mirror of
https://github.com/ggerganov/whisper.cpp.git
synced 2025-06-14 12:58:10 +00:00
talk-llama : sync llama.cpp
ggml-ci
This commit is contained in:
Georgi Gerganov
@ -463,11 +463,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
||||
GGML_ASSERT(hparams.n_expert_used == 0);
|
||||
}
|
||||
|
||||
// zero-out the array hparams
|
||||
std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0);
|
||||
std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
|
||||
std::fill(hparams.n_ff_arr.begin(), hparams.n_ff_arr.end(), 0);
|
||||
|
||||
std::fill(hparams.rope_sections.begin(), hparams.rope_sections.end(), 0);
|
||||
|
||||
std::fill(hparams.swa_layers.begin(), hparams.swa_layers.end(), 0);
|
||||
|
||||
ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer, false);
|
||||
ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer, false);
|
||||
|
||||
@ -571,9 +574,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
||||
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
|
||||
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
|
||||
ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP, hparams.n_moe_layer_step);
|
||||
hparams.n_swa_pattern = 4; // pattern: 3 chunked - 1 full
|
||||
hparams.n_attn_chunk = 8192; // should this be a gguf kv? currently it's the same for Scout and Maverick
|
||||
hparams.n_swa = 1; // TODO @ngxson : this is added to trigger the SWA branch (we store the chunked attn mask in the SWA tensor), will need to clean this up later
|
||||
|
||||
hparams.swa_type = LLAMA_SWA_TYPE_CHUNKED;
|
||||
hparams.n_swa = 8192; // should this be a gguf kv? currently it's the same for Scout and Maverick
|
||||
hparams.set_swa_pattern(4); // pattern: 3 chunked - 1 full
|
||||
|
||||
switch (hparams.n_expert) {
|
||||
case 16: type = LLM_TYPE_17B_16E; break;
|
||||
@ -852,22 +856,17 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
||||
default: type = LLM_TYPE_UNKNOWN;
|
||||
}
|
||||
|
||||
// for backward compatibility ; see: https://github.com/ggerganov/llama.cpp/pull/8931
|
||||
if ((hparams.n_layer == 32 || hparams.n_layer == 40) && hparams.n_ctx_train == 4096) {
|
||||
// default value for Phi-3-mini-4k-instruct and Phi-3-medium-4k-instruct
|
||||
hparams.n_swa = 2047;
|
||||
} else if (hparams.n_layer == 32 && hparams.n_head_kv(0) == 32 && hparams.n_ctx_train == 131072) {
|
||||
// default value for Phi-3-mini-128k-instruct
|
||||
// note: this seems incorrect because the window is bigger than the train context?
|
||||
hparams.n_swa = 262144;
|
||||
} else if (hparams.n_layer == 40 && hparams.n_ctx_train == 131072) {
|
||||
// default value for Phi-3-medium-128k-instruct
|
||||
// note: this seems incorrect because the window is equal to the train context?
|
||||
hparams.n_swa = 131072;
|
||||
}
|
||||
bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
|
||||
if (!found_swa && hparams.n_swa == 0) {
|
||||
throw std::runtime_error("invalid value for sliding_window");
|
||||
const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
|
||||
|
||||
if (found_swa && hparams.n_swa > 0) {
|
||||
LLAMA_LOG_WARN("%s: Phi SWA is currently disabled - results might be suboptimal for some models (see %s)\n",
|
||||
__func__, "https://github.com/ggml-org/llama.cpp/pull/13676");
|
||||
|
||||
// TODO: fix conversion scripts to correctly populate `n_swa` and `n_swa_pattern`
|
||||
hparams.swa_type = LLAMA_SWA_TYPE_NONE;
|
||||
|
||||
hparams.n_swa = 0;
|
||||
hparams.set_swa_pattern(1);
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_PHIMOE:
|
||||
@ -937,8 +936,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
||||
} break;
|
||||
case LLM_ARCH_GEMMA2:
|
||||
{
|
||||
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
|
||||
hparams.n_swa = 4096; // default value of gemma 2
|
||||
hparams.n_swa_pattern = 2;
|
||||
hparams.set_swa_pattern(2);
|
||||
hparams.attn_soft_cap = true;
|
||||
|
||||
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
|
||||
@ -955,7 +955,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
||||
} break;
|
||||
case LLM_ARCH_GEMMA3:
|
||||
{
|
||||
hparams.n_swa_pattern = 6;
|
||||
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
|
||||
hparams.set_swa_pattern(6);
|
||||
|
||||
hparams.rope_freq_base_train_swa = 10000.0f;
|
||||
hparams.rope_freq_scale_train_swa = 1.0f;
|
||||
@ -1039,7 +1040,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
||||
} break;
|
||||
case LLM_ARCH_COHERE2:
|
||||
{
|
||||
hparams.n_swa_pattern = 4;
|
||||
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
|
||||
hparams.set_swa_pattern(4);
|
||||
|
||||
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
|
||||
ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
|
||||
@ -2487,7 +2489,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
|
||||
|
||||
// output
|
||||
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
|
||||
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
|
||||
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
|
||||
// if output is NULL, init from the input tok embed
|
||||
if (output == NULL) {
|
||||
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
|
||||
}
|
||||
|
||||
for (int i = 0; i < n_layer; ++i) {
|
||||
auto & layer = layers[i];
|
||||
@ -4321,7 +4327,7 @@ void llama_model::print_info() const {
|
||||
LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
|
||||
LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot);
|
||||
LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa);
|
||||
LLAMA_LOG_INFO("%s: n_swa_pattern = %u\n", __func__, hparams.n_swa_pattern);
|
||||
LLAMA_LOG_INFO("%s: is_swa_any = %u\n", __func__, hparams.is_swa_any());
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k);
|
||||
LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v);
|
||||
LLAMA_LOG_INFO("%s: n_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il); }, hparams.n_layer).c_str());
|
||||
@ -4489,7 +4495,17 @@ const ggml_tensor * llama_model::get_tensor(const char * name) const {
|
||||
return it->second;
|
||||
}
|
||||
|
||||
ggml_tensor * llama_model::get_rope_factors(uint32_t n_ctx_per_seq, int il) const {
|
||||
float llama_model::get_rope_freq_base (const llama_cparams & cparams, int il) const {
|
||||
return hparams.is_swa(il) ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base;
|
||||
}
|
||||
|
||||
float llama_model::get_rope_freq_scale(const llama_cparams & cparams, int il) const {
|
||||
return hparams.is_swa(il) ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale;
|
||||
}
|
||||
|
||||
ggml_tensor * llama_model::get_rope_factors(const llama_cparams & cparams, int il) const {
|
||||
const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
|
||||
|
||||
// choose long/short freq factors based on the context size
|
||||
if (layers[il].rope_freqs != nullptr) {
|
||||
return layers[il].rope_freqs;
|
||||
@ -4517,22 +4533,13 @@ struct llm_build_llama : public llm_graph_context {
|
||||
// inp_pos - contains the positions
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
|
||||
// temperature tuning
|
||||
ggml_tensor * inp_attn_scale = nullptr;
|
||||
if (arch == LLM_ARCH_LLAMA4) {
|
||||
inp_attn_scale = build_inp_attn_scale();
|
||||
}
|
||||
|
||||
auto * inp_attn = build_attn_inp_kv_unified();
|
||||
|
||||
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) {
|
||||
ggml_tensor * inpSA = inpL;
|
||||
|
||||
bool use_rope = arch == LLM_ARCH_LLAMA4
|
||||
? (il + 1) % hparams.n_no_rope_layer_step != 0
|
||||
: true;
|
||||
|
||||
// norm
|
||||
cur = build_norm(inpL,
|
||||
model.layers[il].attn_norm, NULL,
|
||||
@ -4542,7 +4549,169 @@ struct llm_build_llama : public llm_graph_context {
|
||||
// self-attention
|
||||
{
|
||||
// rope freq factors for llama3; may return nullptr for llama2 and other models
