Skip to content

Support Step3.5/3.7 flash mtp3 #24340

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.

Sign up for GitHub

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
ggerganov merged 16 commits into from
Jun 21, 2026
Merged

Support Step3.5/3.7 flash mtp3 #24340

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
e98e0b3
add mtp_layer_offset + include nextn flags in graph reuse
forforever73 Jun 7, 2026
1208d99
add llama_set_mtp_layer_offset + llama_model_n_nextn_layer API
forforever73 Jun 7, 2026
0b8aa51
offset head select + require all MTP blocks
forforever73 Jun 7, 2026
e0fb9ff
speculative multi-head process()
forforever73 Jun 7, 2026
34f68f5
speculative multi-head draft()
forforever73 Jun 7, 2026
ae013e3
gather outputs via inp_out_ids
forforever73 Jun 8, 2026
1885d8f
cleanup
forforever73 Jun 9, 2026
f4a2c12
fix core
forforever73 Jun 9, 2026
9a0ff26
minor cleanup
forforever73 Jun 9, 2026
2952d83
merged draft_multi_head into draft()
forforever73 Jun 14, 2026
48a7484
mtp rename nextn
forforever73 Jun 14, 2026
2ce24fb
Apply suggestions from code review
forforever73 Jun 14, 2026
9b8f3b6
clean-up comments
forforever73 Jun 14, 2026
7a0a247
fix for multi seq
forforever73 Jun 15, 2026
9858fd2
apply suggestions && chain-heads comment
forforever73 Jun 21, 2026
60b04d3
add a reference for chain_heads discussion
forforever73 Jun 21, 2026
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
137 changes: 102 additions & 35 deletions common/speculative.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -825,7 +825,13 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {

int32_t n_embd = 0;

bool is_mem_shared = false;
// One MTP draft driver, three modes (set once in the ctor):
// is_mem_shared (gemma4): shares the target KV, runs all heads in one graph.
// chain_heads (step35): n_mtp_layers trained heads, one per draft step.
// neither (qwen35 / qwen35moe): a single trained MTP head.
int32_t n_mtp_layers = 1;
bool is_mem_shared = false; // gemma4
bool chain_heads = false; // derived in the ctor: n_mtp_layers > 1 && !is_mem_shared

// Per-sequence cross-batch carryover: pair (h_p, x_{p+1}) at MTP pos p+1.
// The last h-row of one process() call needs the first token of the NEXT
Expand All @@ -840,10 +846,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
std::vector<std::vector<float>> verify_h;
std::vector<int32_t> verify_h_rows;

// Per-seq draft length from the last draft() call, used in accept() to
// roll back ctx_dft's recurrent state past the AR draft's redundant
// pre-advancement before process() mirrored the verify batch.
std::vector<uint16_t> last_n_drafted;
std::vector<int> i_last;
std::vector<std::vector<float>> chain_h;

common_speculative_impl_draft_mtp(const common_params_speculative & params, uint32_t n_seq)
: common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq)
Expand All @@ -856,6 +860,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
n_embd = llama_model_n_embd_out(llama_get_model(ctx_dft));
GGML_ASSERT(n_embd == llama_model_n_embd(llama_get_model(ctx_tgt)) &&
"MTP input row width must match the target h_nextn width");
n_mtp_layers = std::max(1, (int) llama_model_n_layer_nextn(llama_get_model(ctx_dft)));

LOG_INF("%s: adding speculative implementation 'draft-mtp'\n", __func__);
LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling);
Expand Down Expand Up @@ -902,16 +907,25 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);

is_mem_shared = llama_get_ctx_other(ctx_dft) == ctx_tgt;
chain_heads = n_mtp_layers > 1 && !is_mem_shared;

if (chain_heads) {
this->params.n_max = std::min(this->params.n_max, n_mtp_layers);

chain_h.assign(n_seq, {});
for (auto & c : chain_h) {
c.reserve((size_t) (this->params.n_max + 1) * n_embd);
}
}

pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));

i_last.assign(n_seq, -1);
i_batch_beg.assign(n_seq, -1);
i_batch_end.assign(n_seq, -1);

verify_h.assign(n_seq, {});
verify_h_rows.assign(n_seq, 0);

last_n_drafted.assign(n_seq, 0);
}

~common_speculative_impl_draft_mtp() override {
Expand Down Expand Up @@ -1017,9 +1031,34 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
}

