Skip to content

GitLab

Unverified Commit 4987f13d authored by Gabe Goodhart's avatar Gabe Goodhart Committed by GitHub
Browse files

Llama cpp bump (df1b612): granite docling / mamba2 optimizations / multimodal... 


Llama cpp bump (df1b612): granite docling / mamba2 optimizations / multimodal encoding fixes (#12552)

* feat: Bump llama.cpp to df1b612

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* fix(mtmd): Correctly encode text chunks during mtmd tokenization

There can be text chunks that appear interspersed with the image embeddings
that contain template delimiter tokens for some models. These need to be
correctly translated to text tokens.

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* tests: Use MtmdChunk in image_test

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* style: Fix unnecessary conversion linting

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* fix(ggml): Revert changes to ggml_hip.cpp

These changes were done largely by our code assistant and are likely wrong

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* fix: Revert changes in mem_nvml.cpp

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* feat: Update sync point to 1deee0

This brings in several more optimization commits and model support for
EmbeddingGemma

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* feat: Update patches for 1deee0

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* feat: sync for bump to 1deee0

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* fix: Bad patch updates with errant `+`

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* feat: Bump llama.cpp/ggml to 7049736

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

* fix: format-patches after latest bump

Branch: LlamaCPPBump-GraniteDocling
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>

---------
Signed-off-by: default avatarGabe Goodhart <ghart@us.ibm.com>
parent e638f2ac
Hide whitespace changes
Inline Side-by-side
Showing with 517 additions and 68 deletions
Makefile.sync
View file @ 4987f13d
UPSTREAM=https://github.com/ggml-org/llama.cpp.git
WORKDIR=llama/vendor
FETCH_HEAD=364a7a6d4a786e98947c8a90430ea581213c0ba9
FETCH_HEAD=7049736b2dd9011bf819e298b844ebbc4b5afdc9
.PHONY: help
help:
......
llama/build-info.cpp
View file @ 4987f13d
int LLAMA_BUILD_NUMBER = 0;
char const *LLAMA_COMMIT = "364a7a6d4a786e98947c8a90430ea581213c0ba9";
char const *LLAMA_COMMIT = "7049736b2dd9011bf819e298b844ebbc4b5afdc9";
char const *LLAMA_COMPILER = "";
char const *LLAMA_BUILD_TARGET = "";
llama/llama.cpp/common/common.cpp
View file @ 4987f13d
......@@ -1133,6 +1133,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
mparams.use_mlock = params.use_mlock;
mparams.check_tensors = params.check_tensors;
mparams.use_extra_bufts = !params.no_extra_bufts;
mparams.no_host = params.no_host;
if (params.kv_overrides.empty()) {
mparams.kv_overrides = NULL;
......
llama/llama.cpp/common/common.h
View file @ 4987f13d
......@@ -378,7 +378,7 @@ struct common_params {
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
bool cont_batching = true; // insert new sequences for decoding on-the-fly
bool no_perf = false; // disable performance metrics
bool ctx_shift = false; // context shift on infinite text generation
bool ctx_shift = false; // context shift on infinite text generation
bool swa_full = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
bool kv_unified = false; // enable unified KV cache
......@@ -392,6 +392,7 @@ struct common_params {
bool check_tensors = false; // validate tensor data
bool no_op_offload = false; // globally disable offload host tensor operations to device
bool no_extra_bufts = false; // disable extra buffer types (used for weight repacking)
bool no_host = false; // bypass host buffer allowing extra buffers to be used
bool single_turn = false; // single turn chat conversation
......@@ -424,7 +425,8 @@ struct common_params {
int32_t timeout_write = timeout_read; // http write timeout in seconds
int32_t n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
int32_t n_cache_reuse = 0; // min chunk size to reuse from the cache via KV shifting
int32_t n_swa_checkpoints = 3; // max number of SWA checkpoints per slot
int32_t n_ctx_checkpoints = 8; // max number of context checkpoints per slot
int32_t cache_ram_mib = 8192; // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
std::string hostname = "127.0.0.1";
std::string public_path = ""; // NOLINT
......@@ -432,7 +434,7 @@ struct common_params {
std::string chat_template = ""; // NOLINT
bool use_jinja = false; // NOLINT
bool enable_chat_template = true;
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_AUTO;
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
int reasoning_budget = -1;
bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
......
llama/llama.cpp/include/llama.h
View file @ 4987f13d
......@@ -296,6 +296,7 @@ extern "C" {
bool use_mlock; // force system to keep model in RAM
bool check_tensors; // validate model tensor data
bool use_extra_bufts; // use extra buffer types (used for weight repacking)
bool no_host; // bypass host buffer allowing extra buffers to be used
};
// NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
......@@ -543,6 +544,9 @@ extern "C" {
// Returns true if the model is recurrent (like Mamba, RWKV, etc.)
LLAMA_API bool llama_model_is_recurrent(const struct llama_model * model);
// Returns true if the model is hybrid (like Jamba, Granite, etc.)
LLAMA_API bool llama_model_is_hybrid(const struct llama_model * model);
// Returns true if the model is diffusion-based (like LLaDA, Dream, etc.)
LLAMA_API bool llama_model_is_diffusion(const struct llama_model * model);
......@@ -791,8 +795,12 @@ extern "C" {
size_t n_token_capacity,
size_t * n_token_count_out);
// for backwards-compat
#define LLAMA_STATE_SEQ_FLAGS_SWA_ONLY 1
// work only with partial states, such as SWA KV cache or recurrent cache (e.g. Mamba)
#define LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY 1
typedef uint32_t llama_state_seq_flags;
LLAMA_API size_t llama_state_seq_get_size_ext(
......
llama/llama.cpp/src/llama-arch.cpp
View file @ 4987f13d
......@@ -94,12 +94,14 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_SMOLLM3, "smollm3" },
{ LLM_ARCH_OPENAI_MOE, "gpt-oss" },
{ LLM_ARCH_LFM2, "lfm2" },
{ LLM_ARCH_LFM2MOE, "lfm2moe" },
{ LLM_ARCH_DREAM, "dream" },
{ LLM_ARCH_SMALLTHINKER, "smallthinker" },
{ LLM_ARCH_LLADA, "llada" },
{ LLM_ARCH_LLADA_MOE, "llada-moe" },
{ LLM_ARCH_SEED_OSS, "seed_oss" },
{ LLM_ARCH_GROVEMOE, "grovemoe" },
{ LLM_ARCH_APERTUS, "apertus" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
......@@ -219,6 +221,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
{ LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
// sentence-transformers dense modules feature dims
{ LLM_KV_DENSE_2_FEAT_IN, "%s.dense_2_feat_in" },
{ LLM_KV_DENSE_2_FEAT_OUT, "%s.dense_2_feat_out" },
{ LLM_KV_DENSE_3_FEAT_IN, "%s.dense_3_feat_in" },
{ LLM_KV_DENSE_3_FEAT_OUT, "%s.dense_3_feat_out" },
{ LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" },
{ LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" },
......@@ -258,6 +265,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ADAPTER_LORA_PROMPT_PREFIX, "adapter.lora.prompt_prefix" },
{ LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS, "adapter.alora.invocation_tokens" },
{ LLM_KV_XIELU_ALPHA_N, "xielu.alpha_n" },
{ LLM_KV_XIELU_ALPHA_P, "xielu.alpha_p" },
{ LLM_KV_XIELU_BETA, "xielu.beta" },
{ LLM_KV_XIELU_EPS, "xielu.eps" },
// deprecated
{ LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" },
{ LLM_KV_TOKENIZER_SUFFIX_ID, "tokenizer.ggml.suffix_token_id" },
......@@ -1066,6 +1078,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_DENSE_2_OUT, "dense_2" },
{ LLM_TENSOR_DENSE_3_OUT, "dense_3" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
......@@ -2118,6 +2132,32 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_OUTPUT, "output" },
}
},
{
LLM_ARCH_LFM2MOE,
{
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_SHORTCONV_CONV, "blk.%d.shortconv.conv" },
{ LLM_TENSOR_SHORTCONV_INPROJ, "blk.%d.shortconv.in_proj" },
{ LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
{ LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
}
},
{
LLM_ARCH_SMALLTHINKER,
{
......@@ -2139,6 +2179,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }
},
},
{
LLM_ARCH_APERTUS,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_DREAM,
{
......@@ -2249,6 +2308,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_OUTPUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_CLS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_CLS_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_DENSE_2_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
{LLM_TENSOR_DENSE_3_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
{LLM_TENSOR_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
{LLM_TENSOR_DEC_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
{LLM_TENSOR_ENC_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
......@@ -2489,6 +2550,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
case LLM_ARCH_PLAMO2:
case LLM_ARCH_GRANITE_HYBRID:
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
case LLM_ARCH_NEMOTRON_H:
return true;
default:
......
llama/llama.cpp/src/llama-arch.h
View file @ 4987f13d
......@@ -98,12 +98,14 @@ enum llm_arch {
LLM_ARCH_SMOLLM3,
LLM_ARCH_OPENAI_MOE,
LLM_ARCH_LFM2,
LLM_ARCH_LFM2MOE,
LLM_ARCH_DREAM,
LLM_ARCH_SMALLTHINKER,
LLM_ARCH_LLADA,
LLM_ARCH_LLADA_MOE,
LLM_ARCH_SEED_OSS,
LLM_ARCH_GROVEMOE,
LLM_ARCH_APERTUS,
LLM_ARCH_UNKNOWN,
};
......@@ -262,10 +264,21 @@ enum llm_kv {
LLM_KV_SHORTCONV_L_CACHE,
LLM_KV_XIELU_ALPHA_N,
LLM_KV_XIELU_ALPHA_P,
LLM_KV_XIELU_BETA,
LLM_KV_XIELU_EPS,
// deprecated:
LLM_KV_TOKENIZER_PREFIX_ID,
LLM_KV_TOKENIZER_SUFFIX_ID,
LLM_KV_TOKENIZER_MIDDLE_ID,
// sentence-transformers dense layers in and out features
LLM_KV_DENSE_2_FEAT_IN,
LLM_KV_DENSE_2_FEAT_OUT,
LLM_KV_DENSE_3_FEAT_IN,
LLM_KV_DENSE_3_FEAT_OUT,
};
enum llm_tensor {
......@@ -273,6 +286,8 @@ enum llm_tensor {
LLM_TENSOR_TOKEN_EMBD_NORM,
LLM_TENSOR_TOKEN_TYPES,
LLM_TENSOR_POS_EMBD,
LLM_TENSOR_DENSE_2_OUT,
LLM_TENSOR_DENSE_3_OUT,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_ROPE_FREQS,
......
llama/llama.cpp/src/llama-chat.cpp
View file @ 4987f13d
......@@ -590,7 +590,7 @@ int32_t llm_chat_apply_template(
ss << message->content << "<|end_of_text|>\n";
}
if (add_ass) {
ss << "<|start_of_role|>assistant<|end_of_role|>\n";
ss << "<|start_of_role|>assistant<|end_of_role|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {
// GigaChat template
......
llama/llama.cpp/src/llama-context.cpp
View file @ 4987f13d
......@@ -2345,6 +2345,12 @@ llama_context * llama_init_from_model(
return nullptr;
}
if (params.pooling_type != model->hparams.pooling_type) {
//user-specified pooling-type is different from the model default
LLAMA_LOG_WARN("%s: model default pooling_type is [%d], but [%d] was specified\n", __func__,
model->hparams.pooling_type, params.pooling_type);
}
try {
auto * ctx = new llama_context(*model, params);
return ctx;
......
llama/llama.cpp/src/llama-graph.cpp
View file @ 4987f13d
......@@ -1853,6 +1853,23 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
}
void llm_graph_context::build_dense_out(
ggml_tensor * dense_2,
ggml_tensor * dense_3) const {
if (!cparams.embeddings || dense_2 == nullptr || dense_3 == nullptr) {
return;
}
ggml_tensor * cur = res->t_embd_pooled != nullptr ? res->t_embd_pooled : res->t_embd;
GGML_ASSERT(cur != nullptr && "missing t_embd_pooled/t_embd");
cur = ggml_mul_mat(ctx0, dense_2, cur);
cur = ggml_mul_mat(ctx0, dense_3, cur);
cb(cur, "result_embd_pooled", -1);
res->t_embd_pooled = cur;
ggml_build_forward_expand(gf, cur);
}
void llm_graph_context::build_pooling(
ggml_tensor * cls,
ggml_tensor * cls_b,
......
llama/llama.cpp/src/llama-graph.h
View file @ 4987f13d
......@@ -814,6 +814,14 @@ struct llm_graph_context {
ggml_tensor * cls_b,
ggml_tensor * cls_out,
ggml_tensor * cls_out_b) const;
//
// dense (out)
//
void build_dense_out(
ggml_tensor * dense_2,
ggml_tensor * dense_3) const;
};
// TODO: better name
......
llama/llama.cpp/src/llama-hparams.cpp
View file @ 4987f13d
......@@ -140,7 +140,11 @@ uint32_t llama_hparams::n_embd_s() const {
}
bool llama_hparams::is_recurrent(uint32_t il) const {
return recurrent_layer_arr[il];
if (il < n_layer) {
return recurrent_layer_arr[il];
}
GGML_ABORT("%s: il (%u) out of bounds (n_layer: %u)\n", __func__, il, n_layer);
}
uint32_t llama_hparams::n_pos_per_embd() const {
......
llama/llama.cpp/src/llama-hparams.h
View file @ 4987f13d
......@@ -42,7 +42,7 @@ struct llama_hparams {
uint32_t n_embd;
uint32_t n_embd_features = 0;
uint32_t n_layer;
int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
uint32_t n_rot;
uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
......@@ -171,6 +171,18 @@ struct llama_hparams {
uint32_t laurel_rank = 64;
uint32_t n_embd_altup = 256;
// needed for sentence-transformers dense layers
uint32_t dense_2_feat_in = 0; // in_features of the 2_Dense
uint32_t dense_2_feat_out = 0; // out_features of the 2_Dense
uint32_t dense_3_feat_in = 0; // in_features of the 3_Dense
