Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

merged support for gemma4. the e2b, e4b and 26b work, the 31b does not

Browse Source
This commit is contained in:
Concedo
2026-04-03 11:07:46 +08:00
parent 5dee1a1cbb 7c7d6ce5c7
commit 34ad53e950
44 changed files with 2908 additions and 120 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-arch.cpp
+43
View File
@@ -56,6 +56,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_GEMMA2, "gemma2" },
{ LLM_ARCH_GEMMA3, "gemma3" },
{ LLM_ARCH_GEMMA3N, "gemma3n" },
{ LLM_ARCH_GEMMA4, "gemma4" },
{ LLM_ARCH_GEMMA_EMBEDDING, "gemma-embedding" },
{ LLM_ARCH_STARCODER2, "starcoder2" },
{ LLM_ARCH_MAMBA, "mamba" },
@@ -165,6 +166,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_CONTEXT_LENGTH, "%s.context_length" },
{ LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" },
{ LLM_KV_EMBEDDING_LENGTH_OUT, "%s.embedding_length_out" },
{ LLM_KV_EMBEDDING_LENGTH_PER_LAYER, "%s.embedding_length_per_layer_input" },
{ LLM_KV_FEATURES_LENGTH, "%s.features_length" },
{ LLM_KV_BLOCK_COUNT, "%s.block_count" },
{ LLM_KV_LEADING_DENSE_BLOCK_COUNT, "%s.leading_dense_block_count" },
@@ -238,6 +240,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, "%s.attention.indexer.head_count" },
{ LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, "%s.attention.indexer.key_length" },
{ LLM_KV_ATTENTION_INDEXER_TOP_K, "%s.attention.indexer.top_k" },
{ LLM_KV_ATTENTION_SHARED_KV_LAYERS, "%s.attention.shared_kv_layers" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_DIMENSION_COUNT_SWA, "%s.rope.dimension_count_swa" },
@@ -364,6 +367,9 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_ATTN_GATE, "blk.%d.attn_gate" },
{ LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
{ LLM_TENSOR_FFN_POST_NORM_1, "blk.%d.post_ffw_norm_1" },
{ LLM_TENSOR_FFN_POST_NORM_2, "blk.%d.post_ffw_norm_2" },
{ LLM_TENSOR_FFN_PRE_NORM_2, "blk.%d.pre_ffw_norm_2" },
{ LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
{ LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
@@ -373,6 +379,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
{ LLM_TENSOR_LAYER_OUT_SCALE, "blk.%d.layer_output_scale" },
{ LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
{ LLM_TENSOR_POS_EMBD, "position_embd" },
{ LLM_TENSOR_FFN_ACT, "blk.%d.ffn.act" },
@@ -1342,6 +1349,38 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_LAUREL_R,
LLM_TENSOR_LAUREL_POST_NORM,
};
case LLM_ARCH_GEMMA4:
return {
LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_POST_NORM,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_GATE_UP_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_POST_NORM,
LLM_TENSOR_FFN_POST_NORM_1,
LLM_TENSOR_FFN_POST_NORM_2,
LLM_TENSOR_FFN_PRE_NORM_2,
LLM_TENSOR_LAYER_OUT_SCALE,
LLM_TENSOR_PER_LAYER_TOKEN_EMBD,
LLM_TENSOR_PER_LAYER_MODEL_PROJ,
LLM_TENSOR_PER_LAYER_PROJ_NORM,
LLM_TENSOR_PER_LAYER_INP_GATE,
LLM_TENSOR_PER_LAYER_PROJ,
LLM_TENSOR_PER_LAYER_POST_NORM,
};
case LLM_ARCH_GEMMA_EMBEDDING:
return {
LLM_TENSOR_TOKEN_EMBD,
@@ -2654,11 +2693,15 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_ATTN_OUT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_PRE_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_POST_NORM_1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_POST_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_NORM_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_Q_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_K_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_LAYER_OUT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_LAYER_OUT_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_Q_A_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_KV_A_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_SUB_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
src/llama-arch.h
+7
View File
@@ -60,6 +60,7 @@ enum llm_arch {
LLM_ARCH_GEMMA2,
LLM_ARCH_GEMMA3,
LLM_ARCH_GEMMA3N,
LLM_ARCH_GEMMA4,
LLM_ARCH_GEMMA_EMBEDDING,
LLM_ARCH_STARCODER2,
LLM_ARCH_MAMBA,
@@ -169,6 +170,7 @@ enum llm_kv {
