Merge pull request #2 from ggerganov/master

latest updates

.github/workflows/build.yml

@@ -10,10 +10,10 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp']
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
 
 env:
  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}

Makefile

@@ -127,6 +127,7 @@ endif
 
 ifndef LLAMA_NO_K_QUANTS
 	CFLAGS   += -DGGML_USE_K_QUANTS
+	CXXFLAGS += -DGGML_USE_K_QUANTS
 	OBJS     += k_quants.o
 endif
 

SHA256SUMS

@@ -1,6 +1,6 @@
 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d  models/7B/consolidated.00.pth
 666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847  models/7B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q4_0.bin
+ec2f2d1f0dfb73b72a4cbac7fa121abbe04c37ab327125a38248f930c0f09ddf  models/7B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_1.bin
@@ -8,7 +8,7 @@ ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml
 745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08  models/13B/consolidated.00.pth
 d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085  models/13B/consolidated.01.pth
 2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808  models/13B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q4_0.bin
+fad169e6f0f575402cf75945961cb4a8ecd824ba4da6be2af831f320c4348fa5  models/13B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_1.bin
@@ -18,7 +18,7 @@ e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067  models/30B/con
 24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378  models/30B/consolidated.02.pth
 1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b  models/30B/consolidated.03.pth
 7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37  models/30B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q4_0.bin
+d2a441403944819492ec8c2002cc36fa38468149bfb4b7b4c52afc7bd9a7166d  models/30B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_1.bin
@@ -32,7 +32,7 @@ a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78  models/65B/con
 72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b  models/65B/consolidated.06.pth
 d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638  models/65B/consolidated.07.pth
 60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0  models/65B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q4_0.bin
+cde053439fa4910ae454407e2717cc46cc2c2b4995c00c93297a2b52e790fa92  models/65B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_1.bin

examples/CMakeLists.txt

@@ -37,6 +37,7 @@ else()
     add_subdirectory(save-load-state)
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
+    add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -79,34 +79,39 @@ struct ggml_tensor * randomize_tensor_normal(
         int ndims,
         const int64_t ne[],
         struct random_normal_distribution * rnd) {
+    float scale = 1.0; // xavier
     switch (ndims) {
         case 1:
+            scale /= sqrtf(ne[0]);
             for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = frand_normal(rnd);
+                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
             }
             break;
         case 2:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i1 = 0; i1 < ne[1]; i1++) {
                 for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = frand_normal(rnd);
+                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
                 }
             }
             break;
         case 3:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i2 = 0; i2 < ne[2]; i2++) {
                 for (int i1 = 0; i1 < ne[1]; i1++) {
                     for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                     }
                 }
             }
             break;
         case 4:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i3 = 0; i3 < ne[3]; i3++) {
                 for (int i2 = 0; i2 < ne[2]; i2++) {
                     for (int i1 = 0; i1 < ne[1]; i1++) {
                         for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                         }
                     }
                 }
@@ -148,8 +153,8 @@ struct llama_hparams_lora {
     uint32_t n_rot   = 64;
     uint32_t n_lora  = 64;
 
-    bool operator!=(const llama_hparams & other) const {
+    bool operator!=(const llama_hparams_lora & other) const {
-        return memcmp(this, &other, sizeof(llama_hparams));
+        return memcmp(this, &other, sizeof(llama_hparams_lora)) != 0;
     }
 };
 

examples/common.cpp

@@ -331,6 +331,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
 #else
       fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--low-vram" || arg == "-lv") {
+#ifdef GGML_USE_CUBLAS
+            params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
@@ -479,6 +485,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+    fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
@@ -528,6 +535,7 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
     lparams.n_gpu_layers = params.n_gpu_layers;
     lparams.main_gpu     = params.main_gpu;
     memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
+    lparams.low_vram     = params.low_vram;
     lparams.seed         = params.seed;
     lparams.f16_kv       = params.memory_f16;
     lparams.use_mmap     = params.use_mmap;
@@ -632,6 +640,9 @@ void console_set_color(console_state & con_st, console_color_t color) {
             case CONSOLE_COLOR_USER_INPUT:
                 fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_GREEN);
                 break;
+            case CONSOLE_COLOR_ERROR:
+                fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_RED);
+                break;
         }
         con_st.color = color;
         fflush(con_st.out);

examples/common.h

@@ -21,15 +21,16 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    int32_t seed                           = -1;   // RNG seed
+    int32_t seed                            = -1;  // RNG seed
-    int32_t n_threads                      = get_num_physical_cores();
+    int32_t n_threads                       = get_num_physical_cores();
-    int32_t n_predict                      = -1;   // new tokens to predict
+    int32_t n_predict                       = -1;  // new tokens to predict
-    int32_t n_ctx                          = 512;  // context size
+    int32_t n_ctx                           = 512; // context size
-    int32_t n_batch                        = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_batch                         = 512; // batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep                         = 0;    // number of tokens to keep from initial prompt
+    int32_t n_keep                          = 0;   // number of tokens to keep from initial prompt
-    int32_t n_gpu_layers                   = 0;    // number of layers to store in VRAM
+    int32_t n_gpu_layers                    = 0;   // number of layers to store in VRAM
-    int32_t main_gpu                       = 0;    // the GPU that is used for scratch and small tensors
+    int32_t main_gpu                        = 0;   // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
+    bool    low_vram                        = 0;   // if true, reduce VRAM usage at the cost of performance
 
     // sampling parameters
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
@@ -112,7 +113,8 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params);
 enum console_color_t {
     CONSOLE_COLOR_DEFAULT=0,
     CONSOLE_COLOR_PROMPT,
-    CONSOLE_COLOR_USER_INPUT
+    CONSOLE_COLOR_USER_INPUT,
+    CONSOLE_COLOR_ERROR
 };
 
 struct console_state {

examples/main/README.md

@@ -288,5 +288,6 @@ These options provide extra functionality and customization when running the LLa
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/main/main.cpp

@@ -81,6 +81,9 @@ int main(int argc, char ** argv) {
     if (params.n_ctx > 2048) {
         fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);"
                 "expect poor results\n", __func__, params.n_ctx);
+    } else if (params.n_ctx < 8) {
+        fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
     }
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
@@ -328,9 +331,29 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> embd;
 
+    // do one empty run to warm up the model
+    {
+        const std::vector<llama_token> tmp = { llama_token_bos(), };
+        llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
+        llama_reset_timings(ctx);
+    }
+
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (embd.size() > 0) {
+            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            auto max_embd_size = n_ctx - 4;
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int)embd.size() > max_embd_size) {
+                auto skipped_tokens = embd.size() - max_embd_size;
+                console_set_color(con_st, CONSOLE_COLOR_ERROR);
+                printf("<<input too long: skipped %ld token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+                fflush(stdout);
+                embd.resize(max_embd_size);
+            }
+
             // infinite text generation via context swapping
             // if we run out of context:
             // - take the n_keep first tokens from the original prompt (via n_past)

examples/quantize/quantize.cpp

@@ -4,43 +4,135 @@
 
 #include <cstdio>
 #include <cstring>
-#include <map>
+#include <vector>
 #include <string>
 
-static const std::map<std::string, llama_ftype> LLAMA_FTYPE_MAP = {
+struct quant_option {
-  {"q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0},
+    std::string name;
-  {"q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1},
+    llama_ftype ftype;
-  {"q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0},
+    std::string desc;
-  {"q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1},
-  {"q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0},
-  {"q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K},
-  {"q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M},
-  {"q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S},
-  {"q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M},
-  {"q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L},
-  {"q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M},
-  {"q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S},
-  {"q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M},
-  {"q5_K",   LLAMA_FTYPE_MOSTLY_Q5_K_M},
-  {"q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S},
-  {"q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M},
-  {"q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K},
 };
 
-bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::string & ftype_str_out) {
+static const std::vector<struct quant_option> QUANT_OPTIONS = {
-    auto it = LLAMA_FTYPE_MAP.find(ftype_str);
+    {
-    if (it != LLAMA_FTYPE_MAP.end()) {
+        "Q4_0",
-        ftype = it->second;
+        LLAMA_FTYPE_MOSTLY_Q4_0,
-        ftype_str_out = it->first;
+        " 3.50G, +0.2499 ppl @ 7B - small, very high quality loss - legacy, prefer using Q3_K_M",
-        return true;
+    },
+    {
+        "Q4_1",
+        LLAMA_FTYPE_MOSTLY_Q4_1,
+        " 3.90G, +0.1846 ppl @ 7B - small, substantial quality loss - legacy, prefer using Q3_K_L",
+    },
+    {
+        "Q5_0",
+        LLAMA_FTYPE_MOSTLY_Q5_0,
+        " 4.30G, +0.0796 ppl @ 7B - medium, balanced quality - legacy, prefer using Q4_K_M",
+    },
+    {
+        "Q5_1",
+        LLAMA_FTYPE_MOSTLY_Q5_1,
+        " 4.70G, +0.0415 ppl @ 7B - medium, low quality loss - legacy, prefer using Q5_K_M",
+    },
+#ifdef GGML_USE_K_QUANTS
+    {
+        "Q2_K",
+        LLAMA_FTYPE_MOSTLY_Q2_K,
+        " 2.67G, +0.8698 ppl @ 7B - smallest, extreme quality loss - not recommended",
+    },
+    {
+        "Q3_K",
+        LLAMA_FTYPE_MOSTLY_Q3_K_M,
+        "alias for Q3_K_M"
+    },
+    {
+        "Q3_K_S",
+        LLAMA_FTYPE_MOSTLY_Q3_K_S,
+        " 2.75G, +0.5505 ppl @ 7B - very small, very high quality loss",
+    },
+    {
+        "Q3_K_M",
+        LLAMA_FTYPE_MOSTLY_Q3_K_M,
+        " 3.06G, +0.2437 ppl @ 7B - very small, very high quality loss",
+    },
+    {
+        "Q3_K_L",
+        LLAMA_FTYPE_MOSTLY_Q3_K_L,
+        " 3.35G, +0.1803 ppl @ 7B - small, substantial quality loss",
+    },
+    {
+        "Q4_K",
+        LLAMA_FTYPE_MOSTLY_Q4_K_M,
+        "alias for Q4_K_M",
+    },
+    {
+        "Q4_K_S",
+        LLAMA_FTYPE_MOSTLY_Q4_K_S,
+        " 3.56G, +0.1149 ppl @ 7B - small, significant quality loss",
+    },
+    {
+        "Q4_K_M",
+        LLAMA_FTYPE_MOSTLY_Q4_K_M,
+        " 3.80G, +0.0535 ppl @ 7B - medium, balanced quality - *recommended*",
+    },
+    {
+        "Q5_K",
+        LLAMA_FTYPE_MOSTLY_Q5_K_M,
+        "alias for Q5_K_M",
+    },
+    {
+        "Q5_K_S",
+        LLAMA_FTYPE_MOSTLY_Q5_K_S,
+        " 4.33G, +0.0353 ppl @ 7B - large, low quality loss - *recommended*",
+    },
+    {
+        "Q5_K_M",
+        LLAMA_FTYPE_MOSTLY_Q5_K_M,
+        " 4.45G, +0.0142 ppl @ 7B - large, very low quality loss - *recommended*",
+    },
+    {
+        "Q6_K",
+        LLAMA_FTYPE_MOSTLY_Q6_K,
+        " 5.15G, +0.0044 ppl @ 7B - very large, extremely low quality loss",
+    },
+#endif
+    {
+        "Q8_0",
+        LLAMA_FTYPE_MOSTLY_Q8_0,
+        " 6.70G, +0.0004 ppl @ 7B - very large, extremely low quality loss - not recommended",
+    },
+    {
+        "F16",
+        LLAMA_FTYPE_MOSTLY_F16,
+        "13.00G              @ 7B - extremely large, virtually no quality loss - not recommended",
+    },
+    {
+        "F32",
+        LLAMA_FTYPE_ALL_F32,
+        "26.00G              @ 7B - absolutely huge, lossless - not recommended",
+    },
+};
+
+
+bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
+    std::string ftype_str;
+
+    for (auto ch : ftype_str_in) {
+        ftype_str.push_back(std::toupper(ch));
+    }
+    for (auto & it : QUANT_OPTIONS) {
+        if (it.name == ftype_str) {
+            ftype = it.ftype;
+            ftype_str_out = it.name;
+            return true;
+        }
     }
-    // try to parse as an integer
     try {
         int ftype_int = std::stoi(ftype_str);
-        for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
+        for (auto & it : QUANT_OPTIONS) {
-            if (it->second == ftype_int) {
+            if (it.ftype == ftype_int) {
-                ftype = it->second;
+                ftype = it.ftype;
-                ftype_str_out = it->first;
+                ftype_str_out = it.name;
                 return true;
             }
         }
@@ -52,15 +144,15 @@ bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::st
 }
 
 // usage:
-//  ./quantize models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
+//  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
 //
 void usage(const char * executable) {
-    fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.bin [model-quant.bin] type [nthreads]\n", executable);
+    fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.bin [model-quant.bin] type [nthreads]\n\n", executable);
     fprintf(stderr, "  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     fprintf(stderr, "  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
-    fprintf(stderr, "Allowed quantization types:\n");
+    fprintf(stderr, "\nAllowed quantization types:\n");
-    for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
+    for (auto & it : QUANT_OPTIONS) {
-        fprintf(stderr, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
+        printf("  %2d  or  %-6s : %s\n", it.ftype, it.name.c_str(), it.desc.c_str());
     }
     exit(1);
 }

examples/server/README.md

@@ -289,6 +289,7 @@ Test();
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--embedding`: Enable the embedding mode. **Completion function doesn't work in this mode**.
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`;
 -   `--port`: Set the port to listen. Default: `8080`.

examples/server/server.cpp

@@ -405,6 +405,7 @@ void server_print_usage(int /*argc*/, char **argv, const gpt_params &params)
   fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
   fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
   fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+  fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
   fprintf(stderr, "  -m FNAME, --model FNAME\n");
   fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
@@ -537,6 +538,14 @@ bool server_params_parse(int argc, char **argv, server_params &sparams, gpt_para
       }
 #else
       fprintf(stderr, "WARNING: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+    }
+    else if (arg == "--low-vram" || arg == "-lv")
+    {
+#ifdef GGML_USE_CUBLAS
+      params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
     }
     else if (arg == "--main-gpu" || arg == "-mg")

examples/train-text-from-scratch/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET train-text-from-scratch)
+add_executable(${TARGET} train-text-from-scratch.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/train-text-from-scratch/README.md

@@ -0,0 +1,22 @@
+# train-text-from-scratch
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://github.com/brunoklein99/deep-learning-notes/blob/master/shakespeare.txt
+
+# train
+./bin/train-text-from-scratch \
+        --vocab-model ../models/ggml-vocab.bin \
+        --ctx 64 --embd 256 --head 8 --layer 16 \
+        --checkpoint-in  chk-shakespeare-256x16.bin \
+        --checkpoint-out chk-shakespeare-256x16.bin \
+        --model-out ggml-shakespeare-256x16-f32.bin \
+        --train-data "shakespeare.txt" \
+        -t 6 -b 16 -n 32 --seed 1 --adam-iter 16 \
+        --print-details-interval 0 --predict 16 --use-flash
+
+# predict
+./bin/main -m ggml-shakespeare-256x16-f32.bin
+```

ggml-cuda.h

@@ -24,11 +24,14 @@ void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
 void * ggml_cuda_host_malloc(size_t size);
 void   ggml_cuda_host_free(void * ptr);
 
-void   ggml_cuda_load_data(const char * fname, struct ggml_tensor * tensors, size_t offset);
+void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
+
 void   ggml_cuda_free_data(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 void   ggml_cuda_set_main_device(int main_device);
 void   ggml_cuda_set_scratch_size(size_t scratch_size);
+void   ggml_cuda_free_scratch(void);
 bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus

ggml-metal.m

@@ -52,14 +52,18 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_q4_0);
     GGML_METAL_DECL_KERNEL(get_rows_q4_1);
     GGML_METAL_DECL_KERNEL(get_rows_q2_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q3_k);
     GGML_METAL_DECL_KERNEL(get_rows_q4_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q5_k);
     GGML_METAL_DECL_KERNEL(get_rows_q6_k);
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q2_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q3_k_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q5_k_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q6_k_f32);
     GGML_METAL_DECL_KERNEL(rope);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
@@ -86,6 +90,7 @@ struct ggml_metal_context * ggml_metal_init(void) {
 
     ctx->device = MTLCreateSystemDefaultDevice();
     ctx->queue  = [ctx->device newCommandQueue];
+    ctx->n_buffers = 0;
 
     // determine if we can use MPS
     if (MPSSupportsMTLDevice(ctx->device)) {
@@ -152,14 +157,18 @@ struct ggml_metal_context * ggml_metal_init(void) {
         GGML_METAL_ADD_KERNEL(get_rows_q4_0);
         GGML_METAL_ADD_KERNEL(get_rows_q4_1);
         GGML_METAL_ADD_KERNEL(get_rows_q2_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q3_k);
         GGML_METAL_ADD_KERNEL(get_rows_q4_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q5_k);
         GGML_METAL_ADD_KERNEL(get_rows_q6_k);
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q2_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q3_k_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q5_k_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q6_k_f32);
         GGML_METAL_ADD_KERNEL(rope);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
@@ -574,6 +583,15 @@ void ggml_metal_graph_compute(
                                     nth1 = 16;
                                     [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_k_f32];
                                 } break;
+                            case GGML_TYPE_Q3_K:
+                                {
+                                    GGML_ASSERT(ne02 == 1);
+                                    GGML_ASSERT(ne12 == 1);
+
+                                    nth0 = 4;
+                                    nth1 = 16;
+                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_k_f32];
+                                } break;
                             case GGML_TYPE_Q4_K:
                                 {
                                     GGML_ASSERT(ne02 == 1);
@@ -583,6 +601,15 @@ void ggml_metal_graph_compute(
                                     nth1 = 16;
                                     [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
                                 } break;
+                            case GGML_TYPE_Q5_K:
+                                {
+                                    GGML_ASSERT(ne02 == 1);
+                                    GGML_ASSERT(ne12 == 1);
+
+                                    nth0 = 4;
+                                    nth1 = 16;
+                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_k_f32];
+                                } break;
                             case GGML_TYPE_Q6_K:
                                 {
                                     GGML_ASSERT(ne02 == 1);
@@ -619,15 +646,14 @@ void ggml_metal_graph_compute(
                         if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
                             [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                        } else if (src0t == GGML_TYPE_Q2_K) {
+                        }
+                        else if (src0t == GGML_TYPE_Q2_K ||
+                                 src0t == GGML_TYPE_Q3_K ||
+                                 src0t == GGML_TYPE_Q4_K ||
+                                 src0t == GGML_TYPE_Q5_K ||
+                                 src0t == GGML_TYPE_Q6_K) {
                             [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                        } else if (src0t == GGML_TYPE_Q4_K) {
-                            [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                        } else if (src0t == GGML_TYPE_Q6_K) {
-                            [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                         } else {
                             [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -645,7 +671,9 @@ void ggml_metal_graph_compute(
                         case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
                         case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
                         case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_k]; break;
+                        case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_k]; break;
                         case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
+                        case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_k]; break;
                         case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
                         default: GGML_ASSERT(false && "not implemented");
                     }

ggml-metal.metal

@@ -304,34 +304,22 @@ kernel void kernel_mul_mat_q4_0_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
     const int nb = ne00/QK4_0;
 