uint32_t dense_3_feat_out = 0; // out_features of the 3_Dense
// xIELU
std::array<float, LLAMA_MAX_LAYERS> xielu_alpha_n;
std::array<float, LLAMA_MAX_LAYERS> xielu_alpha_p;
std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
// needed by encoder-decoder models (e.g. T5, FLAN-T5)
// ref: https://github.com/ggerganov/llama.cpp/pull/8141
llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
......
llama/llama.cpp/src/llama-kv-cache-iswa.cpp
View file @ 4987f13d
......@@ -220,7 +220,7 @@ bool llama_kv_cache_iswa::get_can_shift() const {
}
void llama_kv_cache_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {
kv_base->state_write(io, seq_id, flags);
}
......@@ -228,7 +228,7 @@ void llama_kv_cache_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id
}
void llama_kv_cache_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {
kv_base->state_read(io, seq_id, flags);
}
......
llama/llama.cpp/src/llama-kv-cache.cpp
View file @ 4987f13d
......@@ -123,11 +123,8 @@ llama_kv_cache::llama_kv_cache(
throw std::runtime_error("failed to create ggml context for kv cache");
}
ggml_tensor * k;
ggml_tensor * v;
k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream);
v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream);
ggml_tensor * k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream);
ggml_tensor * v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream);
ggml_format_name(k, "cache_k_l%d", il);
ggml_format_name(v, "cache_v_l%d", il);
......
llama/llama.cpp/src/llama-memory-hybrid.cpp
View file @ 4987f13d
......@@ -73,7 +73,9 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
// if all tokens are output, split by sequence
ubatch = balloc.split_seq(n_ubatch);
} else {
ubatch = balloc.split_equal(n_ubatch, false);
// TODO: non-sequential equal split can be done if using unified KV cache
// for simplicity, we always use sequential equal split for now
ubatch = balloc.split_equal(n_ubatch, true);
}
if (ubatch.n_tokens == 0) {
......@@ -175,17 +177,17 @@ std::map<ggml_backend_buffer_type_t, size_t> llama_memory_hybrid::memory_breakdo
}
void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
GGML_UNUSED(flags);
mem_attn->state_write(io, seq_id);
mem_recr->state_write(io, seq_id);
if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {
mem_attn->state_write(io, seq_id, flags);
}
mem_recr->state_write(io, seq_id, flags);
}
void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
GGML_UNUSED(flags);
mem_attn->state_read(io, seq_id);
mem_recr->state_read(io, seq_id);
if ((flags & LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY) == 0) {
mem_attn->state_read(io, seq_id, flags);
}
mem_recr->state_read(io, seq_id, flags);
}
llama_kv_cache * llama_memory_hybrid::get_mem_attn() const {
......
llama/llama.cpp/src/llama-memory-recurrent.cpp
View file @ 4987f13d
......@@ -136,6 +136,7 @@ void llama_memory_recurrent::clear(bool data) {
}
bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
//printf("[DEBUG] calling llama_memory_recurrent::seq_rm` with `seq_id=%d, p0=%d, p1=%d`\n", seq_id, p0, p1);
uint32_t new_head = size;
if (p0 < 0) {
......@@ -156,7 +157,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
if (tail_id >= 0) {
const auto & cell = cells[tail_id];
// partial intersection is invalid
if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) {
if ((0 < p0 && p0 < cell.pos) || (0 < p1 && p1 <= cell.pos)) {
//printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
return false;
}
// invalidate tails which will be cleared
......@@ -167,6 +169,7 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
} else {
// seq_id is negative, then the range should include everything or nothing
if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits<llama_pos>::max())) {
//printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: `seq_id` is negative, so returning false\n");
return false;
}
}
......@@ -379,7 +382,9 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
// if all tokens are output, split by sequence
ubatch = balloc.split_seq(n_ubatch);
} else {
ubatch = balloc.split_equal(n_ubatch, false);
// TODO: non-sequential equal split can be done if using unified KV cache
// for simplicity, we always use sequential equal split for now
ubatch = balloc.split_equal(n_ubatch, true);
}
if (ubatch.n_tokens == 0) {
......@@ -856,9 +861,12 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) {
if (dest_seq_id != -1) {
// single sequence
seq_rm(dest_seq_id, -1, -1);
if (cell_count == 0) {
return true;
}
llama_batch_allocr balloc(hparams.n_pos_per_embd());
llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1);
......
llama/llama.cpp/src/llama-model-loader.cpp
View file @ 4987f13d
......@@ -465,6 +465,7 @@ namespace GGUFMeta {
// TODO: this is not very clever - figure out something better
template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
template bool llama_model_loader::get_key_or_arr<std::array<float, 512>>(enum llm_kv kid, std::array<float, 512> & result, uint32_t n, bool required);
template bool llama_model_loader::get_key_or_arr<uint32_t>(const std::string & key, std::array<uint32_t, 512> & result, uint32_t n, bool required);
llama_model_loader::llama_model_loader(
......
llama/llama.cpp/src/llama-model.cpp
View file @ 4987f13d
......@@ -114,6 +114,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_17B_16E: return "17Bx16E (Scout)";
case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)";
case LLM_TYPE_A13B: return "A13B";
case LLM_TYPE_8B_A1B: return "8B.A1B";
case LLM_TYPE_21B_A3B: return "21B.A3B";
case LLM_TYPE_30B_A3B: return "30B.A3B";
case LLM_TYPE_106B_A12B: return "106B.A12B";
......@@ -310,7 +311,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara
}
 
// CPU: ACCEL -> GPU host -> CPU extra -> CPU
static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices, bool use_extra_bufts) {
static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices, bool use_extra_bufts, bool no_host) {
buft_list_t buft_list;
 
// add ACCEL buffer types
......@@ -331,11 +332,13 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
// generally, this will be done using the first device in the list
// a better approach would be to handle this on a weight-by-weight basis using the offload_op
// function of the device to determine if it would benefit from being stored in a host buffer
for (auto * dev : devices) {
ggml_backend_buffer_type_t buft = ggml_backend_dev_host_buffer_type(dev);
if (buft) {
buft_list.emplace_back(dev, buft);
break;
if (!no_host) {
for (auto * dev : devices) {
ggml_backend_buffer_type_t buft = ggml_backend_dev_host_buffer_type(dev);
if (buft) {
buft_list.emplace_back(dev, buft);
break;
}
}
}
 
......@@ -512,9 +515,13 @@ void llama_model::load_hparams(llama_model_loader & ml) {
llm_arch_is_recurrent(ml.get_arch()));
 
std::fill(hparams.rope_sections.begin(), hparams.rope_sections.end(), 0);
std::fill(hparams.swa_layers.begin(), hparams.swa_layers.end(), 0);
 
std::fill(hparams.xielu_alpha_n.begin(), hparams.xielu_alpha_n.end(), 0.0f);
std::fill(hparams.xielu_alpha_p.begin(), hparams.xielu_alpha_p.end(), 0.0f);
std::fill(hparams.xielu_beta.begin(), hparams.xielu_beta.end(), 0.0f);
std::fill(hparams.xielu_eps.begin(), hparams.xielu_eps.end(), 0.0f);
ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer, false);
ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer, false);
 
......@@ -1084,7 +1091,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
}
break;
default: type = LLM_TYPE_UNKNOWN;
}
}
// Load attention parameters
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false);
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
} break;
case LLM_ARCH_GPT2:
{
......@@ -1207,12 +1218,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.set_swa_pattern(6);
 
hparams.causal_attn = false; // embeddings do not use causal attention
hparams.rope_freq_base_train_swa = 10000.0f;
hparams.rope_freq_base_train_swa = 10000.0f;
hparams.rope_freq_scale_train_swa = 1.0f;
 
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
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_POOLING_TYPE, hparams.pooling_type);
ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type);
//applied only if model converted with --sentence-transformers-dense-modules
ml.get_key(LLM_KV_DENSE_2_FEAT_IN, hparams.dense_2_feat_in, false);
ml.get_key(LLM_KV_DENSE_2_FEAT_OUT, hparams.dense_2_feat_out, false);
ml.get_key(LLM_KV_DENSE_3_FEAT_IN, hparams.dense_3_feat_in, false);
ml.get_key(LLM_KV_DENSE_3_FEAT_OUT, hparams.dense_3_feat_out, false);
GGML_ASSERT((hparams.dense_2_feat_in == 0 || hparams.dense_2_feat_in == hparams.n_embd) && "dense_2_feat_in must be equal to n_embd");
GGML_ASSERT((hparams.dense_3_feat_out == 0 || hparams.dense_3_feat_out == hparams.n_embd) && "dense_3_feat_out must be equal to n_embd");
 