LLM_KV_CONTEXT_LENGTH,
LLM_KV_EMBEDDING_LENGTH,
LLM_KV_EMBEDDING_LENGTH_OUT,
LLM_KV_EMBEDDING_LENGTH_PER_LAYER,
LLM_KV_FEATURES_LENGTH,
LLM_KV_BLOCK_COUNT,
LLM_KV_LEADING_DENSE_BLOCK_COUNT,
@@ -242,6 +244,7 @@ enum llm_kv {
LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
LLM_KV_ATTENTION_INDEXER_TOP_K,
LLM_KV_ATTENTION_SHARED_KV_LAYERS,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_DIMENSION_COUNT_SWA,
@@ -369,6 +372,9 @@ enum llm_tensor {
LLM_TENSOR_FFN_GATE_INP_SHEXP,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_POST_NORM,
LLM_TENSOR_FFN_POST_NORM_1,
LLM_TENSOR_FFN_POST_NORM_2,
LLM_TENSOR_FFN_PRE_NORM_2,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
@@ -393,6 +399,7 @@ enum llm_tensor {
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_LAYER_OUT_NORM,
LLM_TENSOR_LAYER_OUT_SCALE,
LLM_TENSOR_POST_ATTN_NORM,
LLM_TENSOR_POST_MLP_NORM,
LLM_TENSOR_PER_LAYER_TOKEN_EMBD, // gemma3n
src/llama-chat.cpp
+22 -4
View File
@@ -60,7 +60,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "exaone4", LLM_CHAT_TEMPLATE_EXAONE_4 },
{ "exaone-moe", LLM_CHAT_TEMPLATE_EXAONE_MOE },
{ "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD },
{ "granite", LLM_CHAT_TEMPLATE_GRANITE },
{ "granite", LLM_CHAT_TEMPLATE_GRANITE_3_X },
{ "granite-4.0", LLM_CHAT_TEMPLATE_GRANITE_4_0 },
{ "gigachat", LLM_CHAT_TEMPLATE_GIGACHAT },
{ "megrez", LLM_CHAT_TEMPLATE_MEGREZ },
{ "yandex", LLM_CHAT_TEMPLATE_YANDEX },
@@ -191,7 +192,10 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
} else if (tmpl_contains("rwkv-world") || tmpl_contains("{{- 'User: ' + message['content']|trim + '\\n\\n' -}}")) {
return LLM_CHAT_TEMPLATE_RWKV_WORLD;
} else if (tmpl_contains("<|start_of_role|>")) {
return LLM_CHAT_TEMPLATE_GRANITE;
if (tmpl_contains("<tool_call>") || tmpl_contains("<tools>")) {
return LLM_CHAT_TEMPLATE_GRANITE_4_0;
}
return LLM_CHAT_TEMPLATE_GRANITE_3_X;
} else if (tmpl_contains("message['role'] + additional_special_tokens[0] + message['content'] + additional_special_tokens[1]")) {
return LLM_CHAT_TEMPLATE_GIGACHAT;
} else if (tmpl_contains("<|role_start|>")) {
@@ -617,8 +621,8 @@ int32_t llm_chat_apply_template(
ss << "Assistant: " << trim(chat[i]->content) << "\n\n";
}
}
} else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) {
// IBM Granite template
} else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE_3_X) {
// IBM Granite 3.x template
for (const auto & message : chat) {
std::string role(message->role);
ss << "<|start_of_role|>" << role << "<|end_of_role|>";
@@ -630,6 +634,20 @@ int32_t llm_chat_apply_template(
if (add_ass) {
ss << "<|start_of_role|>assistant<|end_of_role|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE_4_0) {
// IBM Granite 4.0 template
for (const auto & message : chat) {
std::string role(message->role);
if (role == "assistant_tool_call") {
ss << "<|start_of_role|>assistant<|end_of_role|><|tool_call|>";
} else {
ss << "<|start_of_role|>" << role << "<|end_of_role|>";
}
ss << message->content << "<|end_of_text|>\n";
}
if (add_ass) {
ss << "<|start_of_role|>assistant<|end_of_role|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {
// GigaChat template
bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";
src/llama-chat.h
+2 -1
View File
@@ -39,7 +39,8 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_EXAONE_4,
LLM_CHAT_TEMPLATE_EXAONE_MOE,
LLM_CHAT_TEMPLATE_RWKV_WORLD,
LLM_CHAT_TEMPLATE_GRANITE,
LLM_CHAT_TEMPLATE_GRANITE_3_X,
LLM_CHAT_TEMPLATE_GRANITE_4_0,
LLM_CHAT_TEMPLATE_GIGACHAT,
LLM_CHAT_TEMPLATE_MEGREZ,
LLM_CHAT_TEMPLATE_YANDEX,
src/llama-ext.h
+47 -3
View File
@@ -1,8 +1,8 @@
#pragma once
#include "llama-context.h"
#include "ggml.h"
#include "stdint.h"
#include "llama.h"
#include <cstdint>
// Reserve a new compute graph. It is valid until the next call to llama_graph_reserve.
LLAMA_API struct ggml_cgraph * llama_graph_reserve(
@@ -10,3 +10,47 @@ LLAMA_API struct ggml_cgraph * llama_graph_reserve(
uint32_t n_tokens,
uint32_t n_seqs,
uint32_t n_outputs);
// Get the default ggml_type for a given ftype.