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(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);
|
||||
}
|
||||
|
||||
ggml_tensor * Kcur = build_lora_mm(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);
|
||||
}
|
||||
|
||||
ggml_tensor * Vcur = build_lora_mm(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_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
|
||||
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
|
||||
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
|
||||
|
||||
Qcur = ggml_rope_ext(
|
||||
ctx0, Qcur, inp_pos, rope_factors,
|
||||
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow
|
||||
);
|
||||
|
||||
Kcur = ggml_rope_ext(
|
||||
ctx0, Kcur, 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);
|
||||
cb(Kcur, "Kcur", il);
|
||||
cb(Vcur, "Vcur", il);
|
||||
|
||||
cur = build_attn(inp_attn, gf,
|
||||
model.layers[il].wo, model.layers[il].bo,
|
||||
Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
|
||||
cb(cur, "attn_out", il);
|
||||
}
|
||||
|
||||
if (il == n_layer - 1) {
|
||||
// skip computing output for unused tokens
|
||||
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);
|
||||
}
|
||||
|
||||
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
||||
cb(ffn_inp, "ffn_inp", il);
|
||||
|
||||
// feed-forward network (non-MoE)
|
||||
if (model.layers[il].ffn_gate_inp == nullptr) {
|
||||
|
||||
cur = build_norm(ffn_inp,
|
||||
model.layers[il].ffn_norm, NULL,
|
||||
LLM_NORM_RMS, il);
|
||||
cb(cur, "ffn_norm", il);
|
||||
|
||||
cur = build_ffn(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, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
} else {
|
||||
// MoE branch
|
||||
cur = build_norm(ffn_inp,
|
||||
model.layers[il].ffn_norm, NULL,
|
||||
LLM_NORM_RMS, il);
|
||||
cb(cur, "ffn_norm", il);
|
||||
|
||||
cur = build_moe_ffn(cur,
|
||||
model.layers[il].ffn_gate_inp,
|
||||
model.layers[il].ffn_up_exps,
|
||||
model.layers[il].ffn_gate_exps,
|
||||
model.layers[il].ffn_down_exps,
|
||||
nullptr,
|
||||
n_expert, n_expert_used,
|
||||
LLM_FFN_SILU, true,
|
||||
false, 0.0,
|
||||
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
|
||||
il);
|
||||
cb(cur, "ffn_moe_out", il);
|
||||
}
|
||||
|
||||
cur = ggml_add(ctx0, cur, ffn_inp);
|
||||
cb(cur, "ffn_out", il);
|
||||
|
||||
cur = build_cvec(cur, il);
|
||||
cb(cur, "l_out", il);
|
||||
|
||||
// input for next layer
|
||||
inpL = cur;
|
||||
}
|
||||
|
||||
cur = inpL;
|
||||
|
||||
cur = build_norm(cur,
|
||||
model.output_norm, NULL,
|
||||
LLM_NORM_RMS, -1);
|
||||
|
||||
cb(cur, "result_norm", -1);
|
||||
res->t_embd = cur;
|
||||
|
||||
// lm_head
|
||||
cur = build_lora_mm(model.output, cur);
|
||||
|
||||
cb(cur, "result_output", -1);
|
||||
res->t_logits = cur;
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
}
|
||||
};
|
||||
|
||||
struct llm_build_llama_iswa : public llm_graph_context {
|
||||
llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
|
||||
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);
|
||||
|
||||
ggml_tensor * cur;
|
||||
ggml_tensor * inpL;
|
||||
|
||||
inpL = build_inp_embd(model.tok_embd);
|
||||
|
||||
// inp_pos - contains the positions
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
|
||||
// temperature tuning
|
||||
ggml_tensor * inp_attn_scale = nullptr;
|
||||
inp_attn_scale = build_inp_attn_scale();
|
||||
|
||||
auto * inp_attn = build_attn_inp_kv_unified_iswa();
|
||||
|
||||
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) {
|
||||
ggml_tensor * inpSA = inpL;
|
||||
|
||||
const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
|
||||
|
||||
// norm
|
||||
cur = build_norm(inpL,
|
||||
model.layers[il].attn_norm, NULL,
|
||||
LLM_NORM_RMS, il);
|
||||
cb(cur, "attn_norm", il);
|
||||
|
||||
// self-attention
|
||||
{
|
||||
