ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-28 00:54:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

This works and TG is descent, but PP is low

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-12-12 14:59:33 +00:00
parent 5645be6cfc
commit 328c3ff5e0
5 changed files with 63 additions and 75 deletions
Download Patch File Download Diff File Expand all files Collapse all files

97
src/llama-build-context.cpp
View File

@@ -1224,10 +1224,10 @@ llm_expert_gating_func_type gating_op,
} else {
cur = routed_out;
}
if (cur->ne[1] >= 32) {
cur = ggml_cast(ctx, cur, GGML_TYPE_F16);
cb(cur, "ffn_out_f16", il_cb);
}
//if (cur->ne[1] >= 32) {
// cur = ggml_cast(ctx, cur, GGML_TYPE_F16);
// cb(cur, "ffn_out_f16", il_cb);
//}
ggml_build_forward_expand(graph, routed_out);
results.push_back(cur);
}
@@ -1743,7 +1743,8 @@ ggml_cgraph * llm_build_context::build_llama() {
// self-attention
if (use_rope) {
cur = build_std_attention(gf, inpL, inp_pos, nullptr, this_KQ_mask, nullptr, nullptr, kq_scale, hparams.f_attention_scale, this_n_swa, il);
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, nullptr,
this_KQ_mask, nullptr, nullptr, kq_scale, hparams.f_attention_scale, this_n_swa, il);
}
else {
@@ -1935,7 +1936,8 @@ ggml_cgraph * llm_build_context::build_mistral3() {
auto rope_factors = build_rope_factors(il);
cur = build_std_attention(gf, inpL, inp_pos, rope_factors, KQ_mask, nullptr, inp_attn_scale, kq_scale, hparams.f_attention_scale, 0, il);
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, rope_factors, KQ_mask,
nullptr, inp_attn_scale, kq_scale, hparams.f_attention_scale, 0, il);
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
@@ -3927,7 +3929,7 @@ ggml_cgraph * llm_build_context::build_qwen3moe() {
//cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il);
//cb(cur, "attn_norm", il);
cur = build_std_attention(gf, inpL, inp_pos, nullptr, KQ_mask, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), 0.0f, 0, il);
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, nullptr, KQ_mask, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), 0.0f, 0, il);
if (il == n_layer - 1) {
// skip computing output for unused tokens
@@ -6806,7 +6808,7 @@ ggml_cgraph * llm_build_context::build_glm4_moe() {
// self-attention
if (rope_cache == nullptr) {
cur = build_std_attention(gf, inpL, inp_pos, nullptr, KQ_mask, nullptr, nullptr, kq_scale, 0.0f, 0, il);
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL, inp_pos, nullptr, KQ_mask, nullptr, nullptr, kq_scale, 0.0f, 0, il);
} else {
// Pre-attention norm
cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il);
@@ -7223,54 +7225,28 @@ ggml_cgraph * llm_build_context::build_cohere2() {
struct ggml_tensor * ffn_inp = cur;
// self-attention
{
// rope freq factors for 128k context
struct ggml_tensor * rope_factors = build_rope_factors(il);
cur = build_std_attention(gf, nullptr, cur, inp_pos, nullptr, KQ_mask_l, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), 0.f,
is_sliding ? hparams.n_swa : 0, il, is_sliding);
auto [Qcur, Kcur, Vcur] = llm_build_mul_mat_qkv(gf, cur,
model.layers[il].wqkv, model.layers[il].bqkv,
model.layers[il].wqk, model.layers[il].bqk,
model.layers[il].wq, model.layers[il].bq,
model.layers[il].wk, model.layers[il].bk,
model.layers[il].wv, model.layers[il].bv, nullptr, nullptr, 0.f, il);
if (is_sliding) {
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);
cb(Qcur, "Qcur", il);
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(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_l, n_tokens, kv_head, n_kv, 1.0f / sqrtf(float(n_embd_head)), cb, il, nullptr,
is_sliding ? hparams.n_swa : 0);
}
cur = ggml_add(ctx0, cur, inpL);
cb(cur, "attn_out", 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);
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
}