LLAMA_API ggml_type llama_ftype_get_default_type(llama_ftype ftype);
// Quantization state.
struct quantize_state_impl;
LLAMA_API quantize_state_impl * llama_quant_init(
const llama_model * model,
const llama_model_quantize_params * params);
LLAMA_API void llama_quant_free(quantize_state_impl * qs);
// Descriptor for constructing a mock model for quantization testing.
struct llama_quant_model_desc {
const char * architecture;
uint32_t n_embd;
uint32_t n_ff;
uint32_t n_layer;
uint32_t n_head;
uint32_t n_head_kv;
uint32_t n_expert;
uint32_t n_embd_head_k;
uint32_t n_embd_head_v;
};
// Create a mock model from a metadata descriptor (for testing).
// The returned model must be freed with llama_model_free().
LLAMA_API llama_model * llama_quant_model_from_metadata(const llama_quant_model_desc * desc);
// Returns true if this tensor should be quantized (based on name, dims, params).
LLAMA_API bool llama_quant_tensor_allows_quantization(
const quantize_state_impl * qs,
const ggml_tensor * tensor);
// Compute quantization type assignments for a list of tensors.
// All tensors should be quantizable (use llama_quant_tensor_allows_quantization to filter).
// result_types: caller-allocated array of n_tensors elements, filled with assigned types.
LLAMA_API void llama_quant_compute_types(
quantize_state_impl * qs,
llama_ftype ftype,
ggml_tensor ** tensors,
ggml_type * result_types,
size_t n_tensors);
src/llama-hparams.h
+3
View File
@@ -209,6 +209,9 @@ struct llama_hparams {
// qwen3vl deepstack
uint32_t n_deepstack_layers = 0;
// gemma4 per-layer embedding
uint32_t n_embd_per_layer = 0;
// needed by encoder-decoder models (e.g. T5, FLAN-T5)
// ref: https://github.com/ggml-org/llama.cpp/pull/8141
llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
src/llama-kv-cache-iswa.cpp
+2 -1
View File
@@ -70,8 +70,9 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa);
// note: the SWA cache is never quantized because it is relatively small
kv_swa = std::make_unique<llama_kv_cache>(
model, type_k, type_v,
model, GGML_TYPE_F16, GGML_TYPE_F16,
v_trans, offload, unified, size_swa, n_seq_max, n_pad,
hparams.n_swa, hparams.swa_type, filter_swa, reuse);
}
src/llama-model.cpp
+126 -1
View File
@@ -66,6 +66,7 @@
#include "models/gemma2-iswa.cpp"
#include "models/gemma3.cpp"
#include "models/gemma3n-iswa.cpp"
#include "models/gemma4-iswa.cpp"
#include "models/glm4-moe.cpp"
#include "models/glm4.cpp"
#include "models/gpt2.cpp"
@@ -1375,6 +1376,31 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_GEMMA4:
{
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
uint32_t n_kv_shared_layers = 0;
ml.get_key(LLM_KV_ATTENTION_SHARED_KV_LAYERS, n_kv_shared_layers, false);
hparams.n_layer_kv_from_start = hparams.n_layer - (int32_t)n_kv_shared_layers;
hparams.f_attention_scale = 1.0f; // Gemma4 uses self.scaling = 1.0 (no pre-attn scaling)
ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_EMBEDDING_LENGTH_PER_LAYER, hparams.n_embd_per_layer);
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_SWA, hparams.n_embd_head_k_swa);
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_SWA, hparams.n_embd_head_v_swa);
switch (hparams.n_layer) {
case 35: type = LLM_TYPE_E2B; break;
case 42: type = LLM_TYPE_E4B; break; // to confirm: E4B or E5B?
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_GEMMA_EMBEDDING:
{
hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC;
@@ -4387,6 +4413,100 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.laurel_post_norm = create_tensor(tn(LLM_TENSOR_LAUREL_POST_NORM, "weight", i), {n_embd}, 0);
}
} break;
case LLM_ARCH_GEMMA4:
{
const uint32_t n_embd_per_layer = hparams.n_embd_per_layer;
const int64_t n_ff_exp = hparams.n_ff_exp;
if (n_embd_head_k != n_embd_head_v) {
throw std::runtime_error("Gemma 4 requires n_embd_head_k == n_embd_head_v");
}
if (hparams.n_embd_head_k_swa != hparams.n_embd_head_v_swa) {
throw std::runtime_error("Gemma 4 requires n_embd_head_k_swa == n_embd_head_v_swa");
}
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
if (n_embd_per_layer > 0) {
tok_embd_per_layer = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"), {n_embd_per_layer * n_layer, n_vocab}, 0);
per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight"), {n_embd, n_embd_per_layer * n_layer}, 0);
per_layer_proj_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM, "weight"), {n_embd_per_layer}, 0);
}
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
int rope_freqs_flag = 0;
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
const int64_t n_head = hparams.n_head(i);
const int64_t n_embd_head = hparams.n_embd_head_k(i);
const int64_t n_embd_k = hparams.n_embd_k_gqa(i);
const int64_t n_embd_v = hparams.n_embd_v_gqa(i);
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
// note: use_alternative_attention (v_proj is optional, if it's not present, use k_proj)
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head * n_head}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v}, TENSOR_NOT_REQUIRED);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head * n_head, n_embd}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head}, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
layer.out_scale = create_tensor(tn(LLM_TENSOR_LAYER_OUT_SCALE, "weight", i), {1u}, TENSOR_NOT_REQUIRED);
if (!hparams.is_swa(i)) {
// full_attention layers use rope_freqs for proportional rope
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_embd_head/2}, rope_freqs_flag);
rope_freqs_flag = TENSOR_DUPLICATED;
}
// handle use_double_wide_mlp
int64_t n_ff_cur = hparams.n_ff(i);
// for expert layers, we use normal FFN as shared expert (same as python code)
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff_cur}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff_cur}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff_cur, n_embd}, 0);
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
// MoE router
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
bool has_expert = layer.ffn_gate_inp != nullptr;
// norm
if (has_expert) {
layer.ffn_gate_inp_s = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "scale", i), {n_embd}, 0);
layer.ffn_pre_norm_2 = create_tensor(tn(LLM_TENSOR_FFN_PRE_NORM_2, "weight", i), {n_embd}, 0);
layer.ffn_post_norm_1 = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM_1, "weight", i), {n_embd}, 0);
layer.ffn_post_norm_2 = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM_2, "weight", i), {n_embd}, 0);
// MoE FFN
layer.ffn_gate_up_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_UP_EXPS, "weight", i), {n_embd, n_ff_exp * 2, n_expert}, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
// per-expert scale will be loaded as down_exps_s at the end of the current switch case
}
// per-layer embeddings
if (n_embd_per_layer > 0) {
layer.per_layer_inp_gate = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE, "weight", i), {n_embd, n_embd_per_layer}, 0);
layer.per_layer_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ, "weight", i), {n_embd_per_layer, n_embd}, 0);
layer.per_layer_post_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0);
}
}
} break;
case LLM_ARCH_STARCODER2:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -8391,7 +8511,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
} else {
llama_memory_i::layer_reuse_cb reuse = nullptr;
if (arch == LLM_ARCH_GEMMA3N) {
if (arch == LLM_ARCH_GEMMA3N || arch == LLM_ARCH_GEMMA4) {
reuse = [&](int32_t il) {
if (il >= (int32_t) hparams.n_layer_kv_from_start) {
return (int32_t) hparams.n_layer_kv_from_start - (hparams.is_swa(il) ? 2 : 1);
@@ -8644,6 +8764,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_gemma3n_iswa>(*this, params);
} break;
case LLM_ARCH_GEMMA4:
{
llm = std::make_unique<llm_build_gemma4_iswa>(*this, params);
} break;
case LLM_ARCH_GEMMA_EMBEDDING:
{
llm = std::make_unique<llm_build_gemma_embedding>(*this, params);
@@ -9164,6 +9288,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_GEMMA2:
case LLM_ARCH_GEMMA3:
case LLM_ARCH_GEMMA3N:
case LLM_ARCH_GEMMA4:
case LLM_ARCH_GEMMA_EMBEDDING:
case LLM_ARCH_STARCODER2:
case LLM_ARCH_OPENELM:
src/llama-model.h
+7
View File