switch (hparams.n_layer) {
case 24: type = LLM_TYPE_0_3B; break;
......@@ -2000,13 +2020,28 @@ void llama_model::load_hparams(llama_model_loader & ml) {
for (uint32_t il = 0; il < hparams.n_layer; ++il) {
hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
}
hparams.n_layer_dense_lead = hparams.n_layer;
switch (hparams.n_ff()) {
case 4608: type = LLM_TYPE_350M; break;
case 6912: type = LLM_TYPE_700M; break;
case 8192: type = LLM_TYPE_1_2B; break;
case 10752: type = LLM_TYPE_2_6B; break;
default: type = LLM_TYPE_UNKNOWN;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_LFM2MOE:
{
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);
ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
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;
}
type = LLM_TYPE_8B_A1B;
} break;
case LLM_ARCH_SMALLTHINKER:
{
......@@ -2044,6 +2079,19 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_APERTUS:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key_or_arr(LLM_KV_XIELU_ALPHA_N, hparams.xielu_alpha_n, hparams.n_layer);
ml.get_key_or_arr(LLM_KV_XIELU_ALPHA_P, hparams.xielu_alpha_p, hparams.n_layer);
ml.get_key_or_arr(LLM_KV_XIELU_BETA, hparams.xielu_beta, hparams.n_layer);
ml.get_key_or_arr(LLM_KV_XIELU_EPS, hparams.xielu_eps, hparams.n_layer);
switch (hparams.n_layer) {
case 32: type = LLM_TYPE_8B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
default: throw std::runtime_error("unsupported model architecture");
}
 
......@@ -2077,7 +2125,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, ml.use_mmap ? "true" : "false");
 
// build a list of buffer types for the CPU and GPU devices
pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts);
pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts, params.no_host);
for (auto * dev : devices) {
buft_list_t buft_list = make_gpu_buft_list(dev, split_mode, tensor_split);
// add CPU buffer types as a fallback
......@@ -3407,17 +3455,17 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
} break;
case LLM_ARCH_PLAMO2:
{
// mamba parameters
const uint32_t d_conv = hparams.ssm_d_conv;
const uint32_t d_state = hparams.ssm_d_state;
const uint32_t num_heads = hparams.ssm_dt_rank;
const uint32_t intermediate_size = hparams.ssm_d_inner;
const uint32_t head_dim = intermediate_size / num_heads;
const uint32_t qk_dim = head_dim;
const uint32_t v_dim = head_dim;
const int64_t num_attention_heads = hparams.n_head();
const int64_t q_num_heads = num_attention_heads;
const int64_t dt_dim = std::max(64, int(hparams.n_embd / 16));
 
// attention parameters
const uint32_t qk_dim = hparams.n_embd_head_k;
const uint32_t v_dim = hparams.n_embd_head_v;
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
// output
......@@ -3451,6 +3499,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ssm_b_norm = create_tensor(tn(LLM_TENSOR_SSM_B_NORM, i), {d_state}, 0);
layer.ssm_c_norm = create_tensor(tn(LLM_TENSOR_SSM_C_NORM, i), {d_state}, 0);
} else {
const int64_t num_attention_heads = hparams.n_head(i);
const int64_t q_num_heads = num_attention_heads;
const int64_t num_key_value_heads = hparams.n_head_kv(i);
const int64_t k_num_heads = num_key_value_heads;
const int64_t v_num_heads = num_key_value_heads;
......@@ -3459,8 +3509,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
const int64_t v_proj_dim = v_num_heads * v_dim;
 
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, q_proj_dim + k_proj_dim + v_proj_dim}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {head_dim, num_attention_heads}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {head_dim, k_num_heads}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {qk_dim, num_attention_heads}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {qk_dim, k_num_heads}, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {q_num_heads * v_dim, n_embd}, 0);
}
 
......@@ -3660,6 +3710,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
 
// Dense linear weights
dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.dense_2_feat_out}, TENSOR_NOT_REQUIRED);
dense_3_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_3_OUT, "weight"), {hparams.dense_3_feat_in, n_embd}, TENSOR_NOT_REQUIRED);
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
 
......@@ -4840,11 +4895,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
// NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags);
// Optional tensors
layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
}
}
}
......@@ -5830,6 +5887,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
} break;
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
......@@ -5841,11 +5899,23 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
// ffn is same for transformer and conv layers
const bool is_moe_layer = i >= static_cast<int>(hparams.n_layer_dense_lead);
// ffn/moe is same for transformer and conv layers
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}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
if (is_moe_layer) {
GGML_ASSERT(n_expert && n_expert_used);
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
} else { // dense
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
}
 
