Skip to content

LLM inference in C/C++

C++ ggml

spec : fuse the DFlash encoder into the KV cache injection (#27310)

王金旭662a0b0121a5
diff --git a/common/speculative.cpp b/common/speculative.cpp
index d34d1c9c59..851a47b9a5 100644
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -941,9 +941,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
const int32_t * target_layer_ids = nullptr; // model_dft's extract layer indices
uint32_t target_layer_ids_n = 0;
- // scratch buffer for concatenated target features [n_tokens, n_embd_enc]
- std::vector<float> features_buf;
-
common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq,
common_speculative_type type = COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)
: common_speculative_impl(type, n_seq, params.draft.n_max)
@@ -1011,7 +1008,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
this->n_max = this->params.n_max;
batch = llama_batch_init(llama_n_batch(ctx_dft), 0, n_seq);
- batch_inject = llama_batch_init(llama_n_batch(ctx_dft), n_embd_dec, n_seq);
+ batch_inject = llama_batch_init(llama_n_ubatch(ctx_dft), n_embd_enc, n_seq);
// embd batches on an M-RoPE draft need 4 position rows per token
is_mrope = llama_model_rope_type(model_dft) == LLAMA_ROPE_TYPE_MROPE;
@@ -1137,58 +1134,21 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
for (int32_t offset = 0; offset < n_rows; offset += n_ubatch) {
const int32_t n_chunk = std::min(n_ubatch, n_rows - offset);
- // gather this chunk's target features, interleaved by extract layer
- features_buf.resize((size_t) n_chunk * n_embd_enc);
+ // gather target features per extract layer; the fused decode encodes and
+ // injects them into the K/V cache at the target positions
+ batch_inject.n_tokens = n_chunk;
for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]);
if (!layer) {
GGML_ABORT("DFlash: target layer %d input not extracted.", target_layer_ids[k]);
}
for (int32_t i = 0; i < n_chunk; ++i) {
- float * dst = features_buf.data() + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
+ float * dst = batch_inject.embd + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
const float * src = layer + (size_t) (i_batch_beg[seq_id] + offset + i) * n_embd_tgt;
std::memcpy(dst, src, (size_t) n_embd_tgt * sizeof(float));
}
}
- // fuse extracted features through DFlash encoder
- // M-RoPE drafts read 4 position rows per token from embd batches, so pass them explicitly
- std::vector<llama_pos> enc_pos;
- if (is_mrope) {
- enc_pos.resize((size_t) 4 * n_chunk);
- for (int32_t i = 0; i < n_chunk; ++i) {
- const llama_pos p = batch_in.pos[i_batch_beg[seq_id] + offset + i];
- enc_pos[0 * n_chunk + i] = p;
- enc_pos[1 * n_chunk + i] = p;
- enc_pos[2 * n_chunk + i] = p;
- enc_pos[3 * n_chunk + i] = 0;
- }
- }
-
- llama_batch enc_batch = {
- /*.n_tokens =*/ n_chunk,
- /*.token =*/ nullptr,
- /*.embd =*/ features_buf.data(),
- /*.pos =*/ is_mrope ? enc_pos.data() : nullptr,
- /*.n_seq_id =*/ nullptr,
- /*.seq_id =*/ nullptr,
- /*.logits =*/ nullptr,
- };
-
- int32_t rc = llama_encode(ctx_dft, enc_batch);
- if (rc != 0) {
- LOG_ERR("%s: llama_encode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
- __func__, rc, (int) n_chunk, (int) offset);
- return false;
- }
-
- const float * inp_g = llama_get_embeddings_nextn(ctx_dft);
- GGML_ASSERT(inp_g && "DFlash encoder produced no output.");
-
- // inject the DFlash decoder K/V cache at the tokens' target positions
- batch_inject.n_tokens = n_chunk;
- std::memcpy(batch_inject.embd, inp_g, (size_t) n_chunk * n_embd_dec * sizeof(float));
-
for (int32_t i = 0; i < n_chunk; ++i) {
const llama_pos p = batch_in.pos[i_batch_beg[seq_id] + offset + i];
batch_inject.pos[i] = p;
@@ -1201,7 +1161,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
batch_inject.seq_id[i][0] = seq_id;
batch_inject.logits[i] = false;
}
- rc = llama_decode(ctx_dft, batch_inject);
+ const int32_t rc = llama_decode(ctx_dft, batch_inject);
if (rc != 0) {
LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
__func__, rc, (int) n_chunk, (int) offset);
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 9aed801332..a6d2f264fc 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -1660,7 +1660,9 @@ int llama_context::decode(const llama_batch & batch_inp) {