@@ -270,6 +270,9 @@ struct llama_layer {
struct ggml_tensor * ffn_norm = nullptr;
struct ggml_tensor * ffn_norm_b = nullptr;
struct ggml_tensor * ffn_post_norm = nullptr;
struct ggml_tensor * ffn_post_norm_1 = nullptr; // gemma4
struct ggml_tensor * ffn_post_norm_2 = nullptr; // gemma4
struct ggml_tensor * ffn_pre_norm_2 = nullptr; // gemma4
struct ggml_tensor * layer_out_norm = nullptr;
struct ggml_tensor * layer_out_norm_b = nullptr;
struct ggml_tensor * ffn_norm_exps = nullptr;
@@ -285,6 +288,7 @@ struct llama_layer {
// ff MoE
struct ggml_tensor * ffn_gate_inp = nullptr;
struct ggml_tensor * ffn_gate_inp_s = nullptr; // gemma4
struct ggml_tensor * ffn_gate_exps = nullptr;
struct ggml_tensor * ffn_down_exps = nullptr;
struct ggml_tensor * ffn_up_exps = nullptr;
@@ -483,6 +487,9 @@ struct llama_layer {
struct ggml_tensor * indexer_attn_k = nullptr;
struct ggml_tensor * indexer_attn_q_b = nullptr; // note: for lora a/b, not bias
// gemma4 layer output scale
struct ggml_tensor * out_scale = nullptr;
struct llama_layer_posnet posnet;
struct llama_layer_convnext convnext;
src/llama-quant.cpp
+125 -31
View File
@@ -1,11 +1,11 @@
#include "llama.h"
#include "llama-impl.h"
#include "llama-model.h"
#include "llama-model-loader.h"
#include "llama-ext.h"
#include <algorithm>
#include <cmath>
#include <cstring>
#include <string>
#include <cinttypes>
#include <fstream>
#include <mutex>
@@ -197,6 +197,7 @@ struct quantize_state_impl {
// per-tensor metadata, computed in the preliminary loop and used in the main loop
struct tensor_metadata {
std::string name;
ggml_type target_type;
tensor_category category;
std::string remapped_imatrix_name;
@@ -790,7 +791,7 @@ static bool tensor_requires_imatrix(const char * tensor_name, const ggml_type ds
// given a file type, get the default tensor type
//
static ggml_type llama_ftype_get_default_type(llama_ftype ftype) {
ggml_type llama_ftype_get_default_type(llama_ftype ftype) {
switch (ftype) {
case LLAMA_FTYPE_MOSTLY_Q4_0: return GGML_TYPE_Q4_0;
case LLAMA_FTYPE_MOSTLY_Q4_1: return GGML_TYPE_Q4_1;
@@ -829,16 +830,32 @@ static ggml_type llama_ftype_get_default_type(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ3_S:
case LLAMA_FTYPE_MOSTLY_IQ3_M: return GGML_TYPE_IQ3_S;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
default: return GGML_TYPE_COUNT;
}
}
static void init_quantize_state_counters(quantize_state_impl & qs, std::vector<tensor_metadata> & metadata) {
for (auto & tm : metadata) {
tensor_category cat = tensor_get_category(tm.name);
tm.category = cat;
if (category_is_attn_v(cat)) {
++qs.n_attention_wv;
}
if (cat == tensor_category::OUTPUT) {
qs.has_tied_embeddings = false;
}
}
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)qs.model.hparams.n_layer;
}
//
// main quantization driver
//
static void llama_model_quantize_impl(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
ggml_type default_type;
llama_ftype ftype = params->ftype;
int nthread = params->nthread;
@@ -847,7 +864,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
nthread = std::thread::hardware_concurrency();
}
default_type = llama_ftype_get_default_type(ftype);
ggml_type default_type = llama_ftype_get_default_type(ftype);
if (default_type == GGML_TYPE_COUNT) {
throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
// mmap consistently increases speed on Linux, and also increases speed on Windows with
// hot cache. It may cause a slowdown on macOS, possibly related to free memory.
@@ -966,6 +986,15 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
});
}
// compute tensor metadata once and cache it
std::vector<tensor_metadata> metadata(tensors.size());
for (size_t i = 0; i < tensors.size(); ++i) {
metadata[i].name = ggml_get_name(tensors[i]->tensor);
}
// initialize quantization state counters and metadata categories
init_quantize_state_counters(qs, metadata);
int idx = 0;
uint16_t n_split = 1;
@@ -978,25 +1007,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
std::vector<gguf_context_ptr> ctx_outs(n_split);
ctx_outs[0] = std::move(ctx_out);
// compute tensor metadata once and cache it
std::vector<tensor_metadata> metadata(tensors.size());
// initialize quantization state before preliminary loop (counters for use_more_bits)
{
for (size_t i = 0; i < tensors.size(); ++i) {
const auto cat = tensor_get_category(tensors[i]->tensor->name);
if (category_is_attn_v(cat)) {
++qs.n_attention_wv;
}
if (cat == tensor_category::OUTPUT) {
qs.has_tied_embeddings = false;
}
metadata[i].category = cat; // save and re-use the category while we're at it
}
// these also need to be set to n_layer by default
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)qs.model.hparams.n_layer;
}
// flag for --dry-run
bool will_require_imatrix = false;
@@ -1007,7 +1017,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
for (size_t i = 0; i < tensors.size(); ++i) {
const auto * it = tensors[i];
