vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	README.md
#	scripts/sync-ggml.last
This commit is contained in:
Concedo 2024-02-22 15:49:19 +08:00
commit be696e0da9
10 changed files with 1017 additions and 532 deletions
Download patch file Download diff file Expand all files Collapse all files

13
examples/llava/README.md
View file

@ -63,13 +63,12 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
```console
git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
```
2) Backup your pth/safetensor model files as llava-surgery modifies them
3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
```console
python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
```
- you will find a llava.projector and a llava.clip file in your model directory
4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
```console
mkdir vit
cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
@ -77,18 +76,18 @@ cp ../llava-v1.6-vicuna-7b/llava.projector vit/
curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
```
5) Create the visual gguf model:
4) Create the visual gguf model:
```console
python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
```
- This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
6) Then convert the model to gguf format:
5) Then convert the model to gguf format:
```console
python ./convert.py ../llava-v1.6-vicuna-7b/
python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
```
7) And finally we can run the llava-cli using the 1.6 model version:
6) And finally we can run the llava-cli using the 1.6 model version:
```console
./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
```

12
examples/llava/llava-surgery-v2.py
View file

@ -65,9 +65,7 @@ def clean_vision_tower_from_checkpoint(checkpoint_path):
for name in clip_tensors:
del checkpoint[name]
# Save the updated checkpoint
checkpoint_path = checkpoint_path
save_model(checkpoint, checkpoint_path, file_type)
return True
return False
@ -152,16 +150,6 @@ for name in first_mm_tensors:
if len(projector) > 0:
save_model(projector, f"{args.model}/llava.projector", 'pytorch')
for name in mm_tensors:
del last_checkpoint[name]
for name in first_mm_tensors:
del first_checkpoint[name]
if len(mm_tensors) > 0:
save_model(last_checkpoint, projector_checkpoint_path, file_type)
if len(first_mm_tensors) > 0:
save_model(first_checkpoint, newline_checkpoint_path, file_type)
print("Done!")
print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

12
examples/main/main.cpp
View file

@ -335,6 +335,8 @@ int main(int argc, char ** argv) {
// number of tokens to keep when resetting context
if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
params.n_keep = (int)embd_inp.size();
} else {
params.n_keep += add_bos; // always keep the BOS token
}
// prefix & suffix for instruct mode
@ -384,8 +386,8 @@ int main(int argc, char ** argv) {
}
}
if (params.n_keep > 0) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__);
if (params.n_keep > add_bos) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
@ -541,14 +543,14 @@ int main(int argc, char ** argv) {
break;
}
const int n_left = n_past - params.n_keep - 1;
const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2;
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1);
llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
n_past -= n_discard;

3
examples/server/README.md
View file

@ -41,6 +41,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
- `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
- `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
- `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
## Build
@ -140,6 +141,8 @@ node index.js
- 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available.
- 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slot are currently available.
If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
*Options:*