// for operator_norm
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
......@@ -5950,6 +6020,48 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_up_chexps = create_tensor(tn(LLM_TENSOR_FFN_UP_CHEXPS, "weight", i), { n_embd, n_ff_chexp, n_chunk_expert}, 0);
}
} break;
case LLM_ARCH_APERTUS:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
// output
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);
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
} else {
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
}
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_gqa }, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_gqa }, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
// optional bias tensors
layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_gqa }, TENSOR_NOT_REQUIRED);
layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_gqa }, TENSOR_NOT_REQUIRED);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);
// Q and K layernorms for Apertus
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED);
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
......@@ -6284,7 +6396,7 @@ void llama_model::print_info() const {
LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm);
}
 
if (arch == LLM_ARCH_SMALLTHINKER) {
if (arch == LLM_ARCH_SMALLTHINKER || arch == LLM_ARCH_LFM2MOE) {
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
}
......@@ -7819,6 +7931,8 @@ struct llm_build_bert : public llm_graph_context {
}
 
if (model.layers[il].attn_q_norm) {
Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head*n_head, n_tokens);
Qcur = build_norm(Qcur,
model.layers[il].attn_q_norm,
model.layers[il].attn_q_norm_b,
......@@ -7828,6 +7942,8 @@ struct llm_build_bert : public llm_graph_context {
}
 
if (model.layers[il].attn_k_norm) {
Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head*n_head_kv, n_tokens);
Kcur = build_norm(Kcur,
model.layers[il].attn_k_norm,
model.layers[il].attn_k_norm_b,
......@@ -8210,6 +8326,9 @@ struct llm_build_mpt : public llm_graph_context {
 
// Q/K Layernorm
if (model.layers[il].attn_q_norm) {
Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head*n_head, n_tokens);
Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head*n_head_kv, n_tokens);
Qcur = build_norm(Qcur,
model.layers[il].attn_q_norm,
model.layers[il].attn_q_norm_b,
......@@ -16237,10 +16356,10 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
}
 
ggml_tensor * build_layer_ffn(
ggml_tensor * cur,
ggml_tensor * inpSA,
const llama_model & model,
const int il) {
ggml_tensor * cur,
ggml_tensor * inpSA,
const llama_model & model,
const int il) {
 
// For Granite architectures - scale residual
if (hparams.f_residual_scale) {
......@@ -17811,6 +17930,7 @@ private:
const int64_t n_embd_head_q = hparams.n_embd_head_k;
const int64_t n_embd_head_k = hparams.n_embd_head_k;
const int64_t n_embd_head_v = hparams.n_embd_head_v;
int32_t n_head = hparams.n_head(il);
int32_t n_head_kv = hparams.n_head_kv(il);
 
const int64_t q_offset = 0;
......@@ -18727,6 +18847,8 @@ struct llm_build_lfm2 : public llm_graph_context {
ggml_tensor * inp_out_ids = build_inp_out_ids();
 
for (int il = 0; il < n_layer; ++il) {
const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);
auto * prev_cur = cur;
cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "model.layers.{}.operator_norm", il);
......@@ -18741,7 +18863,16 @@ struct llm_build_lfm2 : public llm_graph_context {
}
 
cur = ggml_add(ctx0, prev_cur, cur);
cur = ggml_add(ctx0, cur, build_feed_forward(cur, il));
auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);
ggml_tensor * ffn_out = is_moe_layer ?
build_moe_feed_forward(ffn_norm_out, il) :
build_dense_feed_forward(ffn_norm_out, il);
cb(ffn_norm_out, "model.layers.{}.ffn_out", il);
cur = ggml_add(ctx0, cur, ffn_out);
}
 
cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1);
......@@ -18756,23 +18887,32 @@ struct llm_build_lfm2 : public llm_graph_context {
ggml_build_forward_expand(gf, cur);
}
 
ggml_tensor * build_feed_forward(ggml_tensor * cur,
int il) const {
cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "model.layers.{}.ffn_norm", il);
ggml_tensor * build_moe_feed_forward(ggml_tensor * cur,
int il) const {
return build_moe_ffn(cur,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
il);
}
 
ggml_tensor * build_dense_feed_forward(ggml_tensor * cur,
int il) const {
GGML_ASSERT(!model.layers[il].ffn_up_b);
GGML_ASSERT(!model.layers[il].ffn_gate_b);
GGML_ASSERT(!model.layers[il].ffn_down_b);
cur = build_ffn(cur,
return build_ffn(cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "model.layers.{}.feed_forward.w2", il);
return cur;
}
 
ggml_tensor * build_attn_block(ggml_tensor * cur,
......@@ -19292,6 +19432,141 @@ struct llm_build_grovemoe : public llm_graph_context {
}
};
 
