merging, do not use

ggml-cuda.cu

@@ -42,19 +42,19 @@ typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y,
 #define QK4_0 32
 #define QR4_0 2
 typedef struct {
-    float   d;              // delta
+    half    d;              // delta
     uint8_t qs[QK4_0 / 2];  // nibbles / quants
 } block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
 
 #define QK4_1 32
 #define QR4_1 2
 typedef struct {
-    float   d;              // delta
+    half    d;              // delta
-    float   m;              // min
+    half    m;              // min
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
-static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
 
 #define QK5_0 32
 #define QR5_0 2
@@ -78,10 +78,10 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) +
 #define QK8_0 32
 #define QR8_0 1
 typedef struct {
-    float   d;              // delta
+    half    d;              // delta
     int8_t  qs[QK8_0];      // quants
 } block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
 
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
 #define CUDA_DMMV_BLOCK_SIZE 32 // dmmv = dequantize_mul_mat_vec

ggml-opencl.cpp

@@ -813,7 +813,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
     if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
         src1->type == GGML_TYPE_F32 &&
         dst->type == GGML_TYPE_F32 &&
-        ((GetQuantsUnshuffled() && GetGPULayers()>0 && ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_CL)) {
+        ((GetQuantsUnshuffled() && ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_CL)) {
         return true;
     }
 

ggml.h

@@ -1094,8 +1094,6 @@ extern "C" {
 
     void SetQuantsUnshuffled(bool unshuffled);
     bool GetQuantsUnshuffled();
-    void SetGPULayers(bool layers);
-    bool GetGPULayers();
 
     GGML_API int ggml_cpu_has_avx        (void);
     GGML_API int ggml_cpu_has_avx2       (void);

gpttype_adapter.cpp

@@ -20,8 +20,8 @@
 #include "gptj_v2.cpp"
 #include "gpt2_v1.cpp"
 #include "gpt2_v2.cpp"
-#include "rwkv.cpp"
+#include "rwkv_v2.cpp"
-#include "neox.cpp"
+#include "neox_v2.cpp"
 
 
 //return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt)

llama.cpp

@@ -412,6 +412,7 @@ enum llama_file_version {
     LLAMA_FILE_VERSION_GGMF_V1, // added version field and scores in vocab
     LLAMA_FILE_VERSION_GGJT_V1, // added padding
     LLAMA_FILE_VERSION_GGJT_V2, // changed quantization format
+    LLAMA_FILE_VERSION_GGJT_V3, // changed Q4 and Q8 quantization format
 };
 
 struct llama_file_loader {
@@ -444,6 +445,8 @@ struct llama_file_loader {
             file_version = LLAMA_FILE_VERSION_GGJT_V1;
         } else if (magic == 'ggjt' && version == 2) {
             file_version = LLAMA_FILE_VERSION_GGJT_V2;
+        } else if (magic == 'ggjt' && version == 3) {
+            file_version = LLAMA_FILE_VERSION_GGJT_V3;
         } else {
             throw format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?",
                          magic, version);
@@ -861,7 +864,8 @@ static const char *llama_file_version_name(llama_file_version version) {
         case LLAMA_FILE_VERSION_GGML: return "'ggml' (old version with low tokenizer quality and no mmap support)";
         case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (old version with no mmap support)";
         case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (pre #1405)";
-        case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (latest)";
+        case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (pre #1508)";
+        case LLAMA_FILE_VERSION_GGJT_V3: return "ggjt v3 (latest)";
     }
 
     return "unknown";
@@ -946,12 +950,19 @@ static void llama_model_load_internal(
         fprintf(stderr, "%s: model size = %s\n",  __func__, llama_model_type_name(model.type));
     }
 
-    if (file_version != LLAMA_FILE_VERSION_GGJT_V2) {
+    if (file_version < LLAMA_FILE_VERSION_GGJT_V2) {
         if (hparams.ftype != LLAMA_FTYPE_ALL_F32     &&
             hparams.ftype != LLAMA_FTYPE_MOSTLY_F16  &&
             hparams.ftype != LLAMA_FTYPE_MOSTLY_Q8_0) {
-            printf("\nLegacy LLAMA GGJT compatability changes triggered.\n");
+            printf("\nLegacy LLAMA GGJT v1 compatability changes triggered.\n");
-            //throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1305)");
+        }
+    }
+
+    if (file_version < LLAMA_FILE_VERSION_GGJT_V3) {
+        if (hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
+            hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_1 ||
+            hparams.ftype == LLAMA_FTYPE_MOSTLY_Q8_0) {
+            printf("\nLegacy LLAMA GGJT v2 compatability changes triggered.\n");
         }
     }
 

llama.h

@@ -19,7 +19,7 @@
 #    define LLAMA_API
 #endif
 
-#define LLAMA_FILE_VERSION           2
+#define LLAMA_FILE_VERSION           3
 #define LLAMA_FILE_MAGIC             'ggjt'
 #define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml'
 #define LLAMA_SESSION_MAGIC          'ggsn'

otherarch/rwkv_v2.cpp

@@ -3,7 +3,7 @@
 
 #include "otherarch.h"
 
-#include "rwkv.h"
+#include "rwkv_v2.h"
 #include "ggml.h"
 
 #include <string>