lookup : minor

* Add API key authentication for enhanced server-client security

* server : to snake_case

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.gitignore

@@ -48,6 +48,7 @@ models-mnt
 /llama-bench
 /llava-cli
 /lookahead
+/lookup
 /main
 /metal
 /perplexity

Makefile

@@ -2,7 +2,7 @@
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -639,6 +639,9 @@ parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 ifdef LLAMA_METAL
 metal: examples/metal/metal.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)

common/common.h

@@ -240,3 +240,4 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
 void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+

common/train.cpp

@@ -71,7 +71,7 @@ void free_random_uniform_distribution(struct random_uniform_distribution * rnd)
 
 struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
     float scale = 1.0f; // xavier
-    switch (tensor->n_dims) {
+    switch (ggml_n_dims(tensor)) {
         case 1:
             scale /= sqrtf((float) tensor->ne[0]);
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
@@ -119,7 +119,7 @@ struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct
 }
 
 struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
-    switch (tensor->n_dims) {
+    switch (ggml_n_dims(tensor)) {
         case 1:
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
                 float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
@@ -183,25 +183,27 @@ float fclamp(const float v, const float min, const float max) {
 }
 
 void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
-    GGML_ASSERT(tensor->n_dims == 1);
     GGML_ASSERT(tensor->ne[0] == ne0);
+    GGML_ASSERT(tensor->ne[1] == 1);
+    GGML_ASSERT(tensor->ne[2] == 1);
+    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
-    GGML_ASSERT(tensor->n_dims == 2);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
+    GGML_ASSERT(tensor->ne[2] == 1);
+    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
-    GGML_ASSERT(tensor->n_dims == 3);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
+    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
-    GGML_ASSERT(tensor->n_dims == 4);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
@@ -225,8 +227,8 @@ int64_t get_example_targets_batch(
     bool                   sample_random_offsets
 ) {
     GGML_ASSERT(samples_count > 0);
-    GGML_ASSERT(tokens_input->n_dims  == 2);
-    GGML_ASSERT(target_probs->n_dims  == 3);
+    GGML_ASSERT(ggml_is_matrix(tokens_input));
+    GGML_ASSERT(ggml_is_3d(target_probs));
     int64_t n_vocab  = target_probs->ne[0];
     int64_t n_tokens = tokens_input->ne[0];
     int64_t n_batch  = tokens_input->ne[1];

examples/CMakeLists.txt

@@ -33,6 +33,7 @@ else()
     add_subdirectory(simple)
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
+    add_subdirectory(lookup)
     add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)

examples/baby-llama/baby-llama.cpp

@@ -1258,9 +1258,9 @@ static struct ggml_tensor * forward_lora(
 }
 
 static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
-    assert(logits->n_dims == 2);
-    assert(probs->n_dims == 2);
-    assert(best_samples->n_dims == 1);
+    assert(ggml_is_matrix(logits));
+    assert(ggml_is_matrix(probs));
+    assert(ggml_is_vector(best_samples));
     assert(logits->ne[1] == best_samples->ne[0]);
     assert(logits->ne[0] == probs->ne[0]);
     assert(logits->ne[1] == probs->ne[1]);
@@ -1292,9 +1292,9 @@ static void sample_softmax_batch(
     struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
     struct ggml_tensor * best_samples
 ) {
-    GGML_ASSERT(best_samples->n_dims == 2);
-    GGML_ASSERT(logits->n_dims == 3);
-    GGML_ASSERT(probs->n_dims == 3);
+    GGML_ASSERT(ggml_is_matrix(best_samples));
+    GGML_ASSERT(ggml_is_3d(logits));
+    GGML_ASSERT(ggml_is_3d(probs));
     int n_tokens = best_samples->ne[0];
     int n_batch  = best_samples->ne[1];
     int n_vocab  = logits->ne[0];
@@ -1334,7 +1334,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
+    assert(ggml_is_matrix(probs));
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
@@ -1386,8 +1386,8 @@ static void get_example_targets(int example_id, struct ggml_tensor * tokens_inpu
 static void get_example_targets_batch(
     struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
 ) {
-    GGML_ASSERT(tokens_input->n_dims == 2);
-    GGML_ASSERT(     targets->n_dims == 3);
+    GGML_ASSERT(ggml_is_matrix(tokens_input));
+    GGML_ASSERT(ggml_is_3d(targets));
     int n_tokens = tokens_input->ne[0];
     int n_batch  = tokens_input->ne[1];
     GGML_ASSERT(n_tokens == targets->ne[1]);

examples/benchmark/benchmark-matmult.cpp

@@ -129,13 +129,13 @@ int main(int argc, char ** argv)  {
     const ggml_type qtype = GGML_TYPE_Q4_1;
 
     size_t ctx_size = 0;
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
-    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
+    ctx_size += ggml_row_size(qtype,         sizex*sizey);
+    ctx_size += ggml_row_size(qtype,         sizex*sizey);
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
+    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
     ctx_size += 1024*1024*16;
 
     printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -427,7 +427,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
+    assert(ggml_is_matrix(probs));
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
@@ -639,7 +639,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
     int ct;
-    switch (gg_weights->n_dims){
+    switch (ggml_n_dims(gg_weights)) {
         case 1:
             ct = 0;
             for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){

examples/finetune/finetune.cpp

@@ -1110,7 +1110,7 @@ static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor,
         name = ggml_get_name(tensor);
     }
     uint32_t name_len = strlen(name);
-    uint32_t nd = tensor->n_dims;
+    uint32_t nd = ggml_n_dims(tensor);
     uint32_t ne[4] = { (uint32_t)tensor->ne[0],
                        (uint32_t)tensor->ne[1],
                        (uint32_t)tensor->ne[2],

examples/gguf/gguf.cpp

@@ -195,7 +195,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
 
             struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
 
-            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
 
             // print first 10 elements
             const float * data = (const float *) cur->data;

examples/llava/clip.cpp

@@ -514,7 +514,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             ctx_size += padded_size;
             if (verbosity >= 3) {
                 printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, padded_size=%zu, offset=%zu\n", __func__, i,
-                       cur->n_dims, cur->name, tensor_size, padded_size, offset);
+                       ggml_n_dims(cur), cur->name, tensor_size, padded_size, offset);
             }
         }
     }
@@ -962,7 +962,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
         }
 