// rope freq factors for llama3; may return nullptr for llama2 and other models
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
@ -4590,7 +4759,7 @@ struct llm_build_llama : public llm_graph_context {
|
||||
cb(Kcur, "Kcur", il);
|
||||
cb(Vcur, "Vcur", il);
|
||||
|
||||
if (arch == LLM_ARCH_LLAMA4 && use_rope && hparams.use_kq_norm) {
|
||||
if (use_rope && hparams.use_kq_norm) {
|
||||
// Llama4TextL2Norm
|
||||
Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
|
||||
Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
|
||||
@ -4616,7 +4785,6 @@ struct llm_build_llama : public llm_graph_context {
|
||||
|
||||
// feed-forward network (non-MoE)
|
||||
if (model.layers[il].ffn_gate_inp == nullptr) {
|
||||
|
||||
cur = build_norm(ffn_inp,
|
||||
model.layers[il].ffn_norm, NULL,
|
||||
LLM_NORM_RMS, il);
|
||||
@ -4629,9 +4797,7 @@ struct llm_build_llama : public llm_graph_context {
|
||||
NULL,
|
||||
LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
|
||||
} else if (arch == LLM_ARCH_LLAMA4) {
|
||||
// llama4 MoE
|
||||
} else {
|
||||
ggml_tensor * ffn_inp_normed = build_norm(ffn_inp,
|
||||
model.layers[il].ffn_norm, NULL,
|
||||
LLM_NORM_RMS, il);
|
||||
@ -4660,26 +4826,6 @@ struct llm_build_llama : public llm_graph_context {
|
||||
|
||||
cur = ggml_add(ctx0, moe_out, shexp_out);
|
||||
cb(cur, "ffn_moe_out_merged", il);
|
||||
|
||||
} else {
|
||||
// MoE branch
|
||||
cur = build_norm(ffn_inp,
|
||||
model.layers[il].ffn_norm, NULL,
|
||||
LLM_NORM_RMS, il);
|
||||
cb(cur, "ffn_norm", il);
|
||||
|
||||
cur = build_moe_ffn(cur,
|
||||
model.layers[il].ffn_gate_inp,
|
||||
model.layers[il].ffn_up_exps,
|
||||
model.layers[il].ffn_gate_exps,
|
||||
model.layers[il].ffn_down_exps,
|
||||
nullptr,
|
||||
n_expert, n_expert_used,
|
||||
LLM_FFN_SILU, true,
|
||||
false, 0.0,
|
||||
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
|
||||
il);
|
||||
cb(cur, "ffn_moe_out", il);
|
||||
}
|
||||
|
||||
cur = ggml_add(ctx0, cur, ffn_inp);
|
||||
@ -4753,7 +4899,7 @@ struct llm_build_deci : public llm_graph_context {
|
||||
} else if (n_head > 0) {
|
||||
// self-attention
|
||||
// rope freq factors for llama3; may return nullptr for llama2 and other models
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
@ -7202,6 +7348,7 @@ struct llm_build_phi2 : public llm_graph_context {
|
||||
}
|
||||
};
|
||||
|
||||
template<bool iswa>
|
||||
struct llm_build_phi3 : public llm_graph_context {
|
||||
llm_build_phi3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
|
||||
const int64_t n_embd_head = hparams.n_embd_head_v;
|
||||
@ -7217,7 +7364,14 @@ struct llm_build_phi3 : public llm_graph_context {
|
||||
// inp_pos - contains the positions
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
|
||||
auto * inp_attn = build_attn_inp_kv_unified();
|
||||
using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
|
||||
inp_attn_type * inp_attn = nullptr;
|
||||
|
||||
if constexpr (iswa) {
|
||||
inp_attn = build_attn_inp_kv_unified_iswa();
|
||||
} else {
|
||||
inp_attn = build_attn_inp_kv_unified();
|
||||
}
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
auto * residual = inpL;
|
||||
@ -7225,7 +7379,7 @@ struct llm_build_phi3 : public llm_graph_context {
|
||||
// self-attention
|
||||
{
|
||||
// rope freq factors for 128k context
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
ggml_tensor* attn_norm_output = build_norm(inpL,
|
||||
model.layers[il].attn_norm,
|
||||
@ -7977,7 +8131,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
ggml_tensor * inpSA = inpL;
|
||||
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// norm
|
||||
cur = build_norm(inpL,
|
||||
@ -8277,8 +8431,8 @@ struct llm_build_gemma : public llm_graph_context {
|
||||
}
|
||||
};
|
||||
|
||||
struct llm_build_gemma2 : public llm_graph_context {
|
||||
llm_build_gemma2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