-    const int8_t m8 = 8;
-
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
 
     device const block_q4_0 * x = (device const block_q4_0 *) src0 + r0*nb;
     device const float      * y = (device const float      *) src1 + r1*ne10;
 
-    const uint nth = tptg.x*tptg.y;
+    const int nth = tptg.x*tptg.y;
-    const uint ith = tptg.y*tpitg.x + tpitg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
 
     const int ix = tpitg.y/4;           // 0 or 1
     const int iy = tpitg.y - 4*ix;      // 0...3
@@ -351,47 +339,32 @@ kernel void kernel_mul_mat_q4_0_f32(
 
         for (int j = 0; j < 4; ++j) {
 
-            acc[0] += yl[j+ 0] * ((int8_t)(xl[j] & 0xF) - m8);
+            acc[0] += yl[j] * (xl[j] & 0xF) + yl[j+16] * (xl[j] >> 4);
-            acc[1] += yl[j+16] * ((int8_t)(xl[j] >>  4) - m8);
+            acc[1] += yl[j] + yl[j+16];
 
         }
 
-        sumf += d * (acc[0] + acc[1]);
+        sumf += d * (acc[0] - 8.f*acc[1]);
     }
 
     sum[ith] = sumf;
 
     //
     // Accumulate the sum from all threads in the threadgroup
-    // This version is slightly faster than the commented out one below,
-    // which I copy-pasted from ggerganov's q4_0 dot product for metal.
     //
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith%4 == 0) {
-        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
+        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
     }
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith%16 == 0) {
-        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
+        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
     }
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith == 0) {
-        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+        for (uint i = 16; i < nth; i += 16) sum[0] += sum[i];
         dst[r1*ne0 + r0] = sum[0];
     }
-
-    //// accumulate the sum from all threads in the threadgroup
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
-    //for (uint i = nth/2; i > 0; i /= 2) {
-    //    if (ith < i) {
-    //        sum[ith] += sum[ith + i];
-    //    }
-    //    threadgroup_barrier(mem_flags::mem_threadgroup);
-    //}
-
-    //if (ith == 0) {
-    //    dst[r1*ne0 + r0] = sum[0];
-    //}
 }
 
 kernel void kernel_mul_mat_q4_1_f32(
@@ -399,20 +372,10 @@ kernel void kernel_mul_mat_q4_1_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
     const int nb = ne00/QK4_1;
@@ -460,11 +423,11 @@ kernel void kernel_mul_mat_q4_1_f32(
     //
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith%4 == 0) {
-        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
+        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
     }
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith%16 == 0) {
-        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
+        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
     }
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith == 0) {
@@ -671,6 +634,15 @@ typedef struct {
     half d;           // super-block scale for quantized scales
     half dmin;        // super-block scale for quantized mins
 } block_q2_k;
+// 84 bytes / block
+
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
+    half d;                    // super-block scale
+} block_q3_k;
+// 110 bytes / block
 
 typedef struct {
     half d;             // super-block scale for quantized scales
@@ -678,6 +650,16 @@ typedef struct {
     uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
     uint8_t qs[QK_K/2];        // 4--bit quants
 } block_q4_k;
+// 144 bytes / block
+
+typedef struct {
+    half d;                      // super-block scale for quantized scales
+    half dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_k;
+// 176 bytes / block
 
 typedef struct {
     uint8_t ql[QK_K/2];      // quants, lower 4 bits
@@ -685,16 +667,19 @@ typedef struct {
     int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
     half d;                  // super-block scale
 } block_q6_k;
+// 210 bytes / block
 
 static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
     uchar4 r;
     if (j < 4) {
-        r[0] = q[j+0] & 63; r[1] = q[j+4] & 63;
+        r[0] = q[j+0] & 63;
-        r[2] = q[j+1] & 63; r[3] = q[j+5] & 63;
+        r[2] = q[j+1] & 63;
+        r[1] = q[j+4] & 63;
+        r[3] = q[j+5] & 63;
     } else {
         r[0] = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
-        r[1] = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
         r[2] = (q[j+5] & 0xF) | ((q[j-3] >> 6) << 4);
+        r[1] = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
         r[3] = (q[j+5] >>  4) | ((q[j+1] >> 6) << 4);
     }
     return r;
@@ -735,10 +720,65 @@ static void dequantize_row_q2_k(device const block_q2_k * x, device float * y, i
     }
 }
 
+static void dequantize_row_q3_k(device const block_q3_k * x, device float * y, int k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
+
+    uint16_t aux[8];
+    thread const int8_t * scales = (thread const int8_t*)aux;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d_all = (float)(x[i].d);
+
+        device const uint8_t * q = x[i].qs;
+        device const uint8_t * h = x[i].hmask;
+        uint8_t m = 1;
+
+        device const uint16_t * a = (device const uint16_t *)x[i].scales;
+        aux[0] = (a[0] & kmask2) | (((a[4] >> 0) & kmask1) << 4);
+        aux[1] = (a[1] & kmask2) | (((a[5] >> 0) & kmask1) << 4);
+        aux[2] = (a[2] & kmask2) | (((a[4] >> 2) & kmask1) << 4);
+        aux[3] = (a[3] & kmask2) | (((a[5] >> 2) & kmask1) << 4);
+        aux[4] = ((a[0] >> 4) & kmask2) | (((a[4] >> 4) & kmask1) << 4);
+        aux[5] = ((a[1] >> 4) & kmask2) | (((a[5] >> 4) & kmask1) << 4);
+        aux[6] = ((a[2] >> 4) & kmask2) | (((a[4] >> 6) & kmask1) << 4);
+        aux[7] = ((a[3] >> 4) & kmask2) | (((a[5] >> 6) & kmask1) << 4);
+
+        int is = 0;
+        float dl;
+        for (int n = 0; n < QK_K; n += 128) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+
+                dl = d_all * (scales[is++] - 32);
+                for (int l = 0; l < 16; ++l) {
+                    *y++ = dl * ((int8_t)((q[l+ 0] >> shift) & 3) - ((h[l+ 0] & m) ? 0 : 4));
+                }
+
+                dl = d_all * (scales[is++] - 32);
+                for (int l = 0; l < 16; ++l) {
+                    *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3) - ((h[l+16] & m) ? 0 : 4));
+                }
+
+                shift += 2;
+                m <<= 1;
+            }
+            q += 32;
+        }
+
+    }
+
+}
+
 static void dequantize_row_q4_k(device const block_q4_k * x, device float * y, int k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
+
     for (int i = 0; i < nb; i++) {
 
         const float d = x[i].d;
@@ -760,6 +800,33 @@ static void dequantize_row_q4_k(device const block_q4_k * x, device float * y, i
     }
 }
 
+static void dequantize_row_q5_k(device const block_q5_k * x, device float * y, int k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+   for (int i = 0; i < nb; i++) {
+
+        const float d = (float)(x[i].d);
+        const float min = (float)(x[i].dmin);
+
+        device const uint8_t * ql = x[i].qs;
+        device const uint8_t * qh = x[i].qh;
+
+        int is = 0;
+        uint8_t u1 = 1, u2 = 2;
+        for (int j = 0; j < QK_K; j += 64) {
+            const uchar4 sc = get_scale_min_k4(is, x[i].scales);
+            const float d1 = d * sc[0]; const float m1 = min * sc[1];
+            const float d2 = d * sc[2]; const float m2 = min * sc[3];
+            for (int l = 0; l < 32; ++l) *y++ = d1 * ((ql[l] & 0xF) + (qh[l] & u1 ? 16 : 0)) - m1;
+            for (int l = 0; l < 32; ++l) *y++ = d2 * ((ql[l]  >> 4) + (qh[l] & u2 ? 16 : 0)) - m2;
+            ql += 32; is += 2;
+            u1 <<= 2; u2 <<= 2;
+        }
+    }
+
+}
+
 static void dequantize_row_q6_k(device const block_q6_k * x, device float * y, int k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
@@ -808,6 +875,22 @@ kernel void kernel_get_rows_q2_k(
                        (device float *) ((device char *)  dst + i*nb1), ne00);
 }
 
+kernel void kernel_get_rows_q3_k(
+        device const  void * src0,
+        device const   int * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb1,
+        uint tpig[[thread_position_in_grid]]) {
+    const int i = tpig;
+    const int r = ((device int32_t *) src1)[i];
+
+    dequantize_row_q3_k(
+            (device const block_q3_k *) ((device char *) src0 + r*nb01),
+                       (device float *) ((device char *)  dst + i*nb1), ne00);
+}
+
 kernel void kernel_get_rows_q4_k(
         device const  void * src0,
         device const   int * src1,
@@ -824,6 +907,22 @@ kernel void kernel_get_rows_q4_k(
                        (device float *) ((device char *)  dst + i*nb1), ne00);
 }
 
+kernel void kernel_get_rows_q5_k(
+        device const  void * src0,
+        device const   int * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb1,
+        uint tpig[[thread_position_in_grid]]) {
+    const int i = tpig;
+    const int r = ((device int32_t *) src1)[i];
+
+    dequantize_row_q5_k(
+            (device const block_q5_k *) ((device char *) src0 + r*nb01),
+                       (device float *) ((device char *)  dst + i*nb1), ne00);
+}
+
 kernel void kernel_get_rows_q6_k(
         device const  void * src0,
         device const   int * src1,
@@ -847,20 +946,10 @@ kernel void kernel_mul_mat_q2_k_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],               // we don't use this for now
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
 