         // quantize only 2D tensors
-        quantize &= (cur->n_dims == 2);
+        quantize &= (ggml_n_dims(cur) == 2);
 
         if (quantize) {
             new_type = type;
@@ -1035,7 +1035,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             fout.put(0);
         }
 
-        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), cur->n_dims, quantize,
+        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
                orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
     }
 

examples/lookup/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET lookup)
+add_executable(${TARGET} lookup.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/lookup/README.md

@@ -0,0 +1,13 @@
+# llama.cpp/examples/lookup
+
+Demonstration of Prompt Lookup Decoding
+
+https://github.com/apoorvumang/prompt-lookup-decoding
+
+The two key parameters for lookup decoding are `max_ngram_size` and `n_draft`. The first, determines how many ngrams to use when searching through the prompt for a match and the second specifies how many subsequent tokens to draft if a match is found.
+
+More info:
+
+https://github.com/ggerganov/llama.cpp/pull/4484
+https://github.com/ggerganov/llama.cpp/issues/4226
+

examples/lookup/lookup.cpp

@@ -0,0 +1,229 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    // maximum n-grams to search for in prompt
+    const int max_ngram_size = 3;
+
+    // length of the candidate / draft sequence, if match is found
+    const int n_draft = 10;
+
+    const bool dump_kv_cache = params.dump_kv_cache;
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("lookup", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+
+    // tokenize the prompt
+    const bool add_bos = llama_should_add_bos_token(model);
+    LOG("add_bos tgt: %d\n", add_bos);
+
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+
+    const int max_context_size     = llama_n_ctx(ctx);
+    const int max_tokens_list_size = max_context_size - 4;
+
+    if ((int) inp.size() > max_tokens_list_size) {
+        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
+        return 1;
+    }
+
+    fprintf(stderr, "\n\n");
+
+    for (auto id : inp) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+    }
+
+    fflush(stderr);
+
+    const int n_input = inp.size();
+
+    const auto t_enc_start = ggml_time_us();
+
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+
+    const auto t_enc_end = ggml_time_us();
+
+    int n_predict = 0;
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    int n_past = inp.size();
+
+    bool has_eos = false;
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+
+    std::vector<llama_token> draft;
+
+    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
+
+    // debug
+    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
+
+    const auto t_dec_start = ggml_time_us();
+
+    while (true) {
+        // debug
+        if (dump_kv_cache) {
+            llama_kv_cache_view_update(ctx, &kvc_view);
+            dump_kv_cache_view_seqs(kvc_view, 40);
+        }
+
+        // print current draft sequence
+        LOG("drafted %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, draft).c_str());
+
+        int i_dft = 0;
+        while (true) {
+            // sample from the target model
+            llama_token id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_dft);
+
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
+
+            const std::string token_str = llama_token_to_piece(ctx, id);
+
+            if (!params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+
+            if (id == llama_token_eos(model)) {
+                has_eos = true;
+            }
+
+            ++n_predict;
+
+            // check if the target token matches the draft
+            if (i_dft < (int) draft.size() && id == draft[i_dft]) {
+                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
+                ++n_accept;
+                ++n_past;
+                ++i_dft;
+                inp.push_back(id);
+
+                if (params.use_color) {
+                    // color accepted draft token
+                    printf("\033[34m%s\033[0m", token_str.c_str());
+                    fflush(stdout);
+                }
+                continue;
+            }
+
+            if (params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+            fflush(stdout);
+
+
+            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
+
+            draft.clear();
+            draft.push_back(id);
+            inp.push_back(id);
+            break;
+        }
+
+        if ((params.n_predict > 0 && n_predict > params.n_predict) || has_eos) {
+            break;
+        }
+
+        // KV cache management
+        // clean the cache of draft tokens that weren't accepted
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
+        llama_batch_clear(batch_tgt);
+        llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
+
+        // generate n_pred tokens through prompt lookup
+        auto prompt_lookup = [&]() -> void {
+            int inp_size = inp.size();
+            for (int ngram_size = max_ngram_size ; ngram_size > 0; --ngram_size){
+                const llama_token * ngram = &inp[inp_size - ngram_size];
+
+                for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) {
+                    bool match = true;
+                    for (int j = 0; j < ngram_size; ++j) {
+                        if (inp[i + j] != ngram[j]) {
+                            match = false;
+                            break;
+                        }
+                    }
+
+                    if (match) {
+                        const int startIdx = i + ngram_size;
+                        const int endIdx = startIdx + n_draft;
+                        if (endIdx < inp_size) {
+                            for (int j = startIdx; j < endIdx; ++j) {
+                                LOG(" - draft candidate %d: %d\n", j, inp[j]);
+                                draft.push_back(inp[j]);
+                                llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true);
+                                ++n_drafted;
+                            }
+                            return;
+                        }
+                    }
+                }
+            }
+            return;
+        };
+
+        prompt_lookup();
+
+        llama_decode(ctx, batch_tgt);
+        ++n_past;
+
+        draft.erase(draft.begin());
+    }
+
+    auto t_dec_end = ggml_time_us();
+
+    LOG_TEE("\n\n");
+
+    LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
+    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
+
+    LOG_TEE("\n");
+    LOG_TEE("n_draft   = %d\n", n_draft);
+    LOG_TEE("n_predict = %d\n", n_predict);
+    LOG_TEE("n_drafted = %d\n", n_drafted);
+    LOG_TEE("n_accept  = %d\n", n_accept);
+    LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    LOG_TEE("\ntarget:\n");
+    llama_print_timings(ctx);
+
+    llama_sampling_free(ctx_sampling);
+    llama_batch_free(batch_tgt);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/server/public/completion.js

@@ -34,7 +34,8 @@ export async function* llama(prompt, params = {}, config = {}) {
     headers: {
       'Connection': 'keep-alive',
       'Content-Type': 'application/json',
-      'Accept': 'text/event-stream'
+      'Accept': 'text/event-stream',
+      ...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
     },
     signal: controller.signal,
   });

examples/server/public/index.html

@@ -235,7 +235,8 @@
       grammar: '',
       n_probs: 0, // no completion_probabilities,
       image_data: [],
-      cache_prompt: true
+      cache_prompt: true,
+      api_key: ''
     })
 
     /* START: Support for storing prompt templates and parameters in browsers LocalStorage */
@@ -790,6 +791,10 @@
             <fieldset>
               ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
             </fieldset>
+            <fieldset>
+              <label for="api_key">API Key</label>
+              <input type="text" name="api_key" value="${params.value.api_key}" placeholder="Enter API key" oninput=${updateParams} />
+            </fieldset>
           </details>
         </form>
       `

examples/server/server.cpp

@@ -36,6 +36,7 @@ using json = nlohmann::json;
 struct server_params
 {
     std::string hostname = "127.0.0.1";
+    std::string api_key;
     std::string public_path = "examples/server/public";
     int32_t port = 8080;
     int32_t read_timeout = 600;
@@ -1953,6 +1954,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
     printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
     printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
+    printf("  --api-key API_KEY     optional api key to enhance server security. If set, requests must include this key for access.\n");
     printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
     printf("  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N   number of slots for process requests (default: %d)\n", params.n_parallel);
@@ -2002,6 +2004,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             sparams.public_path = argv[i];
         }
+        else if (arg == "--api-key")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            sparams.api_key = argv[i];
+        }
         else if (arg == "--timeout" || arg == "-to")
         {
             if (++i >= argc)
@@ -2669,6 +2680,32 @@ int main(int argc, char **argv)
 
     httplib::Server svr;
 
+    // Middleware for API key validation
+    auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
+        // If API key is not set, skip validation
+        if (sparams.api_key.empty()) {
+            return true;
+        }
+
+        // Check for API key in the header
+        auto auth_header = req.get_header_value("Authorization");
+        std::string prefix = "Bearer ";
+        if (auth_header.substr(0, prefix.size()) == prefix) {
+            std::string received_api_key = auth_header.substr(prefix.size());
+            if (received_api_key == sparams.api_key) {
+                return true; // API key is valid
+            }
+        }
+
+        // API key is invalid or not provided
+        res.set_content("Unauthorized: Invalid API Key", "text/plain");
+        res.status = 401; // Unauthorized
+
+        LOG_WARNING("Unauthorized: Invalid API Key", {});
+
+        return false;
+    };
+
     svr.set_default_headers({{"Server", "llama.cpp"},
                              {"Access-Control-Allow-Origin", "*"},
                              {"Access-Control-Allow-Headers", "content-type"}});
@@ -2711,8 +2748,11 @@ int main(int argc, char **argv)
                 res.set_content(data.dump(), "application/json");
             });
 
-    svr.Post("/completion", [&llama](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/completion", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                if (!validate_api_key(req, res)) {
+                    return;
+                }
                 json data = json::parse(req.body);
                 const int task_id = llama.request_completion(data, false, false, -1);
                 if (!json_value(data, "stream", false)) {
@@ -2799,8 +2839,11 @@ int main(int argc, char **argv)
             });
 
     // TODO: add mount point without "/v1" prefix -- how?
-    svr.Post("/v1/chat/completions", [&llama](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/v1/chat/completions", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                if (!validate_api_key(req, res)) {
+                    return;
+                }
                 json data = oaicompat_completion_params_parse(json::parse(req.body));
 
                 const int task_id = llama.request_completion(data, false, false, -1);
@@ -2869,8 +2912,11 @@ int main(int argc, char **argv)
                 }
             });
 
-    svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/infill", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                if (!validate_api_key(req, res)) {
+                    return;
+                }
                 json data = json::parse(req.body);
                 const int task_id = llama.request_completion(data, true, false, -1);
                 if (!json_value(data, "stream", false)) {
@@ -3005,11 +3051,15 @@ int main(int argc, char **argv)
 
     svr.set_error_handler([](const httplib::Request &, httplib::Response &res)
             {
+                if (res.status == 401)
+                {
+                    res.set_content("Unauthorized", "text/plain");
+                }
                 if (res.status == 400)
                 {
                     res.set_content("Invalid request", "text/plain");
                 }
-                else if (res.status != 500)
+                else if (res.status == 404)
                 {
                     res.set_content("File Not Found", "text/plain");
                     res.status = 404;
@@ -3032,11 +3082,15 @@ int main(int argc, char **argv)
     // to make it ctrl+clickable:
     LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
 
-    LOG_INFO("HTTP server listening", {
-                                          {"hostname", sparams.hostname},
-                                          {"port", sparams.port},
-                                      });
+    std::unordered_map<std::string, std::string> log_data;
+    log_data["hostname"] = sparams.hostname;
+    log_data["port"] = std::to_string(sparams.port);
 
+    if (!sparams.api_key.empty()) {
+        log_data["api_key"] = "api_key: ****" + sparams.api_key.substr(sparams.api_key.length() - 4);
+    }
+
+    LOG_INFO("HTTP server listening", log_data);
     // run the HTTP server in a thread - see comment below
     std::thread t([&]()
             {

ggml-cuda.cu

@@ -8898,6 +8898,12 @@ static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, gg
     (void) dst;
 }
 
+static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
+    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+
+    return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
+}
+
 void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
     const int64_t nrows = ggml_nrows(tensor);
 