|
||||
struct llm_build_gemma2_iswa : public llm_graph_context {
|
||||
llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
|
||||
const int64_t n_embd_head = hparams.n_embd_head_k;
|
||||
|
||||
ggml_tensor * cur;
|
||||
@ -8292,7 +8446,7 @@ struct llm_build_gemma2 : public llm_graph_context {
|
||||
// inp_pos - contains the positions
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
|
||||
auto * inp_attn = build_attn_inp_kv_unified();
|
||||
auto * inp_attn = build_attn_inp_kv_unified_iswa();
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
// norm
|
||||
@ -8414,8 +8568,8 @@ struct llm_build_gemma2 : public llm_graph_context {
|
||||
}
|
||||
};
|
||||
|
||||
struct llm_build_gemma3 : public llm_graph_context {
|
||||
llm_build_gemma3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
|
||||
struct llm_build_gemma3_iswa : public llm_graph_context {
|
||||
llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
|
||||
const int64_t n_embd_head = hparams.n_embd_head_k;
|
||||
|
||||
ggml_tensor * cur;
|
||||
@ -8433,13 +8587,11 @@ struct llm_build_gemma3 : public llm_graph_context {
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
|
||||
// TODO: is causal == true correct? might need some changes
|
||||
auto * inp_attn = build_attn_inp_kv_unified();
|
||||
auto * inp_attn = build_attn_inp_kv_unified_iswa();
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
const bool is_swa = hparams.is_swa(il);
|
||||
|
||||
const float freq_base_l = is_swa ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base;
|
||||
const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale;
|
||||
const float freq_base_l = model.get_rope_freq_base (cparams, il);
|
||||
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
|
||||
|
||||
// norm
|
||||
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
|
||||
@ -9016,8 +9168,8 @@ struct llm_build_command_r : public llm_graph_context {
|
||||
}
|
||||
};
|
||||
|
||||
struct llm_build_cohere2 : public llm_graph_context {
|
||||
llm_build_cohere2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
|
||||
struct llm_build_cohere2_iswa : public llm_graph_context {
|
||||
llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
|
||||
const int64_t n_embd_head = hparams.n_embd_head_v;
|
||||
|
||||
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
|
||||
@ -9032,7 +9184,7 @@ struct llm_build_cohere2 : public llm_graph_context {
|
||||
// inp_pos - contains the positions
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
|
||||
auto * inp_attn = build_attn_inp_kv_unified();
|
||||
auto * inp_attn = build_attn_inp_kv_unified_iswa();
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
const bool is_swa = hparams.is_swa(il);
|
||||
@ -9045,7 +9197,7 @@ struct llm_build_cohere2 : public llm_graph_context {
|
||||
// self-attention
|
||||
{
|
||||
// rope freq factors for 128k context
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
@ -9983,7 +10135,7 @@ struct llm_build_deepseek : public llm_graph_context {
|
||||
// self-attention
|
||||
{
|
||||
// rope freq factors for llama3; may return nullptr for llama2 and other models
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
@ -11347,7 +11499,7 @@ struct llm_build_exaone : public llm_graph_context {
|
||||
// self-attention
|
||||
{
|
||||
// rope freq factors for llama3; may return nullptr for llama2 and other models
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
@ -12263,7 +12415,7 @@ struct llm_build_granite : public llm_graph_context {
|
||||
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
|
||||
|
||||
if (use_rope) {
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
Qcur = ggml_rope_ext(
|
||||
ctx0, Qcur, inp_pos, rope_factors,
|
||||
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
||||
@ -12916,7 +13068,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
