ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-22 22:24:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support for Llama-3-Nemotron models (#377)

Browse Source 
* conflict resolution

* Changes to make work and add longrope support

* Changes to n_attention_wv rule

* Untested support of 253B

* DeciLMCausalModel now reads rope_theta from config.json properly

* Remove errant Granite mentions

* Better n_attention_vw rule

* Update vocab.py

---------

Co-authored-by: Yee Man Chan <ymchan@gmail.com>
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
saood06
2025-05-09 02:09:59 -05:00
committed by GitHub
parent 828758ec0d
commit 87bfad8437
6 changed files with 523 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

288
src/llama.cpp
View File

@@ -185,6 +185,7 @@ static std::string format(const char * fmt, ...) {
enum llm_arch {
LLM_ARCH_LLAMA,
LLM_ARCH_LLAMA4,
LLM_ARCH_DECI,
LLM_ARCH_FALCON,
LLM_ARCH_BAICHUAN,
LLM_ARCH_GROK,
@@ -240,6 +241,7 @@ enum llm_arch {
static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_LLAMA, "llama" },
{ LLM_ARCH_LLAMA4, "llama4" },
{ LLM_ARCH_DECI, "deci" },
{ LLM_ARCH_FALCON, "falcon" },
{ LLM_ARCH_GROK, "grok" },
{ LLM_ARCH_GPT2, "gpt2" },
@@ -282,7 +284,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_GLM4, "glm4" },
{ LLM_ARCH_BITNET, "bitnet" },
{ LLM_ARCH_BITNET_25, "bitnet-25" },
{ LLM_ARCH_BITNET_B158, "bitnet-b1.58" },
{ LLM_ARCH_BITNET_B158, "bitnet-b1.58" },
{ LLM_ARCH_T5, "t5" },
{ LLM_ARCH_T5ENCODER, "t5encoder" },
{ LLM_ARCH_JAIS, "jais" },
@@ -632,6 +634,32 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
},
},
{
LLM_ARCH_DECI,
{
{ 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_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
},
},
{
LLM_ARCH_LLAMA4,
{
@@ -2525,6 +2553,7 @@ enum e_model {
MODEL_70B,
MODEL_236B,
MODEL_314B,
MODEL_405B,
MODEL_671B,
MODEL_SMALL,
MODEL_MEDIUM,
@@ -5128,6 +5157,7 @@ static const char * llama_model_type_name(e_model type) {
case MODEL_70B: return "70B";
case MODEL_236B: return "236B";
case MODEL_314B: return "314B";
case MODEL_405B: return "405B";
case MODEL_671B: return "671B";
case MODEL_SMALL: return "0.1B";
case MODEL_MEDIUM: return "0.4B";
@@ -5265,7 +5295,7 @@ static void llm_load_hparams(
ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON || model.arch == LLM_ARCH_BITNET_25 || model.arch == LLM_ARCH_BITNET_B158) {
if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON || model.arch == LLM_ARCH_BITNET_25 || model.arch == LLM_ARCH_BITNET_B158 || model.arch == LLM_ARCH_DECI) {
if (hparams.n_rot != hparams.n_embd_head_k) {
throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
}
@@ -5320,6 +5350,16 @@ static void llm_load_hparams(
if (model.type == MODEL_17B_128E) {
hparams.use_kq_norm = false;
}
} break;
case LLM_ARCH_DECI:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 32: model.type = e_model::MODEL_7B; break;
case 80: model.type = e_model::MODEL_70B; break;
case 162: model.type = e_model::MODEL_405B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_MINICPM:
@@ -6956,6 +6996,76 @@ static bool llm_load_tensors(
}
}
} break;
case LLM_ARCH_DECI:
{
model.tok_embd = create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
// output
model.output_norm = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (model.output == NULL) {
model.output = create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
}
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];
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);
const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);
const int64_t n_embd_gqa = hparams.n_embd_v_gqa(i);
const int64_t n_ff = hparams.n_ff(i);
const int64_t n_head = hparams.n_head(i);
const int64_t n_head_kv = hparams.n_head_kv(i);
if (n_head_kv == 0 && n_head > 0) {
// linear attention for DeciLMCausalModel
layer.attn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.wo = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
}
else if (n_head_kv > 0) {
layer.attn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.wq = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head});
layer.wk = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa});
layer.wv = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa});
layer.wo = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
}
// optional bias tensors
layer.bq = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
layer.bk = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
layer.bv = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
layer.bo = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
if (n_ff > 0) {
layer.ffn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
}
if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
layer.rope_long = create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight"), { n_rot/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
