quantize: options for output and token embedding tensors qtype (#6239)

* quantize: be able to specify the output tensor type

* quantize: be able to specify the token embedding tensor type

---------

Co-authored-by: Iwan Kawrakow <redacted>

diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp
index 7662ec80c5e0fe77e834bf3b1a0604bff74a668b..79e60ea7ba6b9ff3c6186e9a742122f2effca21a 100644 (file)
--- a/examples/quantize/quantize.cpp
+++ b/examples/quantize/quantize.cpp
@@ -189,6 +189,18 @@ static void prepare_imatrix(const std::string& imatrix_file,
     }
 }
 
+static ggml_type parse_ggml_type(const char * arg) {
+    ggml_type result = GGML_TYPE_COUNT;
+    for (int j = 0; j < GGML_TYPE_COUNT; ++j) {
+        auto type = ggml_type(j);
+        const auto * name = ggml_type_name(type);
+        if (name && strcmp(arg, name) == 0) {
+            result = type; break;
+        }
+    }
+    return result;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -203,6 +215,18 @@ int main(int argc, char ** argv) {
     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;
+        } else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.output_tensor_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.token_embedding_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {

diff --git a/llama.cpp b/llama.cpp
index aa6c8924660b095eaaccdaedab4e65eda7eb95a3..eedca802b86a78442f43a9f588881ef317d660ce 100644 (file)
--- a/llama.cpp
+++ b/llama.cpp
@@ -12141,27 +12141,34 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
     // for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings
     // with the quantization of the output tensor
     if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) {
-        int nx = tensor->ne[0];
-        if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
-            new_type = GGML_TYPE_Q8_0;
-        }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
-                 ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
-            new_type = GGML_TYPE_Q5_K;
-        }
-        else if (new_type != GGML_TYPE_Q8_0) {
-            new_type = GGML_TYPE_Q6_K;
+        if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
+            new_type = qs.params->output_tensor_type;
+        } else {
+            int nx = tensor->ne[0];
+            if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
+                new_type = GGML_TYPE_Q8_0;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
+                    ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
+                new_type = GGML_TYPE_Q5_K;
+            }
+            else if (new_type != GGML_TYPE_Q8_0) {
+                new_type = GGML_TYPE_Q6_K;
+            }
         }
     } else if (name == "token_embd.weight") {
-        if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
-            ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
-            new_type = GGML_TYPE_Q2_K;
-        }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
-            new_type = GGML_TYPE_IQ3_S;
-        }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
-            new_type = GGML_TYPE_IQ3_S;
+        if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
+            new_type = qs.params->token_embedding_type;
+        } else {
+            if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
+                new_type = GGML_TYPE_Q2_K;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
+                new_type = GGML_TYPE_IQ3_S;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+                new_type = GGML_TYPE_IQ3_S;
+            }
         }
     } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
                ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
@@ -13051,6 +13058,8 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
     struct llama_model_quantize_params result = {
         /*.nthread                     =*/ 0,
         /*.ftype                       =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
+        /*.output_tensor_type          =*/ GGML_TYPE_COUNT,
+        /*.token_embedding_type        =*/ GGML_TYPE_COUNT,
         /*.allow_requantize            =*/ false,
         /*.quantize_output_tensor      =*/ true,
         /*.only_copy                   =*/ false,

diff --git a/llama.h b/llama.h
index 7e8ac4b62beb56a590024679b090ee6fb849a48f..74f0e56de71c6425ebd6a66c0cc58f13d0b208b1 100644 (file)
--- a/llama.h
+++ b/llama.h
@@ -275,13 +275,15 @@ extern "C" {
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
-        int32_t nthread;             // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;      // quantize to this llama_ftype
-        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor; // quantize output.weight
-        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                   // quantize all tensors to the default type
-        void * imatrix;              // pointer to importance matrix data
+        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;              // quantize to this llama_ftype
+        enum ggml_type output_tensor_type;   // output tensor type
+        enum ggml_type token_embedding_type; // itoken embeddings tensor type
+        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor;         // quantize output.weight
+        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                           // quantize all tensors to the default type
+        void * imatrix;                      // pointer to importance matrix data
     } llama_model_quantize_params;
 
     // grammar types