8000 ggml : group all experts in a single ggml_mul_mat_id by slaren · Pull Request #6505 · ggml-org/llama.cpp · GitHub
[go: up one dir, main page]

Skip to content

ggml : group all experts in a single ggml_mul_mat_id #6505

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 22 commits into from
Apr 18, 2024
Merged
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Merge remote-tracking branch 'origin/master' into sl/moe-rework-2
  • Loading branch information
slaren committed Apr 17, 2024
commit 42003fdc3286d173f0ac06db7e9a84f1d43163e5
199 changes: 0 additions & 199 deletions llama.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -6762,80 +6762,6 @@ struct llm_build_context {
return gf;
}

// REVIEW: will be replaced by https://github.com/ggerganov/llama.cpp/pull/6505
ggml_tensor * build_moe_ffn(ggml_tensor * cur, int32_t n_tokens, llm_ffn_op_type type_op, bool norm_w, int il) {
ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts]
cb(logits, "ffn_moe_logits", il);

ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts]
cb(probs, "ffn_moe_probs", il);

// select experts
ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok]
cb(selected_experts->src[0], "ffn_moe_argsort", il);

ggml_tensor * weights = ggml_get_rows(ctx0,
ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts);
cb(weights, "ffn_moe_weights", il);

weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok]

if (norm_w) {
ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights);
cb(weights_sum, "ffn_moe_weights_sum", il);

weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok]
cb(weights, "ffn_moe_weights_norm", il);
}

// compute expert outputs
ggml_tensor * moe_out = nullptr;

for (int i = 0; i < n_expert_used; ++i) {
ggml_tensor * cur_expert;

ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exps, selected_experts, i, cur);
cb(cur_up, "ffn_moe_up", il);

ggml_tensor * gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exps, selected_experts, i, cur);
cb(gate, "ffn_moe_gate", il);

switch (type_op) {
case LLM_FFN_SILU:
{
gate = ggml_silu(ctx0, gate);
cb(gate, "ffn_moe_silu", il);
} break;
case LLM_FFN_GELU:
{
gate = ggml_gelu(ctx0, gate);
cb(gate, "ffn_moe_gelu", il);
} break;
default:
GGML_ASSERT(false);
}

cur_expert = ggml_mul(ctx0, cur_up, gate);
cb(cur_expert, "ffn_moe_gate_par", il);

cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exps, selected_experts, i, cur_expert); // [n_tokens, n_embd]
cb(cur_expert, "ffn_moe_down", il);

cur_expert = ggml_mul(ctx0, cur_expert,
ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0]));
cb(cur_expert, "ffn_moe_weighted", il);

if (i == 0) {
moe_out = cur_expert;
} else {
moe_out = ggml_add(ctx0, moe_out, cur_expert);
cb(moe_out, "ffn_moe_out", il);
}
}

return moe_out;
}

struct ggml_cgraph * build_baichuan() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);

Expand Down Expand Up @@ -7317,131 +7243,6 @@ struct llm_build_context {

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 = ggml_mul_mat(ctx0, model.output, cur);

// Grok
// multiply logits by output_multiplier_scale of 0.5773502691896257

cur = ggml_scale(ctx0, cur, 0.5773502691896257f);

cb(cur, "result_output", -1);

ggml_build_forward_expand(gf, cur);

return gf;
}

struct ggml_cgraph * build_dbrx() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, 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;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
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();

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, cb, il);
cb(cur, "attn_norm", il);

// self-attention
{
struct ggml_tensor * Qcur = nullptr;
struct ggml_tensor * Kcur = nullptr;
struct ggml_tensor * Vcur = nullptr;

cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);

cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
cb(cur, "wqkv_clamped", il);

Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));

cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);

Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);

Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);

cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, 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();
n_tokens = n_outputs;
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}

struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);

// feed-forward network
// MoE branch
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].attn_out_norm, NULL,
LLM_NORM, cb, il);
cb(cur, "attn_out_norm", il);

cur = build_moe_ffn(cur, n_tokens, LLM_FFN_SILU, true, il);

cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", il);

ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
if (layer_dir != nullptr) {
cur = ggml_add(ctx0, cur, layer_dir);
}
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, cb, -1);
Expand Down
You are viewing a condensed version of this merge commit. You can view the full changes here.
0