|
||||
// self-attention
|
||||
{
|
||||
// rope freq factors for llama3; may return nullptr for llama2 and other models
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
|
||||
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
@ -13044,6 +13196,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
|
||||
case LLM_ARCH_JINA_BERT_V2:
|
||||
case LLM_ARCH_NOMIC_BERT:
|
||||
case LLM_ARCH_NOMIC_BERT_MOE:
|
||||
case LLM_ARCH_WAVTOKENIZER_DEC:
|
||||
{
|
||||
res = nullptr;
|
||||
} break;
|
||||
@ -13058,7 +13211,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
|
||||
GGML_TYPE_F32,
|
||||
GGML_TYPE_F32,
|
||||
cparams.offload_kqv,
|
||||
std::max((uint32_t) 1, cparams.n_seq_max));
|
||||
std::max((uint32_t) 1, cparams.n_seq_max),
|
||||
cparams.n_seq_max);
|
||||
} break;
|
||||
default:
|
||||
{
|
||||
@ -13068,14 +13222,36 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
|
||||
|
||||
LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
|
||||
|
||||
res = new llama_kv_cache_unified(
|
||||
*this,
|
||||
params.type_k,
|
||||
params.type_v,
|
||||
!cparams.flash_attn,
|
||||
cparams.offload_kqv,
|
||||
cparams.n_ctx,
|
||||
padding);
|
||||
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
|
||||
GGML_ASSERT(hparams.is_swa_any());
|
||||
|
||||
res = new llama_kv_cache_unified_iswa(
|
||||
*this,
|
||||
params.type_k,
|
||||
params.type_v,
|
||||
!cparams.flash_attn,
|
||||
cparams.offload_kqv,
|
||||
params.swa_full,
|
||||
cparams.n_ctx,
|
||||
cparams.n_seq_max,
|
||||
cparams.n_batch,
|
||||
padding);
|
||||
} else {
|
||||
GGML_ASSERT(!hparams.is_swa_any());
|
||||
|
||||
res = new llama_kv_cache_unified(
|
||||
*this,
|
||||
nullptr,
|
||||
params.type_k,
|
||||
params.type_v,
|
||||
!cparams.flash_attn,
|
||||
cparams.offload_kqv,
|
||||
cparams.n_ctx,
|
||||
cparams.n_seq_max,
|
||||
padding,
|
||||
hparams.n_swa,
|
||||
hparams.swa_type);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@ -13090,11 +13266,14 @@ llm_graph_result_ptr llama_model::build_graph(
|
||||
|
||||
switch (arch) {
|
||||
case LLM_ARCH_LLAMA:
|
||||
case LLM_ARCH_LLAMA4:
|
||||
case LLM_ARCH_MINICPM:
|
||||
{
|
||||
llm = std::make_unique<llm_build_llama>(*this, params, gf);
|
||||
} break;
|
||||
case LLM_ARCH_LLAMA4:
|
||||
{
|
||||
llm = std::make_unique<llm_build_llama_iswa>(*this, params, gf);
|
||||
} break;
|
||||
case LLM_ARCH_DECI:
|
||||
{
|
||||
llm = std::make_unique<llm_build_deci>(*this, params, gf);
|
||||
@ -13169,7 +13348,11 @@ llm_graph_result_ptr llama_model::build_graph(
|
||||
case LLM_ARCH_PHI3:
|
||||
case LLM_ARCH_PHIMOE:
|
||||
{
|
||||
llm = std::make_unique<llm_build_phi3>(*this, params, gf);
|
||||
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
|
||||
llm = std::make_unique<llm_build_phi3<true>> (*this, params, gf);
|
||||
} else {
|
||||
llm = std::make_unique<llm_build_phi3<false>>(*this, params, gf);
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_PLAMO:
|
||||
{
|
||||
@ -13201,11 +13384,11 @@ llm_graph_result_ptr llama_model::build_graph(
|
||||
} break;
|
||||
case LLM_ARCH_GEMMA2:
|
||||
{
|
||||
llm = std::make_unique<llm_build_gemma2>(*this, params, gf);
|
||||
llm = std::make_unique<llm_build_gemma2_iswa>(*this, params, gf);
|
||||
} break;
|
||||
case LLM_ARCH_GEMMA3:
|
||||
{
|
||||
llm = std::make_unique<llm_build_gemma3>(*this, params, gf);
|
||||
llm = std::make_unique<llm_build_gemma3_iswa>(*this, params, gf);
|
||||
} break;
|
||||
case LLM_ARCH_STARCODER2:
|
||||
{
|
||||
@ -13225,7 +13408,7 @@ llm_graph_result_ptr llama_model::build_graph(
|
||||
} break;
|
||||
case LLM_ARCH_COHERE2:
|
||||
{
|
||||
llm = std::make_unique<llm_build_cohere2>(*this, params, gf);
|
||||
llm = std::make_unique<llm_build_cohere2_iswa>(*this, params, gf);
|
||||
} break;
|
||||
case LLM_ARCH_DBRX:
|
||||
{
|
||||
|
||||
Reference in New Issue
Block a user