ollama/llm/patches/10-quantize-callback.diff
2024-08-27 16:44:32 -07:00

52 lines
2.0 KiB
Diff

From ed941590d59fc07b1ad21d6aa458588e47d1e446 Mon Sep 17 00:00:00 2001
From: Josh Yan <jyan00017@gmail.com>
Date: Wed, 10 Jul 2024 13:39:39 -0700
Subject: [PATCH] quantize progress
---
include/llama.h | 3 +++
src/llama.cpp | 8 ++++++++
2 files changed, 11 insertions(+)
diff --git a/include/llama.h b/include/llama.h
index bb4b05ba..613db68e 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -349,6 +349,9 @@ extern "C" {
bool keep_split; // quantize to the same number of shards
void * imatrix; // pointer to importance matrix data
void * kv_overrides; // pointer to vector containing overrides
+
+ llama_progress_callback quantize_callback; // callback to report quantization progress
+ void * quantize_callback_data; // user data for the callback
} llama_model_quantize_params;
// grammar types
diff --git a/src/llama.cpp b/src/llama.cpp
index 2b9ace28..ac640c02 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -18252,6 +18252,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
const auto tn = LLM_TN(model.arch);
new_ofstream(0);
for (int i = 0; i < ml.n_tensors; ++i) {
+ if (params->quantize_callback){
+ if (!params->quantize_callback(i, params->quantize_callback_data)) {
+ return;
+ }
+ }
+
auto weight = ml.get_weight(i);
struct ggml_tensor * tensor = weight->tensor;
if (weight->idx != cur_split && params->keep_split) {
@@ -18789,6 +18795,8 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
/*.keep_split =*/ false,
/*.imatrix =*/ nullptr,
/*.kv_overrides =*/ nullptr,
+ /*.quantize_callback =*/ nullptr,
+ /*.quantize_callback_data =*/ nullptr,
};
return result;
--
2.39.3 (Apple Git-146)