const int32_t rc = llama_decode(ctx_dft, batch);
if (rc != 0) {
LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (pos=%d)\n", __func__, (int) rc, (int) batch_in.pos[0]);
auto * mem_dft = llama_get_memory(ctx_dft);

bool ok = true;
for (int head = 0; head < n_mtp_layers; ++head) {
if (chain_heads) {
// ref: https://github.com/ggml-org/llama.cpp/pull/24340/changes#r3413498544
for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
if (i_batch_beg[seq_id] < 0) {
continue;
}
llama_memory_seq_rm(mem_dft, seq_id, batch_in.pos[i_batch_beg[seq_id]], -1);
}
llama_set_nextn_layer_offset(ctx_dft, head);
}

const int32_t rc = llama_decode(ctx_dft, batch);
if (rc != 0) {
LOG_ERR("%s: llama_decode(ctx_dft) head=%d failed rc=%d (pos=%d)\n",
__func__, head, (int) rc, (int) batch_in.pos[0]);
ok = false;
break;
}
}

if (chain_heads) {
llama_set_nextn_layer_offset(ctx_dft, 0); // restore default for non-draft decodes
}
Comment on lines +1034 to +1060

@ggerganov ggerganov Jun 15, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't understand the logic here - seems incorrect. Every head iteration will basically erase the result of the previous iteration.

@forforever73 forforever73 Jun 15, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Each head runs a different layer, set_nextn_layer_offset(head) makes graph_mtp build layer n_layer()+head, i.e. 45/46/47. So each head writes its own k_l[il]/v_l[il].
seq_rm here doesn't drop any KV data — it just clears the cell metadata so find_slot hands back the same cells at the same positions for every head. Without it, head 46/47 would land on fresh cells and we'd get duplicate positions in v_cells .
So after the loop those cells hold valid KV for all three MTP layers at once. it's the teacher-forcing catch-up that seeds each head's layer so the next draft() round attends to a correct, target-aligned cache.

@ggerganov ggerganov Jun 15, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, got it. That's interesting.

if (!ok) {
return false;
}
}
Expand Down Expand Up @@ -1054,7 +1093,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
int n_drafting = 0;
std::vector<bool> drafting(n_seq);

const float * h_row = nullptr;
const size_t row_bytes = (size_t) n_embd * sizeof(float);

for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
Expand All @@ -1069,22 +1107,43 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
common_sampler_reset(smpls[seq_id].get());

common_batch_add(batch, dp.id_last, dp.n_past, { seq_id }, true);
std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, pending_h[seq_id].data(), row_bytes);

h_row = pending_h[seq_id].data();
std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
}
i_last[seq_id] = batch.n_tokens - 1;

int ret = llama_decode(ctx_dft, batch);
if (ret != 0) {
LOG_WRN("%s: llama_decode returned %d\n", __func__, ret);
return;
if (chain_heads) {
chain_h[seq_id].assign(pending_h[seq_id].begin(), pending_h[seq_id].end());
}
}

int i = 0;

while (n_drafting > 0) {
int i_batch = 0;
// each step decodes under a different head, i.e. a different decoder layer, and
// KV is per layer. process() filled this layer's KV only for positions < n_past
// (prompt + accepted prefix) — nothing in the draft region yet. so reset the
// draft region (the seq_rm lower bound is n_past, leaving the prompt KV intact)
// and select head i so it rebuilds its own layer's KV there; decoding just the
// latest token would leave its attention reading cells only another head wrote.
if (chain_heads) {
auto * mem_dft = llama_get_memory(ctx_dft);
for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
if (drafting[seq_id]) {
llama_memory_seq_rm(mem_dft, seq_id, dparams[seq_id].n_past, -1);
}
}
llama_set_nextn_layer_offset(ctx_dft, i);
}

int ret = llama_decode(ctx_dft, batch);
if (ret != 0) {
LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret);
break;
}

// rebuild the batch for the next step: the growing-KV paths re-add only the
// new token (the KV already holds the prefix), while chained heads re-add the
// whole prefix at the next head. dropped sequences are simply not re-added.
common_batch_clear(batch);

for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
Expand All @@ -1094,9 +1153,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {

auto * smpl = smpls[seq_id].get();

common_sampler_sample(smpl, ctx_dft, i_batch, true);
h_row = llama_get_embeddings_nextn_ith(ctx_dft, i_batch);
++i_batch;
common_sampler_sample(smpl, ctx_dft, i_last[seq_id], true);
const float * h_row = llama_get_embeddings_nextn_ith(ctx_dft, i_last[seq_id]);

const auto * cur_p = common_sampler_get_candidates(smpl, true);

Expand Down Expand Up @@ -1130,30 +1188,41 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
continue;
}

if (is_mem_shared) {
if (chain_heads) {
// ref: https://github.com/ggml-org/llama.cpp/pull/24340#discussion_r3448031546
chain_h[seq_id].insert(chain_h[seq_id].end(), h_row, h_row + n_embd);

const int n_rows = (int) result.size() + 1; // id_last + tokens drafted so far
for (int t = 0; t < n_rows; ++t) {
const llama_token tok = (t == 0) ? dp.id_last : result[t - 1];
common_batch_add(batch, tok, dp.n_past + t, { seq_id }, t == n_rows - 1);
std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd,
chain_h[seq_id].data() + (size_t) t * n_embd, row_bytes);
}
Comment on lines +1191 to +1201

@ggerganov ggerganov Jun 21, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This seems incorrect - every next draft, we decode all previous tokens again. Why is that? Normally, we should decode just the latest token.

@forforever73 forforever73 Jun 21, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You're right that every draft step re-decodes the whole prefix — that's intentional. This path isn't a normal AR draft; it's an accumulating-batch + per-step-seq_rm flow, because each step runs a different head.

After process(), every head's KV is already filled for the prompt + accepted prefix (positions < n_past), and the per-step seq_rm lower bound is n_past, so we never re-decode the prompt. What gets replayed each step is only the draft region — id_last plus the few tokens drafted so far (capped at n_max=3).

Why replay it under each head: since we switch heads per step, the current head hasn't written its own KV for any draft-region position yet this round. If we only decoded the latest token, its attention over the earlier draft positions would read cells that only another head ever wrote → garbage. So each step we seq_rm the draft region (so find_slot reuses the same slots and positions stay aligned), switch the head, and replay the accumulated prefix so this head fills its own KV.

For contrast, the other two branches reuse a single head across steps, so their prefix KV is already valid and they just append the latest token

@ggerganov ggerganov Jun 21, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah yes. I'm still not used to that approach, but it seems correct. Add a reference to this explanation in the code.

@forforever73 forforever73 Jun 21, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed it's not intuitive, it took me quite a while to design this too :) Added a comment and committed the other suggestions as well.

@ggerganov ggerganov Jun 21, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Add a reference:

diff --git a/common/speculative.cpp b/common/speculative.cpp
index d7a177b7b..f8a6287c2 100644
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -1184,6 +1184,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                 }
 
                 if (chain_heads) {
+                    // ref: https://github.com/ggml-org/llama.cpp/pull/24340#discussion_r3448031546
                     chain_h[seq_id].insert(chain_h[seq_id].end(), h_row, h_row + n_embd);
 
                     const int n_rows = (int) result.size() + 1; // id_last + tokens drafted so far

} else if (is_mem_shared) {
// note: with shared memory (e.g. Gemma4 assistants) we use the same position for all draft tokens
// ref: https://github.com/huggingface/transformers/blob/effde20942e3f82a1b97449f60b3a48c5ff96145/docs/source/en/model_doc/gemma4_assistant.md?plain=1#L36-L37
common_batch_add(batch, id, dp.n_past, { seq_id }, true);
std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
} else {
common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
}
std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
}

if (batch.n_tokens == 0) {
break;
i_last[seq_id] = batch.n_tokens - 1;
}

// evaluate the drafted tokens on the draft model
ret = llama_decode(ctx_dft, batch);
if (ret != 0) {
LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret);
if (batch.n_tokens == 0) {
break;
}

++i;
}

if (chain_heads) {
llama_set_nextn_layer_offset(ctx_dft, 0); // restore default for non-draft decodes
}

for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
auto & dp = dparams[seq_id];
if (!dp.drafting) {
Expand All @@ -1163,8 +1232,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
if (dp.result->size() < (size_t) params.n_min) {
dp.result->clear();
}

last_n_drafted[seq_id] = (uint16_t) dp.result->size();
}
}

Expand Down Expand Up @@ -1777,7 +1844,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,

bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr;
bool has_mtp = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP)) && params.draft.ctx_dft != nullptr;
bool has_draft_mtp = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP)) && params.draft.ctx_dft != nullptr;