@@ -8947,8 +8953,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
 
         // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
         if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
-                * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
 
         char * buf;
@@ -9485,8 +9490,7 @@ static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_t
 
     if (ggml_is_quantized(tensor->type)) {
         if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
-                * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
     }
 

ggml.c

@@ -1997,12 +1997,6 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
     return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN);
 }
 
-size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
-
-    return (nrows_split*tensor->ne[0]*ggml_type_size(tensor->type))/ggml_blck_size(tensor->type);
-}
-
 int ggml_blck_size(enum ggml_type type) {
     return type_traits[type].blck_size;
 }
@@ -2011,8 +2005,13 @@ size_t ggml_type_size(enum ggml_type type) {
     return type_traits[type].type_size;
 }
 
-float ggml_type_sizef(enum ggml_type type) {
-    return ((float)(type_traits[type].type_size))/type_traits[type].blck_size;
+size_t ggml_row_size(enum ggml_type type, int64_t ne) {
+    assert(ne % ggml_blck_size(type) == 0);
+    return ggml_type_size(type)*ne/ggml_blck_size(type);
+}
+
+double ggml_type_sizef(enum ggml_type type) {
+    return ((double)(type_traits[type].type_size))/type_traits[type].blck_size;
 }
 
 const char * ggml_type_name(enum ggml_type type) {
@@ -2049,24 +2048,37 @@ size_t ggml_element_size(const struct ggml_tensor * tensor) {
     return ggml_type_size(tensor->type);
 }
 
-static inline bool ggml_is_scalar(const struct ggml_tensor * tensor) {
+bool ggml_is_scalar(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[0] == 1 && tensor->ne[1] == 1 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
-static inline bool ggml_is_vector(const struct ggml_tensor * tensor) {
+bool ggml_is_vector(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[1] == 1 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
-static inline bool ggml_is_matrix(const struct ggml_tensor * tensor) {
+bool ggml_is_matrix(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
+bool ggml_is_3d(const struct ggml_tensor * tensor) {
+    return tensor->ne[3] == 1;
+}
+
+int ggml_n_dims(const struct ggml_tensor * tensor) {
+    for (int i = GGML_MAX_DIMS - 1; i >= 1; --i) {
+        if (tensor->ne[i] > 1) {
+            return i + 1;
+        }
+    }
+    return 1;
+}
+
 static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -2473,7 +2485,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         view_src   = view_src->view_src;
     }
 
-    size_t data_size = ggml_type_size(type)*(ne[0]/ggml_blck_size(type));
+    size_t data_size = ggml_row_size(type, ne[0]);
     for (int i = 1; i < n_dims; i++) {
         data_size *= ne[i];
     }
@@ -2516,7 +2528,6 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.type         =*/ type,
         /*.backend      =*/ GGML_BACKEND_CPU,
         /*.buffer       =*/ NULL,
-        /*.n_dims       =*/ n_dims,
         /*.ne           =*/ { 1, 1, 1, 1 },
         /*.nb           =*/ { 0, 0, 0, 0 },
         /*.op           =*/ GGML_OP_NONE,
@@ -2623,7 +2634,7 @@ struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) {
 }
 
 struct ggml_tensor * ggml_dup_tensor(struct ggml_context * ctx, const struct ggml_tensor * src) {
-    return ggml_new_tensor(ctx, src->type, src->n_dims, src->ne);
+    return ggml_new_tensor(ctx, src->type, GGML_MAX_DIMS, src->ne);
 }
 
 static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
@@ -3072,7 +3083,7 @@ struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char *
 struct ggml_tensor * ggml_view_tensor(
         struct ggml_context * ctx,
         struct ggml_tensor  * src) {
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src, 0);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, GGML_MAX_DIMS, src->ne, src, 0);
     ggml_format_name(result, "%s (view)", src->name);
 
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
@@ -3230,10 +3241,10 @@ static struct ggml_tensor * ggml_add_cast_impl(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, type, a->n_dims, a->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
 
     result->op   = GGML_OP_ADD;
-    result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne) : NULL;
+    result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne) : NULL;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3602,12 +3613,12 @@ struct ggml_tensor * ggml_sum_rows(
         is_node = true;
     }
 
-    int64_t ne[4] = {1,1,1,1};
-    for (int i=1; i<a->n_dims; ++i) {
+    int64_t ne[GGML_MAX_DIMS] = { 1 };
+    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
         ne[i] = a->ne[i];
     }
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, a->n_dims, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne);
 
     result->op   = GGML_OP_SUM_ROWS;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3628,8 +3639,8 @@ struct ggml_tensor * ggml_mean(
         is_node = true;
     }
 
-    int64_t ne[GGML_MAX_DIMS] = { 1, a->ne[1], a->ne[2], a->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, ne);
+    int64_t ne[4] = { 1, a->ne[1], a->ne[2], a->ne[3] };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     result->op   = GGML_OP_MEAN;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3651,8 +3662,7 @@ struct ggml_tensor * ggml_argmax(
         is_node = true;
     }
 
-    int64_t ne[GGML_MAX_DIMS] = { a->ne[1], 1, 1, 1 };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, a->n_dims, ne);
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]);
 
     result->op   = GGML_OP_ARGMAX;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3675,7 +3685,7 @@ struct ggml_tensor * ggml_repeat(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
 
     result->op   = GGML_OP_REPEAT;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3702,7 +3712,7 @@ struct ggml_tensor * ggml_repeat_back(
         return a;
     }
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
 
     result->op   = GGML_OP_REPEAT_BACK;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -4078,7 +4088,7 @@ struct ggml_tensor * ggml_mul_mat(
     }
 
     const int64_t ne[4] = { a->ne[1], b->ne[1], b->ne[2], b->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(a->n_dims, b->n_dims), ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     result->op   = GGML_OP_MUL_MAT;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -4112,7 +4122,7 @@ struct ggml_tensor * ggml_mul_mat_id(
     }
 
     const int64_t ne[4] = { as[0]->ne[1], b->ne[1], b->ne[2], b->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(as[0]->n_dims, b->n_dims), ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     ggml_set_op_params_i32(result, 0, id);
     ggml_set_op_params_i32(result, 1, n_as);
@@ -4150,7 +4160,7 @@ struct ggml_tensor * ggml_out_prod(
 
