llama : remove memory_f16 and kv_f16 flags

Author: ggerganov

common/common.cpp

@@ -278,8 +278,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.yarn_beta_slow = std::stof(argv[i]);
-        } else if (arg == "--memory-f32") {
-            params.memory_f16 = false;
         } else if (arg == "--top-p") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -804,8 +802,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     printf("  --no-penalize-nl      do not penalize newline token\n");
-    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
     printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
     printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
@@ -948,7 +944,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
     cparams.mul_mat_q         = params.mul_mat_q;
     cparams.seed              = params.seed;
-    cparams.f16_kv            = params.memory_f16;
     cparams.logits_all        = params.logits_all;
     cparams.embedding         = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
@@ -1375,7 +1370,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
-    fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
     fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
     fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);

common/common.h

@@ -98,7 +98,6 @@ struct gpt_params {
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
-    bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode

examples/quantize-stats/quantize-stats.cpp

@@ -321,7 +321,6 @@ int main(int argc, char ** argv) {
         auto cparams = llama_context_default_params();
         cparams.n_ctx      = 256;
         cparams.seed       = 1;
-        cparams.f16_kv     = false;
 
         ctx = llama_new_context_with_model(model, cparams);
 

examples/server/server.cpp

@@ -2108,10 +2108,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.yarn_beta_slow = std::stof(argv[i]);
         }
-        else if (arg == "--memory-f32" || arg == "--memory_f32")
-        {
-            params.memory_f16 = false;
-        }
         else if (arg == "--threads" || arg == "-t")
         {
             if (++i >= argc)

llama.cpp

@@ -8583,7 +8583,6 @@ struct llama_context_params llama_context_default_params() {
         /*.type_k                      =*/ GGML_TYPE_F16,
         /*.type_v                      =*/ GGML_TYPE_F16,
         /*.mul_mat_q                   =*/ true,
-        /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
         /*.embedding                   =*/ false,
         /*.offload_kqv                 =*/ true,
@@ -8737,8 +8736,6 @@ struct llama_context * llama_new_context_with_model(
     ctx->rng = std::mt19937(params.seed);
     ctx->logits_all = params.logits_all;
 
-    //const ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
-
     const ggml_type type_k = params.type_k;
     const ggml_type type_v = params.type_v;
 

llama.h

@@ -196,7 +196,6 @@ extern "C" {
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool mul_mat_q;   // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
-        bool f16_kv;      // use fp16 for KV cache, fp32 otherwise
         bool logits_all;  // the llama_eval() call computes all logits, not just the last one
         bool embedding;   // embedding mode only
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU