mirrors/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-05-12 02:46:36 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

remove mllama patch

Browse source
This commit is contained in:
Michael Yang 2025-04-25 15:54:42 -07:00
parent 0d6e35d3c6
commit f8586c6b2b
35 changed files with 37 additions and 2444 deletions
Download patch file Download diff file Expand all files Collapse all files

25
llama/llama.cpp/src/llama-context.cpp vendored
View file

@ -809,7 +809,7 @@ float * llama_context::get_logits_ith(int32_t i) {
throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
}
return logits + j*model.hparams.n_vocab;
return logits + j*model.vocab.n_tokens();
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
#ifndef NDEBUG
@ -930,10 +930,6 @@ void llama_context::set_warmup(bool value) {
cparams.warmup = value;
}
void llama_context::set_cross_attn(bool value) {
cparams.cross_attn = value;
}
void llama_context::set_adapter_lora(
llama_adapter_lora * adapter,
float scale) {
@ -1009,7 +1005,7 @@ int llama_context::encode(llama_batch & inp_batch) {
const int64_t n_embd = hparams.n_embd;
sbatch.from_batch(batch, batch.n_embd, /* simple_split */ true, /* logits_all */ true);
sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
@ -1149,9 +1145,10 @@ int llama_context::decode(llama_batch & inp_batch) {
const llama_batch & batch = batch_allocr.batch;
const auto & vocab = model.vocab;
const auto & hparams = model.hparams;
const int32_t n_vocab = hparams.n_vocab;
const int32_t n_vocab = vocab.n_tokens();
const int64_t n_tokens_all = batch.n_tokens;
const int64_t n_embd = hparams.n_embd;
@ -1199,7 +1196,7 @@ int llama_context::decode(llama_batch & inp_batch) {
const bool logits_all = n_outputs_all == n_tokens_all;
sbatch.from_batch(batch, batch.n_embd,
sbatch.from_batch(batch, n_embd,
/* simple_split */ !kv_self->recurrent,
/* logits_all */ logits_all);
@ -1436,11 +1433,12 @@ int llama_context::decode(llama_batch & inp_batch) {
int32_t llama_context::output_reserve(int32_t n_outputs) {
const auto & hparams = model.hparams;
const auto & vocab = model.vocab;
const int64_t n_outputs_max = std::max<int64_t>(n_outputs, n_seq_max());
const auto n_batch = cparams.n_batch;
const auto n_vocab = hparams.n_vocab;
const auto n_vocab = vocab.n_tokens();
const auto n_embd = hparams.n_embd;
// TODO: use a per-batch flag for logits presence instead
@ -1508,7 +1506,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
void llama_context::output_reorder() {
auto & out_ids = sbatch.out_ids;
if (!out_ids.empty()) {
const uint32_t n_vocab = model.hparams.n_vocab;
const uint32_t n_vocab = model.vocab.n_tokens();
const uint32_t n_embd = model.hparams.n_embd;
GGML_ASSERT((size_t) n_outputs == out_ids.size());
@ -2015,7 +2013,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
{
LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.hparams.n_vocab);
const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.vocab.n_tokens());
io.write(&logits_size, sizeof(logits_size));
@ -2198,7 +2196,6 @@ llama_context_params llama_context_default_params() {
/*.offload_kqv =*/ true,
/*.flash_attn =*/ false,
/*.no_perf =*/ true,
/*.cross_attn =*/ false,
/*.abort_callback =*/ nullptr,
/*.abort_callback_data =*/ nullptr,
};
@ -2326,10 +2323,6 @@ void llama_set_warmup(llama_context * ctx, bool warmup) {
ctx->set_warmup(warmup);
}
void llama_set_cross_attention(struct llama_context * ctx, bool cross_attention) {
ctx->set_cross_attn(cross_attention);
}
void llama_synchronize(llama_context * ctx) {
ctx->synchronize();
}