rename scale

developer_document.md

@@ -8,7 +8,7 @@ For simplicity, we take [polyglot](https://huggingface.co/EleutherAI/polyglot-ko
 
 Firstly, we need to add its temp buffer in its [related model-arch header file](neural_speed/models/gptneox/gptneox.h) and [re-compile](README.md#Install).
 ```diff
-static const model_scratch gptneox_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch gptneox_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 44:
       return {2048ull * MB, 2048ull * MB, 4096ull * MB};
@@ -167,7 +167,7 @@ and update [model_name_to_arch()](neural_speed/models/model_utils/model_types.h#
 +  NEW_MODEL_13B,
 +};
 
-+static const model_scratch new_model_mem_req(int n_layers, float enlarge_scale = 1.0f) {
++static const model_scratch new_model_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
 +  switch (n_layers) {
 +    case N:
 +      return {8192ull * MB, 8192ull * MB, 8192ull * MB};
@@ -390,7 +390,7 @@ We recommend to use continuous batching way since it has no padding effect and c
 +     ne_view_2d(ctx0, KQV_merged_contiguous, head_size * n_head, attn_sl * attn_bs, head_size * n_head * ne_element_size(KQV_merged_contiguous), ne_element_size(KQV_merged_contiguous) * off_sl)));
 +  off_sl += head_size * n_head * attn_sl * attn_bs;
 ```
->Note: You can set larger [`NE_MAX_NODES`](neural_speed/core/ne.h#43) and [`model_scratch_enlarge_scale`](neural_speed/models/llama/llama.h#29) values if out of memory when the inputs' batch size becomes larger.
+>Note: You can set larger [`NE_MAX_NODES`](neural_speed/core/ne.h#43) and [`model_scratch_size_ratio`](neural_speed/models/llama/llama.h#29) values if out of memory when the inputs' batch size becomes larger.
 
 ## 2.3.	Application
 - Q4_0 quant : We can quantize the model generated by convert by adding a quant layer class to quantize it into an int4 low-bit file, so as to obtain better inference performance. Register quant layer class in your new_model_utils.cpp, just like [gptneox_utils.cpp](neural_speed/models/gptneox/gptneox_utils.cpp#L163), replace `gptneox_quant_layer` to your `new_model_quant_layer`.

neural_speed/application/main_pybind.cpp

@@ -84,7 +84,7 @@ void init_gpt_params(gpt_params* params, const std::string& model_path, int max_
                      bool early_stopping = false, int n_keep = 0, int n_discard = -1, bool shift_roped_k = false,
                      int batch_size = 1, model_vocab::id pad_token = -1, const std::string& memory_dtype = "auto",
                      bool continuous_batching = true, const int& max_request_num = MODEL_MAX_REQUEST_NUM,
-                     const float& model_scratch_enlarge_scale = 1.0f) {
+                     const float& model_scratch_size_ratio = 1.0f) {
   MODEL_ASSERT(params != nullptr);
 #ifdef MODEL_NAME
   params->model_name = MODEL_NAME;
@@ -126,7 +126,7 @@ void init_gpt_params(gpt_params* params, const std::string& model_path, int max_
   params->min_new_tokens = min_new_tokens;
   params->length_penalty = length_penalty;
   params->do_early_stopping = early_stopping;
-  params->model_scratch_enlarge_scale = model_scratch_enlarge_scale;
+  params->model_scratch_size_ratio = model_scratch_size_ratio;
 
   printf(
       "beam_size: %d, do_sample: %d, top_k: %d, top_p: %f, continuous_batching: %d, max_request_num: %d, "
@@ -142,7 +142,7 @@ class ModelServer {
               int top_k, float top_p, float temperature, int min_new_tokens, float length_penalty, bool early_stopping,
               int n_keep, int n_discard, bool shift_roped_k, int batch_size, model_vocab::id pad_token,
               const std::string& memory_dtype, bool continuous_batching, const int& max_request_num,
-              const float& model_scratch_enlarge_scale, const std::string& policy, bool print_log,
+              const float& model_scratch_size_ratio, const std::string& policy, bool print_log,
               const std::function<void()>& init_cb)
       : response(response),
         waiting(),
@@ -161,7 +161,7 @@ class ModelServer {
           this->InitServerParams(model_path, max_new_tokens, n_batch, ctx_size, seed, threads, repetition_penalty,
                                  num_beams, do_sample, top_k, top_p, temperature, min_new_tokens, length_penalty,
                                  early_stopping, n_keep, n_discard, shift_roped_k, batch_size, pad_token, memory_dtype,
-                                 true, max_request_num, model_scratch_enlarge_scale);
+                                 true, max_request_num, model_scratch_size_ratio);
           Cont_batch_gen_scheduler scheduler(this->params, policy, print_log ? 0 : 1);
           std::vector<sequence> added_seqs;
           while (running) {
@@ -263,11 +263,11 @@ class ModelServer {
                         float temperature, int min_new_tokens, float length_penalty, bool early_stopping, int n_keep,
                         int n_discard, bool shift_roped_k, int batch_size, model_vocab::id pad_token,
                         const std::string& memory_dtype, bool continuous_batching, const int& max_request_num,
-                        const float& model_scratch_enlarge_scale) {
+                        const float& model_scratch_size_ratio) {
     init_gpt_params(&params, model_path, max_new_tokens, n_batch, ctx_size, seed, threads, repetition_penalty,
                     num_beams, do_sample, top_k, top_p, temperature, min_new_tokens, length_penalty, early_stopping,
                     n_keep, n_discard, shift_roped_k, batch_size, pad_token, memory_dtype, continuous_batching,
-                    max_request_num, model_scratch_enlarge_scale);
+                    max_request_num, model_scratch_size_ratio);
     if (cont_batching_model_archs.count(params.model_arch) == 0) {
       fprintf(stderr, "\nERROR: ModelServer only supports gpt-j, llama!\n");
       running = false;
@@ -325,7 +325,7 @@ class Model {
                   float repetition_penalty, int num_beams, bool do_sample, int top_k, float top_p, float temperature,
                   int min_new_tokens, float length_penalty, bool early_stopping, int n_keep, int n_discard,
                   bool shift_roped_k, int batch_size, model_vocab::id pad_token, const std::string& memory_dtype,
-                  bool continuous_batching, const int& max_request_num, const float& model_scratch_enlarge_scale);
+                  bool continuous_batching, const int& max_request_num, const float& model_scratch_size_ratio);
   void reinit();
   std::vector<std::vector<model_token>> generate(const std::vector<std::vector<model_token>>& input_ids);
   // deprecated API
@@ -419,11 +419,11 @@ void Model::init_model(const std::string& model_path, int max_new_tokens, int n_
                        float temperature, int min_new_tokens, float length_penalty, bool early_stopping, int n_keep,
                        int n_discard, bool shift_roped_k, int batch_size, model_vocab::id pad_token,
                        const std::string& memory_dtype, bool continuous_batching, const int& max_request_num,
-                       const float& model_scratch_enlarge_scale) {
+                       const float& model_scratch_size_ratio) {
   init_gpt_params(&params, model_path, max_new_tokens, n_batch, ctx_size, seed, threads, repetition_penalty, num_beams,
                   do_sample, top_k, top_p, temperature, min_new_tokens, length_penalty, early_stopping, n_keep,
                   n_discard, shift_roped_k, batch_size, pad_token, memory_dtype, continuous_batching, max_request_num,
-                  model_scratch_enlarge_scale);
+                  model_scratch_size_ratio);
 
   n_past = 0;
   n_total = 0;
@@ -922,7 +922,7 @@ PYBIND11_MODULE(mixtral_cpp, m)
            py::arg("n_keep") = 0, py::arg("n_discard") = -1, py::arg("shift_roped_k") = false,
            py::arg("batch_size") = 1, py::arg("pad_token") = -1, py::arg("memory_dtype") = "auto",
            py::arg("continuous_batching") = true, py::arg("max_request_num") = MODEL_MAX_REQUEST_NUM,
-           py::arg("model_scratch_enlarge_scale") = 1.0f)
+           py::arg("model_scratch_size_ratio") = 1.0f)
       .def("generate", &Model::generate, "Generate token with input ids", py::arg("input_ids"))
       .def("evaluate", &Model::evaluate, "Evaluate token with input ids and output logits",
            py::arg("input_ids") = std::vector<std::vector<model_token>>{}, py::arg("logits_all") = false)
@@ -962,7 +962,7 @@ PYBIND11_MODULE(mixtral_cpp, m)
            py::arg("length_penalty") = 1.0, py::arg("early_stopping") = false, py::arg("n_keep") = 0,
            py::arg("n_discard") = -1, py::arg("shift_roped_k") = false, py::arg("batch_size") = 1,
            py::arg("pad_token") = -1, py::arg("memory_dtype") = "auto", py::arg("continuous_batching") = true,
-           py::arg("max_request_num") = MODEL_MAX_REQUEST_NUM, py::arg("model_scratch_enlarge_scale") = 1.0f,
+           py::arg("max_request_num") = MODEL_MAX_REQUEST_NUM, py::arg("model_scratch_size_ratio") = 1.0f,
            py::arg("policy") = "fcfs", py::arg("print_log") = false,
            py::arg("init_cb") = std::function<void()>{[]() {}})
       .def("issueQuery", &ModelServer::issueQuery, "desc placeholder", py::arg("qs"))

neural_speed/models/baichuan/baichuan.h

@@ -23,7 +23,7 @@ enum baichuan_model {
   BAICHUAN_13B,
 };
 
-static const model_scratch baichuan_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch baichuan_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 40:
       return {8192ull * MB, 8192ull * MB, 8192ull * MB};

neural_speed/models/baichuan/baichuan_utils.cpp

@@ -75,7 +75,7 @@ void BAICHUAN::init(const char* path_model, model_context* ctx, int n_gpu_layer_
   n_embd = hparams.n_embd;
   n_vocab = hparams.n_vocab;
   n_layer = hparams.n_layer;
-  scratch = baichuan_mem_req(n_layer, lctx.model_scratch_enlarge_scale);
+  scratch = baichuan_mem_req(n_layer, lctx.model_scratch_size_ratio);
   model.scratchs = scratch;
 }
 

neural_speed/models/bloom/bloom.h

@@ -23,7 +23,7 @@ enum bloom_model {
   BLOOM_7B,
 };
 
-static const model_scratch bloom_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch bloom_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 30:
       return {4 * 2048ull * MB, 4 * 2048ull * MB, 4 * 4096ull * MB};

neural_speed/models/bloom/bloom_utils.cpp

@@ -73,7 +73,7 @@ void BLOOM::init(const char* path_model, model_context* ctx, int n_gpu_layer_, b
   n_embd = hparams.n_embd;
   n_vocab = hparams.n_vocab;
   n_layer = hparams.n_layer;
-  scratch = bloom_mem_req(n_layer, lctx.model_scratch_enlarge_scale);
+  scratch = bloom_mem_req(n_layer, lctx.model_scratch_size_ratio);
   model.scratchs = scratch;
 }
 

neural_speed/models/chatglm/chatglm.h

@@ -23,7 +23,7 @@ enum chatglm_model {
   CHATGLM_6B,
 };
 
-static const model_scratch chatglm_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch chatglm_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 28:
       return {2048ull * MB, 2048ull * MB, 4096ull * MB};

neural_speed/models/chatglm/chatglm2.h

@@ -23,7 +23,7 @@ enum chatglm2_model {
   CHATGLM2_6B,
 };
 
-static const model_scratch chatglm_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch chatglm_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 28:
       return {4096ull * MB, 4096ull * MB, 8192ull * MB};

neural_speed/models/chatglm/chatglm2_utils.cpp

@@ -78,7 +78,7 @@ void CHATGLM2::init(const char* path_model, model_context* ctx, int n_gpu_layer_
   n_embd = hparams.n_embd;
   n_vocab = hparams.n_vocab;
   n_layer = hparams.n_layer;
-  scratch = chatglm_mem_req(n_layer, lctx.model_scratch_enlarge_scale);
+  scratch = chatglm_mem_req(n_layer, lctx.model_scratch_size_ratio);
   model.scratchs = scratch;
 }
 

neural_speed/models/chatglm/chatglm_utils.cpp

@@ -72,7 +72,7 @@ void CHATGLM::init(const char* path_model, model_context* ctx, int n_gpu_layer_,
   n_embd = hparams.n_embd;
   n_vocab = hparams.n_vocab;
   n_layer = hparams.n_layer;
-  scratch = chatglm_mem_req(n_layer, lctx.model_scratch_enlarge_scale);
+  scratch = chatglm_mem_req(n_layer, lctx.model_scratch_size_ratio);
   model.scratchs = scratch;
 }
 

neural_speed/models/falcon/falcon.h

@@ -23,7 +23,7 @@ enum falcon_model {
   FALCON_7B,
 };
 
-static const model_scratch falcon_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch falcon_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 32:
       return {2048ull * MB, 2048ull * MB, 4096ull * MB};

neural_speed/models/falcon/falcon_utils.cpp

@@ -77,7 +77,7 @@ void FALCON::init(const char* path_model, model_context* ctx, int n_gpu_layer_,
   n_vocab = hparams.n_vocab;
   n_layer = hparams.n_layer;
   n_head_kv = hparams.n_head_kv;
-  scratch = falcon_mem_req(n_layer, lctx.model_scratch_enlarge_scale);
+  scratch = falcon_mem_req(n_layer, lctx.model_scratch_size_ratio);
   model.scratchs = scratch;
 }
 

neural_speed/models/gptj/gptj.h

@@ -26,14 +26,14 @@ enum gptj_model {
   GPTJ_65B,
 };
 
-static const model_scratch gptj_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch gptj_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 28:
       // should be enough for batch=8 * beam=4
       return {
-          static_cast<unsigned long long>(enlarge_scale * 3072) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 2048) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 3072) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 3072) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 2048) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 3072) * MB,
       };
     default:
       MODEL_ASSERT(false);

neural_speed/models/gptj/gptj_utils.cpp

@@ -75,7 +75,7 @@ void GPTJ::init(const char* path_model, model_context* ctx, int n_gpu_layer_, bo
   n_embd = hparams.n_embd;
   n_vocab = hparams.n_vocab;
   n_layer = hparams.n_layer;
-  scratch = gptj_mem_req(n_layer, lctx.model_scratch_enlarge_scale);
+  scratch = gptj_mem_req(n_layer, lctx.model_scratch_size_ratio);
   model.scratchs = scratch;
 }
 

neural_speed/models/gptneox/gptneox.h

@@ -23,7 +23,7 @@ enum gptneox_model {
   GPTNEOX_7B,
 };
 
-static const model_scratch gptneox_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch gptneox_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 44:
       return {2048ull * MB, 2048ull * MB, 4096ull * MB};

neural_speed/models/gptneox/gptneox_utils.cpp

@@ -74,7 +74,7 @@ void GPTNEOX::init(const char* path_model, model_context* ctx, int n_gpu_layer_,
   n_embd = hparams.n_embd;
   n_vocab = hparams.n_vocab;
   n_layer = hparams.n_layer;
-  scratch = gptneox_mem_req(n_layer, lctx.model_scratch_enlarge_scale);
+  scratch = gptneox_mem_req(n_layer, lctx.model_scratch_size_ratio);
   model.scratchs = scratch;
 }
 

neural_speed/models/llama/llama.h

@@ -26,37 +26,37 @@ enum llama_model {
   LLAMA_65B,
 };
 
-static const model_scratch llama_mem_req(int n_layers, float enlarge_scale = 1.0f) {
+static const model_scratch llama_mem_req(int n_layers, float scratch_size_ratio = 1.0f) {
   switch (n_layers) {
     case 32:
       return {
-          static_cast<unsigned long long>(enlarge_scale * 1024) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 1024) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 1608) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 1024) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 1024) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 1608) * MB,
       };
     case 40:
       return {
-          static_cast<unsigned long long>(enlarge_scale * 512) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 512) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 1608) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 512) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 512) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 1608) * MB,
       };
     case 48:
       return {
-          static_cast<unsigned long long>(enlarge_scale * 512) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 512) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 2366) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 512) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 512) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 2366) * MB,
       };
     case 60:
       return {
-          static_cast<unsigned long long>(enlarge_scale * 512) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 512) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 3124) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 512) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 512) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 3124) * MB,
       };
     case 80:
       return {
-          static_cast<unsigned long long>(enlarge_scale * 2048) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 2048) * MB,
-          static_cast<unsigned long long>(enlarge_scale * 10240) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 2048) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 2048) * MB,
+          static_cast<unsigned long long>(scratch_size_ratio * 10240) * MB,
       };
     default:
       MODEL_ASSERT(false);

neural_speed/models/llama/llama_utils.cpp

@@ -81,7 +81,7 @@ void Llama::init(const char* path_model, model_context* ctx, int n_gpu_layer_, b
   n_head = hparams.n_head;
   n_expert = hparams.n_experts;
   n_expert_used = hparams.n_experts_used;
-  scratch = llama_mem_req(n_layer, lctx.model_scratch_enlarge_scale);
+  scratch = llama_mem_req(n_layer, lctx.model_scratch_size_ratio);
   model.scratchs = scratch;
 }
 

neural_speed/models/model_utils/model_config.h

@@ -103,7 +103,7 @@ struct gpt_params {
   float length_penalty = 1.0f;     // exponential penalty to the length in beam search generation
   bool do_early_stopping = false;  // early stopping in beam search generation
 
-  float model_scratch_enlarge_scale = 1.0f;  // model memory scratch enlarge scale
+  float model_scratch_size_ratio = 1.0f;  // model memory scratch enlarge scale
 };
 
 bool gpt_params_parse(int argc, char** argv, gpt_params& params);

neural_speed/models/model_utils/model_types.h

@@ -320,7 +320,7 @@ struct model_context {
 
   size_t mem_per_token = 0;
 
-  float model_scratch_enlarge_scale = 1.0f;  // model memory scratch enlarge scale
+  float model_scratch_size_ratio = 1.0f;  // model memory scratch enlarge scale
 
   // decode output (3-dimensional array: [batch_size] [n_tokens] [n_vocab])
   std::vector<float> logits;
@@ -441,7 +441,7 @@ struct model_context_params {
   // global generation config
   generation_config gen_conf;
   // model memory scratch enlarge scale
-  float model_scratch_enlarge_scale;
+  float model_scratch_size_ratio;
 
   // called with a progress value between 0 and 1, pass nullptr to disable
   model_progress_callback progress_callback;

neural_speed/models/model_utils/model_utils.cpp

@@ -188,7 +188,7 @@ struct model_context_params model_context_default_params() {
       /*cont_batching                =*/true,
       /*.max_request_num             =*/1,
       /*.gen_conf                    =*/generation_config(),
-      /*model_scratch_enlarge_scale  =*/1.0f,
+      /*model_scratch_size_ratio  =*/1.0f,
       /*.progress_callback           =*/nullptr,
       /*.progress_callback_user_data =*/nullptr,
   };
@@ -911,7 +911,7 @@ struct model_context* model_init_from_file(const char* path_model, struct model_
   }
   ctx->cont_batching = params.cont_batching;
   ctx->generation_conf = params.gen_conf;
-  ctx->model_scratch_enlarge_scale = params.model_scratch_enlarge_scale;
+  ctx->model_scratch_size_ratio = params.model_scratch_size_ratio;
   const model_archs arch = params.arch;
 
   // the type so that kv-cache allocated according to this type must be large enough
@@ -1284,7 +1284,7 @@ struct model_context* model_init_from_gpt_params(const gpt_params& params) {
   lparams.gen_conf.min_new_tokens = params.min_new_tokens;
   lparams.gen_conf.length_penalty = params.length_penalty;
   lparams.gen_conf.do_early_stopping = params.do_early_stopping;
-  lparams.model_scratch_enlarge_scale = params.model_scratch_enlarge_scale;
+  lparams.model_scratch_size_ratio = params.model_scratch_size_ratio;
 
   NE_ASSERT(("Start size cannot be greater than the maximum context size!", lparams.n_keep < lparams.n_ctx));