struct llm_build_apertus : public llm_graph_context {
llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv();
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
ggml_tensor * inp_out_ids = build_inp_out_ids();
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);
cb(cur, "attn_norm", il);
// self-attention
{
ggml_tensor * rope_factors = model.get_rope_factors(cparams, 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 * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
cb(Kcur, "Kcur_normed", il);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur_pos", il);
cb(Kcur, "Kcur_pos", il);
cb(Vcur, "Vcur_pos", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// feed-forward network with xIELU activation
{
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, nullptr,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
// Up projection
ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
cb(up, "ffn_up", il);
float alpha_n_val = hparams.xielu_alpha_n[il];
float alpha_p_val = hparams.xielu_alpha_p[il];
float beta_val = hparams.xielu_beta[il];
float eps_val = hparams.xielu_eps[il];
// Apply xIELU activation
ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
cb(activated, "ffn_xielu", il);
// Down projection
cur = build_lora_mm(model.layers[il].ffn_down, activated);
cb(cur, "ffn_down", il);
}
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "ffn_out", 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);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
};
llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
llama_memory_i * res;
 
......@@ -19811,6 +20086,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
llm = std::make_unique<llm_build_falcon_h1>(*this, params);
} break;
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
{
llm = std::make_unique<llm_build_lfm2>(*this, params);
} break;
......@@ -19826,6 +20102,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_grovemoe>(*this, params);
} break;
case LLM_ARCH_APERTUS:
{
llm = std::make_unique<llm_build_apertus>(*this, params);
} break;
default:
GGML_ABORT("fatal error");
}
......@@ -19833,6 +20113,12 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
// add on pooling layer
llm->build_pooling(cls, cls_b, cls_out, cls_out_b);
 
// if the gguf model was converted with --sentence-transformers-dense-modules
// there will be two additional dense projection layers
// dense linear projections are applied after pooling
// TODO: move reranking logic here and generalize
llm->build_dense_out(dense_2_out_layers, dense_3_out_layers);
return llm->res->get_gf();
}
 
......@@ -19857,6 +20143,7 @@ llama_model_params llama_model_default_params() {
/*.use_mlock =*/ false,
/*.check_tensors =*/ false,
/*.use_extra_bufts =*/ true,
/*.no_host =*/ false,
};
 
return result;
......@@ -20029,10 +20316,12 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_OPENAI_MOE:
case LLM_ARCH_HUNYUAN_DENSE:
case LLM_ARCH_LFM2:
case LLM_ARCH_LFM2MOE:
case LLM_ARCH_SMALLTHINKER:
case LLM_ARCH_GLM4_MOE:
case LLM_ARCH_SEED_OSS:
case LLM_ARCH_GROVEMOE:
case LLM_ARCH_APERTUS:
return LLAMA_ROPE_TYPE_NEOX;
 
case LLM_ARCH_QWEN2VL:
......@@ -20143,6 +20432,10 @@ bool llama_model_is_recurrent(const llama_model * model) {
return llm_arch_is_recurrent(model->arch);
}
 
bool llama_model_is_hybrid(const llama_model * model) {
return llm_arch_is_hybrid(model->arch);
}
bool llama_model_is_diffusion(const llama_model * model) {
return llm_arch_is_diffusion(model->arch);
}
......
llama/llama.cpp/src/llama-model.h
View file @ 4987f13d
......@@ -108,6 +108,7 @@ enum llm_type {
LLM_TYPE_17B_16E, // llama4 Scout
LLM_TYPE_17B_128E, // llama4 Maverick
LLM_TYPE_A13B,
LLM_TYPE_8B_A1B, // lfm2moe
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_30B_A3B,
LLM_TYPE_106B_A12B, // GLM-4.5-Air
......@@ -381,6 +382,12 @@ struct llama_layer {
// openai-moe
struct ggml_tensor * attn_sinks = nullptr;
// xIELU activation parameters for Apertus
struct ggml_tensor * ffn_act_alpha_n = nullptr;
struct ggml_tensor * ffn_act_alpha_p = nullptr;
struct ggml_tensor * ffn_act_beta = nullptr;
struct ggml_tensor * ffn_act_eps = nullptr;
struct ggml_tensor * bskcn_tv = nullptr;
struct llama_layer_posnet posnet;
......@@ -434,6 +441,12 @@ struct llama_model {
std::vector<llama_layer> layers;
//Dense linear projections for SentenceTransformers models like embeddinggemma
// For Sentence Transformers models structure see
// https://sbert.net/docs/sentence_transformer/usage/custom_models.html#structure-of-sentence-transformer-models
struct ggml_tensor * dense_2_out_layers = nullptr;
struct ggml_tensor * dense_3_out_layers = nullptr;
llama_model_params params;
// gguf metadata
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment