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

tests : add support for qwen3 SSM archs (#24031)

Browse Source 
* tests : add support for qwen3 SSM archs

* arch : add LLM_KV_ATTENTION_RECURRENT_LAYERS

* cont : naming + TODOs
This commit is contained in:
Georgi Gerganov
2026-06-03 10:15:27 +03:00
committed by GitHub
parent d545a2a993
commit 06938ac129
25 changed files with 109 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/falcon-h1.cpp
+1 -1
View File
@@ -11,7 +11,7 @@ void llama_model_falcon_h1::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
std::fill(hparams.recurrent_layer_arr.begin(), hparams.recurrent_layer_arr.end(), true);
std::fill(hparams.is_recr_impl.begin(), hparams.is_recr_impl.end(), true);
switch (hparams.n_layer) {
case 36:
src/models/gemma4.cpp
+1 -1
View File
@@ -2,7 +2,7 @@
void llama_model_gemma4::load_arch_hparams(llama_model_loader & ml) {
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer);
uint32_t n_kv_shared_layers = 0;
ml.get_key(LLM_KV_ATTENTION_SHARED_KV_LAYERS, n_kv_shared_layers, false);
src/models/granite-hybrid.cpp
+3 -3
View File
@@ -20,7 +20,7 @@ void llama_model_granite_hybrid::load_arch_hparams(llama_model_loader & ml) {
// A layer is recurrent IFF the n_head_kv value is set to 0
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0;
hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;
}
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -71,7 +71,7 @@ void llama_model_granite_hybrid::load_arch_tensors(llama_model_loader &) {
// norm
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
if (hparams.is_recurrent(i)) {
if (hparams.is_recr(i)) {
// ssm layers
layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
@@ -158,7 +158,7 @@ llama_model_granite_hybrid::graph::graph(const llama_model & model, const llm_gr
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
if (hparams.is_recurrent(il)) {
if (hparams.is_recr(il)) {
// ssm layer //
cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
} else {
src/models/jamba.cpp
+1 -1
View File
@@ -9,7 +9,7 @@ void llama_model_jamba::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0;
hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;
}
switch (hparams.n_layer) {
src/models/kimi-linear.cpp
+3 -3
View File
@@ -15,7 +15,7 @@ void llama_model_kimi_linear::load_arch_hparams(llama_model_loader & ml) {
// Mark KDA layers as recurrent using n_head_kv pattern (like Jamba)
// Set n_head_kv = 0 for KDA layers (recurrent), n_head_kv = n_head for MLA layers (attention)
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0; // KDA layers are recurrent
hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0; // KDA layers are recurrent
}
// MoE parameters - Kimi uses moe_intermediate_size = 1024
@@ -53,7 +53,7 @@ void llama_model_kimi_linear::load_arch_tensors(llama_model_loader &) {
const int64_t n_embd_head_v_kda = hparams.n_embd_head_kda;
const int64_t ssm_d_conv = hparams.ssm_d_conv;
if (hparams.is_recurrent(i)) {
if (hparams.is_recr(i)) {
// Conv1d weights: try 4D first, then 3D (quantization may remove trailing 1)
// 4D: [d_conv, 1, d_inner, 1], 3D: [d_conv, 1, d_inner]
layer.ssm_q_conv = create_tensor(tn(LLM_TENSOR_SSM_CONV1D_Q, "weight", i), {ssm_d_conv, 1, n_embd_head_k_kda * n_head, 1}, TENSOR_NOT_REQUIRED);
@@ -285,7 +285,7 @@ llama_model_kimi_linear::graph::graph(const llama_model & model, const llm_graph
ggml_build_forward_expand(gf, cur);
if (hparams.is_recurrent(il)) {
if (hparams.is_recr(il)) {
// === KDA Layer (Kimi Delta Attention) with Recurrent State ===
// Reference: vLLM kda.py
const auto * mctx_cur = inp_rs->mctx;
src/models/lfm2.cpp
+5 -5
View File
@@ -6,7 +6,7 @@ void llama_model_lfm2::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_SHORTCONV_L_CACHE, hparams.n_shortconv_l_cache);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
for (uint32_t il = 0; il < hparams.n_layer; ++il) {
hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
hparams.is_recr_impl[il] = hparams.n_head_kv(il) == 0;
}
