n_rot of q/k, pass the other half, and concat.
struct ggml_tensor * qrot = ggml_cont(ctx0, ggml_view_3d(
ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
ggml_element_size(tmpq) * n_embd_head,
ggml_element_size(tmpq) * n_embd_head * n_head,
0
));
cb(qrot, "qrot", il);
struct ggml_tensor * krot = ggml_cont(ctx0, ggml_view_3d(
ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
ggml_element_size(tmpk) * n_embd_head,
ggml_element_size(tmpk) * n_embd_head * n_head_kv,
0
));
cb(krot, "krot", il);
// get the second half of tmpq, e.g tmpq[n_rot:, :, :]
struct ggml_tensor * qpass = ggml_view_3d(
ctx0, tmpq, (n_embd_head - hparams.n_rot), n_head, n_tokens,
ggml_element_size(tmpq) * n_embd_head,
ggml_element_size(tmpq) * n_embd_head * n_head,
ggml_element_size(tmpq) * hparams.n_rot
);
cb(qpass, "qpass", il);
struct ggml_tensor * kpass = ggml_view_3d(
ctx0, tmpk, (n_embd_head - hparams.n_rot), n_head_kv, n_tokens,
ggml_element_size(tmpk) * (n_embd_head),
ggml_element_size(tmpk) * (n_embd_head) * n_head_kv,
ggml_element_size(tmpk) * hparams.n_rot
);
cb(kpass, "kpass", il);
struct ggml_tensor * qrotated = ggml_rope_custom(
ctx0, qrot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(qrotated, "qrotated", il);
struct ggml_tensor * krotated = ggml_rope_custom(
ctx0, krot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(krotated, "krotated", il);
// ggml currently only supports concatenation on dim=2
// so we need to permute qrot, qpass, concat, then permute back.
qrotated = ggml_cont(ctx0, ggml_permute(ctx0, qrotated, 2, 1, 0, 3));
cb(qrotated, "qrotated", il);
krotated = ggml_cont(ctx0, ggml_permute(ctx0, krotated, 2, 1, 0, 3));
cb(krotated, "krotated", il);
qpass = ggml_cont(ctx0, ggml_permute(ctx0, qpass, 2, 1, 0, 3));
cb(qpass, "qpass", il);
kpass = ggml_cont(ctx0, ggml_permute(ctx0, kpass, 2, 1, 0, 3));
cb(kpass, "kpass", il);
struct ggml_tensor * Qcur = ggml_concat(ctx0, qrotated, qpass);
cb(Qcur, "Qcur", il);
struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass);
cb(Kcur, "Kcur", il);
struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 2, 1, 0, 3));
cb(Q, "Q", il);
Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3));
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, hparams, kv_self,
model.layers[il].wo, NULL,
Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 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,
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, NULL,
model.layers[il].ffn_gate, NULL,
model.layers[il].ffn_down, 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,
model.output_norm_b,
LLM_NORM, 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;
}
};
//
@ -5055,6 +5341,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_mpt();
} break;
case LLM_ARCH_STABLELM:
{
result = llm.build_stablelm();
} break;
default:
GGML_ASSERT(false);
}
@ -5232,7 +5522,8 @@ static int llama_decode_internal(
model.arch == LLM_ARCH_FALCON ||
model.arch == LLM_ARCH_REFACT ||
model.arch == LLM_ARCH_MPT ||
model.arch == LLM_ARCH_STARCODER;
model.arch == LLM_ARCH_STARCODER ||
model.arch == LLM_ARCH_STABLELM;
const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3;
if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) {
@ -8467,7 +8758,7 @@ struct llama_context * llama_new_context_with_model(
{
static const size_t tensor_alignment = 32;
// the compute buffer is used to store the tensor and graph structs, while the allocator buffer is used for the tensor data
ctx->buf_compute.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead());
ctx->buf_compute.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead());
// create measure allocator
ctx->alloc = ggml_allocr_new_measure(tensor_alignment);
@ -8856,8 +9147,8 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
if (kv_buf_size) {
const size_t elt_size = ggml_element_size(kv_self.k);
ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
ggml_cgraph gf{};
ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true });
ggml_cgraph * gf = ggml_new_graph(cpy_ctx);
ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0);
@ -8875,9 +9166,9 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
kv_head, n_embd, n_layer,
elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d));
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k3d, kout3d));
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v3d, vout3d));
ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);
ggml_free(cpy_ctx);
@ -8984,8 +9275,8 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
const size_t elt_size = ggml_element_size(kv_self.k);
ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
ggml_cgraph gf{};
ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true });
ggml_cgraph * gf = ggml_new_graph(cpy_ctx);
ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
kin3d->data = (void *) inp;
@ -9003,9 +9294,9 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
kv_head, n_embd, n_layer,
elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, vin3d, v3d));
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin3d, k3d));
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin3d, v3d));
ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);
ggml_free(cpy_ctx);
}