const struct ggml_tensor * tensor = it->tensor;
const std::string name = ggml_get_name(tensor);
uint16_t i_split = params->keep_split ? it->idx : 0;
if (!ctx_outs[i_split]) {
@@ -1036,7 +1045,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
" - offending tensor: %s\n"
" - target type: %s\n"
"============================================================================\n\n",
name.c_str(), ggml_type_name(metadata[i].target_type));
metadata[i].name.c_str(), ggml_type_name(metadata[i].target_type));
throw std::runtime_error("this quantization requires an imatrix!");
}
}
@@ -1109,7 +1118,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
new_ofstream(weight.idx);
}
const std::string name = ggml_get_name(tensor);
const size_t tensor_size = ggml_nbytes(tensor);
if (!params->dry_run) {
@@ -1240,9 +1248,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
total_size_new += new_size;
// update the gguf meta data as we go
gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);
gguf_set_tensor_type(ctx_outs[cur_split].get(), metadata[i].name.c_str(), new_type);
GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), metadata[i].name.c_str())) == new_size);
gguf_set_tensor_data(ctx_outs[cur_split].get(), metadata[i].name.c_str(), new_data);
// write tensor data + padding
fout.write((const char *) new_data, new_size);
@@ -1307,3 +1315,89 @@ uint32_t llama_model_quantize(
return 0;
}
//
// Helper functions for external tools exposed in llama-ext.h
//
quantize_state_impl * llama_quant_init(
const llama_model * model,
const llama_model_quantize_params * params) {
return new quantize_state_impl(*model, params);
}
void llama_quant_free(quantize_state_impl * qs) {
delete qs;
}
llama_model * llama_quant_model_from_metadata(const llama_quant_model_desc * desc) {
struct llama_model_params mparams = llama_model_default_params();
auto * model = new llama_model(mparams);
model->arch = llm_arch_from_string(desc->architecture);
// infer llm_type: only LLM_TYPE_70B matters for quantization logic
if (model->arch == LLM_ARCH_LLAMA && desc->n_layer == 80 && desc->n_head != desc->n_head_kv) {
model->type = LLM_TYPE_70B;
}
model->hparams.n_embd = desc->n_embd;
model->hparams.n_embd_head_k_full = desc->n_embd_head_k;
model->hparams.n_embd_head_v_full = desc->n_embd_head_v;
model->hparams.n_layer = desc->n_layer;
model->hparams.n_expert = desc->n_expert;
for (uint32_t i = 0; i < desc->n_layer; i++) {
model->hparams.n_head_arr[i] = desc->n_head;
model->hparams.n_head_kv_arr[i] = desc->n_head_kv;
model->hparams.n_ff_arr[i] = desc->n_ff;
}
return model;
}
bool llama_quant_tensor_allows_quantization(
const quantize_state_impl * qs,
const ggml_tensor * tensor) {
return tensor_allows_quantization(qs->params, qs->model.arch, tensor);
}
void llama_quant_compute_types(
quantize_state_impl * qs,
llama_ftype ftype,
ggml_tensor ** tensors,
ggml_type * result_types,
size_t n_tensors) {
// reset per-computation state
qs->n_attention_wv = 0;
qs->n_ffn_down = 0;
qs->n_ffn_gate = 0;
qs->n_ffn_up = 0;
qs->i_attention_wv = 0;
qs->i_ffn_down = 0;
qs->i_ffn_gate = 0;
qs->i_ffn_up = 0;
qs->n_fallback = 0;
qs->has_imatrix = false;
qs->has_tied_embeddings = true;
// build metadata from tensor names
std::vector<tensor_metadata> metadata(n_tensors);
for (size_t i = 0; i < n_tensors; i++) {
metadata[i].name = ggml_get_name(tensors[i]);
}
// initialize counters and categories
init_quantize_state_counters(*qs, metadata);
// use a local copy of params with the requested ftype
llama_model_quantize_params local_params = *qs->params;
local_params.ftype = ftype;
ggml_type default_type = llama_ftype_get_default_type(ftype);
// compute types
for (size_t i = 0; i < n_tensors; i++) {
result_types[i] = llama_tensor_get_type(*qs, &local_params, tensors[i], default_type, metadata[i]);
}
}
src/llama-vocab.cpp
+13
View File
@@ -2099,6 +2099,18 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
special_sep_id = LLAMA_TOKEN_NULL;
special_pad_id = 3; // <|plamo:pad|>
special_mask_id = LLAMA_TOKEN_NULL;
} else if (tokenizer_model == "gemma4") {
type = LLAMA_VOCAB_TYPE_SPM;
// default special tokens (to be read from GGUF)
special_bos_id = LLAMA_TOKEN_NULL;
special_eos_id = LLAMA_TOKEN_NULL;
special_unk_id = LLAMA_TOKEN_NULL;
special_sep_id = LLAMA_TOKEN_NULL;
special_pad_id = LLAMA_TOKEN_NULL;
special_mask_id = LLAMA_TOKEN_NULL;
tokenizer_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
} else {
throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str()));
}
@@ -2729,6 +2741,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "[EOS]" // Kimi-K2
|| t.first == "<|end_of_text|>"
|| t.first == "<end_of_utterance>" // smoldocling
|| t.first == "<turn|>" // gemma4
|| t.first == "<｜end▁of▁sentence｜>" // deepseek-ocr
) {
special_eog_ids.insert(t.second);
src/models/gemma4-iswa.cpp
+311
View File
@@ -0,0 +1,311 @@
#include "models.h"