135
examples/server/server.cpp
View file

@ -1395,6 +1395,46 @@ struct llama_server_context
case TASK_TYPE_NEXT_RESPONSE: {
// do nothing
} break;
case TASK_TYPE_SLOTS_DATA: {
json slots_data = json::array();
int n_idle_slots = 0;
int n_processing_slots = 0;
for (llama_client_slot &slot: slots) {
if (slot.available()) {
n_idle_slots++;
} else {
n_processing_slots++;
}
json slot_data = get_formated_generation(slot);
slot_data["id"] = slot.id;
slot_data["task_id"] = slot.task_id;
slot_data["state"] = slot.state;
slot_data["prompt"] = slot.prompt;
slot_data["next_token"] = {
{"has_next_token", slot.has_next_token},
{"n_remain", slot.n_remaining},
{"num_tokens_predicted", slot.n_decoded},
{"stopped_eos", slot.stopped_eos},
{"stopped_word", slot.stopped_word},
{"stopped_limit", slot.stopped_limit},
{"stopping_word", slot.stopping_word},
};
slots_data.push_back(slot_data);
}
LOG_TEE("task %i - slots data: idle=%i processing=%i\n", task.id, n_idle_slots, n_processing_slots);
task_result res;
res.id = task.id;
res.multitask_id = task.multitask_id;
res.stop = true;
res.error = false;
res.result_json = {
{ "idle", n_idle_slots },
{ "processing", n_processing_slots },
{ "slots", slots_data }
};
queue_results.send(res);
} break;
}
}
@ -1448,14 +1488,15 @@ struct llama_server_context
if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
{
// Shift context
const int n_left = system_tokens.size() + slot.n_past - slot.params.n_keep - 1;
const int n_keep = slot.params.n_keep + add_bos_token;
const int n_left = system_tokens.size() + slot.n_past - n_keep;
const int n_discard = n_left / 2;
LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard);
llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1);
llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, system_tokens.size() + slot.n_past, -n_discard);
LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, n_keep, n_left, n_discard);
llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++)
for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++)
{
slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
}
@ -1468,7 +1509,7 @@ struct llama_server_context
LOG_VERBOSE("context shift", {
{ "n_ctx", n_ctx },
{ "n_keep", params.n_keep },
{ "n_keep", n_keep },
{ "n_left", n_left },
});
}
@ -2558,34 +2599,38 @@ int main(int argc, char **argv)
server_state current_state = state.load();
switch(current_state) {
case SERVER_STATE_READY: {
int available_slots = 0;
int processing_slots = 0;
for (llama_client_slot &slot: llama.slots) {
if (slot.available()) {
available_slots++;
} else {
processing_slots++;
}
// request slots data using task queue
task_server task;
task.id = llama.queue_tasks.get_new_id();
task.type = TASK_TYPE_SLOTS_DATA;
task.target_id = -1;
llama.queue_results.add_waiting_task_id(task.id);
llama.queue_tasks.post(task);
// get the result
task_result result = llama.queue_results.recv(task.id);
llama.queue_results.remove_waiting_task_id(task.id);
int n_idle_slots = result.result_json["idle"];
int n_processing_slots = result.result_json["processing"];
json health = {
{"status", "ok"},
{"slots_idle", n_idle_slots},
{"slots_processing", n_processing_slots}};
res.status = 200; // HTTP OK
if (sparams.slots_endpoint && req.has_param("include_slots")) {
health["slots"] = result.result_json["slots"];
}
if (available_slots > 0) {
json health = {
{"status", "ok"},
{"slots_idle", available_slots},
{"slots_processing", processing_slots}};
res.set_content(health.dump(), "application/json");
res.status = 200; // HTTP OK
} else {
json health = {
{"status", "no slot available"},
{"slots_idle", available_slots},
{"slots_processing", processing_slots}};
res.set_content(health.dump(), "application/json");
if (n_idle_slots == 0) {
health["status"] = "no slot available";
if (req.has_param("fail_on_no_slot")) {
res.status = 503; // HTTP Service Unavailable
} else {
res.status = 200; // HTTP OK
}
}
res.set_content(health.dump(), "application/json");
break;
}
case SERVER_STATE_LOADING_MODEL:
@ -2601,26 +2646,20 @@ int main(int argc, char **argv)
if (sparams.slots_endpoint) {
svr.Get("/slots", [&](const httplib::Request&, httplib::Response& res) {
json slots;
for (llama_client_slot & slot : llama.slots) {
json slot_data = llama.get_formated_generation(slot);
slot_data["id"] = slot.id;
slot_data["task_id"] = slot.task_id;
slot_data["state"] = slot.state;
slot_data["prompt"] = slot.prompt;
slot_data["next_token"] = {
{"has_next_token", slot.has_next_token},
{"n_remain", slot.n_remaining},
{"num_tokens_predicted", slot.n_decoded},
{"stopped_eos", slot.stopped_eos},
{"stopped_word", slot.stopped_word},
{"stopped_limit", slot.stopped_limit},
{"stopping_word", slot.stopping_word},
};
// request slots data using task queue
task_server task;
task.id = llama.queue_tasks.get_new_id();
task.type = TASK_TYPE_SLOTS_DATA;
task.target_id = -1;
slots.push_back(slot_data);
}
res.set_content(slots.dump(), "application/json");
llama.queue_results.add_waiting_task_id(task.id);
llama.queue_tasks.post(task);
// get the result
task_result result = llama.queue_results.recv(task.id);
llama.queue_results.remove_waiting_task_id(task.id);
res.set_content(result.result_json["slots"].dump(), "application/json");
res.status = 200; // HTTP OK
});
}

3
examples/server/utils.hpp
View file

@ -49,7 +49,8 @@ enum server_state {
enum task_type {
TASK_TYPE_COMPLETION,
TASK_TYPE_CANCEL,
TASK_TYPE_NEXT_RESPONSE
TASK_TYPE_NEXT_RESPONSE,
TASK_TYPE_SLOTS_DATA
};
struct task_server {

1150
ggml.c
View file

File diff suppressed because it is too large Load diff

15
gguf-py/gguf/constants.py
View file

@ -111,6 +111,7 @@ class MODEL_ARCH(IntEnum):
ORION = auto()
INTERNLM2 = auto()
MINICPM = auto()
GEMMA = auto()
class MODEL_TENSOR(IntEnum):
@ -167,6 +168,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.ORION: "orion",
MODEL_ARCH.INTERNLM2: "internlm2",
MODEL_ARCH.MINICPM: "minicpm",
MODEL_ARCH.GEMMA: "gemma",
}
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@ -511,6 +513,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.GEMMA: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_NORM,
],
# TODO
}

