Add ability to hide imatrix details in llama-quantize (#329)

Co-authored-by: Iwan Kawrakow <[email protected]>

Author: ikawrakow

examples/quantize/quantize.cpp

@@ -142,11 +142,12 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--attn-q-type] [--attn-k-type] [--attn-v-type] [--attn-qkv-type] [--attn-output-type] [--ffn-gate-type] [--ffn-down-type] [--ffn-up-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--hide-imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--attn-q-type] [--attn-k-type] [--attn-v-type] [--attn-qkv-type] [--attn-output-type] [--ffn-gate-type] [--ffn-down-type] [--ffn-up-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
     printf("  --imatrix file_name: use data in file_name as importance matrix for quant optimizations\n");
+    printf("  --hide-imatrix: do not store imatrix details in the quantized model\n");
     printf("  --include-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor.\n");
@@ -337,6 +338,8 @@ int main(int argc, char ** argv) {
 
     std::vector<std::string> repack_patterns;
 
+    bool hide_imatrix = false;
+
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
             params.quantize_output_tensor = false;
@@ -429,6 +432,8 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--hide-imatrix") == 0) {
+            hide_imatrix = true;
         } else if (strcmp(argv[arg_idx], "--include-weights") == 0) {
             if (arg_idx < argc-1) {
                 included_weights.emplace_back(argv[++arg_idx]);
@@ -469,15 +474,23 @@ int main(int argc, char ** argv) {
             llama_model_kv_override kvo;
             std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_FILE);
             kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
-            strncpy(kvo.val_str, imatrix_file.c_str(), 127);
+            if (hide_imatrix) {
+                strncpy(kvo.val_str, "top_secret", 127);
+            } else {
+                strncpy(kvo.val_str, imatrix_file.c_str(), 127);
+            }
             kvo.val_str[127] = '\0';
             kv_overrides.emplace_back(std::move(kvo));
         }
         if (!imatrix_dataset.empty()) {
             llama_model_kv_override kvo;
             std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_DATASET);
             kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
-            strncpy(kvo.val_str, imatrix_dataset.c_str(), 127);
+            if (hide_imatrix) {
+                strncpy(kvo.val_str, "top_secret", 127);
+            } else {
+                strncpy(kvo.val_str, imatrix_dataset.c_str(), 127);
+            }
             kvo.val_str[127] = '\0';
             kv_overrides.emplace_back(std::move(kvo));
         }
@@ -486,15 +499,23 @@ int main(int argc, char ** argv) {
             llama_model_kv_override kvo;
             std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_N_ENTRIES);
             kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
-            kvo.val_i64 = imatrix_data.size();
+            if (hide_imatrix) {
+                kvo.val_i64 = 0;
+            } else {
+                kvo.val_i64 = imatrix_data.size();
+            }
             kv_overrides.emplace_back(std::move(kvo));
         }
 
         if (m_last_call > 0) {
             llama_model_kv_override kvo;
             std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_N_CHUNKS);
             kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
-            kvo.val_i64 = m_last_call;
+            if (hide_imatrix) {
+                kvo.val_i64 = 0;
+            } else {
+                kvo.val_i64 = m_last_call;
+            }
             kv_overrides.emplace_back(std::move(kvo));
         }
     }