layer.rope_short = create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_rot/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
}
else {
layer.rope_freqs = create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
}
if (n_ff > 0) {
layer.ffn_gate = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
layer.ffn_down = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
layer.ffn_up = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
}
// optional MLP bias
layer.ffn_gate_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
layer.ffn_down_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
layer.ffn_up_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
}
} break;
case LLM_ARCH_LLAMA4:
{
model.tok_embd = create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -10485,6 +10595,168 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_deci() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
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, lctx, hparams, batch, model.tok_embd, cb);
// 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();
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
const int64_t n_head_kv = hparams.n_head_kv(il);
const int64_t n_head = hparams.n_head(il);
const int64_t n_ff = hparams.n_ff(il);
if (n_head == 0) { // attention-free layer of Llama-3_1-Nemotron-51B
cur = inpL;
} else {
// norm
cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
}
if (n_head > 0 && n_head_kv == 0) { // "linear attention" of Llama-3_1-Nemotron-51B
cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur);
cb(cur, "wo", il);
} else if (n_head > 0) {
// self-attention
// rope freq factors for llama3; may return nullptr for llama2 and other models
struct ggml_tensor * rope_factors = build_rope_factors(il);
// 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, rope_factors,
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, rope_factors,
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, kq_scale, cb, il);
}
if (il == n_layer - 1) {
// skip computing output for unused tokens
struct ggml_tensor * inp_out_ids = build_inp_out_ids();
n_tokens = n_outputs;
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
// FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
if (n_ff == 0) {
continue;
}
if (hparams.f_residual_scale) {
cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
}
// modified to support attention-free layer of Llama-3_1-Nemotron-51B
struct ggml_tensor * ffn_inp = cur;
if (n_head > 0) {
ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
}
// feed-forward network
if (model.layers[il].ffn_gate_inp == nullptr) {
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, model.layers[il].ffn_up_b, NULL,
model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
cb(cur, "ffn_out", il);
}
if (hparams.f_residual_scale) {
cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
}
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", il);
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
cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
if (hparams.f_logit_scale) {
cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
}
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_baichuan() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
@@ -16477,6 +16749,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_llama();
} break;
case LLM_ARCH_DECI:
{
result = llm.build_deci();
} break;
case LLM_ARCH_BAICHUAN:
{
result = llm.build_baichuan();
@@ -18997,8 +19273,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// - qs.n_attention_wv == 0 for Mamba models
// - qs.n_attention_wv == model.hparams.n_layer for Transformer models
// - qs.n_attention_wv == 3 * model.hparams.n_layer for Encoder-Decoder models
// - model.arch == LLM_ARCH_DECI for Deci-Nemotron models
//
GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer || qs.n_attention_wv == 3 * (int)model.hparams.n_layer) && "n_attention_wv is unexpected");
GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer || qs.n_attention_wv == 3 * (int)model.hparams.n_layer || model.arch == LLM_ARCH_DECI) && "n_attention_wv is unexpected");
size_t total_size_org = 0;
size_t total_size_new = 0;
@@ -20298,6 +20575,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
// use what we call a normal RoPE, operating on pairs of consecutive head values
case LLM_ARCH_LLAMA:
case LLM_ARCH_DECI:
case LLM_ARCH_LLAMA4:
case LLM_ARCH_BAICHUAN:
case LLM_ARCH_STARCODER:
@@ -20371,6 +20649,10 @@ int32_t llama_n_layer(const struct llama_model * model) {
return model->hparams.n_layer;
}
int32_t llama_n_head(const struct llama_model * model) {
return model->hparams.n_head();
}
float llama_rope_freq_scale_train(const struct llama_model * model) {
return model->hparams.rope_freq_scale_train;
}