     // a is broadcastable to b for ne[2] and ne[3] -> use b->ne[2] and b->ne[3]
     const int64_t ne[4] = { a->ne[0], b->ne[0], b->ne[2], b->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(a->n_dims, b->n_dims), ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     result->op   = GGML_OP_OUT_PROD;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -4435,7 +4445,7 @@ struct ggml_tensor * ggml_reshape(
         //GGML_ASSERT(false);
     }
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a, 0);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b->ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -4813,7 +4823,7 @@ struct ggml_tensor * ggml_diag(
     }
 
     const int64_t ne[4] = { a->ne[0], a->ne[0], a->ne[2], a->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, MAX(a->n_dims, 2), ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, 4, ne);
 
     result->op   = GGML_OP_DIAG;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5460,7 +5470,7 @@ struct ggml_tensor * ggml_pool_1d(
         is_node = true;
     }
 
-    const int64_t ne[3] = {
+    const int64_t ne[2] = {
         ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
         a->ne[1],
     };
@@ -5579,7 +5589,7 @@ struct ggml_tensor * ggml_argsort(
         enum ggml_sort_order  order) {
     bool is_node = false;
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, a->n_dims, a->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
 
@@ -5626,7 +5636,7 @@ struct ggml_tensor * ggml_flash_attn(
     }
 
     //struct ggml_tensor * result = ggml_dup_tensor(ctx, q);
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, q->n_dims, q->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, q->ne);
 
     int32_t t = masked ? 1 : 0;
     ggml_set_op_params(result, &t, sizeof(t));
@@ -5659,7 +5669,7 @@ struct ggml_tensor * ggml_flash_ff(
     }
 
     //struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne);
 
     result->op   = GGML_OP_FLASH_FF;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5775,7 +5785,6 @@ struct ggml_tensor * ggml_win_part(
     const int np  = npx*npy;
 
     const int64_t ne[4] = { a->ne[0], w, w, np, };
-
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { npx, npy, w };
@@ -9571,16 +9580,11 @@ static bool ggml_compute_forward_mul_mat_use_blas(
 }
 #endif
 
-// off1 = offset in i11 and i1
-// cne1 = ne11 and ne1
-// in a normal matrix multiplication, off1 = 0 and cne1 = ne1
-// during GGML_TASK_INIT, the full src1 is converted regardless of off1 and cne1
 static void ggml_compute_forward_mul_mat(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
-              struct ggml_tensor * dst,
-              int64_t off1, int64_t cne1) {
+              struct ggml_tensor * dst) {
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
@@ -9648,9 +9652,9 @@ static void ggml_compute_forward_mul_mat(
                 const int64_t i03 = i13/r3;
                 const int64_t i02 = i12/r2;
 
-                const void  * x = (char *)            src0->data +             i02*nb02 + i03*nb03;
-                const float * y = (float *) ((char *) src1->data + off1*nb11 + i12*nb12 + i13*nb13);
-                      float * d = (float *) ((char *)  dst->data + off1*nb1  + i12*nb2  + i13*nb3);
+                const void  * x = (char *)            src0->data + i02*nb02 + i03*nb03;
+                const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
+                      float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
 
                 if (type != GGML_TYPE_F32) {
                             float * const wdata    = params->wdata;
@@ -9667,7 +9671,7 @@ static void ggml_compute_forward_mul_mat(
                 }
 
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
-                         cne1, ne01, ne10,
+                          ne1, ne01, ne10,
                          1.0f,    y, ne10,
                                   x, ne00,
                          0.0f,    d, ne01);
@@ -9683,7 +9687,7 @@ static void ggml_compute_forward_mul_mat(
     if (params->type == GGML_TASK_INIT) {
         if (src1->type != vec_dot_type) {
             char * wdata = params->wdata;
-            const size_t row_size = ne10*ggml_type_size(vec_dot_type)/ggml_blck_size(vec_dot_type);
+            const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
             assert(params->wsize >= ne11*ne12*ne13*row_size);
             assert(src1->type == GGML_TYPE_F32);
@@ -9706,10 +9710,10 @@ static void ggml_compute_forward_mul_mat(
     }
 
     const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-    const size_t row_size = ne10*ggml_type_size(vec_dot_type)/ggml_blck_size(vec_dot_type);
+    const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-    const int64_t nr0 = ne01;           // src0 rows
-    const int64_t nr1 = cne1*ne12*ne13; // src1 rows
+    const int64_t nr0 = ne01;          // src0 rows
+    const int64_t nr1 = ne1*ne12*ne13; // src1 rows
 
     //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
@@ -9751,9 +9755,9 @@ static void ggml_compute_forward_mul_mat(
     for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
         for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
             for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
-                const int64_t i13 = (ir1/(ne12*cne1));
-                const int64_t i12 = (ir1 - i13*ne12*cne1)/cne1;
-                const int64_t i11 = (ir1 - i13*ne12*cne1 - i12*cne1) + off1;
+                const int64_t i13 = (ir1/(ne12*ne1));
+                const int64_t i12 = (ir1 - i13*ne12*ne1)/ne1;
+                const int64_t i11 = (ir1 - i13*ne12*ne1 - i12*ne1);
 
                 // broadcast src0 into src1
                 const int64_t i03 = i13/r3;
@@ -9793,28 +9797,191 @@ static void ggml_compute_forward_mul_mat(
 
 static void ggml_compute_forward_mul_mat_id(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
+        const struct ggml_tensor * ids,
         const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        // during GGML_TASK_INIT the entire src1 is converted to vec_dot_type
-        ggml_compute_forward_mul_mat(params, dst->src[2], src1, dst, 0, dst->ne[1]);
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[2]; // only for GGML_TENSOR_BINARY_OP_LOCALS
 
-    const struct ggml_tensor * ids = src0;
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const enum ggml_type type = src0->type;
+
+    const bool src1_cont = ggml_is_contiguous(src1);
+
+    ggml_vec_dot_t    const vec_dot               = type_traits[type].vec_dot;
+    enum ggml_type    const vec_dot_type          = type_traits[type].vec_dot_type;
+    ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    // row groups
     const int id   = ggml_get_op_params_i32(dst, 0);
     const int n_as = ggml_get_op_params_i32(dst, 1);
 
-    for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-        const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
+    char * wdata_src1_end = (src1->type == vec_dot_type) ?
+            (char *) params->wdata :
+            (char *) params->wdata + GGML_PAD(ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t));
 
-        GGML_ASSERT(row_id >= 0 && row_id < n_as);
+    int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
+    int64_t * matrix_rows       = matrix_row_counts + n_as;     // [n_as][ne11]
 
-        const struct ggml_tensor * src0_row = dst->src[row_id + 2];
-        ggml_compute_forward_mul_mat(params, src0_row, src1, dst, i01, 1);
+    #define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne11 + (i1)]
+
+   if (params->type == GGML_TASK_INIT) {
+        char * wdata = params->wdata;
+        if (src1->type != vec_dot_type) {
+            const size_t row_size = ggml_row_size(vec_dot_type, ne10);
+
+            assert(params->wsize >= ne11*ne12*ne13*row_size);
+            assert(src1->type == GGML_TYPE_F32);
+
+            for (int64_t i13 = 0; i13 < ne13; ++i13) {
+                for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                    for (int64_t i11 = 0; i11 < ne11; ++i11) {
+                        from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
+                        wdata += row_size;
+                    }
+                }
+            }
+        }
+
+        // initialize matrix_row_counts
+        GGML_ASSERT(wdata == wdata_src1_end);
+        memset(matrix_row_counts, 0, n_as*sizeof(int64_t));
+
+        // group rows by src0 matrix
+        for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+            const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
+
+            GGML_ASSERT(row_id >= 0 && row_id < n_as);
+            MMID_MATRIX_ROW(row_id, matrix_row_counts[row_id]) = i01;
+            matrix_row_counts[row_id] += 1;
+        }
+
+        return;
     }
+
+    if (params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    // compute each matrix multiplication in sequence
+    for (int cur_a = 0; cur_a < n_as; ++cur_a) {
+        const int64_t cne1 = matrix_row_counts[cur_a];
+
+        if (cne1 == 0) {
+            continue;
+        }
+
+        const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
+
+        const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+        const size_t row_size = ggml_row_size(vec_dot_type, ne10);
+
+        const int64_t nr0 = ne01;           // src0 rows
+        const int64_t nr1 = cne1*ne12*ne13; // src1 rows
+
+        //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
+
+        // distribute the thread work across the inner or outer loop based on which one is larger
+
+        const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
+        const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
+
+        const int64_t ith0 = ith % nth0;
+        const int64_t ith1 = ith / nth0;
+
+        const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
+        const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
+
+        const int64_t ir010 = dr0*ith0;
+        const int64_t ir011 = MIN(ir010 + dr0, nr0);
+
+        const int64_t ir110 = dr1*ith1;
+        const int64_t ir111 = MIN(ir110 + dr1, nr1);
+
+        //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
+
+        // threads with no work simply yield (not sure if it helps)
+        if (ir010 >= ir011 || ir110 >= ir111) {
+            sched_yield();
+            continue;
+        }
+
+        assert(ne12 % ne02 == 0);
+        assert(ne13 % ne03 == 0);
+
+        // block-tiling attempt
+        const int64_t blck_0 = 16;
+        const int64_t blck_1 = 16;
+
+        // attempt to reduce false-sharing (does not seem to make a difference)
+        float tmp[16];
+
+        for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
+            for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
+                for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
+                    const int64_t  i13 = (ir1/(ne12*cne1)); // Note: currently, src1 is always a matrix
+                    const int64_t  i12 = (ir1 - i13*ne12*cne1)/cne1;
+                    const int64_t _i11 = (ir1 - i13*ne12*cne1 - i12*cne1);
+                    const int64_t  i11 = MMID_MATRIX_ROW(cur_a, _i11);
+
+                    // broadcast src0 into src1
+                    const int64_t i03 = i13/r3;
+                    const int64_t i02 = i12/r2;
+
+                    const int64_t i1 = i11;
+                    const int64_t i2 = i12;
+                    const int64_t i3 = i13;
+
+                    const char * src0_row = (const char *) src0_cur->data + (0 + i02*nb02 + i03*nb03);
+
+                    // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
+                    //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
+                    //       the original src1 data pointer, so we should index using the indices directly
+                    // TODO: this is a bit of a hack, we should probably have a better way to handle this
+                    const char * src1_col = (const char *) wdata +
+                        (src1_cont || src1->type != vec_dot_type
+                        ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
+                        : (i11*nb11 + i12*nb12 + i13*nb13));
+
+                    float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+
+                    //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                    //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
+                    //}
+
+                    for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                        vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                    }
+                    memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
+                }
+            }
+        }
+    }
+
+    #undef MMID_MATRIX_ROW
 }
 
 // ggml_compute_forward_out_prod
@@ -14182,7 +14349,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_MUL_MAT:
             {
-                ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor, 0, tensor->ne[1]);
+                ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
@@ -14558,7 +14725,7 @@ static struct ggml_tensor * ggml_recompute_graph_node(
         return replacements->vals[i];
     }
 
-    struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, node->n_dims, node->ne);
+    struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, GGML_MAX_DIMS, node->ne);
 
