if (ubatch->token && n_pos_per_embd == 4) {
// in case we're using M-RoPE with text tokens, convert the 1D positions to 4D
// the 3 first dims are the same, and 4th dim is all 0
std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd);
// copy the first dimension
for (int i = 0; i < n_tokens; ++i) {
pos_data[ i] = ubatch->pos[i];
pos_data[ n_tokens + i] = ubatch->pos[i];
pos_data[2 * n_tokens + i] = ubatch->pos[i];
pos_data[3 * n_tokens + i] = 0; // 4th dim is 0
}
ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
} else {
ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_embd*ggml_element_size(pos));
}
}
}
bool llm_graph_input_pos::can_reuse(const llm_graph_params & params) {
bool res = true;
float * data = (float *) mean->data;
memset(mean->data, 0, n_tokens*n_seqs_unq*ggml_element_size(mean));
std::vector<uint64_t> sums(n_seqs_unq, 0);
for (int i = 0; i < n_tokens; i += n_seq_tokens) {
for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
const llama_seq_id seq_id = ubatch->seq_id[i][s];
const int32_t seq_idx = ubatch->seq_idx[seq_id];
sums[seq_idx] += ubatch->n_seq_tokens;
}
}
std::vector<float> div(n_seqs_unq, 0.0f);
for (int s = 0; s < n_seqs_unq; ++s) {
const uint64_t sum = sums[s];
if (sum > 0) {
div[s] = 1.0f/float(sum);
}
}
for (int i = 0; i < n_tokens; i += n_seq_tokens) {
for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
const llama_seq_id seq_id = ubatch->seq_id[i][s];
const int32_t seq_idx = ubatch->seq_idx[seq_id];
if (s_copy) {
GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
int32_t * data = (int32_t *) s_copy->data;
// assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
for (uint32_t i = 0; i < n_rs; ++i) {
data[i] = mctx->s_copy(i);
}
}
}
res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
//res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
res &= self_kq_mask->ne[0] == mctx->get_n_kv();
res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
res &= mctx->get_supports_set_rows(); // TODO: tmp
res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
//res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
res &= self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
//res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv();
res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
res &= mctx->get_base()->get_supports_set_rows(); // TODO: tmp
if (gate) {
switch (type_gate) {
case LLM_FFN_SEQ:
{
cur = build_lora_mm(gate, tmp);
cb(cur, "ffn_gate", il);
} break;
case LLM_FFN_PAR:
{
cur = build_lora_mm(gate, cur);
cb(cur, "ffn_gate", il);
} break;
}
if (gate_b) {
cur = ggml_add(ctx0, cur, gate_b);
cb(cur, "ffn_gate_b", il);
}
if (gate_s) {
cur = ggml_mul(ctx0, cur, gate_s);
cb(cur, "ffn_gate_s", il);
}
} else {
cur = tmp;
}
switch (type_op) {
case LLM_FFN_SILU:
if (gate && type_gate == LLM_FFN_PAR) {
cur = ggml_swiglu_split(ctx0, cur, tmp);
cb(cur, "ffn_swiglu", il);
type_gate = LLM_FFN_SEQ;
} else {
cur = ggml_silu(ctx0, cur);
cb(cur, "ffn_silu", il);
} break;
case LLM_FFN_GELU:
if (gate && type_gate == LLM_FFN_PAR) {
cur = ggml_geglu_split(ctx0, cur, tmp);
cb(cur, "ffn_geglu", il);
type_gate = LLM_FFN_SEQ;
} else {
cur = ggml_gelu(ctx0, cur);
cb(cur, "ffn_gelu", il);
if (act_scales != NULL) {
cur = ggml_div(ctx0, cur, act_scales);
cb(cur, "ffn_act", il);
}
} break;
case LLM_FFN_RELU:
if (gate && type_gate == LLM_FFN_PAR) {
cur = ggml_reglu_split(ctx0, cur, tmp);
cb(cur, "ffn_reglu", il);
type_gate = LLM_FFN_SEQ;
} else {
cur = ggml_relu(ctx0, cur);
cb(cur, "ffn_relu", il);
} break;
case LLM_FFN_RELU_SQR:
{
cur = ggml_relu(ctx0, cur);
cb(cur, "ffn_relu", il);
cur = ggml_sqr(ctx0, cur);
cb(cur, "ffn_sqr(relu)", il);
} break;
case LLM_FFN_SWIGLU:
{
cur = ggml_swiglu(ctx0, cur);
cb(cur, "ffn_swiglu", il);
} break;
case LLM_FFN_GEGLU:
{
cur = ggml_geglu(ctx0, cur);
cb(cur, "ffn_geglu", il);
} break;
case LLM_FFN_REGLU:
{
cur = ggml_reglu(ctx0, cur);
cb(cur, "ffn_reglu", il);
} break;
default:
GGML_ABORT("fatal error");
}
if (gate && type_gate == LLM_FFN_PAR) {
