Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-13 22:08:55 +01:00
Code Issues Packages Projects Releases Wiki Activity

llama : skip token bounds check when evaluating embeddings (#9437)

Browse Source
This commit is contained in:
slaren 2024-09-11 17:52:13 +02:00 committed by GitHub
parent 8db003a19d
commit 1b28061400
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 18 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
src/llama.cpp
View File

@ -16076,19 +16076,21 @@ static int llama_decode_internal(
return -1; return -1;
} }
for (uint32_t i = 0; i < n_tokens_all; ++i) {
if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= lctx.model.vocab.n_vocab) {
LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch_all.token[i]);
return -1;
}
}
const auto & model = lctx.model; const auto & model = lctx.model;
const auto & hparams = model.hparams; const auto & hparams = model.hparams;
const auto & cparams = lctx.cparams; const auto & cparams = lctx.cparams;
GGML_ASSERT((!batch_all.token && batch_all.embd) || (batch_all.token && !batch_all.embd)); // NOLINT GGML_ASSERT((!batch_all.token && batch_all.embd) || (batch_all.token && !batch_all.embd)); // NOLINT
if (batch_all.token) {
for (uint32_t i = 0; i < n_tokens_all; ++i) {
if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= model.vocab.n_vocab) {
LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch_all.token[i]);
return -1;
}
}
}
GGML_ASSERT(n_tokens_all <= cparams.n_batch); GGML_ASSERT(n_tokens_all <= cparams.n_batch);
GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens"); GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens");
@ -16375,19 +16377,21 @@ static int llama_encode_internal(
return -1; return -1;
} }
for (uint32_t i = 0; i < n_tokens; ++i) {
if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= lctx.model.vocab.n_vocab) {
LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch.token[i]);
return -1;
}
}
const auto & model = lctx.model; const auto & model = lctx.model;
const auto & hparams = model.hparams; const auto & hparams = model.hparams;
const auto & cparams = lctx.cparams; const auto & cparams = lctx.cparams;
GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
if (batch.token) {
for (uint32_t i = 0; i < n_tokens; ++i) {
if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= model.vocab.n_vocab) {
LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch.token[i]);
return -1;
}
}
}
// micro-batching is not possible for non-causal encoding, so we process the batch in a single shot // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens"); GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");