vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge and update

Browse source
This commit is contained in:
Henri Vasserman 2023-08-09 00:36:11 +03:00
commit 28046d1e52
No known key found for this signature in database
GPG key ID: 2995FC0F58B1A986
37 changed files with 6347 additions and 2237 deletions
Download patch file Download diff file Expand all files Collapse all files

24
examples/baby-llama/baby-llama.cpp
View file

@ -8,6 +8,12 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#ifdef LLAMA_DEFAULT_RMS_EPS
static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
#else
static const float rms_norm_eps = 5e-6f;
#endif
float frand() {
return (float)rand()/(float)RAND_MAX;
}
@ -562,7 +568,7 @@ struct ggml_tensor * forward(
// norm
{
// cur shape [n_embd,N,1,1]
cur = ggml_rms_norm(ctx0, inpL);
cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
// cur = attention_norm*cur
cur = ggml_mul(ctx0,
@ -685,7 +691,7 @@ struct ggml_tensor * forward(
// norm
{
// cur shape [n_embd,N,1,1]
cur = ggml_rms_norm(ctx0, inpFF);
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
// cur = ffn_norm*cur
// cur shape [n_embd,N,1,1]
@ -729,7 +735,7 @@ struct ggml_tensor * forward(
{
// inpL shape [n_embd,N,1,1]
inpL = ggml_rms_norm(ctx0, inpL);
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
// inpL = norm*inpL
// inpL shape [n_embd,N,1,1]
@ -817,7 +823,7 @@ struct ggml_tensor * forward_batch(
// norm
{
// cur shape [n_embd,N*n_batch,1,1]
cur = ggml_rms_norm(ctx0, inpL);
cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
assert_shape_2d(cur, n_embd, N*n_batch);
// cur = attention_norm*cur
@ -981,7 +987,7 @@ struct ggml_tensor * forward_batch(
// norm
{
// cur shape [n_embd,N*n_batch,1,1]
cur = ggml_rms_norm(ctx0, inpFF);
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
assert_shape_2d(cur, n_embd, N*n_batch);
// cur = ffn_norm*cur
@ -1034,7 +1040,7 @@ struct ggml_tensor * forward_batch(
{
// inpL shape [n_embd,N*n_batch,1,1]
inpL = ggml_rms_norm(ctx0, inpL);
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
assert_shape_2d(inpL, n_embd, N*n_batch);
// inpL = norm*inpL
@ -1104,7 +1110,7 @@ struct ggml_tensor * forward_lora(
// norm
{
// cur shape [n_embd,N,1,1]
cur = ggml_rms_norm(ctx0, inpL);
cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
// cur = attention_norm*cur
cur = ggml_mul(ctx0,
@ -1251,7 +1257,7 @@ struct ggml_tensor * forward_lora(
// norm
{
// cur shape [n_embd,N,1,1]
cur = ggml_rms_norm(ctx0, inpFF);
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
// cur = ffn_norm*cur
// cur shape [n_embd,N,1,1]
@ -1295,7 +1301,7 @@ struct ggml_tensor * forward_lora(
{
// inpL shape [n_embd,N,1,1]
inpL = ggml_rms_norm(ctx0, inpL);
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
// inpL = norm*inpL
// inpL shape [n_embd,N,1,1]