Expand Down Expand Up @@ -1815,7 +1882,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
if (has_draft_eagle3) {
configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params));
}
if (has_mtp) {
if (has_draft_mtp) {
configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, params));
}
}
Expand Down
17 changes: 9 additions & 8 deletions include/llama.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -558,14 +558,15 @@ extern "C" {
LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
LLAMA_API enum llama_rope_type llama_model_rope_type(const struct llama_model * model);

LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_embd_out (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_head (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_swa (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_ctx_train (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_embd_out (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_layer_nextn(const struct llama_model * model);
LLAMA_API int32_t llama_model_n_head (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model);
LLAMA_API int32_t llama_model_n_swa (const struct llama_model * model);

// Get the model's RoPE frequency scaling factor
LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);
Expand Down
8 changes: 8 additions & 0 deletions src/llama-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1156,6 +1156,10 @@ void llama_context::set_embeddings_layer_inp(uint32_t lid, bool enable) {
sched_need_reserve = true;
}

void llama_context::set_nextn_layer_offset(int32_t offset) {
cparams.nextn_layer_offset = offset;
}

void llama_context::set_causal_attn(bool value) {
LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value);

Expand Down Expand Up @@ -3699,6 +3703,10 @@ void llama_set_embeddings_layer_inp(llama_context * ctx, uint32_t lid, bool valu
ctx->set_embeddings_layer_inp(lid, value);
}

void llama_set_nextn_layer_offset(llama_context * ctx, int32_t offset) {
ctx->set_nextn_layer_offset(offset);
}

llama_memory_t llama_get_memory(const struct llama_context * ctx) {
if (!ctx) {
return nullptr;
Expand Down
1 change: 1 addition & 0 deletions src/llama-context.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,6 +115,7 @@ struct llama_context {
void set_embeddings (bool value);
void set_embeddings_nextn(bool value, bool masked);
void set_embeddings_layer_inp(uint32_t lid, bool enable);
void set_nextn_layer_offset(int32_t offset);
void set_causal_attn(bool value);
void set_warmup(bool value);

Expand Down
2 changes: 2 additions & 0 deletions src/llama-cparams.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,8 @@ struct llama_cparams {
int32_t n_threads; // number of threads to use for generation
int32_t n_threads_batch; // number of threads to use for batch processing

int32_t nextn_layer_offset = 0;

float rope_freq_base;
float rope_freq_scale;

Expand Down
5 changes: 5 additions & 0 deletions src/llama-ext.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,6 +95,11 @@ LLAMA_API llama_memory_breakdown llama_get_memory_breakdown(const struct llama_c
// If masked == false, output the embeddings for all tokens in the batch regardless of batch.logits
LLAMA_API void llama_set_embeddings_nextn(struct llama_context * ctx, bool value, bool masked);

// Select which appended NextN block the DECODER_MTP graph runs (offset past
// the trunk: il = n_layer() + offset). Used by the speculative NextN driver to
// chain multiple trained NextN heads. Default 0 (first head).
LLAMA_API void llama_set_nextn_layer_offset(struct llama_context * ctx, int32_t offset);

// mirrors:
// LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
LLAMA_API float * llama_get_embeddings_nextn(struct llama_context * ctx);
Expand Down
11 changes: 9 additions & 2 deletions src/llama-graph.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -682,9 +682,16 @@ struct llm_graph_params {
}
}

// TODO: https://github.com/ggml-org/llama.cpp/pull/24340#discussion_r3448035248
if (cparams.nextn_layer_offset != other.cparams.nextn_layer_offset) {
return false;
}

return
cparams.embeddings == other.cparams.embeddings &&
cparams.causal_attn == other.cparams.causal_attn &&
cparams.embeddings == other.cparams.embeddings &&
cparams.embeddings_nextn == other.cparams.embeddings_nextn &&
cparams.embeddings_nextn_masked == other.cparams.embeddings_nextn_masked &&
cparams.causal_attn == other.cparams.causal_attn &&
arch == other.arch &&
gtype == other.gtype &&
cvec == other.cvec &&
Expand Down
4 changes: 4 additions & 0 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2312,6 +2312,10 @@ int32_t llama_model_n_layer(const llama_model * model) {
return model->hparams.n_layer();
}

int32_t llama_model_n_layer_nextn(const llama_model * model) {
return model->hparams.n_layer_nextn;
}

int32_t llama_model_n_head(const llama_model * model) {
return model->hparams.n_head();
}
Expand Down
Loading
Loading