cur = ggml_mul(ctx0, cur, tmp);
cb(cur, "ffn_gate_par", il);
}
if (down) {
cur = build_lora_mm(down, cur);
if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
// GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
}
if (down_b) {
cb(cur, "ffn_down", il);
}
if (down_b) {
cur = ggml_add(ctx0, cur, down_b);
}
if (down_s) {
cur = ggml_mul(ctx0, cur, down_s);
cb(cur, "ffn_down_s", il);
}
// add experts selection bias - introduced in DeepSeek V3
// leave probs unbiased as it's later used to get expert weights
ggml_tensor * selection_probs = probs;
if (exp_probs_b != nullptr) {
selection_probs = ggml_add(ctx0, probs, exp_probs_b);
cb(selection_probs, "ffn_moe_probs_biased", il);
}
// llama4 doesn't have exp_probs_b, and sigmoid is only used after top_k
// see: https://github.com/meta-llama/llama-models/blob/699a02993512fb36936b1b0741e13c06790bcf98/models/llama4/moe.py#L183-L198
if (arch == LLM_ARCH_LLAMA4) {
selection_probs = logits;
}
// order the views before the adds
for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
cur_experts[i] = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]);
ggml_build_forward_expand(gf, cur_experts[i]);
}
// aggregate experts
// note: here we explicitly use hparams.n_expert_used instead of n_expert_used
// to avoid potentially a large number of add nodes during warmup
// ref: https://github.com/ggml-org/llama.cpp/pull/14753
ggml_tensor * moe_out = cur_experts[0];
for (uint32_t i = 1; i < hparams.n_expert_used; ++i) {
moe_out = ggml_add(ctx0, moe_out, cur_experts[i]);
}
if (hparams.n_expert_used == 1) {
// avoid returning a non-contiguous tensor
moe_out = ggml_cont(ctx0, moe_out);
}
// For Granite architecture
if (hparams.f_embedding_scale != 0.0f) {
cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale);
}
cb(cur, "inp_embd", -1);
res->add_input(std::move(inp));
return cur;
}
ggml_tensor * llm_graph_context::build_inp_pos() const {
auto inp = std::make_unique<llm_graph_input_pos>(hparams.n_pos_per_embd());
auto & cur = inp->pos;
cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, (int64_t)n_tokens*hparams.n_pos_per_embd());
ggml_set_input(cur);
res->add_input(std::move(inp));
return cur;
}
ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
auto & cur = inp->attn_scale;
// this need to be 1x1xN for broadcasting
cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens);
ggml_set_input(cur);
res->add_input(std::move(inp));
return cur;
}
ggml_tensor * llm_graph_context::build_inp_out_ids() const {
// note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls,
// but this would make the graph topology depend on the number of output tokens, which can interere with
// features that require constant topology such as pipline parallelism
// ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471
//if (n_outputs < n_tokens) {
// return nullptr;
//}
auto inp = std::make_unique<llm_graph_input_out_ids>(hparams, cparams, n_outputs);
auto & cur = inp->out_ids;
cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs);
ggml_set_input(cur);
res->add_input(std::move(inp));
return cur;
}
ggml_tensor * llm_graph_context::build_inp_mean() const {
auto inp = std::make_unique<llm_graph_input_mean>(cparams);
auto & cur = inp->mean;
cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, ubatch.n_seqs_unq);
ggml_set_input(cur);
res->add_input(std::move(inp));
return cur;
}
ggml_tensor * llm_graph_context::build_inp_cls() const {
auto inp = std::make_unique<llm_graph_input_cls>(cparams);
auto & cur = inp->cls;
cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_seqs_unq);
ggml_set_input(cur);
res->add_input(std::move(inp));
return cur;
}
ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
auto inp = std::make_unique<llm_graph_input_cross_embd>(cross);
auto & cur = inp->cross_embd;
// if we have the output embeddings from the encoder, use them directly
// TODO: needs more work to be correct, for now just use the tensor shape
//if (cross->t_embd) {
// cur = ggml_view_tensor(ctx0, cross->t_embd);
