mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

model: Add Mimo v2.5 model support (#22493)

Browse Source 
* add mimo-v2.5 support

* mimo-v2.5: fix modify_tensors row split

* mimi-v2.5: forgot `add_attn_value_scale` plumbing

* mimi-v2.5: fix tp dequant to detect tp rows

* mimo-v2.5: fix TP iteration to be descending

* mimo-v2.5: fix comment

* mimo-v2.5: retain fused qkv

* mimo-v2.5: missed the attn_value scale during merge

* mimo-v2.5: fused QKV needs contiguous for scaling attention value

* mimo-v2.5: move `speech_embeddings.` to TextModel filter_tensors

* Update src/llama-hparams.h

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

* Update src/models/mimo2.cpp

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

* Update src/models/mimo2.cpp

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

* Update convert_hf_to_gguf.py

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

* Update convert_hf_to_gguf.py

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

* Update src/models/mimo2.cpp

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

* mimo-v2.5: include MTP weights in gguf

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
AesSedai
2026-05-07 04:21:58 -07:00
committed by GitHub
parent f4b5a2ee91
commit 8e52631d55
9 changed files with 235 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-arch.cpp
+1
View File
@@ -232,6 +232,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, "%s.attention.sliding_window_pattern" },
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
{ LLM_KV_ATTENTION_OUTPUT_SCALE, "%s.attention.output_scale" },
{ LLM_KV_ATTENTION_VALUE_SCALE, "%s.attention.value_scale" },
{ LLM_KV_ATTENTION_TEMPERATURE_LENGTH, "%s.attention.temperature_length" },
{ LLM_KV_ATTENTION_TEMPERATURE_SCALE, "%s.attention.temperature_scale" },
{ LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
src/llama-arch.h
+1
View File
@@ -236,6 +236,7 @@ enum llm_kv {
LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
LLM_KV_ATTENTION_SCALE,
LLM_KV_ATTENTION_OUTPUT_SCALE,
LLM_KV_ATTENTION_VALUE_SCALE,
LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
LLM_KV_ATTENTION_TEMPERATURE_SCALE,
LLM_KV_ATTENTION_KEY_LENGTH_MLA,
src/llama-hparams.h
+2
View File
@@ -166,6 +166,8 @@ struct llama_hparams {
float f_attn_out_scale = 0.0f;
uint32_t attn_temp_length = 0;
float f_attn_value_scale = 0.0f;
bool causal_attn = true;
bool use_alibi = false;
bool attn_soft_cap = false;
src/llama-model-saver.cpp
+1
View File
@@ -268,6 +268,7 @@ void llama_model_saver::add_kv_from_model() {
// add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, ???);
add_kv(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale);
add_kv(LLM_KV_ATTENTION_OUTPUT_SCALE, hparams.f_attn_out_scale);
add_kv(LLM_KV_ATTENTION_VALUE_SCALE, hparams.f_attn_value_scale);
add_kv(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.attn_temp_length);
add_kv(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale);
add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl);
src/llama-model.cpp
+1
View File
@@ -1671,6 +1671,7 @@ void llama_model::print_info() const {
LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias);
LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale);
LLAMA_LOG_INFO("%s: f_attn_scale = %.1e\n", __func__, hparams.f_attention_scale);
LLAMA_LOG_INFO("%s: f_attn_value_scale = %.4f\n", __func__, hparams.f_attn_value_scale);
LLAMA_LOG_INFO("%s: n_ff = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str());
LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert);
LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used);
src/models/mimo2.cpp
+79 -26
View File
@@ -10,7 +10,16 @@ void llama_model_mimo2::load_arch_hparams(llama_model_loader & ml) {
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);
switch (hparams.n_layer) {
float value_scale = 0.0f;
if (ml.get_key(LLM_KV_ATTENTION_VALUE_SCALE, value_scale, false) && value_scale != 1.0f) {
hparams.f_attn_value_scale = value_scale;
}
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");
hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
switch (hparams.n_layer - hparams.nextn_predict_layers) {
case 48: type = LLM_TYPE_310B_A15B; break;
default: type = LLM_TYPE_UNKNOWN;
}
@@ -25,32 +34,45 @@ void llama_model_mimo2::load_arch_tensors(llama_model_loader &) {
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
const uint32_t n_nextn = hparams.nextn_predict_layers;
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);
uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);
uint32_t n_head = hparams.n_head(i);
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_v * n_head, n_embd }, 0);
// NextN/MTP layers (the last n_nextn blocks) are preserved but disabled pending support
const bool is_nextn = (n_nextn > 0) && (static_cast<uint32_t>(i) >= n_layer - n_nextn);
const int skip = is_nextn ? TENSOR_SKIP : 0;
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, TENSOR_NOT_REQUIRED);
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, skip);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_v * n_head, n_embd }, skip);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, skip);
layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, TENSOR_NOT_REQUIRED | skip);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, skip);
// non-MoE branch
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED | skip);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, TENSOR_NOT_REQUIRED | skip);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED | skip);
// MoE branch
int64_t n_ff_exp = hparams.n_ff_exp;
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED | skip);
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED | skip);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED | skip);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED | skip);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | skip);
if (is_nextn) {
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, skip);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), {n_embd}, skip);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), {n_embd}, skip);
layer.layer_out_norm = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, skip);
}
}
}
@@ -68,7 +90,12 @@ llama_model_mimo2::graph::graph(const llama_model & model, const llm_graph_param
auto * inp_attn = build_attn_inp_kv_iswa();
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
const float v_scale = hparams.f_attn_value_scale;
// The last hparams.nextn_predict_layers blocks are MTP heads, currently inactive
const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
for (int il = 0; il < n_transformer_layers; ++il) {
ggml_tensor * inpSA = inpL;
uint32_t n_head_l = hparams.n_head(il);
@@ -83,19 +110,39 @@ llama_model_mimo2::graph::graph(const llama_model & model, const llm_graph_param
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
ggml_tensor * Qcur;
ggml_tensor * Kcur;
ggml_tensor * Vcur;
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
if (model.layers[il].wqkv) {
// Fused qkv_proj - Q/K share head_dim_k, V uses head_dim_v
ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
cb(qkv, "wqkv", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
const size_t row_k = ggml_row_size(qkv->type, n_embd_head_k);
const size_t row_v = ggml_row_size(qkv->type, n_embd_head_v);
const size_t row_full = qkv->nb[1];
const size_t k_off = row_k * n_head_l;
const size_t v_off = k_off + row_k * n_head_kv_l;
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_l, n_tokens, row_k, row_full, 0);
Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv_l, n_tokens, row_k, row_full, k_off);
Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv_l, n_tokens, row_v, row_full, v_off);
} else {
// Split path
Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
}
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,
@@ -118,9 +165,15 @@ llama_model_mimo2::graph::graph(const llama_model & model, const llm_graph_param
cur = build_attn(inp_attn,
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), il);
cb(cur, "attn_out", il);
if (v_scale) {
cur = ggml_scale(ctx0, cur, v_scale);
cb(cur, "attn_out_scaled", il);
}
}
if (il == n_layer - 1 && inp_out_ids) {
if (il == n_transformer_layers - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}