struct ggml_tensor * attn_out = cur;
auto attn_out = cur;
// feed-forward network
{
cur = llm_build_ffn(ctx0, lctx, nullptr, ffn_inp, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate,
cur = llm_build_ffn(ctx0, lctx, nullptr, ffn_inp, 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);
}
cb, il, gf);
cb(cur, "ffn_out", il);
// add together residual + FFN + self-attention
cur = ggml_add(ctx0, cur, inpL);
cur = ggml_add(ctx0, cur, attn_out);
cur = lctx.cvec.apply_to(ctx0, cur, il);
cb(cur, "l_out", il);
@@ -7280,6 +7256,9 @@ ggml_cgraph * llm_build_context::build_cohere2() {
}
cur = inpL;
if (cur->type != GGML_TYPE_F32) {
cur = ggml_cast(ctx0, cur, GGML_TYPE_F32);
}
cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM, cb, -1);
cb(cur, "result_norm", -1);
@@ -9308,12 +9287,14 @@ ggml_cgraph * llm_build_context::llama_build_graph(
return result;
}
ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tensor * input, ggml_tensor * inp_pos, ggml_tensor * rope_factors_in,
ggml_tensor * KQ_mask, ggml_tensor * sinks, ggml_tensor * inp_attn_scale, float KQ_scale, float f_attn_scale, int n_swa, int il) {
ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tensor * the_attn_norm,
ggml_tensor * input, ggml_tensor * inp_pos, ggml_tensor * rope_factors_in,
ggml_tensor * KQ_mask, ggml_tensor * sinks, ggml_tensor * inp_attn_scale, float KQ_scale, float f_attn_scale,
int n_swa, int il, bool do_rope) {
if (!model.layers[il].wqkv && !model.layers[il].wqk && cparams.flash_attn &&
model.layers[il].wq->extra && model.layers[il].wk->extra && model.layers[il].wv->extra && model.layers[il].wo->extra) {
if (kv_self.k_l[il]->extra && kv_self.v_l[il]->extra) {
ggml_split_tensor_t * attn_norm = model.layers[il].attn_norm ? (ggml_split_tensor_t *)model.layers[il].attn_norm->extra : nullptr;
ggml_split_tensor_t * attn_norm = the_attn_norm ? (ggml_split_tensor_t *)the_attn_norm->extra : nullptr;
auto wq = (ggml_split_tensor_t *)model.layers[il].wq->extra;
auto wk = (ggml_split_tensor_t *)model.layers[il].wk->extra;
auto wv = (ggml_split_tensor_t *)model.layers[il].wv->extra;
@@ -9374,10 +9355,12 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
auto extra = (ggml_split_tensor_t *)model.layers[il].rope_freqs->extra;
rope_factors = extra->splits[id];
}
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);
if (do_rope) {
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);
cb(Kcur, "Kcur", il_cb);
if (inp_attn_scale) {
@@ -9477,9 +9460,9 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
cur = ggml_add(ctx0, cur, bo->splits[id]);
cb(cur, "kqv_wo_biased", il_cb);
}
if (cur->ne[1] >= 32) {
cur = ggml_cast(ctx0, cur, GGML_TYPE_F16);
}
//if (cur->ne[1] >= 32) {
// cur = ggml_cast(ctx0, cur, GGML_TYPE_F16);
//}
ggml_build_forward_expand(gf, cur);
attn.push_back(cur);
}
@@ -9497,8 +9480,8 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
}
auto cur = input;
if (model.layers[il].attn_norm) {
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il);
if (the_attn_norm) {
cur = llm_build_norm(ctx0, cur, hparams, the_attn_norm, NULL, LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
}
@@ -9508,10 +9491,12 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
model.layers[il].wq, model.layers[il].bq, model.layers[il].wk, model.layers[il].bk, model.layers[il].wv, model.layers[il].bv,
model.layers[il].attn_q_norm, model.layers[il].attn_k_norm, f_attn_scale, il);
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors_in, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
if (do_rope) {
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors_in, 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_in, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
Kcur = ggml_rope_ext( ctx0, Kcur, inp_pos, rope_factors_in, 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);