const int64_t n_vocab = vocab.n_tokens();
const bool mtp_embd = cparams.ctx_type == LLAMA_CONTEXT_TYPE_MTP && batch_inp.embd;
- const int64_t n_embd = mtp_embd ? hparams.n_embd_out() : hparams.n_embd_inp();
+ // DFlash embd batches carry the fused target features at the encoder input width
+ const bool dflash_embd = model.arch == LLM_ARCH_DFLASH && batch_inp.embd;
+ const int64_t n_embd = mtp_embd ? hparams.n_embd_out() : dflash_embd ? hparams.n_embd_inp_enc() : hparams.n_embd_inp();
// when computing embeddings, all tokens are output
const bool output_all = cparams.embeddings;
diff --git a/src/models/dflash.cpp b/src/models/dflash.cpp
index 037723a902..036bcc14aa 100644
--- a/src/models/dflash.cpp
+++ b/src/models/dflash.cpp
@@ -257,9 +257,10 @@ std::unique_ptr<llm_graph_context> llama_model_dflash::build_arch_graph(const ll
template <>
ggml_tensor * llama_model_dflash::graph<true>::build_inp_embd_enc() const {
- auto inp_target = std::make_unique<llm_graph_input_embd>(hparams.n_embd_inp_enc());
+ const int64_t n_embd_inp = hparams.n_embd_inp_enc();
+ auto inp_target = std::make_unique<llm_graph_input_embd>(n_embd_inp);
- inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp_enc(), n_tokens);
+ inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd_inp, n_tokens);
ggml_set_input(inp_target->embd);
ggml_tensor * cur = inp_target->embd;
@@ -567,6 +568,7 @@ static void build_dflash2_selector(llm_graph_context & g, const llama_model & mo
// * token batch -> noise-block diffusion: attend over [committed, MASK...] to generate draft tokens
template <>
llama_model_dflash::graph<false>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+ const int64_t n_embd_inp = hparams.n_embd_inp_enc();
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
@@ -602,16 +604,21 @@ llama_model_dflash::graph<false>::graph(const llama_model & model, const llm_gra
// KV cache injection
if (ubatch.embd) {
- auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
+ auto inp = std::make_unique<llm_graph_input_embd>(n_embd_inp);
- inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+ inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd_inp, n_tokens);
ggml_set_input(inp->embd);
- ggml_tensor * inp_g = inp->embd;
- cb(inp_g, "inp_g_embeddings", -1);
+ ggml_tensor * inp_target = inp->embd;
+ cb(inp_target, "inp_target_features", -1);
res->add_input(std::move(inp));
+ // fuse the target features through the encoder
+ ggml_tensor * inp_g = build_lora_mm(model.fc, inp_target, model.fc_s);
+ inp_g = build_norm(inp_g, model.output_norm_enc, NULL, LLM_NORM_RMS, -1);
+ cb(inp_g, "inp_g_embeddings", -1);
+
for (int il = 0; il < n_layer; ++il) {
const auto & layer = model.layers[il];
@@ -823,6 +830,7 @@ llama_model_dflash::graph<false>::graph(const llama_model & model, const llm_gra
// * token batch -> noise block through 3 full DSV4 stages (hc + MLA + MoE), markov + confidence heads
llama_model_dflash::graph_dsv4::graph_dsv4(const llama_model & model, const llm_graph_params & params) :
llama_model_deepseek4::graph(params) {
+ const int64_t n_embd_inp = hparams.n_embd_inp_enc();
const int64_t n_embd_head = hparams.n_embd_head_k();
const int64_t n_embd_head_rope = hparams.n_rot();
const int64_t n_embd_head_nope = n_embd_head - n_embd_head_rope;
@@ -833,16 +841,21 @@ llama_model_dflash::graph_dsv4::graph_dsv4(const llama_model & model, const llm_
// KV cache injection: fused target features from the encoder
if (ubatch.embd) {
- auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
+ auto inp = std::make_unique<llm_graph_input_embd>(n_embd_inp);
- inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+ inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd_inp, n_tokens);
ggml_set_input(inp->embd);
- ggml_tensor * inp_g = inp->embd;
- cb(inp_g, "inp_g_embeddings", -1);
+ ggml_tensor * inp_target = inp->embd;
+ cb(inp_target, "inp_target_features", -1);
res->add_input(std::move(inp));
+ // fuse the target features through the encoder
+ ggml_tensor * inp_g = build_lora_mm(model.fc, inp_target, model.fc_s);
+ inp_g = build_norm(inp_g, model.output_norm_enc, nullptr, LLM_NORM_RMS, -1);
+ cb(inp_g, "inp_g_embeddings", -1);
+
for (int il = 0; il < n_layer; ++il) {
const auto & layer = model.layers[il];