42
examples/common.cpp
View file

@ -177,6 +177,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
break;
}
params.n_gqa = std::stoi(argv[i]);
} else if (arg == "-eps" || arg == "--rms-norm-eps") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rms_norm_eps = std::stof(argv[i]);
} else if (arg == "--rope-freq-base") {
if (++i >= argc) {
invalid_param = true;
@ -346,7 +352,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
#ifdef GGML_USE_CUBLAS
params.main_gpu = std::stoi(argv[i]);
#else
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
#endif
} else if (arg == "--tensor-split" || arg == "-ts") {
if (++i >= argc) {
@ -370,13 +376,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
}
}
#else
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
#endif // GGML_USE_CUBLAS
} else if (arg == "--mul-mat-q" || arg == "-mmq") {
#ifdef GGML_USE_CUBLAS
params.mul_mat_q = true;
#else
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n");
#endif // GGML_USE_CUBLAS
} else if (arg == "--low-vram" || arg == "-lv") {
#ifdef GGML_USE_CUBLAS
params.low_vram = true;
#else
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
#endif // GGML_USE_CUBLAS
} else if (arg == "--no-mmap") {
params.use_mmap = false;
@ -396,8 +408,14 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
params.antiprompt.push_back(argv[i]);
} else if (arg == "--perplexity") {
params.perplexity = true;
} else if (arg == "--perplexity-lines") {
params.perplexity_lines = true;
} else if (arg == "--hellaswag") {
params.hellaswag = true;
} else if (arg == "--hellaswag-tasks") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.hellaswag_tasks = std::stoi(argv[i]);
} else if (arg == "--ignore-eos") {
params.logit_bias[llama_token_eos()] = -INFINITY;
} else if (arg == "--no-penalize-nl") {
@ -426,6 +444,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
exit(0);
} else if (arg == "--random-prompt") {
params.random_prompt = true;
} else if (arg == "--in-prefix-bos") {
params.input_prefix_bos = true;
} else if (arg == "--in-prefix") {
if (++i >= argc) {
invalid_param = true;
@ -511,6 +531,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stdout, " not supported with --interactive or other interactive options\n");
fprintf(stdout, " --prompt-cache-ro if specified, uses the prompt cache but does not update it.\n");
fprintf(stdout, " --random-prompt start with a randomized prompt.\n");
fprintf(stdout, " --in-prefix-bos prefix BOS to user inputs, preceding the `--in-prefix` string\n");
fprintf(stdout, " --in-prefix STRING string to prefix user inputs with (default: empty)\n");
fprintf(stdout, " --in-suffix STRING string to suffix after user inputs with (default: empty)\n");
fprintf(stdout, " -f FNAME, --file FNAME\n");
@ -519,6 +540,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stdout, " -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx);
fprintf(stdout, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stdout, " -gqa N, --gqa N grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
fprintf(stdout, " -eps N, --rms-norm-eps N rms norm eps (TEMP!!! use 1e-5 for LLaMAv2) (default: %.1e)\n", params.rms_norm_eps);
fprintf(stdout, " --top-k N top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
fprintf(stdout, " --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
fprintf(stdout, " --tfs N tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
@ -549,8 +571,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stdout, " not recommended: doubles context memory required and no measurable increase in quality\n");
fprintf(stdout, " --temp N temperature (default: %.1f)\n", (double)params.temp);
fprintf(stdout, " --perplexity compute perplexity over each ctx window of the prompt\n");
fprintf(stdout, " --perplexity-lines compute perplexity over each line of the prompt\n");
fprintf(stdout, " --keep number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
fprintf(stdout, " --hellaswag compute HellaSwag score over random tasks from datafile supplied with -f\n");
fprintf(stdout, " --hellaswag-tasks N number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
fprintf(stdout, " --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
fprintf(stdout, " --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
if (llama_mlock_supported()) {
fprintf(stdout, " --mlock force system to keep model in RAM rather than swapping or compressing\n");
@ -568,6 +591,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stdout, " how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
fprintf(stdout, " -mg i, --main-gpu i the GPU to use for scratch and small tensors\n" );
fprintf(stdout, " -lv, --low-vram don't allocate VRAM scratch buffer\n" );
fprintf(stdout, " -mmq, --mul-mat-q use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
fprintf(stdout, " Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
fprintf(stdout, " is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
#endif
fprintf(stdout, " --mtest compute maximum memory usage\n");
fprintf(stdout, " --export export the computation graph to 'llama.ggml'\n");
@ -615,10 +641,12 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
lparams.n_ctx = params.n_ctx;
lparams.n_batch = params.n_batch;
lparams.n_gqa = params.n_gqa;
lparams.rms_norm_eps = params.rms_norm_eps;
lparams.n_gpu_layers = params.n_gpu_layers;
lparams.main_gpu = params.main_gpu;
lparams.tensor_split = params.tensor_split;
lparams.low_vram = params.low_vram;
lparams.mul_mat_q = params.mul_mat_q;
lparams.seed = params.seed;
lparams.f16_kv = params.memory_f16;
lparams.use_mmap = params.use_mmap;

31
examples/common.h
View file

@ -22,18 +22,19 @@
int32_t get_num_physical_cores();
struct gpt_params {
uint32_t seed = -1; // RNG seed
uint32_t seed = -1; // RNG seed
int32_t n_threads = get_num_physical_cores();
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 512; // context size
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_gqa = 1; // grouped-query attention factor (TODO: move to hparams)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_gpu_layers = 0; // number of layers to store in VRAM
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
float tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 512; // context size
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_gqa = 1; // grouped-query attention factor (TODO: move to hparams)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_gpu_layers = 0; // number of layers to store in VRAM
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
float tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS; // rms norm epsilon
float rope_freq_base = 10000.0f; // RoPE base frequency
float rope_freq_scale = 1.0f; // RoPE frequency scaling factor
@ -69,7 +70,11 @@ struct gpt_params {
std::string lora_adapter = ""; // lora adapter path
std::string lora_base = ""; // base model path for the lora adapter
bool low_vram = false; // if true, reduce VRAM usage at the cost of performance
bool hellaswag = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
size_t hellaswag_tasks = 400; // number of tasks to use when computing the HellaSwag score
bool low_vram = false; // if true, reduce VRAM usage at the cost of performance
bool mul_mat_q = false; // if true, use experimental mul_mat_q kernels
bool memory_f16 = true; // use f16 instead of f32 for memory kv
bool random_prompt = false; // do not randomize prompt if none provided
bool use_color = false; // use color to distinguish generations and inputs
@ -81,10 +86,10 @@ struct gpt_params {
bool interactive_first = false; // wait for user input immediately
bool multiline_input = false; // reverse the usage of `\`
bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix
bool instruct = false; // instruction mode (used for Alpaca models)
bool penalize_nl = true; // consider newlines as a repeatable token
bool perplexity = false; // compute perplexity over the prompt
bool perplexity_lines = false; // compute perplexity over each line of the prompt
bool use_mmap = true; // use mmap for faster loads
bool use_mlock = false; // use mlock to keep model in memory
bool mem_test = false; // compute maximum memory usage