// return cur;
//}
const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
const auto n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
ggml_set_input(cur);
res->add_input(std::move(inp));
return cur;
}
ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
auto inp = std::make_unique<llm_graph_input_pos_bucket>(hparams);
auto & cur = inp->pos_bucket;
cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens);
ggml_set_input(cur);
// split the batch into streams if needed
const auto n_stream = k->ne[3];
q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream);
q = ggml_permute(ctx0, q, 0, 2, 1, 3);
k = ggml_permute(ctx0, k, 0, 2, 1, 3);
v = ggml_permute(ctx0, v, 0, 2, 1, 3);
const auto n_kv = k->ne[1];
ggml_tensor * cur;
// TODO: replace hardcoded padding with ggml-provided padding
if (cparams.flash_attn && (n_kv % 256 == 0) && kq_b == nullptr) {
GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");
if (v_trans) {
v = ggml_transpose(ctx0, v);
}
// this can happen when KV cache is not used (e.g. an embedding model with non-causal attn)
if (k->type == GGML_TYPE_F32) {
k = ggml_cast(ctx0, k, GGML_TYPE_F16);
}
if (v->type == GGML_TYPE_F32) {
v = ggml_cast(ctx0, v, GGML_TYPE_F16);
}
cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
if (v_mla) {
#if 0
// v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens.
// However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient.
cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
cur = ggml_mul_mat(ctx0, v_mla, cur);
#else
// It's preferable to do the calculation as a matrix-matrix multiplication with n_tokens in dimension 1.
// The permutations are noops and only change how the tensor data is interpreted.
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_mul_mat(ctx0, v_mla, cur);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs.
#endif
}
cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
} else {
ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
// note: this op tends to require high floating point range
// while for some models F16 is enough, for others it is not, so we default to F32 here
ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
if (arch == LLM_ARCH_GROK) {
// need to do the following:
// multiply by attn_output_multiplyer of 0.08838834764831845
// and then :
// kq = 30 * tanh(kq / 30)
// before the softmax below
kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f));
kq = ggml_scale(ctx0, kq, 30);
}
if (hparams.attn_soft_cap) {
kq = ggml_scale(ctx0, kq, 1.0f / hparams.f_attn_logit_softcapping);
kq = ggml_tanh (ctx0, kq);
kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
}
if (kq_b) {
kq = ggml_add(ctx0, kq, kq_b);
}
kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
ggml_soft_max_add_sinks(kq, sinks);
if (!v_trans) {
// note: avoid this branch
v = ggml_cont(ctx0, ggml_transpose(ctx0, v));
}
ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
// for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA
if (v_mla) {
kqv = ggml_mul_mat(ctx0, v_mla, kqv);
}
cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
// recombine streams
cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
if (!cparams.offload_kqv) {
// all nodes between the KV store and the attention output are run on the CPU
ggml_backend_sched_set_tensor_backend(sched, cur, backend_cpu);
}
}
ggml_build_forward_expand(gf, cur);
return cur;
}
llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() const {
auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
// note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
ggml_set_input(inp->kq_mask);
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(gf, q_cur);
ggml_build_forward_expand(gf, k_cur);
ggml_build_forward_expand(gf, v_cur);
const auto & kq_mask = inp->get_kq_mask();
// [TAG_NO_CACHE_PAD]
// TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
assert(!ubatch.equal_seqs());
ggml_tensor * q = q_cur;
ggml_tensor * k = k_cur;
ggml_tensor * v = v_cur;
ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_kv_unified * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * v_mla,
float kq_scale,
int il) const {
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(gf, q_cur);
ggml_build_forward_expand(gf, k_cur);
ggml_build_forward_expand(gf, v_cur);
const auto * mctx_cur = inp->mctx;
// store to KV cache
{
const auto & k_idxs = inp->get_k_idxs();
const auto & v_idxs = inp->get_v_idxs();
if (wo) {
cur = build_lora_mm(wo, cur);
if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
// GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
}
ggml_tensor * llm_graph_context::build_attn_with_sinks(
llm_graph_input_attn_kv_unified_iswa * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * v_mla,
ggml_tensor * sinks,
float kq_scale,
int il) const {
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(gf, q_cur);
if (k_cur) {
ggml_build_forward_expand(gf, k_cur);
}
if (v_cur) {
ggml_build_forward_expand(gf, v_cur);
}
const auto * mctx_iswa = inp->mctx;
const bool is_swa = hparams.is_swa(il);
const auto * mctx_cur = is_swa ? mctx_iswa->get_swa() : mctx_iswa->get_base();
// optionally store to KV cache
if (k_cur) {
const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs();
ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_cross * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * v_mla,
float kq_scale,
int il) const {
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(gf, q_cur);
ggml_build_forward_expand(gf, k_cur);
ggml_build_forward_expand(gf, v_cur);
const auto & kq_mask = inp->get_kq_mask_cross();
ggml_tensor * q = q_cur;
ggml_tensor * k = k_cur;
ggml_tensor * v = v_cur;
// TODO: maybe separate the inner implementation into a separate function
// like with the non-sliding window equivalent
// once sliding-window hybrid caches are a thing.
llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
const auto * mctx_cur = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
{
const auto n_kv = mctx_cur->get_base()->get_n_kv();
ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, rs_size);
// Clear a single state which will then be copied to the other cleared states.
// Note that this is a no-op when the view is zero-sized.
ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0));
ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));
// copy states
// NOTE: assuming the copy destinations are ALL contained between rs_head and rs_head + n_rs
// {state_size, rs_size} -> {state_size, n_seqs}
ggml_tensor * output_states = get_state_rows(ctx0, states, state_copy_main);
ggml_build_forward_expand(gf, output_states);
// copy extra states which won't be changed further (between n_seqs and n_rs)
ggml_tensor * states_extra = ggml_get_rows(ctx0, states, state_copy_extra);
ggml_build_forward_expand(gf,
ggml_cpy(ctx0,
states_extra,
ggml_view_1d(ctx0, s, state_size*(n_rs - n_seqs), (rs_head + n_seqs)*state_size*ggml_element_size(s))));
switch (pooling_type) {
case LLAMA_POOLING_TYPE_NONE:
{
cur = inp;
} break;
case LLAMA_POOLING_TYPE_MEAN:
{
ggml_tensor * inp_mean = build_inp_mean();
cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean);
} break;
case LLAMA_POOLING_TYPE_CLS:
case LLAMA_POOLING_TYPE_LAST:
{
ggml_tensor * inp_cls = build_inp_cls();
cur = ggml_get_rows(ctx0, inp, inp_cls);
} break;
case LLAMA_POOLING_TYPE_RANK:
{
ggml_tensor * inp_cls = build_inp_cls();
inp = ggml_get_rows(ctx0, inp, inp_cls);
if (cls) {
// classification head
// https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
cur = ggml_mul_mat(ctx0, cls, inp);
if (cls_b) {
cur = ggml_add(ctx0, cur, cls_b);
}
cur = ggml_tanh(ctx0, cur);
int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) {
// TODO move to hparams if a T5 variant appears that uses a different value
const int64_t max_distance = 128;
if (bidirectional) {
n_buckets >>= 1;
}
const int64_t max_exact = n_buckets >> 1;
int32_t relative_position = x - y;
int32_t relative_bucket = 0;
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.