hparams.n_layer_dense_lead = hparams.n_layer;
switch (hparams.n_ff()) {
@@ -19,7 +19,7 @@ void llama_model_lfm2::load_arch_hparams(llama_model_loader & ml) {
if (const auto is_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); is_swa && hparams.n_swa > 0) {
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
for (uint32_t il = 0; il < hparams.n_layer; ++il) {
hparams.swa_layers[il] = !hparams.recurrent_layer_arr[il];
hparams.is_swa_impl[il] = !hparams.is_recr_impl[il];
}
}
}
@@ -59,7 +59,7 @@ void llama_model_lfm2::load_arch_tensors(llama_model_loader &) {
// for operator_norm
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
if (!hparams.is_recurrent(i)) {
if (!hparams.is_recr(i)) {
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
GGML_ASSERT(n_embd_v_gqa == n_embd_k_gqa);
@@ -235,8 +235,8 @@ llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_
cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "model.layers.{}.operator_norm", il);
cur = hparams.is_recurrent(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);
cur = hparams.is_recr(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
src/models/lfm2moe.cpp
+2 -2
View File
@@ -10,7 +10,7 @@ void llama_model_lfm2moe::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func);
for (uint32_t il = 0; il < hparams.n_layer; ++il) {
hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
hparams.is_recr_impl[il] = hparams.n_head_kv(il) == 0;
}
switch (hparams.n_layer) {
@@ -55,7 +55,7 @@ void llama_model_lfm2moe::load_arch_tensors(llama_model_loader &) {
// for operator_norm
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
if (!hparams.is_recurrent(i)) {
if (!hparams.is_recr(i)) {
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
GGML_ASSERT(n_embd_v_gqa == n_embd_k_gqa);
src/models/llama4.cpp
+2 -1
View File
@@ -15,7 +15,8 @@ void llama_model_llama4::load_arch_hparams(llama_model_loader & ml) {
hparams.n_attn_temp_floor_scale = 8192;
hparams.f_attn_temp_scale = 0.1f;
hparams.f_attn_temp_offset = 1.0f;
uint32_t swa_period = 4; // pattern: 3 chunked - 1 full
uint32_t swa_period = 4; // pattern: 3 chunked - 1 full
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
hparams.set_swa_pattern(swa_period);
src/models/mellum.cpp
+1 -1
View File
@@ -13,7 +13,7 @@ void llama_model_mellum::load_arch_hparams(llama_model_loader & ml) {
if (res) {
hparams.set_swa_pattern(swa_period);
} else {
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer);
}
hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;
src/models/mimo2.cpp
+2 -1
View File
@@ -8,7 +8,8 @@ void llama_model_mimo2::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer);
float value_scale = 0.0f;
if (ml.get_key(LLM_KV_ATTENTION_VALUE_SCALE, value_scale, false) && value_scale != 1.0f) {
src/models/nemotron-h.cpp
+3 -3
View File
@@ -10,7 +10,7 @@ void llama_model_nemotron_h::load_arch_hparams(llama_model_loader & ml) {
// A layer is recurrent IFF the n_head_kv value is set to 0 and
// the n_ff value is set to 0
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = (hparams.n_head_kv(i) == 0 && hparams.n_ff(i) == 0);
hparams.is_recr_impl[i] = (hparams.n_head_kv(i) == 0 && hparams.n_ff(i) == 0);
}
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -62,7 +62,7 @@ void llama_model_nemotron_h::load_arch_tensors(llama_model_loader &) {
// all blocks use the attn norm
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
if (hparams.is_recurrent(i)) {
if (hparams.is_recr(i)) {
// ssm layers
layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
@@ -143,7 +143,7 @@ llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
if (hparams.is_recurrent(il)) {
if (hparams.is_recr(il)) {
// ssm layer //
cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
} else if (hparams.n_ff(il) == 0) {
src/models/plamo2.cpp
+3 -3
View File
@@ -12,7 +12,7 @@ void llama_model_plamo2::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0;
hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;
}
switch (hparams.n_layer) {
