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

llama: Support MiniCPM-1B (with & w/o longrope) (#10559)

Browse source
This commit is contained in:
JFLFY2255 2024-12-04 17:42:50 +08:00 • committed by GitHub
parent 2759916d86
commit 8d0cfd554a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 61 additions and 183 deletions
Download patch file Download diff file Expand all files Collapse all files

175
src/llama.cpp
View file

@ -1036,6 +1036,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
{ LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" },
{ LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
@ -1683,9 +1685,10 @@ struct LLM_TN {
//
static const std::map<llama_rope_scaling_type, const char *> LLAMA_ROPE_SCALING_TYPES = {
{ LLAMA_ROPE_SCALING_TYPE_NONE, "none" },
{ LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" },
{ LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" },
{ LLAMA_ROPE_SCALING_TYPE_NONE, "none" },
{ LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" },
{ LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" },
{ LLAMA_ROPE_SCALING_TYPE_LONGROPE, "longrope" },
};
static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) {
@ -5580,8 +5583,12 @@ static void llm_load_hparams(
case LLM_ARCH_MINICPM:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale);
ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale);
ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
switch (hparams.n_layer) {
case 52: model.type = e_model::MODEL_1B; break;
case 40: model.type = e_model::MODEL_2B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
@ -7065,7 +7072,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
}
if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
if (model.arch == LLM_ARCH_MINICPM || model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale);
LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@ -7690,7 +7697,13 @@ static bool llm_load_tensors(
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
}
else {
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
}
if (n_expert == 0) {
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
@ -13497,153 +13510,6 @@ 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_model_max_nodes(model), 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, lctx, hparams, ubatch, model.tok_embd, cb);
// 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 = build_inp_pos();
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
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 = llm_build_lora_mm(lctx, 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 = llm_build_lora_mm(lctx, 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 = llm_build_lora_mm(lctx, 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_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, lctx, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
// skip computing output for unused tokens
struct ggml_tensor * inp_out_ids = build_inp_out_ids();
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
// 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, lctx, cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, 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);
cur = lctx.cvec.apply_to(ctx0, cur, il);
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 = llm_build_lora_mm(lctx, ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_minicpm3() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
@ -16742,6 +16608,7 @@ static struct ggml_cgraph * llama_build_graph(
switch (model.arch) {
case LLM_ARCH_LLAMA:
case LLM_ARCH_MINICPM:
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
{
@ -16825,10 +16692,6 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_internlm2();
} break;
case LLM_ARCH_MINICPM:
{
result = llm.build_minicpm();
} break;
case LLM_ARCH_MINICPM3:
{
result = llm.build_minicpm3();