@@ -875,7 +964,6 @@ kernel void kernel_mul_mat_q2_k_f32(
     const int nth = tptg.x*tptg.y;
     const int ith = tptg.y*tpitg.x + tpitg.y;
 
-
     const int tid = tpitg.y;    // 0...16
     const int il  = tid/4;      // 0...3
     const int ir  = tid%4;      // 0...3
@@ -885,35 +973,54 @@ kernel void kernel_mul_mat_q2_k_f32(
     const int n   = 8;
     const int is  = 4*il + (n*ir)/16;
 
+    const int y_offset = 64*il + n*ir;
+    const int q_offset = 32*ip + n*ir;
+
     sum[ith] = 0.0f;
 
     float sumf = 0;
     for (int i = tpitg.x; i < nb; i += tptg.x) {
 
-        device const uint8_t * q = x[i].qs + 32*ip + n*ir;
+        device const uint8_t * q = x[i].qs + q_offset;
         device const uint8_t * scales = x[i].scales + is;
 
         uint8_t d1 = scales[0] & 0xF;
-        uint8_t m1 = scales[0] >>  4;
         uint8_t d2 = scales[2] & 0xF;
+        uint8_t m1 = scales[0] >>  4;
         uint8_t m2 = scales[2] >>  4;
 
-        device const float   * y = yy + i*QK_K + 64*il + n*ir;
+        device const float   * y = yy + i*QK_K + y_offset;
+
+        //float4 s = {0.f, 0.f, 0.f, 0.f};
+        float2 s = {0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < n; ++l) {
+            s[0] += y[l+ 0] * ((q[l] >> shift1) & 3);
+            s[1] += y[l+32] * ((q[l] >> shift2) & 3);
+            smin += y[l+ 0] * m1 + y[l+32] * m2;
+        }
 
         const float dall = (float)x[i].d;
         const float dmin = (float)x[i].dmin;
 
-        float4 s = {0.f, 0.f, 0.f, 0.f};
+        sumf += dall * (s[0] * d1 + s[1] * d2) - dmin * smin;
-        for (int l = 0; l < n; ++l) {
-            s[0] += y[l+ 0] * ((q[l] >> shift1) & 3); s[1] += y[l+ 0];
-            s[2] += y[l+32] * ((q[l] >> shift2) & 3); s[3] += y[l+32];
-        }
-        sumf += dall * (s[0] * d1 + s[2] * d2) - dmin * (s[1] * m1 + s[3] * m2);
-
 
     }
     sum[ith] = sumf;
 
+    //int mask1 = (ith%4 == 0);
+    //int mask2 = (ith%16 == 0);
+
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    //for (int i = 1; i < 4; ++i) sum[ith] += mask1 * sum[ith + i];
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    //for (int i = 4; i < 16; i += 4) sum[ith] += mask2 * sum[ith + i];
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    //if (ith == 0) {
+    //    for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+    //    dst[r1*ne0 + r0] = sum[0];
+    //}
+
     //
     // Accumulate the sum from all threads in the threadgroup
     // This version is slightly faster than the commented out one below,
@@ -932,19 +1039,109 @@ kernel void kernel_mul_mat_q2_k_f32(
         for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
         dst[r1*ne0 + r0] = sum[0];
     }
+}
 
-    //// accumulate the sum from all threads in the threadgroup
+kernel void kernel_mul_mat_q3_k_f32(
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
+        device const  void * src0,
-    //for (uint i = nth/2; i > 0; i /= 2) {
+        device const float * src1,
-    //    if (ith < i) {
+        device       float * dst,
-    //        sum[ith] += sum[ith + i];
+        constant   int64_t & ne00,
-    //    }
+        constant   int64_t & ne10,
-    //    threadgroup_barrier(mem_flags::mem_threadgroup);
+        constant   int64_t & ne0,
-    //}
+        constant   int64_t & ne1,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2  tptg[[threads_per_threadgroup]]) {
+
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
+
+    const uint8_t m3 = 3;
+    const int8_t  m4 = 4;
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+
+    device const block_q3_k * x = (device const block_q3_k *) src0 + r0*nb;
+    device const float     * yy = (device const float      *) src1 + r1*ne10;
+
+    const int nth = tptg.x*tptg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
+
+    const int tid = tpitg.y;        // expecting 16
+    const int ip  = tid/8;          // 0 or 1
+    const int il  = tid/2 - 4*ip;   // 0...3
+    const int ir  = tid%2;
+    const int n   = 8;
+    const int l0  = n*ir;
+
+    const uint8_t m = 1 << (4*ip + il);
+
+    const int shift = 2*il;
+
+    const uint16_t s_shift1 = 4*ip;
+    const uint16_t s_shift2 = s_shift1 + 2*(il/2);
+    const int ik = 4 + (il%2);
+
+    const int q_offset = 32*ip + l0;
+    const int y_offset = 128*ip + 32*il + l0;
+
+    //float sumf = 0;
+    float sumf1 = 0, sumf2 = 0;
+    for (int i = tpitg.x; i < nb; i += tptg.x) {
+
+        const float d_all = (float)(x[i].d);
+
+        device const uint8_t * q = x[i].qs + q_offset;
+        device const uint8_t * h = x[i].hmask + l0;
+        device const float   * y = yy + i * QK_K + y_offset;
+
+        device const uint16_t * a = (device const uint16_t *)x[i].scales;
+        const char2 scales = as_type<char2>((uint16_t)(((a[il] >> s_shift1) & kmask2) | (((a[ik] >> s_shift2) & kmask1) << 4)));
+
+        float s = 0;
+        for (int l = 0; l < n; ++l) {
+            s += y[l+ 0] * ((int8_t)((q[l+ 0] >> shift) & m3) - ((h[l+ 0] & m) ? 0 : m4));
+        }
+        float d = d_all * s;
+        sumf1 += d * scales[0];
+        sumf2 += d;
+        //sumf += d_all * s * (scales[0] - 32);
+
+        s = 0;
+        for (int l = 0; l < n; ++l) {
+            s += y[l+16] * ((int8_t)((q[l+16] >> shift) & m3) - ((h[l+16] & m) ? 0 : m4));
+        }
+        d = d_all * s;
+        sumf1 += d * scales[1];
+        sumf2 += d;
+        //sumf += d_all * s * (scales[1] - 32);
+
+    }
+
+    //sum[ith] = sumf;
+    sum[ith] = sumf1 - 32.f*sumf2;
+
+    //
+    // Accumulate the sum from all threads in the threadgroup
+    //
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%4 == 0) {
+        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%16 == 0) {
+        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith == 0) {
+        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+        dst[r1*ne0 + r0] = sum[0];
+    }
 
-    //if (ith == 0) {
-    //    dst[r1*ne0 + r0] = sum[0];
-    //}
 }
 
 kernel void kernel_mul_mat_q4_k_f32(
@@ -952,23 +1149,17 @@ kernel void kernel_mul_mat_q4_k_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],               // we don't use this for now
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
 
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
     const int nb = ne00/QK_K;
 
     const int64_t r0 = tgpig.x;
@@ -977,37 +1168,55 @@ kernel void kernel_mul_mat_q4_k_f32(
     device const block_q4_k * x = (device const block_q4_k *) src0 + r0*nb;
     device const float     * yy = (device const float      *) src1 + r1*ne10;
 
-    const uint nth = tptg.x*tptg.y;
+    const int nth = tptg.x*tptg.y;
-    const uint ith = tptg.y*tpitg.x + tpitg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
 
     const int tid = tpitg.y;   // 0...16
     const int il  = tid/4;     // 0...3
-    const int ir  = tid%4;     // 0...3
+    const int ir  = tid - 4*il;// 0...3
-    const int n   = 8;
+    const int n   = 4;
-    const int is  = 2*il;
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
 
     sum[ith] = 0.0f;
 
+    uchar2 sc1, sc2, sc3, sc4;
+
     float sumf = 0;
     for (int i = tpitg.x; i < nb; i += tptg.x) {
 
-        device const uint8_t * q = (x + i)->qs + 32*il + n*ir;
+        device const uint8_t * q1 = (x + i)->qs + q_offset;
-        device const float   * y = yy + i*QK_K + 64*il + n*ir;
+        device const uint8_t * q2 = q1 + 64;
-        device const uint8_t * scales = (x + i)->scales;
+        device const float   * y1 = yy + i*QK_K + y_offset;
+        device const float   * y2 = y1 + 128;
 
         const float dall = (float)((x + i)->d);
         const float dmin = (float)((x + i)->dmin);
 
-        const uchar4 sc = get_scale_min_k4(is, scales);
+        device const uint16_t * a = (device const uint16_t *)(x + i)->scales;
+        sc1 = as_type<uchar2>((uint16_t)(a[im+0] & kmask1));
+        sc2 = as_type<uchar2>((uint16_t)(a[im+2] & kmask1));
+        sc3 = as_type<uchar2>((uint16_t)(((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2)));
+        sc4 = as_type<uchar2>((uint16_t)(((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2)));
 
         float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
         for (int l = 0; l < n; ++l) {
-            s[0] += y[l+ 0] * (q[l] & 0xF); s[1] += y[l+ 0];
+
-            s[2] += y[l+32] * (q[l] >>  4); s[3] += y[l+32];
+            s[0] += y1[l] * (q1[l] & 0xF); s[1] += y1[l+32] * (q1[l] >> 4);
+            s[2] += y2[l] * (q2[l] & 0xF); s[3] += y2[l+32] * (q2[l] >> 4);
+            smin += y1[l] * sc2[0] + y1[l+32] * sc2[1] + y2[l] * sc4[0] + y2[l+32] * sc4[1];
+
         }
-        sumf += dall * (s[0] * sc[0] + s[2] * sc[2]) - dmin * (s[1] * sc[1] + s[3] * sc[3]);
+        sumf += dall * (s[0] * sc1[0] + s[1] * sc1[1] + s[2] * sc3[0] + s[3] * sc3[1]) - dmin * smin;
 
