github/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2025-06-25 06:53:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

chatllm: do not attempt to serialize incompatible state (#1742)

Browse Source
This commit is contained in:
Jared Van Bortel 2023-12-12 11:45:03 -05:00 committed by GitHub
parent 9481762802
commit 0600f551b3
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 18 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
gpt4all-backend/llamamodel.cpp
View File

@ -163,6 +163,10 @@ bool LLamaModel::loadModel(const std::string &modelPath)
d_ptr->ctx_params.seed = params.seed; d_ptr->ctx_params.seed = params.seed;
d_ptr->ctx_params.f16_kv = params.memory_f16; d_ptr->ctx_params.f16_kv = params.memory_f16;
// The new batch API provides space for n_vocab*n_tokens logits. Tell llama.cpp early
// that we want this many logits so the state serializes consistently.
d_ptr->ctx_params.logits_all = true;
d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency()); d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
d_ptr->ctx_params.n_threads = d_ptr->n_threads; d_ptr->ctx_params.n_threads = d_ptr->n_threads;
d_ptr->ctx_params.n_threads_batch = d_ptr->n_threads; d_ptr->ctx_params.n_threads_batch = d_ptr->n_threads;

1
gpt4all-chat/chat.cpp
View File

@ -435,7 +435,6 @@ bool Chat::deserialize(QDataStream &stream, int version)
if (!m_chatModel->deserialize(stream, version)) if (!m_chatModel->deserialize(stream, version))
return false; return false;
if (!deserializeKV || discardKV)
m_llmodel->setStateFromText(m_chatModel->text()); m_llmodel->setStateFromText(m_chatModel->text());
emit chatModelChanged(); emit chatModelChanged();

18
gpt4all-chat/chatllm.cpp
View File

@ -863,11 +863,11 @@ bool ChatLLM::deserialize(QDataStream &stream, int version, bool deserializeKV,
if (!discardKV) if (!discardKV)
m_state = qUncompress(compressed); m_state = qUncompress(compressed);
} else { } else {
if (!discardKV) if (!discardKV) {
stream >> m_state; stream >> m_state;
else { } else {
QByteArray state; QByteArray state;
stream >> m_state; stream >> state;
} }
} }
@ -912,7 +912,7 @@ void ChatLLM::restoreState()
stream >> context; stream >> context;
chatGPT->setContext(context); chatGPT->setContext(context);
m_state.clear(); m_state.clear();
m_state.resize(0); m_state.squeeze();
return; return;
} }
@ -923,10 +923,16 @@ void ChatLLM::restoreState()
if (m_state.isEmpty()) if (m_state.isEmpty())
return; return;
m_processedSystemPrompt = true; if (m_llModelInfo.model->stateSize() == m_state.size()) {
m_llModelInfo.model->restoreState(static_cast<const uint8_t*>(reinterpret_cast<void*>(m_state.data()))); m_llModelInfo.model->restoreState(static_cast<const uint8_t*>(reinterpret_cast<void*>(m_state.data())));
m_processedSystemPrompt = true;
} else {
qWarning() << "restoring state from text because" << m_llModelInfo.model->stateSize() << "!=" << m_state.size() << "\n";
m_restoreStateFromText = true;
}
m_state.clear(); m_state.clear();
m_state.resize(0); m_state.squeeze();
} }
void ChatLLM::processSystemPrompt() void ChatLLM::processSystemPrompt()