204
llama.cpp
View file

@ -232,6 +232,7 @@ enum llm_arch {
LLM_ARCH_ORION,
LLM_ARCH_INTERNLM2,
LLM_ARCH_MINICPM,
LLM_ARCH_GEMMA,
LLM_ARCH_UNKNOWN,
};
@ -258,6 +259,7 @@ static std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_ORION, "orion" },
{ LLM_ARCH_INTERNLM2, "internlm2" },
{ LLM_ARCH_MINICPM, "minicpm" },
{ LLM_ARCH_GEMMA, "gemma" },
};
enum llm_kv {
@ -784,6 +786,22 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
},
},
{
LLM_ARCH_GEMMA,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ 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_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_UNKNOWN,
{
@ -2806,13 +2824,7 @@ struct llama_model_loader {
std::vector<no_init<uint8_t>> read_buf;
for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) {
struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i));
if (!cur) {
// some tensors may be allocated in a different context
continue;
}
for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) {
if (progress_callback) {
if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) {
return false;
@ -3289,6 +3301,16 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_GEMMA:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 18: model.type = e_model::MODEL_2B; break;
case 28: model.type = e_model::MODEL_7B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
default: (void)0;
}
@ -3766,7 +3788,7 @@ static bool llm_load_tensors(
}
// create one context per buffer type
size_t ctx_size = ggml_tensor_overhead()*ml.n_tensors;
size_t ctx_size = ggml_tensor_overhead()*(ml.n_tensors + 1); // +1 for models where tok_embd is duplicated as output
std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
for (auto & it : buft_layer_count) {
struct ggml_init_params params = {
@ -3904,6 +3926,7 @@ static bool llm_load_tensors(
} else {
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
ml.n_created--; // artificial tensor
ml.size_data += ggml_nbytes(model.output);
}
}
@ -4432,6 +4455,40 @@ static bool llm_load_tensors(
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
}
} break;
case LLM_ARCH_GEMMA:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
// output
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading
ml.n_created--; // artificial tensor
ml.size_data += ggml_nbytes(model.output);
const int64_t n_ff = hparams.n_ff;
const int64_t n_embd_head_k = hparams.n_embd_head_k;
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
for (uint32_t i = 0; i < n_layer; ++i) {
ggml_context * ctx_layer = ctx_for_layer(i);
ggml_context * ctx_split = ctx_for_layer_split(i);
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa});
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa});
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
@ -7438,6 +7495,113 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_gemma() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_embd_head_k = hparams.n_embd_head_k;
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
cb(inpL, "inp_embd", -1);
inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
cb(inpL, "inp_scaled", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
cb(inp_pos, "inp_pos", -1);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
cb(KQ_mask, "KQ_mask", -1);
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
// norm
cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos,
n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
cb(Qcur, "Qcur_scaled", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos,
n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il);
cb(cur, "kqv_out", il);
}
struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
cb(sa_out, "sa_out", il);
cur = llm_build_norm(ctx0, sa_out, hparams,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il);
// feed-forward network
{
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, NULL,
model.layers[il].ffn_gate, NULL,
model.layers[il].ffn_down, NULL,
NULL,
LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
cb(cur, "ffn_out", il);
}
cur = ggml_add(ctx0, cur, sa_out);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = llm_build_norm(ctx0, cur, hparams,
model.output_norm, NULL,
LLM_NORM_RMS, cb, -1);
cb(cur, "result_norm", -1);
// lm_head
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
};
static struct ggml_cgraph * llama_build_graph(
@ -7546,6 +7710,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_minicpm();
} break;
case LLM_ARCH_GEMMA:
{
result = llm.build_gemma();
} break;
default:
GGML_ASSERT(false);
}
@ -12317,18 +12485,19 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
data_ctx->write(&kv_used, sizeof(kv_used));
if (kv_buf_size) {
const size_t elt_size = ggml_element_size(kv_self.k_l[0]);
std::vector<uint8_t> tmp_buf;
for (int il = 0; il < (int) n_layer; ++il) {
tmp_buf.resize(elt_size*n_embd_k_gqa*kv_head);
size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
tmp_buf.resize(k_size);
ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size());
data_ctx->write(tmp_buf.data(), tmp_buf.size());
// v is not contiguous, copy row by row
tmp_buf.resize(elt_size*kv_head);
size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head);
size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx);
tmp_buf.resize(v_row_size);
for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*elt_size*n_ctx, tmp_buf.size());
ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*v_row_stride, tmp_buf.size());
data_ctx->write(tmp_buf.data(), tmp_buf.size());
}
}
@ -12430,17 +12599,16 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
if (kv_buf_size) {
GGML_ASSERT(kv_self.total_size() == kv_buf_size);
const size_t elt_size = ggml_element_size(kv_self.k_l[0]);
for (int il = 0; il < (int) n_layer; ++il) {
size_t k_size = elt_size*n_embd_k_gqa*kv_head;
size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size);
inp += k_size;
// v is not contiguous, copy row by row
size_t v_row_size = elt_size*kv_head;
size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head);
size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx);
for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*elt_size*n_ctx, v_row_size);
ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*v_row_stride, v_row_size);
inp += v_row_size;
}
}

2
llama.h
View file

@ -710,7 +710,7 @@ extern "C" {
/// Apply chat template. Inspired by hf apply_chat_template() on python.
/// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
/// NOTE: This function only support some known jinja templates. It is not a jinja parser.
/// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
/// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
/// @param chat Pointer to a list of multiple llama_chat_message
/// @param n_msg Number of llama_chat_message in this chat