4
examples/main/README.md
View file

@ -202,9 +202,9 @@ Example usage: `--top-p 0.95`
- `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. The method adjusts the logits (token probabilities) by raising them to the power of the parameter z. A higher value of z (e.g., 2.0) will further suppress less likely tokens from the tail of the distribution, while a value of 1.0 disables the effect of TFS. By setting the parameter z, you can control how much the probabilities of less likely tokens are reduced.
Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens, and thus disables the effect of TFS.
Example usage: `--tfs 2.0`
Example usage: `--tfs 0.95`
### Locally Typical Sampling

47
examples/main/main.cpp
View file

@ -325,6 +325,10 @@ int main(int argc, char ** argv) {
}
}
if (params.input_prefix_bos) {
fprintf(stderr, "Input prefix with BOS\n");
}
if (!params.input_prefix.empty()) {
fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
}
@ -633,16 +637,6 @@ int main(int argc, char ** argv) {
last_n_tokens.push_back(id);
}
// replace end of text token with newline token when in interactive mode
if (id == llama_token_eos() && params.interactive && !params.instruct) {
id = llama_token_newline.front();
if (params.antiprompt.size() != 0) {
// tokenize and inject first reverse prompt
const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
}
}
// add it to the context
embd.push_back(id);
@ -708,11 +702,34 @@ int main(int argc, char ** argv) {
}
}
// deal with end of text token in interactive mode
if (last_n_tokens.back() == llama_token_eos()) {
if (params.interactive) {
if (params.antiprompt.size() != 0) {
// tokenize and inject first reverse prompt
const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
is_antiprompt = true;
}
is_interacting = true;
printf("\n");
console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
fflush(stdout);
} else if (params.instruct) {
is_interacting = true;
}
}
if (n_past > 0 && is_interacting) {
if (params.instruct) {
printf("\n> ");
}
if (params.input_prefix_bos) {
embd_inp.push_back(llama_token_bos());
}
std::string buffer;
if (!params.input_prefix.empty()) {
buffer += params.input_prefix;
@ -776,13 +793,9 @@ int main(int argc, char ** argv) {
}
// end of text token
if (!embd.empty() && embd.back() == llama_token_eos()) {
if (params.instruct) {
is_interacting = true;
} else {
fprintf(stderr, " [end of text]\n");
break;
}
if (!embd.empty() && embd.back() == llama_token_eos() && !(params.instruct || params.interactive)) {
fprintf(stderr, " [end of text]\n");
break;
}
// In interactive mode, respect the maximum number of tokens and drop back to user input when reached.

185
examples/perplexity/perplexity.cpp
View file

@ -121,8 +121,23 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
printf("\n");
}
void perplexity_lines(llama_context * ctx, const gpt_params & params) {
// Calculates perplexity over each line of the prompt
void hellaswag_score(llama_context * ctx, const gpt_params & params) {
// Calculates hellaswag score (acc_norm) from prompt
//
// Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
// All used data fields are preprocessed as in https://github.com/EleutherAI/lm-evaluation-harness/blob/df3da98c5405deafd519c2ddca52bb7c3fe36bef/lm_eval/tasks/hellaswag.py#L62-L68
//
// All 10042 tasks should be extracted to keep the results standardized like other implementations.
//
// Datafile layout:
// ['??'] denotes json fields
// 6 lines per task:
// ['activity_label'] + ": " +['ctx'] - The first part of the query, the context
// ['label'] - The index the best common sense ending aka gold ending
// ['endings'][0] - Endings added to the first part of the query
// ['endings'][1]
// ['endings'][2]
// ['endings'][3]
std::vector<std::string> prompt_lines;
std::istringstream strstream(params.prompt);
@ -132,63 +147,149 @@ void perplexity_lines(llama_context * ctx, const gpt_params & params) {
prompt_lines.push_back(line);
}
if( prompt_lines.size() % 6 != 0) {
fprintf(stderr, "%s : number of lines in prompt not a multiple of 6.\n", __func__);
return;
}
size_t hs_task_count = prompt_lines.size()/6;
fprintf(stderr, "%s : loaded %lu tasks from prompt.\n", __func__, hs_task_count);
// This is needed as usual for LLaMA models
bool prepend_bos = true;
// Number of tasks to use when computing the score
if ( params.hellaswag_tasks < hs_task_count ) {
hs_task_count = params.hellaswag_tasks;
}
// The tasks should be randomized so the score stabilizes quickly.
bool randomize_tasks = true;
// The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now
std::mt19937 rng(1);
// Dataholder for hellaswag tasks
struct hs_data_t {
std::string context;
size_t gold_ending_idx;
std::string ending[4];
size_t ending_logprob_count[4];
double ending_logprob[4];
};
fprintf(stderr, "%s : selecting %lu %s tasks.\n", __func__, hs_task_count, (randomize_tasks?"randomized":"the first") );
// Select and read data from prompt lines
hs_data_t *hs_data = new hs_data_t[hs_task_count];
for (size_t i=0; i < hs_task_count; i++) {
size_t idx = i;
// Select a random example of those left in the prompt
if (randomize_tasks) {
std::uniform_int_distribution<size_t> dist(0, prompt_lines.size()/6-1 ) ;
idx = dist(rng);
}
hs_data[i].context = prompt_lines[idx*6];
hs_data[i].gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
for (size_t j=0; j < 4; j++) {
hs_data[i].ending[j] = " " + prompt_lines[idx*6+2+j];
}
// Delete the selected random example from the prompt
if (randomize_tasks) {
prompt_lines.erase( std::next(prompt_lines.begin(),idx*6) , std::next(prompt_lines.begin(),idx*6+6) );
}
}
fprintf(stderr, "%s : calculating hellaswag score over selected tasks.\n", __func__);
printf("\ntask\tacc_norm\n");
double acc = 0.0f;
const int n_vocab = llama_n_vocab(ctx);
int counttotal = 0;
size_t n_lines = prompt_lines.size();
for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
double nll = 0.0;
// Tokenize the context to count tokens
std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, prepend_bos);
size_t context_size = context_embd.size();
fprintf(stderr, "%s: calculating perplexity over %lu lines\n", __func__, n_lines);
for (size_t ending_idx=0;ending_idx<4;ending_idx++) {
printf("\nLine\tPPL line\tPPL cumulative\n");
// Tokenize the query
std::vector<int> query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[ending_idx], prepend_bos);
size_t query_size = query_embd.size();
for (size_t i = 0; i < n_lines; ++i) {
// Stop if query wont fit the ctx window
if (query_size > (size_t)params.n_ctx) {
fprintf(stderr, "%s : number of tokens in query %lu > n_ctxl\n", __func__, query_size);
return;
}
// Tokenize and insert BOS at start
std::vector<int> batch_embd = ::llama_tokenize(ctx, prompt_lines[i], true);
// Speedup small evaluations by evaluating atleast 32 tokens
if (query_size < 32) {
query_embd.resize(32);
}
size_t batch_size = batch_embd.size();
// Evaluate the query
if (llama_eval(ctx, query_embd.data(), query_embd.size(), 0, params.n_threads)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return;
}
// Stop if line is too long
if( batch_size > (size_t)params.n_ctx ) {
fprintf(stderr, "%s : tokens in line %lu > n_ctxl\n", __func__, i);
return;
const auto query_logits = llama_get_logits(ctx);
std::vector<float> logits;
logits.insert(logits.end(), query_logits, query_logits + query_size * n_vocab);
hs_data[task_idx].ending_logprob_count[ending_idx] = 0;
hs_data[task_idx].ending_logprob[ending_idx] = 0.0f;
// Calculate the logprobs over the ending
for (size_t j = context_size-1; j < query_size - 1; j++) {
// Calculate probability of next token, given the previous ones.
const std::vector<float> tok_logits(
logits.begin() + (j + 0) * n_vocab,
logits.begin() + (j + 1) * n_vocab);
const float prob = softmax(tok_logits)[query_embd[ j + 1]];
hs_data[task_idx].ending_logprob[ending_idx] += std::log(prob);
hs_data[task_idx].ending_logprob_count[ending_idx]++;
}
// Calculate the mean token logprob for acc_norm
hs_data[task_idx].ending_logprob[ending_idx] /= hs_data[task_idx].ending_logprob_count[ending_idx];
// printf("task %lu, ending %lu, whole_len %lu, context_len %lu, ending_logprob_count %lu, ending_logprob %.4f\n",
// task_idx,ending_idx,whole_size,context_size, hs_data[task_idx].ending_logprob_count[ending_idx], hs_data[task_idx].ending_logprob[ending_idx] );
}
if (llama_eval(ctx, batch_embd.data(), batch_size, 0, params.n_threads)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return;
// Find the ending with maximum logprob
size_t ending_logprob_max_idx = -1;
double ending_logprob_max_val = -INFINITY;
for (size_t j=0; j < 4; j++) {
if (hs_data[task_idx].ending_logprob[j] > ending_logprob_max_val) {
ending_logprob_max_idx = j;
ending_logprob_max_val = hs_data[task_idx].ending_logprob[j];
}
}
const auto batch_logits = llama_get_logits(ctx);
std::vector<float> logits;
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
// printf("max logprob ending idx %lu, gold ending idx %lu\n", ending_logprob_max_idx, hs_data[task_idx].gold_ending_idx);
double nllline = 0.0;
int countline = 0;
// Perplexity over second half of the line
for (size_t j = batch_size/2; j < batch_size - 1; ++j) {
// Calculate probability of next token, given the previous ones.
const std::vector<float> tok_logits(
logits.begin() + (j + 0) * n_vocab,
logits.begin() + (j + 1) * n_vocab);
const float prob = softmax(tok_logits)[batch_embd[ j + 1]];
nllline += -std::log(prob);
++countline;
// If the gold ending got the maximum logprobe add one accuracy point
if (ending_logprob_max_idx == hs_data[task_idx].gold_ending_idx) {
acc += 1.0;
}
nll += nllline;
counttotal += countline;
// perplexity is e^(average negative log-likelihood)
printf("%lu\t%.8lf\t%.8lf\n", i + 1, std::exp(nllline/countline), std::exp(nll / counttotal) );
// Print the accumulated accuracy mean x 100
printf("%li\t%.8lf\n",task_idx+1, acc/double(task_idx+1)*100.0);
fflush(stdout);
}
delete [] hs_data;
printf("\n");
}
@ -240,8 +341,8 @@ int main(int argc, char ** argv) {
params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
}
if (params.perplexity_lines) {
perplexity_lines(ctx, params);
if (params.hellaswag) {
hellaswag_score(ctx, params);
} else {
perplexity(ctx, params);
}

1
examples/save-load-state/save-load-state.cpp
View file

@ -26,6 +26,7 @@ int main(int argc, char ** argv) {
auto lparams = llama_context_default_params();
lparams.n_ctx = params.n_ctx;
lparams.n_gqa = params.n_gqa;
lparams.seed = params.seed;
lparams.f16_kv = params.memory_f16;
lparams.use_mmap = params.use_mmap;

26
examples/server-llama2-13B.sh Normal file
View file

@ -0,0 +1,26 @@
#!/bin/bash
set -e
cd "$(dirname "$0")/.." || exit
# Specify the model you want to use here:
MODEL="${MODEL:-./models/llama-2-13b-chat.ggmlv3.q5_K_M.bin}"
PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat-system.txt}
# Adjust to the number of CPU cores you want to use.
N_THREAD="${N_THREAD:-12}"
# Note: you can also override the generation options by specifying them on the command line:
GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 4096 --batch-size 1024}"
# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
./server $GEN_OPTIONS \
--model "$MODEL" \
--threads "$N_THREAD" \
--rope-freq-scale 1.0 \
"$@"
# I used this to test the model with mps, but omitted it from the general purpose. If you want to use it, just specify it on the command line.
# -ngl 1 \

2
examples/server/README.md
View file

@ -163,7 +163,7 @@ node .
`content`: Set the text to tokenize.
Note that the special `BOS` token is not added in fron of the text and also a space character is not inserted automatically as it is for `/completion`.
Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.

109
examples/server/chat-llama2.sh Normal file
View file

@ -0,0 +1,109 @@
#!/bin/bash
API_URL="${API_URL:-http://127.0.0.1:8080}"
CHAT=(
"Hello, Assistant."
"Hello. How may I help you today?"
)
INSTRUCTION="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
trim() {
shopt -s extglob
set -- "${1##+([[:space:]])}"
printf "%s" "${1%%+([[:space:]])}"
}
trim_trailing() {
shopt -s extglob
printf "%s" "${1%%+([[:space:]])}"
}
format_prompt() {
if [[ "${#CHAT[@]}" -eq 0 ]]; then
echo -n "[INST] <<SYS>>\n${INSTRUCTION}\n<</SYS>>"
else
LAST_INDEX=$(( ${#CHAT[@]} - 1 ))
echo -n "${CHAT[$LAST_INDEX]}\n[INST] $1 [/INST]"
fi
}
tokenize() {
curl \
--silent \
--request POST \
--url "${API_URL}/tokenize" \
--header "Content-Type: application/json" \
--data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
| jq '.tokens[]'
}
N_KEEP=$(tokenize "[INST] <<SYS>>\n${INSTRUCTION}\n<</SYS>>" | wc -l)
chat_completion() {
PROMPT="$(trim_trailing "$(format_prompt "$1")")"
DATA="$(echo -n "$PROMPT" | jq -Rs --argjson n_keep $N_KEEP '{
prompt: .,
temperature: 0.2,
top_k: 40,
top_p: 0.9,
n_keep: $n_keep,
n_predict: 1024,
stop: ["[INST]"],
stream: true
}')"
# Create a temporary file to hold the Python output
TEMPFILE=$(mktemp)
exec 3< <(curl \
--silent \
--no-buffer \
--request POST \
--url "${API_URL}/completion" \
--header "Content-Type: application/json" \
--data-raw "${DATA}")
python -c "
import json
import sys
answer = ''
while True:
line = sys.stdin.readline()
if not line:
break
if line.startswith('data: '):
json_content = line[6:].strip()
content = json.loads(json_content)['content']
sys.stdout.write(content)
sys.stdout.flush()
answer += content
answer = answer.rstrip('\n')
# Write the answer to the temporary file
with open('$TEMPFILE', 'w') as f:
f.write(answer)
" <&3
exec 3<&-
# Read the answer from the temporary file
ANSWER=$(cat $TEMPFILE)
# Clean up the temporary file
rm $TEMPFILE
printf "\n"
CHAT+=("$1" "$(trim "$ANSWER")")
}
while true; do
echo -en "\033[0;32m" # Green color
read -r -e -p "> " QUESTION
echo -en "\033[0m" # Reset color
chat_completion "${QUESTION}"
done

1788
examples/server/index.html.hpp
View file

File diff suppressed because it is too large Load diff

139
examples/server/public/index.html
View file

@ -3,12 +3,11 @@
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1" />
<meta name="color-scheme" content="light dark">
<title>llama.cpp - chat</title>
<style>
body {
background-color: #fff;
color: #000;
font-family: system-ui;
font-size: 90%;
max-width: 600px;
@ -103,6 +102,36 @@
margin: 0;
}
fieldset.two {
display: grid;
grid-template: "a a";
gap: 1em;
}
fieldset.three {
display: grid;
grid-template: "a a a";
gap: 1em;
}
details {
border: 1px solid #aaa;
border-radius: 4px;
padding: 0.5em 0.5em 0;
margin-top: 0.5em;
}
summary {
font-weight: bold;
margin: -0.5em -0.5em 0;
padding: 0.5em;
cursor: pointer;
}
details[open] {
padding: 0.5em;
}
textarea {
padding: 5px;
flex-grow: 1;
@ -122,7 +151,7 @@
fieldset label {
margin: 0.5em 0;
/*display: block;*/
display: block;
}
header, footer {
@ -152,21 +181,21 @@
})
const params = signal({
n_predict: 400,
top_k: 40,
top_p: 0.95,
tfs_z: 1.0,
typical_p: 1.0,
temperature: 0.7,
repeat_penalty: 1.18,
frequency_penalty: 0.0,
presence_penalty: 0.0,
repeat_last_n: 256,
mirostat: 0,
mirostat_tau: 5.0,
mirostat_eta: 0.1,
cfg_scale: 4.0,
frequency_penalty: 0.0, // 0.0 = disabled
mirostat_eta: 0.1, // learning rate
mirostat_tau: 5, // target entropy
mirostat: 0, // 0/1/2
n_predict: 400,
penalize_nl: true,
presence_penalty: 0.0, // 0.0 = disabled
repeat_last_n: 256, // 0 = disable penalty, -1 = context size
repeat_penalty: 1.18, // 1.0 = disabled
temperature: 0.7,
tfs_z: 1.0, // 1.0 = disabled
top_k: 40, // <= 0 to use vocab size
top_p: 0.5, // 1.0 = disabled
typical_p: 1.0, // 1.0 = disabled
})
const llamaStats = signal(null)
@ -305,29 +334,48 @@
`
}
const updateSession = (el) => session.value = { ...session.value, [el.target.name]: el.target.value }
const updateParams = (el) => params.value = { ...params.value, [el.target.name]: el.target.value }
const updateParamsFloat = (el) => params.value = { ...params.value, [el.target.name]: parseFloat(el.target.value) }
const updateParamsInt = (el) => params.value = { ...params.value, [el.target.name]: Math.floor(parseFloat(el.target.value)) }
const updateArray = (el) => {
const [name, index] = el.target.name.split(".")
const newarr = session.value[name].map((v, i) => i == index ? el.target.value : v).filter(x => x !== "")
session.value = { ...session.value, [name]: newarr }
}
const appendArray = () => session.value = { ...session.value, stop: [...session.value.stop, ""] }
const ParamSlider = ({param, min, max, step, children}) => {
const updateParamsFloat = (el) => params.value = { ...params.value, [param]: parseFloat(el.target.value) }
return html`
<div>
<label for="${param}"><code>${param}</code></label>
<label for="${param}">${children}</label>
<input type="range" id="${param}" min="${min}" max="${max}" step="${step}" name="${param}" value="${params.value[param]}" oninput=${updateParamsFloat} />
<span>${params.value[param]}</span>
<span>${children}</span>
</div>
`
}
const ConfigForm = (props) => {
const updateSession = (el) => session.value = { ...session.value, [el.target.name]: el.target.value }
const updateParams = (el) => params.value = { ...params.value, [el.target.name]: el.target.value }
const updateParamsFloat = (el) => params.value = { ...params.value, [el.target.name]: parseFloat(el.target.value) }
const appendArray = () => session.value = { ...session.value, stop: [...session.value.stop, ""] }
const updateArray = (el) => {
const [name, index] = el.target.name.split(".")
const newarr = session.value[name].map((v, i) => i == index ? el.target.value : v).filter(x => x !== "")
session.value = { ...session.value, [name]: newarr }
}
const FloatField = ({label, max, min, name, step, value}) => {
return html`
<div>
<label for="${name}">${label}</label>
<input type="range" id="${name}" min="${min}" max="${max}" step="${step}" name="${name}" value="${value}" oninput=${updateParamsFloat} />
<span>${value}</span>
</div>
`
};
const IntField = ({label, max, min, name, value}) => {
return html`
<div>
<label for="${name}">${label}</label>
<input type="range" id="${name}" min="${min}" max="${max}" name="${name}" value="${value}" oninput=${updateParamsInt} />
<span>${value}</span>
</div>
`
};
const ConfigForm = (props) => {
return html`
<form>
<fieldset>
@ -356,7 +404,9 @@
<textarea type="text" name="system_cfg" value="${session.value.system_cfg}" rows=4 oninput=${updateSession}/>
</div>
`}
</fieldset>
<fieldset class="two">
<${ParamSlider} min=1 max=1000 step=1 param=n_predict>Predict N tokens<//>
<${ParamSlider} min=0 max=1000 step=1 param=repeat_last_n>Penalize last N tokens<//>
${params.value.repeat_last_n > 0 && html`
@ -365,18 +415,26 @@
<${ParamSlider} min=0 max=4 step=0.01 param=presence_penalty>Penalize tokens not present in prompt<//>
`}
<${ParamSlider} min=0 max=2 step=0.01 param=temperature>Temperature<//>
${params.value.temperature > 0 && html`
<div>
<input id=mirostat_0 type=radio name=mirostat checked=${params.value.mirostat == 0} value=0 oninput=${updateParamsFloat} />
<label for=mirostat_0>Temperature</label>
</fieldset>
<input id=mirostat_1 type=radio name=mirostat checked=${params.value.mirostat == 1} value=1 oninput=${updateParamsFloat} />
<label for=mirostat_1>Mirostat v1</label>
${params.value.temperature > 0 && html`
<fieldset class="three">
<label>
<input type=radio name=mirostat checked=${params.value.mirostat == 0} value=0 oninput=${updateParamsFloat} />
Temperature
</label>
<input id=mirostat_2 type=radio name=mirostat checked=${params.value.mirostat == 2} value=2 oninput=${updateParamsFloat} />
<label for=mirostat_2>Mirostat v2</label>
</div>
<label><input type=radio name=mirostat checked=${params.value.mirostat == 1} value=1 oninput=${updateParamsFloat} />
Mirostat v1
</label>
<label>
<input type=radio name=mirostat checked=${params.value.mirostat == 2} value=2 oninput=${updateParamsFloat} />
Mirostat v2
</label>
</fieldset>
<fieldset class="two">
${params.value.mirostat == 0 && html`
<${ParamSlider} min=1 max=1000 step=1 param=top_k>Top K<//>
<${ParamSlider} min=0 max=1 step=0.01 param=tfs_z>Tail free sampling<//>
@ -387,14 +445,17 @@
<${ParamSlider} min=0 max=1 step=0.01 param=mirostat_eta>Mirostat eta, learning rate<//>
<${ParamSlider} min=0 max=1000 step=1 param=mirostat_tau>Mirostat tau, target entropy<//>
`}
</fieldset>
`}
</fieldset>
</form>
`
}
// poor mans markdown replacement
const Markdownish = (params) => {
const md = params.text
.replace(/&/g, '&amp;')
.replace(/</g, '&lt;')
.replace(/>/g, '&gt;')
.replace(/^#{1,6} (.*)$/gim, '<h3>$1</h3>')
.replace(/\*\*(.*?)\*\*/g, '<strong>$1</strong>')
.replace(/__(.*?)__/g, '<strong>$1</strong>')

