From fa7c39540cff0ce6f6d245baeb15ac1ebe6cdd69 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?M=2E=20Yusuf=20Sar=C4=B1g=C3=B6z?= <yusufsarigoz@gmail.com>
Date: Sat, 12 Aug 2023 15:55:58 +0300
Subject: [PATCH] gguf : start implementing quantization (WIP)

---
 gguf-llama.cpp | 15 +++++++++++----
 gguf-util.h    | 12 ++++++++----
 2 files changed, 19 insertions(+), 8 deletions(-)

diff --git a/gguf-llama.cpp b/gguf-llama.cpp
index f1755fef5..eecefc0f6 100644
--- a/gguf-llama.cpp
+++ b/gguf-llama.cpp
@@ -525,6 +525,11 @@ struct ggml_context * ctx_data = NULL;
 
         // TODO make keysconstants in header
         // TODO: read all hparams from file
+        int q_ver_idx = gguf_find_key (gguf_ctx, "general.quantization_version");
+        if (q_ver_idx != -1) {
+            hparams.ftype = gguf_get_val_u32(gguf_ctx, q_ver_idx);
+        }
+
         hparams.n_vocab = read_n_vocab();
         hparams.n_ctx   = read_u32("llama.context_length");
         hparams.n_embd  = read_u32("llama.embedding_length");
@@ -738,27 +743,29 @@ struct gguf_file_saver {
         info_offset  = file.tell();
         size_t count = gguf_get_data_offset(fl->gguf_ctx) - info_offset;
         file.write_zeros(count);
-        printf("info_offset = %zu\n", info_offset);
         file.seek(info_offset, SEEK_SET);
         GGML_ASSERT(info_offset == file.tell());
     }
 
-    size_t write_tensor_info(llama_load_tensor & tensor) {
+    size_t write_tensor_info(llama_load_tensor & tensor, enum ggml_type type) {
         size_t total_written = 0;
         file.seek(info_offset, SEEK_SET);
         GGML_ASSERT(info_offset == file.tell());
         total_written += file.write_str(tensor.name);
+printf("total_written = %zu, name = %s\n", total_written, tensor.name.c_str());
 
         int32_t n_dims = tensor.ne.size();
         total_written += file.write_i32(n_dims);
         for (int32_t i = 0; i < n_dims; ++i) {
-            total_written += file.write_i32(i);
+            total_written += file.write_i32(tensor.ne[i]);
         }
 
+        total_written += file.write_i32(type);
         total_written += file.write_u64(tensor_offset);
         info_offset   += total_written;
 
         file.seek(0, SEEK_END);
+        printf("total_written = %zu\n", total_written);
 
         return total_written;
     }
@@ -781,7 +788,7 @@ struct gguf_file_saver {
             default: GGML_ASSERT(false);
         }
         
-        write_tensor_info(tensor);
+        write_tensor_info(tensor, new_type);
         file.write_raw(new_data, new_size);
         size_t padded_size = GGML_PAD(new_size, GGUF_DEFAULT_ALIGNMENT); // TODO: handle custom alignment
         size_t pad = padded_size - new_size;
diff --git a/gguf-util.h b/gguf-util.h
index 17f9dc968..ed7d53f69 100644
--- a/gguf-util.h
+++ b/gguf-util.h
@@ -109,18 +109,22 @@ struct gguf_file {
     size_t write_str(const std::string & val) {
         size_t total_written = 0;
         const int32_t n = val.size();
-        total_written += fwrite((const char *) &n, sizeof(n), 1, fp);
-        total_written += fwrite(val.c_str(), n, 1, fp);
+        fwrite((const char *) &n, sizeof(n), 1, fp);
+        total_written += sizeof(n);
+        fwrite(val.c_str(), n, 1, fp);
+        total_written += n;
 
         return total_written;
     }
 
     size_t write_i32(int32_t val) {
-        return fwrite((const char *) &val, sizeof(val), 1, fp);
+        fwrite((const char *) &val, sizeof(val), 1, fp);
+        return sizeof(val);
     }
 
     size_t write_u64(size_t val) {
-        return fwrite((const char *) &val, sizeof(val), 1, fp);
+        fwrite((const char *) &val, sizeof(val), 1, fp);
+        return sizeof(val);
     }
 
     void write_raw(const void * data, size_t size) {