Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

hparams : refactor hparams.n_layer (#24060)

Browse Source 
* hparams : refactor hparams.n_layer

* cont : remove `n_layer_kv()`, use n_layer_all instead

* cont : type consistency

* pi : update SYSTEM.md

* models : fix Step3.5 MTP

* cont : remove duplicate switch cases

* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`

* cont : fix nextn layer count handling

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Georgi Gerganov
2026-06-05 11:09:36 +03:00
committed by GitHub
co-authored by Sigbjørn Skjæret
parent 3ecfb150a4
commit 7acb4e8cd2
129 changed files with 412 additions and 431 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/qwen35moe.cpp
+14 -19
View File
@@ -16,22 +16,20 @@ void llama_model_qwen35moe::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
// NextN/MTP (Qwen3.5/3.6): extra decoder block appended beyond the main stack
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_impl");
// Mark recurrent layers (linear attention layers). MTP layers are dense
// attention-only and must be flagged non-recurrent.
if (!ml.get_key_or_arr(LLM_KV_ATTENTION_RECURRENT_LAYERS, hparams.is_recr_impl, hparams.n_layer, false)) {
const uint32_t n_main = hparams.n_layer - hparams.nextn_predict_layers;
if (!ml.get_key_or_arr(LLM_KV_ATTENTION_RECURRENT_LAYERS, hparams.is_recr_impl, hparams.n_layer_all, false)) {
uint32_t full_attn_interval = 4;
ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.is_recr_impl[i] = (i < n_main) && ((i + 1) % full_attn_interval != 0);
for (uint32_t i = 0; i < hparams.n_layer_all; ++i) {
hparams.is_recr_impl[i] = (i < hparams.n_layer()) && ((i + 1) % full_attn_interval != 0);
}
}
switch (hparams.n_layer - hparams.nextn_predict_layers) {
switch (hparams.n_layer()) {
case 40: type = LLM_TYPE_35B_A3B; break;
case 48: type = LLM_TYPE_122B_A10B; break;
case 60: type = LLM_TYPE_397B_A17B; break;
@@ -42,9 +40,7 @@ void llama_model_qwen35moe::load_arch_hparams(llama_model_loader & ml) {
void llama_model_qwen35moe::load_arch_tensors(llama_model_loader & ml) {
LLAMA_LOAD_LOCALS;
const uint32_t n_main = n_layer - hparams.nextn_predict_layers;
const bool mtp_only = (hparams.nextn_predict_layers > 0) &&
(ml.get_weight("blk.0.attn_norm.weight") == nullptr);
const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
@@ -145,10 +141,10 @@ void llama_model_qwen35moe::load_arch_tensors(llama_model_loader & ml) {
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd }, TENSOR_NOT_REQUIRED);
};
for (int i = 0; i < (int) n_main; ++i) {
for (int i = 0; i < n_layer; ++i) {
load_block_trunk(i, trunk_flags);
}
for (int i = (int) n_main; i < n_layer; ++i) {
for (int i = n_layer; i < n_layer_all; ++i) {
load_block_mtp(i);
}
}
@@ -182,8 +178,7 @@ llama_model_qwen35moe::graph::graph(const llama_model & model, const llm_graph_p
ggml_tensor * inp_out_ids = build_inp_out_ids();
// MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.
const int n_transformer_layers = n_layer - (int) hparams.nextn_predict_layers;
for (int il = 0; il < n_transformer_layers; ++il) {
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
@@ -200,7 +195,7 @@ llama_model_qwen35moe::graph::graph(const llama_model & model, const llm_graph_p
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
}
if (il == n_transformer_layers - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
@@ -555,13 +550,13 @@ ggml_tensor * llama_model_qwen35moe::graph::build_layer_ffn(ggml_tensor * cur, c
// LLM_GRAPH_TYPE_DECODER_MTP draft head for Qwen3.5/3.6 MoE
llama_model_qwen35moe::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
: llm_graph_context(params) {
GGML_ASSERT(hparams.nextn_predict_layers > 0 && "QWEN35MOE MTP requires nextn_predict_layers > 0");
GGML_ASSERT(hparams.nextn_predict_layers == 1 && "QWEN35MOE MTP currently only supports a single MTP block");
GGML_ASSERT(hparams.n_layer_nextn > 0 && "QWEN35MOE MTP requires n_layer_nextn > 0");
GGML_ASSERT(hparams.n_layer_nextn == 1 && "QWEN35MOE MTP currently only supports a single MTP block");
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
const int il = (int) hparams.n_layer - (int) hparams.nextn_predict_layers;
const int il = hparams.n_layer();
const auto & layer = model.layers[il];
GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");