20
examples/server/server.cpp
View file

@ -613,6 +613,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
fprintf(stdout, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stdout, " -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx);
fprintf(stdout, " -gqa N, --gqa N grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
fprintf(stdout, " -eps N, --rms-norm-eps N rms norm eps (TEMP!!! use 1e-5 for LLaMAv2) (default: %.1e)\n", params.rms_norm_eps);
fprintf(stdout, " --rope-freq-base N RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
fprintf(stdout, " --rope-freq-scale N RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
fprintf(stdout, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
@ -632,6 +633,9 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
fprintf(stdout, " how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
fprintf(stdout, " -mg i, --main-gpu i the GPU to use for scratch and small tensors\n");
fprintf(stdout, " -lv, --low-vram don't allocate VRAM scratch buffer\n");
fprintf(stdout, " -mmq, --mul-mat-q use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
fprintf(stdout, " Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
fprintf(stdout, " is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
#endif
fprintf(stdout, " -m FNAME, --model FNAME\n");
fprintf(stdout, " model path (default: %s)\n", params.model.c_str());
@ -708,6 +712,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
break;
}
params.n_gqa = std::stoi(argv[i]);
} else if (arg == "-eps" || arg == "--rms-norm-eps") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rms_norm_eps = std::stof(argv[i]);
} else if (arg == "--rope-freq-base") {
if (++i >= argc) {
invalid_param = true;
@ -769,13 +779,19 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
}
}
#else
LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.", {});
LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {});
#endif // GGML_USE_CUBLAS
} else if (arg == "--low-vram" || arg == "-lv") {
#ifdef GGML_USE_CUBLAS
params.low_vram = true;
#else
fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {});
#endif // GGML_USE_CUBLAS
} else if (arg == "--mul-mat-q" || arg == "-mmq") {
#ifdef GGML_USE_CUBLAS
params.mul_mat_q = true;
#else
LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n", {});
#endif // GGML_USE_CUBLAS
} else if (arg == "--main-gpu" || arg == "-mg") {
if (++i >= argc) {