     }
+
     sum[ith] = sumf;
 
     //
@@ -1043,25 +1252,114 @@ kernel void kernel_mul_mat_q4_k_f32(
     //}
 }
 
+kernel void kernel_mul_mat_q5_k_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne10,
+        constant   int64_t & ne0,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2  tptg[[threads_per_threadgroup]]) {
+
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+
+    device const block_q5_k * x = (device const block_q5_k *) src0 + r0*nb;
+    device const float     * yy = (device const float      *) src1 + r1*ne10;
+
+    const int nth = tptg.x*tptg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
+
+    const int tid = tpitg.y;   // 0...16
+    const int il  = tid/4;     // 0...3
+    const int ir  = tid - 4*il;// 0...3
+    const int n   = 4;
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    const uint8_t hm1 = 1u << (2*im);
+    const uint8_t hm2 = hm1 << 1;
+    const uint8_t hm3 = hm1 << 4;
+    const uint8_t hm4 = hm2 << 4;
+
+    uchar2 sc1, sc2, sc3, sc4;
+
+    float sumf = 0;
+    for (int i = tpitg.x; i < nb; i += tptg.x) {
+
+        device const uint8_t * q1 = (x + i)->qs + q_offset;
+        device const uint8_t * q2 = q1 + 64;
+        device const uint8_t * qh = (x + i)->qh + l0;
+        device const float   * y1 = yy + i*QK_K + y_offset;
+        device const float   * y2 = y1 + 128;
+
+        const float dall = (float)((x + i)->d);
+        const float dmin = (float)((x + i)->dmin);
+
+        device const uint16_t * a = (device const uint16_t *)(x + i)->scales;
+        sc1 = as_type<uchar2>((uint16_t)(a[im+0] & kmask1));
+        sc2 = as_type<uchar2>((uint16_t)(a[im+2] & kmask1));
+        sc3 = as_type<uchar2>((uint16_t)(((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2)));
+        sc4 = as_type<uchar2>((uint16_t)(((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2)));
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < n; ++l) {
+
+            s[0] += y1[l+ 0] * ((q1[l] & 0xF) + (qh[l] & hm1 ? 16 : 0));
+            s[1] += y1[l+32] * ((q1[l] >>  4) + (qh[l] & hm2 ? 16 : 0));
+            s[2] += y2[l+ 0] * ((q2[l] & 0xF) + (qh[l] & hm3 ? 16 : 0));
+            s[3] += y2[l+32] * ((q2[l] >>  4) + (qh[l] & hm4 ? 16 : 0));
+            smin += y1[l] * sc2[0] + y1[l+32] * sc2[1] + y2[l] * sc4[0] + y2[l+32] * sc4[1];
+
+        }
+        sumf += dall * (s[0] * sc1[0] + s[1] * sc1[1] + s[2] * sc3[0] + s[3] * sc3[1]) - dmin * smin;
+
+    }
+    sum[ith] = sumf;
+
+    //
+    // Accumulate the sum from all threads in the threadgroup
+    //
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%4 == 0) {
+        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%16 == 0) {
+        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith == 0) {
+        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+        dst[r1*ne0 + r0] = sum[0];
+    }
+
+}
+
 kernel void kernel_mul_mat_q6_k_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],               // we don't use this for now
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
 
@@ -1078,24 +1376,29 @@ kernel void kernel_mul_mat_q6_k_f32(
     device const block_q6_k * x = (device const block_q6_k *) src0 + r0*nb;
     device const float     * yy = (device const float      *) src1 + r1*ne10;
 
-    const uint nth = tptg.x*tptg.y;
+    const int nth = tptg.x*tptg.y;
-    const uint ith = tptg.y*tpitg.x + tpitg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
 
-    const int step = QK_K / tptg.y;     // we expect this to be 16
+    // Note: we absolutely assume that tptg.y = 16 and QK_K = 256!
-    const int iqs  = step * tpitg.y;    // 0...240 in steps of 16
+    const int iqs  = 16 * tpitg.y;
     const int ip   = iqs / 128;         // 0 or 1
     const int il   = (iqs - 128*ip)/16; // 0...7
     const int n    = 4;
-    const int is   = 8*ip + (n*il)/16;
+    const int l0   = n*il;
+    const int is   = 8*ip + l0/16;
+
+    const int y_offset = 128*ip + l0;
+    const int q_offset_l = 64*ip + l0;
+    const int q_offset_h = 32*ip + l0;
 
     float sumf = 0;
     for (int i = tpitg.x; i < nb; i += tptg.x) {
 
-        device const uint8_t * ql = x[i].ql + 64*ip + n*il;
+        device const uint8_t * ql = x[i].ql + q_offset_l;
-        device const uint8_t * qh = x[i].qh + 32*ip + n*il;
+        device const uint8_t * qh = x[i].qh + q_offset_h;
         device const int8_t  * sc = x[i].scales + is;
 
-        device const float * y = yy + i * QK_K + 128*ip + n*il;
+        device const float * y = yy + i * QK_K + y_offset;
 
         const float dall = x[i].d;
 

ggml-opencl.cpp

@@ -1167,7 +1167,7 @@ size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct g
     return 0;
 }
 
-void ggml_cl_transform_tensor(ggml_tensor * tensor) {
+void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     const int64_t ne0 = tensor->ne[0];
     const int64_t ne1 = tensor->ne[1];
     const int64_t ne2 = tensor->ne[2];
@@ -1179,6 +1179,7 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
     size_t q_size;
     cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
 
+    tensor->data = data;
     // copy tensor to device
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
@@ -1190,35 +1191,5 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
     CL_CHECK(clFinish(queue));
 
     tensor->data = dst;
-    tensor->backend = GGML_BACKEND_GPU;
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
-}
-
-void ggml_cl_load_data(const char * fname, struct ggml_tensor * tensor, const size_t offset) {
-    cl_int err;
-    FILE * fp = fopen(fname, "rb");
-
-    const size_t size = ggml_nbytes(tensor);
-
-    cl_mem dst;
-    CL_CHECK((dst = clCreateBuffer(context, CL_MEM_READ_ONLY, size, nullptr, &err), err));
-    void * buf_host = malloc(size);
-
-#ifdef _WIN32
-    int ret = _fseeki64(fp, (__int64) offset, SEEK_SET);
-#else
-    int ret = fseek(fp, (long) offset, SEEK_SET);
-#endif
-    GGML_ASSERT(ret == 0); // same
-
-    size_t ret2 = fread(buf_host, size, 1, fp);
-    if (ret2 != 1) {
-        fprintf(stderr, "unexpectedly reached end of file");
-        exit(1);
-    }
-
-    clEnqueueWriteBuffer(queue, dst, CL_TRUE, 0, size, buf_host, 0, nullptr, nullptr);
-
-    tensor->data = dst;
-    free(buf_host);
-    fclose(fp);
 }

ggml-opencl.h

@@ -18,8 +18,7 @@ void   ggml_cl_host_free(void * ptr);
 
 void ggml_cl_free_data(const struct ggml_tensor* tensor);
 
-void ggml_cl_transform_tensor(struct ggml_tensor * tensor);
+void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
-void ggml_cl_load_data(const char * fname, struct ggml_tensor * tensor, size_t offset);
 
 #ifdef  __cplusplus
 }

ggml.h

@@ -296,6 +296,7 @@ extern "C" {
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
         GGML_OP_REPEAT,
+        GGML_OP_REPEAT_BACK,
         GGML_OP_ABS,
         GGML_OP_SGN,
         GGML_OP_NEG,
@@ -309,6 +310,7 @@ extern "C" {
         GGML_OP_RMS_NORM_BACK,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
         GGML_OP_SET,
@@ -324,6 +326,7 @@ extern "C" {
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
         GGML_OP_SOFT_MAX,
+        GGML_OP_SOFT_MAX_BACK,
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
@@ -333,10 +336,14 @@ extern "C" {
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
+        GGML_OP_FLASH_ATTN_BACK,
 
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_CROSS_ENTROPY_LOSS,
+        GGML_OP_CROSS_ENTROPY_LOSS_BACK,
+
         GGML_OP_COUNT,
     };
 
@@ -478,6 +485,7 @@ extern "C" {
 
     GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
 
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
@@ -574,6 +582,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add1_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_acc(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -645,6 +658,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_repeat_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_abs(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -698,14 +716,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // A: m rows, n columns
+    // A: n columns, m rows
-    // B: p rows, n columns (i.e. we transpose it internally)
+    // B: n columns, p rows  (i.e. we transpose it internally)
     // result is m columns, p rows
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // A: m columns, n rows,
+    // B: p columns, n rows,
+    // result is m columns, p rows
+    GGML_API struct ggml_tensor * ggml_out_prod(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     //
     // operations on tensors without backpropagation
     //
@@ -916,6 +942,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_soft_max_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // rotary position embedding
     // if mode & 1 == 1, skip n_past elements
     // if mode & 2 == 1, GPT-NeoX style
@@ -982,6 +1019,14 @@ extern "C" {
             struct ggml_tensor  * v,
             bool                  masked);
 
+    GGML_API struct ggml_tensor * ggml_flash_attn_back(
+           struct ggml_context * ctx,
+           struct ggml_tensor  * q,
+           struct ggml_tensor  * k,
+           struct ggml_tensor  * v,
+           struct ggml_tensor  * d,
+           bool                  masked);
+
     GGML_API struct ggml_tensor * ggml_flash_ff(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1005,6 +1050,19 @@ extern "C" {
             struct ggml_tensor          * b,
                    ggml_binary_op_f32_t   fun);
 
+    // loss function
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b);
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+            struct ggml_tensor          * c);
+
     //
     // automatic differentiation
     //
@@ -1099,6 +1157,8 @@ extern "C" {
         struct {
             int n_iter;
 
+            float sched; // schedule multiplier (fixed, decay or warmup)
+            float decay; // weight decay for AdamW, use 0.0f to disable
             float alpha; // learning rate
             float beta1;
             float beta2;
@@ -1123,6 +1183,49 @@ extern "C" {
         } lbfgs;
     };
 
+    struct ggml_opt_context {
+        struct ggml_context * ctx;
+        struct ggml_opt_params params;
+
+        int iter;
+        int64_t nx; // number of parameter elements
+
+        bool just_initialized;
+
+        struct {
+            struct ggml_tensor * x;  // view of the parameters
+            struct ggml_tensor * g1; // gradient
+            struct ggml_tensor * g2; // gradient squared
+            struct ggml_tensor * m;  // first moment
+            struct ggml_tensor * v;  // second moment
+            struct ggml_tensor * mh; // first moment hat
+            struct ggml_tensor * vh; // second moment hat
+            struct ggml_tensor * pf; // past function values
+            float fx_best;
+            float fx_prev;
+            int n_no_improvement;
+        } adam;
+
+        struct {
+            struct ggml_tensor * x;    // current parameters
+            struct ggml_tensor * xp;   // previous parameters
+            struct ggml_tensor * g;    // current gradient
+            struct ggml_tensor * gp;   // previous gradient
+            struct ggml_tensor * d;    // search direction
+            struct ggml_tensor * pf;   // past function values
+            struct ggml_tensor * lmal; // the L-BFGS memory alpha
+            struct ggml_tensor * lmys; // the L-BFGS memory ys
+            struct ggml_tensor * lms;  // the L-BFGS memory s
+            struct ggml_tensor * lmy;  // the L-BFGS memory y
+            float fx_best;
+            float step;
+            int j;
+            int k;
+            int end;
+            int n_no_improvement;
+        } lbfgs;
+    };
+
     GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
 
     // optimize the function defined by the tensor f
@@ -1131,6 +1234,27 @@ extern "C" {
             struct ggml_opt_params params,
             struct ggml_tensor * f);
 
+    // initialize optimizer context
+    GGML_API void ggml_opt_init(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_opt_params params,
+            int64_t nx);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume_g(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f,
+            struct ggml_cgraph * gf,
+            struct ggml_cgraph * gb);
+
     //
     // quantization
     //

llama.cpp

@@ -165,6 +165,11 @@ struct llama_kv_cache {
         if (ctx) {
             ggml_free(ctx);
         }
+
+#ifdef GGML_USE_CUBLAS
+        ggml_cuda_free_data(k);
+        ggml_cuda_free_data(v);
+#endif // GGML_USE_CUBLAS
     }
 };
 
@@ -210,6 +215,7 @@ struct llama_model {
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cuda_free_data(tensors_by_name[i].second);
         }
+        ggml_cuda_free_scratch();
 #elif defined(GGML_USE_CLBLAST)
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cl_free_data(tensors_by_name[i].second);
@@ -707,6 +713,9 @@ struct llama_model_loader {
 
     struct ggml_tensor * get_tensor_for(llama_load_tensor & lt, ggml_backend backend) {
         struct ggml_tensor * tensor;
+        if (backend != GGML_BACKEND_CPU) {
+            ggml_set_no_alloc(ggml_ctx, true);
+        }
         if (lt.ne.size() == 2) {
             tensor = ggml_new_tensor_2d(ggml_ctx, lt.type, lt.ne.at(0), lt.ne.at(1));
         } else {
@@ -716,6 +725,9 @@ struct llama_model_loader {
         ggml_set_name(tensor, lt.name.c_str());
         LLAMA_ASSERT(lt.ggml_tensor == NULL); // if this fails, we called get_tensor twice on the same tensor
 
+        if (backend != GGML_BACKEND_CPU) {
+            ggml_set_no_alloc(ggml_ctx, use_mmap);
+        }
         tensor->backend = backend;
         lt.ggml_tensor = tensor;
         num_ggml_tensors_created++;
@@ -731,6 +743,7 @@ struct llama_model_loader {
     void load_all_data(llama_progress_callback progress_callback, void *  progress_callback_user_data, llama_mlock * lmlock) {
         size_t data_size = 0;
         size_t prefetch_size = 0;
+        size_t lock_size = 0;
         for (const llama_load_tensor & lt : tensors_map.tensors) {
             data_size += lt.size;
             if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
@@ -740,11 +753,6 @@ struct llama_model_loader {
 
         if (use_mmap) {
             mapping.reset(new llama_mmap(&file_loaders.at(0)->file, prefetch_size));
-            if (!lmlock) {
-                // Don't call the callback since the actual loading will be lazy
-                // and we can't measure it.
-                progress_callback = NULL;
-            }
             if (lmlock) {
                 lmlock->init(mapping->addr);
             }
@@ -752,20 +760,49 @@ struct llama_model_loader {
 
         size_t done_size = 0;
         for (llama_load_tensor & lt : tensors_map.tensors) {
-            if (lt.ggml_tensor->backend != GGML_BACKEND_CPU) {
-                continue;
-            }
             if (progress_callback) {
                 progress_callback((float) done_size / data_size, progress_callback_user_data);
             }
             LLAMA_ASSERT(lt.ggml_tensor); // unused tensors should have been caught by load_data already
             lt.data = (uint8_t *) lt.ggml_tensor->data;
-            load_data_for(lt);
+
-            lt.ggml_tensor->data = lt.data;
+            // allocate temp buffer if not using mmap
-            done_size += lt.size;
+            if (!use_mmap && lt.data == NULL) {
-            if (use_mmap && lmlock) {
+                GGML_ASSERT(lt.ggml_tensor->backend != GGML_BACKEND_CPU);
-                lmlock->grow_to(done_size);
+                lt.data = (uint8_t*)malloc(ggml_nbytes(lt.ggml_tensor));
             }
+
+            load_data_for(lt);
+
+            switch(lt.ggml_tensor->backend) {
+                case GGML_BACKEND_CPU:
+                    lt.ggml_tensor->data = lt.data;
+                    if (use_mmap && lmlock) {
+                        lock_size += lt.size;
+                        lmlock->grow_to(lock_size);
+                    }
+                    break;
+#if defined(GGML_USE_CUBLAS)
+                case GGML_BACKEND_GPU:
+                case GGML_BACKEND_GPU_SPLIT:
+                    ggml_cuda_transform_tensor(lt.data, lt.ggml_tensor);
+                    if (!use_mmap) {
+                        free(lt.data);
+                    }
+                    break;
+#elif defined(GGML_USE_CLBLAST)
+                case GGML_BACKEND_GPU:
+                    ggml_cl_transform_tensor(lt.data, lt.ggml_tensor);
+                    if (!use_mmap) {
+                        free(lt.data);
+                    }
+                    break;
+#endif
+                default:
+                    continue;
+            }
+
+            done_size += lt.size;
         }
     }
 
@@ -836,7 +873,8 @@ static bool kv_cache_init(
         const struct llama_hparams & hparams,
              struct llama_kv_cache & cache,
                          ggml_type   wtype,
-                               int   n_ctx) {
+                               int   n_ctx,
+                               int   n_gpu_layers) {
     const int n_embd  = hparams.n_embd;
     const int n_layer = hparams.n_layer;
 
@@ -862,6 +900,15 @@ static bool kv_cache_init(
     ggml_set_name(cache.k, "cache_k");
     ggml_set_name(cache.v, "cache_v");
 
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer + 1) {
+        ggml_cuda_assign_buffers_no_scratch(cache.v);
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        ggml_cuda_assign_buffers_no_scratch(cache.k);
+    }
+#endif // GGML_USE_CUBLAS
+
     return true;
 }
 
@@ -872,6 +919,7 @@ struct llama_context_params llama_context_default_params() {
         /*.gpu_layers                  =*/ 0,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ {0},
+        /*.low_vram                    =*/ false,
         /*.seed                        =*/ -1,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
@@ -980,6 +1028,7 @@ static void llama_model_load_internal(
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
+        bool low_vram,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1005,6 +1054,12 @@ static void llama_model_load_internal(
             case 40: model.type = e_model::MODEL_13B; break;
             case 60: model.type = e_model::MODEL_30B; break;
             case 80: model.type = e_model::MODEL_65B; break;
+            default:
+                {
+                    if (hparams.n_layer < 32) {
+                        model.type = e_model::MODEL_7B;
+                    }
+                } break;
         }
 
         hparams.n_ctx = n_ctx;
@@ -1100,18 +1155,34 @@ static void llama_model_load_internal(
         ml->ggml_ctx = ctx;
 
         model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab}, GGML_BACKEND_CPU);
-        model.norm           = ml->get_tensor("norm.weight",           {n_embd},          GGML_BACKEND_CPU);
 
         // "output" tensor
         {
+            ggml_backend backend_norm;
             ggml_backend backend_output;
             if (n_gpu_layers > int(n_layer)) { // NOLINT
+                // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#else
+                backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
                 backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
             } else {
+                backend_norm = GGML_BACKEND_CPU;
                 backend_output = GGML_BACKEND_CPU;
             }
 
+            model.norm   = ml->get_tensor("norm.weight",   {n_embd},          backend_norm);
             model.output = ml->get_tensor("output.weight", {n_embd, n_vocab}, backend_output);
+            if (backend_norm == GGML_BACKEND_GPU) {
+                vram_weights += ggml_nbytes(model.norm);
+            }
+            if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                vram_weights += ggml_nbytes(model.output);
+            }
         }
 
         const int i_gpu_start = n_layer - n_gpu_layers;
@@ -1141,7 +1212,7 @@ static void llama_model_load_internal(
             if (backend == GGML_BACKEND_GPU) {
                 vram_weights +=
                     ggml_nbytes(layer.attention_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk)             +
-                    ggml_nbytes(layer.wv)             + ggml_nbytes(layer.wo) + ggml_nbytes(layer.attention_norm) +
+                    ggml_nbytes(layer.wv)             + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) +
                     ggml_nbytes(layer.w1)             + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3);
             }
         }
@@ -1169,23 +1240,49 @@ static void llama_model_load_internal(
                 mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
 
         (void) vram_scratch;
+        (void) n_batch;
 #ifdef GGML_USE_CUBLAS
-        vram_scratch = n_batch * MB;
+        if (low_vram) {
-        ggml_cuda_set_scratch_size(vram_scratch);
+            fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
-        if (n_gpu_layers > 0) {
+            ggml_cuda_set_scratch_size(0); // disable scratch
-            fprintf(stderr, "%s: allocating batch_size x 1 MB = %ld MB VRAM for the scratch buffer\n",
+        } else {
-                    __func__, vram_scratch / MB);
+            vram_scratch = n_batch * MB;
+            ggml_cuda_set_scratch_size(vram_scratch);
+            if (n_gpu_layers > 0) {
+                fprintf(stderr, "%s: allocating batch_size x 1 MB = %ld MB VRAM for the scratch buffer\n",
+                        __func__, vram_scratch / MB);
+            }
         }
 #endif // GGML_USE_CUBLAS
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
-        fprintf(stderr, "%s: offloading %d layers to GPU\n", __func__, n_gpu);
+        fprintf(stderr, "%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
         if (n_gpu_layers > (int) hparams.n_layer) {
-            fprintf(stderr, "%s: offloading output layer to GPU\n", __func__);
+            fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__);
         }
+        size_t vram_kv_cache = 0;
+        if (n_gpu_layers > (int) hparams.n_layer + 1) {
+            if (low_vram) {
+                fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
+            } else {
+                fprintf(stderr, "%s: offloading v cache to GPU\n", __func__);
+                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+            }
+        }
+        if (n_gpu_layers > (int) hparams.n_layer + 2) {
+            if (low_vram) {
+                fprintf(stderr, "%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
+            } else {
+                fprintf(stderr, "%s: offloading k cache to GPU\n", __func__);
+                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+            }
+        }
+        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
+        fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n",
+                __func__, std::min(n_gpu_layers, max_offloadable_layers), hparams.n_layer + 3);
         fprintf(stderr, "%s: total VRAM used: %zu MB\n",
-                __func__, (vram_weights + vram_scratch + MB - 1) / MB); // round up
+                __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up
 #else
         (void) n_gpu_layers;
 #endif
@@ -1196,58 +1293,15 @@ static void llama_model_load_internal(
         model.tensors_by_name.emplace_back(lt.name, lt.ggml_tensor);
     }
 
-    ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &lctx.model.mlock_mmap : NULL);
+    (void) tensor_split;
-
 #if defined(GGML_USE_CUBLAS)
     {
         ggml_cuda_set_tensor_split(tensor_split);
-
-        size_t done_size = 0;
-        size_t data_size = 0;
-        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
-            data_size += lt.size;
-            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
-                done_size += lt.size;
-            }
-        }
-        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
-            ggml_backend backend = lt.ggml_tensor->backend;
-            if (backend != GGML_BACKEND_GPU && backend != GGML_BACKEND_GPU_SPLIT) {
-                continue;
-            }
-            if (progress_callback) {
-                progress_callback((float) done_size / data_size, progress_callback_user_data);
-            }
-            ggml_cuda_load_data(fname.c_str(), lt.ggml_tensor, lt.shards.at(0).file_off);
-            done_size += lt.size;
-        }
     }
-#elif defined(GGML_USE_CLBLAST)
-    {
-        size_t done_size = 0;
-        size_t data_size = 0;
-        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
-            data_size += lt.size;
-            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
-                done_size += lt.size;
-            }
-        }
-        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
-            if (lt.ggml_tensor->backend != GGML_BACKEND_GPU) {
-                continue;
-            }
-            if (progress_callback) {
-                progress_callback((float) done_size / data_size, progress_callback_user_data);
-            }
-            ggml_cl_load_data(fname.c_str(), lt.ggml_tensor, lt.shards.at(0).file_off);
-            done_size += lt.size;
-        }
-    }
-#else
-    (void) n_batch;
-    (void) tensor_split;
 #endif
 
+    ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &lctx.model.mlock_mmap : NULL);
+
     if (progress_callback) {
         progress_callback(1.0f, progress_callback_user_data);
     }
@@ -1267,6 +1321,7 @@ static bool llama_model_load(
         int n_gpu_layers,
         int main_gpu,
         float * tensor_split,
+        bool low_vram,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1274,7 +1329,7 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, lctx, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, memory_type,
+        llama_model_load_internal(fname, lctx, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, low_vram, memory_type,
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
@@ -1350,12 +1405,33 @@ static bool llama_eval_internal(
     const int i_gpu_start = n_layer - n_gpu_layers;
     (void) i_gpu_start;
 
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    //
+    // with the low VRAM option VRAM scratch is disabled in llama_load_model_internal
+    // in that case ggml_cuda_assign_buffers has no effect
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+        if (n_gpu_layers > n_layer) {
+            offload_func_nr = ggml_cuda_assign_buffers;
+        }
+        if (n_gpu_layers > n_layer + 1) {
+            offload_func_v  = ggml_cuda_assign_buffers;
+        }
+        if (n_gpu_layers > n_layer + 2) {
+            offload_func_kq = ggml_cuda_assign_buffers;
+        }
+#endif // GGML_USE_CUBLAS
+
     for (int il = 0; il < n_layer; ++il) {
         offload_func_t offload_func = llama_nop;
 
 #ifdef GGML_USE_CUBLAS
         if (il >= i_gpu_start) {
-            offload_func = ggml_cuda_assign_buffers; // sets the output backend to GPU
+            offload_func = ggml_cuda_assign_buffers;
         }
 #endif // GGML_USE_CUBLAS
 
@@ -1378,31 +1454,42 @@ static bool llama_eval_internal(
         // self-attention
         {
             // compute Q and K and RoPE them
-            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-            // offload_func(tmpq);
-            ggml_set_name(tmpq, "tmpq");
-
             struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-            // offload_func(tmpk);
+            offload_func_kq(tmpk);
             ggml_set_name(tmpk, "tmpk");
 
+            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            offload_func_kq(tmpq);
+            ggml_set_name(tmpq, "tmpq");
+
             struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            offload_func_kq(Kcur);
             ggml_set_name(Kcur, "Kcur");
 
             struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 
             // store key and value to memory
             {
                 // compute the transposed [N, n_embd] V matrix
-                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), n_embd, N));
+
+                struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                offload_func_v(tmpv);
+                ggml_set_name(tmpv, "tmpv");
+
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd, N));
+                offload_func_v(Vcur);
                 ggml_set_name(Vcur, "Vcur");
 
                 struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                offload_func_kq(k);
                 ggml_set_name(k, "k");
+
                 struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
                         (   n_ctx)*ggml_element_size(kv_self.v),
                         (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+                offload_func_v(v);
                 ggml_set_name(v, "v");
 
                 // important: storing RoPE-ed version of K in the KV cache!
@@ -1414,6 +1501,7 @@ static bool llama_eval_internal(
                 ggml_permute(ctx0,
                         Qcur,
                         0, 2, 1, 3);
+            offload_func_kq(Q);
             ggml_set_name(Q, "Q");
 
             struct ggml_tensor * K =
@@ -1422,10 +1510,12 @@ static bool llama_eval_internal(
                             ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
                             n_embd/n_head, n_head, n_past + N),
                         0, 2, 1, 3);
+            offload_func_kq(K);
             ggml_set_name(K, "K");
 
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
             ggml_set_name(KQ, "KQ");
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
@@ -1434,14 +1524,17 @@ static bool llama_eval_internal(
 
             // KQ_scaled shape [n_past + N, N, n_head, 1]
             struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
             ggml_set_name(KQ_scaled, "KQ_scaled");
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+            offload_func_kq(KQ_masked);
             ggml_set_name(KQ_masked, "KQ_masked");
 
             // KQ = soft_max(KQ_masked)
             struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+            offload_func_v(KQ_soft_max);
             ggml_set_name(KQ_soft_max, "KQ_soft_max");
 
             // split cached V into n_head heads
@@ -1451,10 +1544,12 @@ static bool llama_eval_internal(
                         n_ctx*ggml_element_size(kv_self.v),
                         n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head,
                         il*n_ctx*ggml_element_size(kv_self.v)*n_embd);
+            offload_func_v(V);
             ggml_set_name(V, "V");
 
 #if 1
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
             ggml_set_name(KQV, "KQV");
 #else
             // make V contiguous in memory to speed up the matmul, however we waste time on the copy
@@ -1466,12 +1561,14 @@ static bool llama_eval_internal(
 
             // KQV_merged = KQV.permute(0, 2, 1, 3)
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
             ggml_set_name(KQV_merged, "KQV_merged");
 
             // cur = KQV_merged.contiguous().view(n_embd, N)
             cur = ggml_cpy(ctx0,
                     KQV_merged,
                     ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+            offload_func_v(cur);
             ggml_set_name(cur, "KQV_merged_contiguous");
 
             // projection (no bias)
@@ -1483,7 +1580,6 @@ static bool llama_eval_internal(
         }
 
         lctx.use_buf(ctx0, 1);
-        //ggml_cuda_set_scratch(1);
 
         struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
         offload_func(inpFF);
@@ -1541,32 +1637,24 @@ static bool llama_eval_internal(
     }
 
     lctx.use_buf(ctx0, 0);
-    //ggml_cuda_set_scratch(0);
 
     // used at the end to optionally extract the embeddings
     struct ggml_tensor * embeddings = NULL;
 
-    offload_func_t offload_func = llama_nop;
-
-#ifdef GGML_USE_CUBLAS
-        if (n_gpu_layers > n_layer) {
-            offload_func = ggml_cuda_assign_buffers; // sets the output backend to GPU
-        }
-#endif // GGML_USE_CUBLAS
 
     // norm
     {
         cur = ggml_rms_norm(ctx0, inpL);
-        offload_func(cur);
+        offload_func_nr(cur);
         ggml_set_name(cur, "rms_norm_inpL");
 
         cur = ggml_rms_norm(ctx0, cur);
-        offload_func(cur);
+        offload_func_nr(cur);
         ggml_set_name(cur, "rms_norm_after");
 
         // cur = cur*norm(broadcasted)
         cur = ggml_mul(ctx0, cur, model.norm);
-        offload_func(cur);
+        offload_func_nr(cur);
         ggml_set_name(cur, "result_norm");
 
         embeddings = cur;
@@ -2174,6 +2262,10 @@ llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_tok
         return -log2f(candidate.p) > *mu;
     }));
 
+    if (candidates->size == 0) {
+        candidates->size = 1;
+    }
+
     // Normalize the probabilities of the remaining words
     llama_sample_softmax(ctx, candidates);
 
@@ -2311,7 +2403,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_0: quantized_type = GGML_TYPE_Q5_0; break;
         case LLAMA_FTYPE_MOSTLY_Q5_1: quantized_type = GGML_TYPE_Q5_1; break;
         case LLAMA_FTYPE_MOSTLY_Q8_0: quantized_type = GGML_TYPE_Q8_0; break;
+        case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break;
+        case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break;
 
+#ifdef GGML_USE_K_QUANTS
         // K-quants
         case LLAMA_FTYPE_MOSTLY_Q2_K:   quantized_type = GGML_TYPE_Q2_K; break;
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:
@@ -2322,6 +2417,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_K_S:
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
         case LLAMA_FTYPE_MOSTLY_Q6_K:   quantized_type = GGML_TYPE_Q6_K; break;
+#endif
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
 
@@ -2333,6 +2429,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                                                                             /*vocab_only*/ false));
     llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), params->ftype);
 
+#ifdef GGML_USE_K_QUANTS
     int n_attention_wv    = 0;
     int n_feed_forward_w2 = 0;
     for (auto& tensor : model_loader->tensors_map.tensors) {
@@ -2346,6 +2443,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
     int i_attention_wv = 0;
     int i_feed_forward_w2 = 0;
+#endif
 
     size_t total_size_org = 0;
     size_t total_size_new = 0;
@@ -2371,12 +2469,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
         // quantize only 2D tensors
         quantize &= (tensor.ne.size() == 2);
-
+        quantize &= params->quantize_output_tensor || tensor.name != "output.weight";
-        // uncomment this to keep the output layer in FP16
+        quantize &= quantized_type != tensor.type;
-        if (!params->quantize_output_tensor && tensor.name == "output.weight") {
-           quantize = false;
-        }
-        quantize = quantize && quantized_type != tensor.type;
 
         enum ggml_type new_type;
         void * new_data;
@@ -2390,31 +2484,28 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             printf("size = %8.3f MB\n", tensor.size/1024.0/1024.0);
         } else {
             new_type = quantized_type;
-            // TODO: temporary disabled until Metal / OpenCL support is available
+#ifdef GGML_USE_K_QUANTS
-            //       ref: https://github.com/ggerganov/llama.cpp/issues/1711
+            if (tensor.name == "output.weight") {
-            //if (tensor.name == "output.weight") {
+               new_type = GGML_TYPE_Q6_K;
-            //    new_type = GGML_TYPE_Q6_K;
+            } else if (tensor.name.find("attention.wv.weight") != std::string::npos) {
-            //}
-            if (tensor.name.find("attention.wv.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
                 else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                          (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8 ||
                          (i_attention_wv - n_attention_wv/8)%3 == 2)) new_type = GGML_TYPE_Q6_K;
                 ++i_attention_wv;
-            }
+            } else if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
-            if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
                 else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                          (i_feed_forward_w2 < n_feed_forward_w2/8 || i_feed_forward_w2 >= 7*n_feed_forward_w2/8 ||
                          (i_feed_forward_w2 - n_feed_forward_w2/8)%3 == 2)) new_type = GGML_TYPE_Q6_K;
                 ++i_feed_forward_w2;
-            }
+            } else if (tensor.name.find("attention.wo.weight") != std::string::npos) {
-            if (tensor.name.find("attention.wo.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
             }
+#endif
 
             float * f32_data;
             size_t nelements = tensor.ne.at(0) * tensor.ne.at(1);
@@ -2554,8 +2645,8 @@ struct llama_context * llama_init_from_file(
 
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
-    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_batch, params.n_gpu_layers,
+    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_batch, params.n_gpu_layers, params.main_gpu,
-                params.main_gpu, params.tensor_split, memory_type, params.use_mmap, params.use_mlock,
+                params.tensor_split, params.low_vram, memory_type, params.use_mmap, params.use_mlock,
                 params.vocab_only, params.progress_callback, params.progress_callback_user_data)) {
         fprintf(stderr, "%s: failed to load model\n", __func__);
         llama_free(ctx);
@@ -2564,7 +2655,7 @@ struct llama_context * llama_init_from_file(
 
     // reserve memory for context buffers
     if (!params.vocab_only) {
-        if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx)) {
+        if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) {
             fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
             return nullptr;
@@ -3301,6 +3392,19 @@ int llama_n_embd(const struct llama_context * ctx) {
     return ctx->model.hparams.n_embd;
 }
 
+int llama_get_vocab(
+        const struct llama_context * ctx,
+        const char * * strings,
+        float  * scores,
+        int capacity) {
+    int n = std::min(capacity, (int) ctx->vocab.id_to_token.size());
+    for (int i = 0; i<n; ++i) {
+        strings[i] = ctx->vocab.id_to_token[i].tok.c_str();
+        scores[i]  = ctx->vocab.id_to_token[i].score;
+    }
+    return n;
+}
+
 float * llama_get_logits(struct llama_context * ctx) {
     return ctx->logits.data();
 }

llama.h

@@ -77,6 +77,7 @@ extern "C" {
         int n_gpu_layers;                      // number of layers to store in VRAM
         int main_gpu;                          // the GPU that is used for scratch and small tensors
         float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+        bool low_vram;                         // if true, reduce VRAM usage at the cost of performance
         int seed;                              // RNG seed, -1 for random
 
         bool f16_kv;     // use fp16 for KV cache
@@ -220,6 +221,14 @@ extern "C" {
     LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
     LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
+    // Get the vocabulary as output parameters.
+    // Returns number of results.
+    LLAMA_API int llama_get_vocab(
+            const struct llama_context * ctx,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
     // Can be mutated in order to change the probabilities of the next token

tests/test-grad0.c

@@ -5,7 +5,7 @@
 #include <stdlib.h>
 #include <assert.h>
 
-#define MAX_NARGS 2
+#define MAX_NARGS 3
 
 #undef MIN
 #undef MAX
@@ -1090,6 +1090,25 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // cross_entropy_loss
+        {
+            const int nargs = 1;
+
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+
+            for (int ndims = 1; ndims <= 3; ++ndims) {
+                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor(ctx0, ndims, ne2, 0.0f, 1.0f);
+                ggml_set_param(ctx0, x[0]);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
+
+                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-1f, 1e-2f, INFINITY);
+                // finite differences regularly fails!
+            }
+        }
+
         // rope
         {
             const int nargs = 1;
@@ -1124,6 +1143,45 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // flash_attn
+        {
+            const int nargs = 3;
+
+            int64_t ne2[4];
+
+            get_random_dims(ne2, 4);
+            int64_t D = ne2[0];
+            int64_t N = ne2[1];
+            int64_t M = ne2[2] + N;
+            int64_t B = ne2[3];
+
+            for (int masked = 0; masked <= 1; ++masked) {
+                for (int ndims = 2; ndims <= 4; ++ndims) {
+                    int64_t neq[4] = { D, N, B, ne[3] };
+                    int64_t nek[4] = { D, M, B, ne[3] };
+                    int64_t nev[4] = { M, D, B, ne[3] };
+                    if (ndims == 2) {
+                        neq[2] = 1; neq[3] = 1;
+                        nek[2] = 1; nek[3] = 1;
+                        nev[2] = 1; nev[3] = 1;
+                    } else if (ndims == 3) {
+                        neq[3] = 1;
+                        nek[3] = 1;
+                        nev[3] = 1;
+                    }
+                    x[0] = get_random_tensor(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    ggml_set_param(ctx0, x[0]);
+                    ggml_set_param(ctx0, x[1]);
+                    ggml_set_param(ctx0, x[2]);
+
+                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                    check_gradient("flash_attn", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                }
+            }
+        }
         ggml_free(ctx0);
     }