AI/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2024-10-01 01:06:10 -04:00
Code Issues Packages Projects Releases Wiki Activity

embed4all: small fixes related to nomic client local embeddings (#2213)

Browse Source 
* actually submit larger batches with increased n_ctx
* fix crash when llama_tokenize returns no tokens

Signed-off-by: Jared Van Bortel <jared@nomic.ai>
This commit is contained in:
Jared Van Bortel 2024-04-12 10:54:15 -04:00 committed by GitHub
parent 1e4c62027b
commit 459289b94c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 16 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
gpt4all-backend/llamamodel.cpp
View File

@ -325,7 +325,7 @@ bool LLamaModel::loadModel(const std::string &modelPath, int n_ctx, int ngl)
bool isEmbedding = is_embedding_arch(llama_model_arch(d_ptr->model));
const int n_ctx_train = llama_n_ctx_train(d_ptr->model);
if (isEmbedding) {
d_ptr->ctx_params.n_batch = n_ctx_train;
d_ptr->ctx_params.n_batch = n_ctx;
} else {
if (n_ctx > n_ctx_train) {
std::cerr << "warning: model was trained on only " << n_ctx_train << " context tokens ("
@ -734,7 +734,7 @@ void LLamaModel::embedInternal(
) {
typedef std::vector<LLModel::Token> TokenString;
static constexpr int32_t atlasMaxLength = 8192;
static constexpr int chunkOverlap = 8; // Atlas overlaps n_batch-sized chunks of input by 8 tokens
static constexpr int chunkOverlap = 8; // Atlas overlaps chunks of input by 8 tokens
const llama_token bos_token = llama_token_bos(d_ptr->model);
const llama_token eos_token = llama_token_eos(d_ptr->model);
@ -751,8 +751,12 @@ void LLamaModel::embedInternal(
tokens.resize(text.length()+4);
int32_t n_tokens = llama_tokenize(d_ptr->model, text.c_str(), text.length(), tokens.data(), tokens.size(), wantBOS, false);
assert(useEOS == (eos_token != -1 && tokens[n_tokens - 1] == eos_token));
tokens.resize(n_tokens - useEOS); // erase EOS/SEP
if (n_tokens) {
assert(useEOS == (eos_token != -1 && tokens[n_tokens - 1] == eos_token));
tokens.resize(n_tokens - useEOS); // erase EOS/SEP
} else {
tokens.clear();
}
};
// tokenize the texts
@ -786,9 +790,14 @@ void LLamaModel::embedInternal(
tokenize(prefix + ':', prefixTokens, true);
}
// n_ctx_train: max sequence length of model (RoPE scaling not implemented)
const uint32_t n_ctx_train = llama_n_ctx_train(d_ptr->model);
// n_batch (equals n_ctx): max tokens per call to llama_decode (one more more sequences)
const uint32_t n_batch = llama_n_batch(d_ptr->ctx);
const uint32_t max_len = n_batch - (prefixTokens.size() + useEOS); // minus BOS/CLS and EOS/SEP
if (chunkOverlap >= max_len) {
// effective sequence length minus prefix and SEP token
const uint32_t max_len = std::min(n_ctx_train, n_batch) - (prefixTokens.size() + useEOS);
if (max_len <= chunkOverlap) {
throw std::logic_error("max chunk length of " + std::to_string(max_len) + " is smaller than overlap of " +
std::to_string(chunkOverlap) + " tokens");
}

2
gpt4all-bindings/python/setup.py
View File

@ -68,7 +68,7 @@ def get_long_description():
setup(
name=package_name,
version="2.4.0",
version="2.4.1",
description="Python bindings for GPT4All",
long_description=get_long_description(),
long_description_content_type="text/markdown",