CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-29 08:41:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

context : remove logits_all flag (#13284)

Browse Source 
* context : remove logits_all flag

ggml-ci

* llama : remove logits_all flag + reorder llama_context_params

ggml-ci
This commit is contained in:
Georgi Gerganov
2025-05-08 14:26:50 +03:00
committed by GitHub
parent 70a6991edf
commit 51fb96b1ff
9 changed files with 13 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
common/arg.cpp
View File

@@ -2097,13 +2097,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.cache_type_v = kv_cache_type_from_str(value);
}
).set_env("LLAMA_ARG_CACHE_TYPE_V"));
add_opt(common_arg(
{"--perplexity", "--all-logits"},
string_format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"),
[](common_params & params) {
params.logits_all = true;
}
).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
add_opt(common_arg(
{"--hellaswag"},
"compute HellaSwag score over random tasks from datafile supplied with -f",

1
common/common.cpp
View File

@@ -1096,7 +1096,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
cparams.n_threads = params.cpuparams.n_threads;
cparams.n_threads_batch = params.cpuparams_batch.n_threads == -1 ?
params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
cparams.logits_all = params.logits_all;
cparams.embeddings = params.embedding;
cparams.rope_scaling_type = params.rope_scaling_type;
cparams.rope_freq_base = params.rope_freq_base;

1
common/common.h
View File

@@ -324,7 +324,6 @@ struct common_params {
bool ctx_shift = true; // context shift on inifinite text generation
bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix
bool logits_all = false; // return logits for all tokens in the batch
bool use_mmap = true; // use mmap for faster loads
bool use_mlock = false; // use mlock to keep model in memory
bool verbose_prompt = false; // print prompt tokens before generation

14
include/llama.h
View File

@@ -351,19 +351,17 @@ extern "C" {
enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
// Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
// TODO: move at the end of the struct
bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
bool embeddings; // if true, extract embeddings (together with logits)
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
bool flash_attn; // whether to use flash attention [EXPERIMENTAL]
bool no_perf; // whether to measure performance timings
// Abort callback
// if it returns true, execution of llama_decode() will be aborted
// currently works only with CPU execution
ggml_abort_callback abort_callback;
void * abort_callback_data;
// Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
bool embeddings; // if true, extract embeddings (together with logits)
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
bool flash_attn; // whether to use flash attention [EXPERIMENTAL]
bool no_perf; // whether to measure performance timings
};
// model quantization parameters

9
src/llama-context.cpp
View File

@@ -116,8 +116,6 @@ llama_context::llama_context(
__func__, n_ctx_per_seq, hparams.n_ctx_train);
}
logits_all = params.logits_all;
if (!hparams.vocab_only) {
// GPU backends
for (auto * dev : model.devices) {
@@ -890,7 +888,7 @@ int llama_context::decode(llama_batch & inp_batch) {
for (uint32_t i = 0; i < n_tokens_all; ++i) {
n_outputs_all += batch.logits[i] != 0;
}
} else if (logits_all || embd_pooled) {
} else if (embd_pooled) {
n_outputs_all = n_tokens_all;
} else {
// keep last output only
@@ -1853,13 +1851,12 @@ llama_context_params llama_context_default_params() {
/*.cb_eval_user_data =*/ nullptr,
/*.type_k =*/ GGML_TYPE_F16,
/*.type_v =*/ GGML_TYPE_F16,
/*.logits_all =*/ false,
/*.abort_callback =*/ nullptr,
/*.abort_callback_data =*/ nullptr,
/*.embeddings =*/ false,
/*.offload_kqv =*/ true,
/*.flash_attn =*/ false,
/*.no_perf =*/ true,
/*.abort_callback =*/ nullptr,
/*.abort_callback_data =*/ nullptr,
};
return result;

3
src/llama-context.h
View File

@@ -187,9 +187,6 @@ private:
std::unique_ptr<llama_memory_i> memory;
// TODO: remove
bool logits_all = false;
// decode output (2-dimensional array: [n_outputs][n_vocab])
size_t logits_size = 0; // capacity (of floats) for logits
float * logits = nullptr;

1
tools/imatrix/imatrix.cpp
View File

@@ -585,7 +585,6 @@ int main(int argc, char ** argv) {
params.out_file = "imatrix.dat" ;
params.n_ctx = 512;
params.logits_all = true;
params.escape = false;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {

8
tools/main/main.cpp
View File

@@ -99,14 +99,6 @@ int main(int argc, char ** argv) {
console::init(params.simple_io, params.use_color);
atexit([]() { console::cleanup(); });
if (params.logits_all) {
LOG_ERR("************\n");
LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.embedding) {
LOG_ERR("************\n");
LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);

6
tools/perplexity/perplexity.cpp
View File

@@ -1554,7 +1554,10 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
if (int(batch_indeces.size()) != num_answers) {
batch_indeces.resize(num_answers);
}
for (int s = 0; s < num_answers; ++s) batch_indeces[s] = s0 + s;
for (int s = 0; s < num_answers; ++s) {
batch_indeces[s] = s0 + s;
}
for (size_t i = 0; i < cur_task.common_prefix; ++i) {
//llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
@@ -1970,7 +1973,6 @@ int main(int argc, char ** argv) {
common_params params;
params.n_ctx = 512;
params.logits_all = true;
params.escape = false;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {