Command line switch to use F16 for memory_k and memory_v (refactor of g…

…gerganov#154) (ggerganov#294)

* Use F16 for memory_k and memory_v

* add command line switch to use f16 instead of f32 for memory k+v

---------

Co-authored-by: Ty Everett <ty@tyweb.us>

main.cpp

@@ -86,7 +86,7 @@ struct llama_model {
 };
 
 // load the model's weights from a file
-bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab & vocab, int n_ctx) {
+bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab & vocab, int n_ctx, ggml_type memory_type = GGML_TYPE_F32) {
     fprintf(stderr, "%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
 
     std::vector<char> f_buf(1024*1024);
@@ -207,8 +207,8 @@ bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab
         ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w2
         ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w3
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_v
 
         ctx_size += (5 + 10*n_layer)*256; // object overhead
 
@@ -293,8 +293,8 @@ bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab
         const int n_mem      = n_layer*n_ctx;
         const int n_elements = n_embd*n_mem;
 
-        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
-        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+        model.memory_k = ggml_new_tensor_1d(ctx, memory_type, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, memory_type, n_elements);
 
         const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
 
@@ -814,8 +814,9 @@ int main(int argc, char ** argv) {
 
     // load the model
     {
+        const ggml_type memory_type = params.memory_f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
         const int64_t t_start_us = ggml_time_us();
-        if (!llama_model_load(params.model, model, vocab, params.n_ctx)) {
+        if (!llama_model_load(params.model, model, vocab, params.n_ctx, memory_type)) {
             fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
             return 1;
         }

utils.cpp

@@ -49,6 +49,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.top_k = std::stoi(argv[++i]);
         } else if (arg == "-c" || arg == "--ctx_size") {
             params.n_ctx = std::stoi(argv[++i]);
+        } else if (arg == "--memory_f16") {
+            params.memory_f16 = true;
         } else if (arg == "--top_p") {
             params.top_p = std::stof(argv[++i]);
         } else if (arg == "--temp") {
@@ -104,6 +106,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --repeat_last_n N     last n tokens to consider for penalize (default: %d)\n", params.repeat_last_n);
     fprintf(stderr, "  --repeat_penalty N    penalize repeat sequence of tokens (default: %.1f)\n", params.repeat_penalty);
     fprintf(stderr, "  -c N, --ctx_size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  --memory_f16          use f16 instead of f32 for memory key+value\n");
     fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", params.temp);
     fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  -m FNAME, --model FNAME\n");

utils.h

@@ -18,6 +18,7 @@ struct gpt_params {
     int32_t n_predict = 128; // new tokens to predict
     int32_t repeat_last_n = 64;  // last n tokens to penalize
     int32_t n_ctx = 512; //context size
+    bool memory_f16 = false; // use f16 instead of f32 for memory kv
 
     // sampling parameters
     int32_t top_k = 40;