llm_build_gemma4_iswa::llm_build_gemma4_iswa(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params),
model(model),
n_embd_per_layer(model.hparams.n_embd_per_layer) {
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
// important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
cb(inpL, "inp_scaled", -1);
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
// TODO: is causal == true correct? might need some changes
auto * inp_attn = build_attn_inp_kv_iswa();
// inp_per_layer shape: [n_embd_per_layer, n_tokens, n_layer]
ggml_tensor * inp_per_layer = nullptr;
if (model.tok_embd_per_layer) {
inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
}
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
const int64_t n_embd_head = hparams.n_embd_head_k(il);
GGML_ASSERT(n_embd_head == hparams.n_embd_head_v(il));
const int64_t n_head = hparams.n_head(il);
const int64_t n_head_kv = hparams.n_head_kv(il);
const float freq_base_l = model.get_rope_freq_base(cparams, il);
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
const int n_rot_l = hparams.n_rot(il);
// norm
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_tensor * freq_factors = nullptr;
if (!hparams.is_swa(il)) {
// full_attention layers use rope_freqs for proportional rope
freq_factors = model.layers[il].rope_freqs;
}
// Q projection (shared for both non-KV and KV layers)
// this is to mirror Gemma4Attention in pytorch code
ggml_tensor * Qcur;
{
Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, freq_factors, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur_pos", il);
}
// self-attention
if (hparams.has_kv(il)) {
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = model.layers[il].wv
? build_lora_mm(model.layers[il].wv, cur)
: Kcur; // if v_proj is not present, use Kcur as Vcur
cb(Vcur, "Vcur", il);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);
cb(Kcur, "Kcur_normed", il);
cb(Vcur, "Vcur_normed", il);
Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, freq_factors, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur_pos", il);
cur = build_attn(inp_attn, model.layers[il].wo,
nullptr, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
hparams.f_attention_scale, il);
} else {
// reuse KV cache of earlier layers
cur = build_attn(inp_attn,
model.layers[il].wo, nullptr,
Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
}
// TODO @ngxson : strip unused token right after the last KV layer to speed up prompt processing
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
}
cur = build_norm(cur,
model.layers[il].attn_post_norm, nullptr,
LLM_NORM_RMS, il);
cb(cur, "attn_post_norm", il);
ggml_tensor * attn_out = ggml_add(ctx0, cur, inpL);
cb(attn_out, "attn_out", il);
// feed-forward network
const bool is_moe_layer = model.layers[il].ffn_gate_inp != nullptr;
if (is_moe_layer) {
// MLP (shared exp)
ggml_tensor * cur_mlp = build_norm(attn_out,
model.layers[il].ffn_norm, nullptr,
LLM_NORM_RMS, il);
cb(cur_mlp, "ffn_norm_1", il);
cur_mlp = build_ffn(cur_mlp,
model.layers[il].ffn_up, nullptr, nullptr,
model.layers[il].ffn_gate, nullptr, nullptr,
model.layers[il].ffn_down, nullptr, nullptr,
nullptr,
LLM_FFN_GELU, LLM_FFN_PAR, il);
cur_mlp = build_norm(cur_mlp,
model.layers[il].ffn_post_norm_1, nullptr,
LLM_NORM_RMS, il);
cb(cur_mlp, "ffn_mlp", il);
// Expert FFN
ggml_tensor * cur_moe = build_norm(attn_out,
model.layers[il].ffn_pre_norm_2, nullptr,
LLM_NORM_RMS, il);
cb(cur_moe, "ffn_norm_2", il);
// custom MoE logits calculation (router operates on attn_out, not cur)
ggml_tensor * tmp = ggml_rms_norm(ctx0, attn_out, hparams.f_norm_rms_eps);
tmp = ggml_scale(ctx0, tmp, 1.0f / sqrtf((float) n_embd));
tmp = ggml_mul(ctx0, tmp, model.layers[il].ffn_gate_inp_s);
ggml_tensor * logits = build_lora_mm(model.layers[il].ffn_gate_inp, tmp); // [n_expert, n_tokens]
cb(logits, "ffn_moe_logits", il);
cur_moe = build_moe_ffn(cur_moe,
nullptr, // gate_inp
nullptr, // up_exps
nullptr, // gate_exps
model.layers[il].ffn_down_exps,
nullptr, // exp_probs_b (not used for gemma4)
n_expert, n_expert_used,
LLM_FFN_GELU, true,
1.0f,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il, logits,
model.layers[il].ffn_gate_up_exps,
nullptr, // up_exps_s
nullptr, // gate_exps_s
model.layers[il].ffn_down_exps_s);
cur_moe = build_norm(cur_moe,
model.layers[il].ffn_post_norm_2, nullptr,
LLM_NORM_RMS, il);
cb(cur_moe, "ffn_moe", il);
cur = ggml_add(ctx0, cur_mlp, cur_moe);
cb(cur, "ffn_moe_combined", il);
} else {
cur = build_norm(attn_out,
model.layers[il].ffn_norm, nullptr,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
cur = build_ffn(cur,
model.layers[il].ffn_up, nullptr, nullptr,
model.layers[il].ffn_gate, nullptr, nullptr,
model.layers[il].ffn_down, nullptr, nullptr,
nullptr,
LLM_FFN_GELU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
}
cur = build_norm(cur,
model.layers[il].ffn_post_norm, nullptr,
LLM_NORM_RMS, -1);
cb(cur, "ffn_post_norm", il);
// residual connection
cur = ggml_add(ctx0, cur, attn_out);
// per-layer embedding
if (inp_per_layer) {
ggml_tensor * pe_in = cur;
cb(cur, "pe_in", il);
cur = build_lora_mm(model.layers[il].per_layer_inp_gate, cur); // [n_embd_per_layer, n_tokens]
cur = ggml_gelu(ctx0, cur);
ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il); // [n_embd_per_layer, n_tokens]
// TODO @ngxson : improve this
if (il == n_layer - 1 && inp_out_ids) {
inp_this_layer = ggml_get_rows(ctx0, inp_this_layer, inp_out_ids);
}
cur = ggml_mul(ctx0, cur, inp_this_layer);
cur = build_lora_mm(model.layers[il].per_layer_proj, cur); // [n_embd, n_tokens]
cur = build_norm(cur, model.layers[il].per_layer_post_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "per_layer_embd_out", il);
// residual connection
cur = ggml_add(ctx0, pe_in, cur);
}
// layer_scalar
if (model.layers[il].out_scale) {
cur = ggml_mul(ctx0, cur, model.layers[il].out_scale);
cb(cur, "out_scaled", il);
}
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = build_norm(cur,
model.output_norm, nullptr,
LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// lm_head
cur = build_lora_mm(model.output, cur);
if (hparams.f_final_logit_softcapping) {
cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
cur = ggml_tanh(ctx0, cur);
cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
}
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
// get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
ggml_tensor * llm_build_gemma4_iswa::view_2d_slice(ggml_tensor * x, int idx) {
GGML_ASSERT(idx < (int) x->ne[2]);
return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], ggml_row_size(x->type, x->ne[0]),
idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
}
// equivalent to get_per_layer_inputs() in python code
// output shape: [n_embd_per_layer, n_layer, n_tokens]
ggml_tensor * llm_build_gemma4_iswa::get_per_layer_inputs() {
auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
ggml_tensor * inp_per_layer;
if (ubatch.token) {
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
ggml_set_input(inp->tokens);
res->t_inp_tokens = inp->tokens;
inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_per_layer, n_layer, n_tokens);
inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_per_layer));
cb(inp_per_layer, "inp_per_layer_selected", -1);
res->add_input(std::move(inp));
} else {
// Vision embedding path: use padding token (ID=0) embedding
// TODO: verify if this is the correct behavior in transformers implementation
const int64_t embd_size = model.tok_embd_per_layer->ne[0]; // n_embd_per_layer * n_layer
// Extract and dequantize padding token embedding (row 0)
ggml_tensor * padding = ggml_view_1d(ctx0, model.tok_embd_per_layer, embd_size, 0);
inp_per_layer = ggml_cast(ctx0, padding, GGML_TYPE_F32);
// Reshape to [n_embd_per_layer, n_layer, 1]
inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_per_layer, n_layer, 1);
cb(inp_per_layer, "inp_per_layer_vision", -1);
}
return inp_per_layer;
}
// equivalent to project_per_layer_inputs() in python code
// this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
// inputs_embeds shape: [n_embd, n_tokens]
// inp_per_layer shape: [n_embd_per_layer, n_layer, n_tokens] (from get_per_layer_inputs)
// output shape: [n_embd_per_layer, n_tokens, n_layer]
ggml_tensor * llm_build_gemma4_iswa::project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) {
const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);
const float per_layer_input_scale = 1.0f / sqrtf(2.0f);
ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds);
per_layer_proj = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale);
per_layer_proj = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_per_layer, n_layer, n_tokens);
per_layer_proj = build_norm(per_layer_proj, model.per_layer_proj_norm, nullptr, LLM_NORM_RMS,
-1); // [n_embd_per_layer, n_layer, n_tokens]
cb(per_layer_proj, "per_layer_proj", -1);
inp_per_layer = ggml_add(ctx0, per_layer_proj, inp_per_layer);
inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
cb(inp_per_layer, "inp_per_layer", -1);
// permute to shape: [n_embd_per_layer, n_tokens, n_layer]
inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));
return inp_per_layer;
}
src/models/models.h
+11
View File
@@ -266,6 +266,17 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
ggml_tensor * altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il);
};
struct llm_build_gemma4_iswa : public llm_graph_context {
const llama_model & model;
const int64_t n_embd_per_layer;
llm_build_gemma4_iswa(const llama_model & model, const llm_graph_params & params);
ggml_tensor * view_2d_slice(ggml_tensor * x, int idx);
ggml_tensor * get_per_layer_inputs();
ggml_tensor * project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer);
};
struct llm_build_gemma_embedding : public llm_graph_context {
llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params);
};