@@ -54,7 +54,7 @@ void llama_model_plamo2::load_arch_tensors(llama_model_loader &) {
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
bool is_mamba_layer = hparams.is_recurrent(i);
bool is_mamba_layer = hparams.is_recr(i);
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
@@ -128,7 +128,7 @@ llama_model_plamo2::graph::graph(const llama_model & model, const llm_graph_para
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
// check if this layer is Mamba or Attention
const bool is_mamba_layer = hparams.is_recurrent(il);
const bool is_mamba_layer = hparams.is_recr(il);
if (is_mamba_layer) {
// PLaMo-2 Mamba layer
src/models/qwen35.cpp
+5 -4
View File
@@ -18,12 +18,13 @@ void llama_model_qwen35::load_arch_hparams(llama_model_loader & ml) {
// 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;
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.recurrent_layer_arr[i] = (i < n_main) && ((i + 1) % full_attn_interval != 0);
hparams.is_recr_impl[i] = (i < n_main) && ((i + 1) % full_attn_interval != 0);
}
}
@@ -69,7 +70,7 @@ void llama_model_qwen35::load_arch_tensors(llama_model_loader & ml) {
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, flags);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, flags);
if (!hparams.is_recurrent(il)) {
if (!hparams.is_recr(il)) {
// Attention layers
create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, flags);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, flags);
@@ -168,7 +169,7 @@ llama_model_qwen35::graph::graph(const llama_model & model, const llm_graph_para
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
if (hparams.is_recr(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, il);
} else {
src/models/qwen35moe.cpp
+5 -4
View File
@@ -21,12 +21,13 @@ void llama_model_qwen35moe::load_arch_hparams(llama_model_loader & ml) {
// 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;
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.recurrent_layer_arr[i] = (i < n_main) && ((i + 1) % full_attn_interval != 0);
hparams.is_recr_impl[i] = (i < n_main) && ((i + 1) % full_attn_interval != 0);
}
}
@@ -75,7 +76,7 @@ void llama_model_qwen35moe::load_arch_tensors(llama_model_loader & ml) {
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, flags);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, flags);
if (!hparams.is_recurrent(il)) {
if (!hparams.is_recr(il)) {
// Attention layers
create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, flags);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, flags);
@@ -191,7 +192,7 @@ llama_model_qwen35moe::graph::graph(const llama_model & model, const llm_graph_p
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
if (hparams.is_recr(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, il);
} else {
src/models/qwen3next.cpp
+4 -4
View File
@@ -14,11 +14,11 @@ void llama_model_qwen3next::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
// Mark recurrent layers (linear attention layers)
{
if (!ml.get_key_or_arr(LLM_KV_ATTENTION_RECURRENT_LAYERS, hparams.is_recr_impl, hparams.n_layer, 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.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0);
hparams.is_recr_impl[i] = ((i + 1) % full_attn_interval != 0);
}
}
@@ -68,7 +68,7 @@ void llama_model_qwen3next::load_arch_tensors(llama_model_loader &) {
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0);
if (!hparams.is_recurrent(i)) {
if (!hparams.is_recr(i)) {
// Attention layers
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
@@ -129,7 +129,7 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
if (hparams.is_recr(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, il);
} else {
src/models/step35.cpp
+3 -1
View File
@@ -22,7 +22,9 @@ void llama_model_step35::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer);
ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_EXP, hparams.swiglu_clamp_exp, hparams.n_layer, false);
ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_SHEXP, hparams.swiglu_clamp_shexp, hparams.n_layer, false);