32
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -16,6 +16,8 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
struct random_normal_distribution {
std::mt19937 gen;
std::normal_distribution<float> rd;
@ -439,7 +441,7 @@ struct ggml_tensor * forward(
// norm
{
// cur shape [n_embd,N,1,1]
cur = ggml_rms_norm(ctx0, inpL);
cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
// cur = attention_norm*cur
cur = ggml_mul(ctx0,
@ -562,7 +564,7 @@ struct ggml_tensor * forward(
// norm
{
// cur shape [n_embd,N,1,1]
cur = ggml_rms_norm(ctx0, inpFF);
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
// cur = ffn_norm*cur
// cur shape [n_embd,N,1,1]
@ -606,7 +608,7 @@ struct ggml_tensor * forward(
{
// inpL shape [n_embd,N,1,1]
inpL = ggml_rms_norm(ctx0, inpL);
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
// inpL = norm*inpL
// inpL shape [n_embd,N,1,1]
@ -694,7 +696,7 @@ struct ggml_tensor * forward_batch(
// norm
{
// cur shape [n_embd,N*n_batch,1,1]
cur = ggml_rms_norm(ctx0, inpL);
cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
assert_shape_2d(cur, n_embd, N*n_batch);
// cur = attention_norm*cur
@ -857,7 +859,7 @@ struct ggml_tensor * forward_batch(
// norm
{
// cur shape [n_embd,N*n_batch,1,1]
cur = ggml_rms_norm(ctx0, inpFF);
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
assert_shape_2d(cur, n_embd, N*n_batch);
// cur = ffn_norm*cur
@ -910,7 +912,7 @@ struct ggml_tensor * forward_batch(
{
// inpL shape [n_embd,N*n_batch,1,1]
inpL = ggml_rms_norm(ctx0, inpL);
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
assert_shape_2d(inpL, n_embd, N*n_batch);
// inpL = norm*inpL
@ -979,7 +981,7 @@ struct ggml_tensor * forward_batch_wo_cache(
// norm
{
// cur shape [n_embd,N*n_batch,1,1]
cur = ggml_rms_norm(ctx0, inpL);
cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
assert_shape_2d(cur, n_embd, N*n_batch);
// cur = attention_norm*cur
@ -1085,7 +1087,7 @@ struct ggml_tensor * forward_batch_wo_cache(
// norm
{
// cur shape [n_embd,N*n_batch,1,1]
cur = ggml_rms_norm(ctx0, inpFF);
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
assert_shape_2d(cur, n_embd, N*n_batch);
// cur = ffn_norm*cur
@ -1138,7 +1140,7 @@ struct ggml_tensor * forward_batch_wo_cache(
{
// inpL shape [n_embd,N*n_batch,1,1]
inpL = ggml_rms_norm(ctx0, inpL);
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
assert_shape_2d(inpL, n_embd, N*n_batch);
// inpL = norm*inpL
@ -1203,7 +1205,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn(
// norm
{
cur = ggml_rms_norm(ctx0, inpL);
cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
assert_shape_2d(cur, n_embd, N*n_batch);
// cur = attention_norm*cur
@ -1267,7 +1269,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn(
{
// norm
{
cur = ggml_rms_norm(ctx0, inpFF);
cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
assert_shape_2d(cur, n_embd, N*n_batch);
// cur = ffn_norm*cur
@ -1311,7 +1313,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn(
// norm
{
inpL = ggml_rms_norm(ctx0, inpL);
inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
assert_shape_2d(inpL, n_embd, N*n_batch);
// inpL = norm*inpL
@ -1603,7 +1605,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
struct my_llama_layer & layer = model->layers[il];
// tensors with values necessary for backward pass are in persistent buf(-1)
// other tensors with buf(0) and buf(1) are only temporary needed, and their memory reused after layer is completed.
use_buf(-1); struct ggml_tensor * t02 = expand(gf, ggml_rms_norm (ctx0, cur)); assert_shape_2d(t02, n_embd, N*n_batch);
use_buf(-1); struct ggml_tensor * t02 = expand(gf, ggml_rms_norm (ctx0, cur, rms_norm_eps)); assert_shape_2d(t02, n_embd, N*n_batch);
use_buf( 0); struct ggml_tensor * t03 = expand(gf, ggml_repeat (ctx0, layer.attention_norm, t02)); assert_shape_2d(t03, n_embd, N*n_batch);
use_buf(-1); struct ggml_tensor * t04 = expand(gf, ggml_mul (ctx0, t02, t03)); assert_shape_2d(t04, n_embd, N*n_batch);
use_buf(-1); struct ggml_tensor * t05 = expand(gf, ggml_mul_mat (ctx0, layer.wq, t04)); assert_shape_2d(t05, n_embd, N*n_batch);
@ -1623,7 +1625,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
use_buf(-1); struct ggml_tensor * t19 = expand(gf, ggml_reshape_2d (ctx0, t18, n_embd, N*n_batch)); assert_shape_2d(t19, n_embd, N*n_batch);
use_buf( 0); struct ggml_tensor * t20 = expand(gf, ggml_mul_mat (ctx0, layer.wo, t19)); assert_shape_2d(t20, n_embd, N*n_batch);
use_buf(-1); struct ggml_tensor * t21 = expand(gf, ggml_add (ctx0, t20, cur)); assert_shape_2d(t21, n_embd, N*n_batch);
use_buf(-1); struct ggml_tensor * t22 = expand(gf, ggml_rms_norm (ctx0, t21)); assert_shape_2d(t22, n_embd, N*n_batch);
use_buf(-1); struct ggml_tensor * t22 = expand(gf, ggml_rms_norm (ctx0, t21, rms_norm_eps)); assert_shape_2d(t22, n_embd, N*n_batch);
use_buf( 0); struct ggml_tensor * t23 = expand(gf, ggml_repeat (ctx0, layer.ffn_norm, t22)); assert_shape_2d(t23, n_embd, N*n_batch);
use_buf(-1); struct ggml_tensor * t24 = expand(gf, ggml_mul (ctx0, t23, t22)); assert_shape_2d(t24, n_embd, N*n_batch);
use_buf(-1); struct ggml_tensor * t25 = expand(gf, ggml_mul_mat (ctx0, layer.w3, t24)); assert_shape_2d(t25, n_ff, N*n_batch);
@ -1666,7 +1668,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
}
clr_buf(0);
use_buf(0);
struct ggml_tensor * t31 = expand(gf, ggml_rms_norm (ctx0, cur)); assert_shape_2d(t31, n_embd, N*n_batch);
struct ggml_tensor * t31 = expand(gf, ggml_rms_norm (ctx0, cur, rms_norm_eps)); assert_shape_2d(t31, n_embd, N*n_batch);
struct ggml_tensor * t32 = expand(gf, ggml_repeat (ctx0, model->norm, t31)); assert_shape_2d(t32, n_embd, N*n_batch);
struct ggml_tensor * t33 = expand(gf, ggml_mul (ctx0, t32, t31)); assert_shape_2d(t33, n_embd, N*n_batch);
use_buf(-1);