     // insert clone into replacements
     GGML_ASSERT(replacements->set.keys[i] == NULL); // assert that we don't overwrite
@@ -15982,7 +16149,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
-                // FIXME: blas
                 n_tasks = n_threads;
             } break;
         case GGML_OP_OUT_PROD:
@@ -16311,25 +16477,21 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     } else
 #endif
                     if (node->src[1]->type != vec_dot_type) {
-                        cur = ggml_type_size(vec_dot_type)*ggml_nelements(node->src[1])/ggml_blck_size(vec_dot_type);
+                        cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
                     }
                 } break;
             case GGML_OP_MUL_MAT_ID:
                 {
-                    const struct ggml_tensor * a = node->src[2];
-                    const struct ggml_tensor * b = node->src[1];
-                    const enum ggml_type vec_dot_type = type_traits[a->type].vec_dot_type;
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-                    if (ggml_compute_forward_mul_mat_use_blas(a, b, node)) {
-                        if (a->type != GGML_TYPE_F32) {
-                            // here we need memory just for single 2D matrix from src0
-                            cur = ggml_type_size(GGML_TYPE_F32)*(a->ne[0]*a->ne[1]);
-                        }
-                    } else
-#endif
-                    if (b->type != vec_dot_type) {
-                        cur = ggml_type_size(vec_dot_type)*ggml_nelements(b)/ggml_blck_size(vec_dot_type);
+                    const struct ggml_tensor * src0 = node->src[2];
+                    const struct ggml_tensor * src1 = node->src[1];
+                    const enum ggml_type vec_dot_type = type_traits[src0->type].vec_dot_type;
+                    if (src1->type != vec_dot_type) {
+                        cur = ggml_row_size(vec_dot_type, ggml_nelements(src1));
                     }
+                    const int n_as = ggml_get_op_params_i32(node, 1);
+                    cur = GGML_PAD(cur, sizeof(int64_t));        // align
+                    cur += n_as * sizeof(int64_t);               // matrix_row_counts
+                    cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows
                 } break;
             case GGML_OP_OUT_PROD:
                 {
@@ -16559,7 +16721,7 @@ static void ggml_graph_export_leaf(const struct ggml_tensor * tensor, FILE * fou
     fprintf(fout, "%-6s %-12s %8d %" PRId64 " %" PRId64 " %" PRId64 " %" PRId64 " %16zu %16zu %16zu %16zu %16p %32s\n",
             ggml_type_name(tensor->type),
             ggml_op_name  (tensor->op),
-            tensor->n_dims,
+            ggml_n_dims(tensor),
             ne[0], ne[1], ne[2], ne[3],
             nb[0], nb[1], nb[2], nb[3],
             tensor->data,
@@ -16574,7 +16736,7 @@ static void ggml_graph_export_node(const struct ggml_tensor * tensor, const char
             arg,
             ggml_type_name(tensor->type),
             ggml_op_name  (tensor->op),
-            tensor->n_dims,
+            ggml_n_dims(tensor),
             ne[0], ne[1], ne[2], ne[3],
             nb[0], nb[1], nb[2], nb[3],
             tensor->data,
@@ -16664,11 +16826,9 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
 
                 const uint32_t type   = tensor->type;
                 const uint32_t op     = tensor->op;
-                const uint32_t n_dims = tensor->n_dims;
 
                 fwrite(&type,   sizeof(uint32_t), 1, fout);
                 fwrite(&op,     sizeof(uint32_t), 1, fout);
-                fwrite(&n_dims, sizeof(uint32_t), 1, fout);
 
                 for (int j = 0; j < GGML_MAX_DIMS; ++j) {
                     const uint64_t ne = tensor->ne[j];
@@ -16698,11 +16858,9 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
 
                 const uint32_t type   = tensor->type;
                 const uint32_t op     = tensor->op;
-                const uint32_t n_dims = tensor->n_dims;
 
                 fwrite(&type,   sizeof(uint32_t), 1, fout);
                 fwrite(&op,     sizeof(uint32_t), 1, fout);
-                fwrite(&n_dims, sizeof(uint32_t), 1, fout);
 
                 for (int j = 0; j < GGML_MAX_DIMS; ++j) {
                     const uint64_t ne = tensor->ne[j];
@@ -16874,12 +17032,10 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
         {
             uint32_t type;
             uint32_t op;
-            uint32_t n_dims;
 
             for (uint32_t i = 0; i < n_leafs; ++i) {
                 type   = *(const uint32_t *) ptr; ptr += sizeof(type);
                 op     = *(const uint32_t *) ptr; ptr += sizeof(op);
-                n_dims = *(const uint32_t *) ptr; ptr += sizeof(n_dims);
 
                 int64_t ne[GGML_MAX_DIMS];
                 size_t  nb[GGML_MAX_DIMS];
@@ -16895,7 +17051,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
                     nb[j] = nb_cur;
                 }
 
-                struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, n_dims, ne);
+                struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, GGML_MAX_DIMS, ne);
 
                 tensor->op = (enum ggml_op) op;
 
@@ -16912,7 +17068,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
                 ptr += ggml_nbytes(tensor);
 
-                fprintf(stderr, "%s: loaded leaf %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor));
+                fprintf(stderr, "%s: loaded leaf %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
             }
         }
 
@@ -16922,12 +17078,10 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
         {
             uint32_t type;
             uint32_t op;
-            uint32_t n_dims;
 
             for (uint32_t i = 0; i < n_nodes; ++i) {
                 type   = *(const uint32_t *) ptr; ptr += sizeof(type);
                 op     = *(const uint32_t *) ptr; ptr += sizeof(op);
-                n_dims = *(const uint32_t *) ptr; ptr += sizeof(n_dims);
 
                 enum ggml_op eop = (enum ggml_op) op;
 
@@ -16998,7 +17152,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
                         } break;
                     default:
                         {
-                            tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, n_dims, ne);
+                            tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, GGML_MAX_DIMS, ne);
 
                             tensor->op = eop;
                         } break;
@@ -17017,7 +17171,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
                 result->nodes[i] = tensor;
 
-                fprintf(stderr, "%s: loaded node %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor));
+                fprintf(stderr, "%s: loaded node %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
             }
         }
     }
@@ -17155,7 +17309,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
             fprintf(fp, "(%s)|", ggml_type_name(node->type));
         }
 
-        if (node->n_dims == 2) {
+        if (ggml_is_matrix(node)) {
             fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], ggml_op_symbol(node->op));
         } else {
             fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], ggml_op_symbol(node->op));
@@ -17422,7 +17576,7 @@ static enum ggml_opt_result ggml_opt_adam(
             int64_t i = 0;
             for (int p = 0; p < np; ++p) {
                 const int64_t ne = ggml_nelements(ps[p]);
-                const float p_decay = ((ps[p]->n_dims >= decay_min_ndim) ? decay : 0.0f) * sched;
+                const float p_decay = ((ggml_n_dims(ps[p]) >= decay_min_ndim) ? decay : 0.0f) * sched;
                 for (int64_t j = 0; j < ne; ++j) {
                     float x  = ggml_get_f32_1d(ps[p], j);
                     float g_ = g[i]*gnorm;
@@ -18696,7 +18850,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 return NULL;
             }
 
-            const size_t size_cur = (ne*ggml_type_size(info->type))/ggml_blck_size(info->type);
+            const size_t size_cur = ggml_row_size(info->type, ne);
 
             ctx->size += GGML_PAD(size_cur, ctx->alignment);
         }
@@ -19200,8 +19354,8 @@ void gguf_add_tensor(
         ctx->infos[idx].ne[i] = 1;
     }
 
-    ctx->infos[idx].n_dims = tensor->n_dims;
-    for (int i = 0; i < tensor->n_dims; i++) {
+    ctx->infos[idx].n_dims = ggml_n_dims(tensor);
+    for (uint32_t i = 0; i < ctx->infos[idx].n_dims; i++) {
         ctx->infos[idx].ne[i] = tensor->ne[i];
     }
 

ggml.h

@@ -502,7 +502,6 @@ extern "C" {
 
         struct ggml_backend_buffer * buffer;
 
-        int     n_dims;
         int64_t ne[GGML_MAX_DIMS]; // number of elements
         size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
                                    // nb[0] = ggml_type_size(type)
@@ -534,7 +533,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[12];
+        char padding[8];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -639,11 +638,14 @@ extern "C" {
     GGML_API int64_t ggml_nrows       (const struct ggml_tensor * tensor);
     GGML_API size_t  ggml_nbytes      (const struct ggml_tensor * tensor);
     GGML_API size_t  ggml_nbytes_pad  (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
-    GGML_API size_t  ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split);
 
-    GGML_API int     ggml_blck_size (enum ggml_type type);
-    GGML_API size_t  ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
-    GGML_API float   ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
+    GGML_API int    ggml_blck_size(enum ggml_type type);
+    GGML_API size_t ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
+    GGML_API size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
+
+    GGML_DEPRECATED(
+    GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
+    "use ggml_row_size() instead");
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
@@ -662,6 +664,11 @@ 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);
+    GGML_API bool ggml_is_scalar    (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_vector    (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_matrix    (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_3d        (const struct ggml_tensor * tensor);
+    GGML_API int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
 
     GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 

llama.cpp

@@ -1555,7 +1555,7 @@ static bool llama_kv_cache_init(
     cache.cells.clear();
     cache.cells.resize(n_ctx);
 
-    cache.buf.resize(n_elements*(ggml_type_sizef(ktype) + ggml_type_sizef(vtype)) + 2u*n_layer*ggml_tensor_overhead());
+    cache.buf.resize(ggml_row_size(ktype, n_elements) + ggml_row_size(vtype, n_elements) + 2u*n_layer*ggml_tensor_overhead());
     memset(cache.buf.data, 0, cache.buf.size);
 
     struct ggml_init_params params;
@@ -3822,8 +3822,8 @@ static void llm_build_k_shift(
             ggml_rope_custom_inplace(ctx,
                     ggml_view_3d(ctx, kv.k_l[il],
                         n_embd_head, n_head_kv, n_ctx,
-                        ggml_type_sizef(kv.k_l[il]->type)*n_embd_head,
-                        ggml_type_sizef(kv.k_l[il]->type)*n_embd_gqa,
+                        ggml_row_size(kv.k_l[il]->type, n_embd_head),
+                        ggml_row_size(kv.k_l[il]->type, n_embd_gqa),
                         0),
                     K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow);
@@ -3852,7 +3852,7 @@ static void llm_build_kv_store(
     cb(v_cur_t, "v_cur_t", il);
 
     struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_gqa,
-            (ggml_type_sizef(kv.k_l[il]->type)*n_embd_gqa)*kv_head);
+            (ggml_row_size(kv.k_l[il]->type, n_embd_gqa))*kv_head);
     cb(k_cache_view, "k_cache_view", il);
 
     struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_gqa,
@@ -4011,8 +4011,8 @@ static struct ggml_tensor * llm_build_kqv(
     struct ggml_tensor * k =
         ggml_view_3d(ctx, kv.k_l[il],
                 n_embd_head, n_kv, n_head_kv,
-                ggml_type_sizef(kv.k_l[il]->type)*n_embd_gqa,
-                ggml_type_sizef(kv.k_l[il]->type)*n_embd_head,
+                ggml_row_size(kv.k_l[il]->type, n_embd_gqa),
+                ggml_row_size(kv.k_l[il]->type, n_embd_head),
                 0);
     cb(k, "k", il);
 
@@ -8471,7 +8471,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'?
 
         // quantize only 2D tensors
-        quantize &= (tensor->n_dims == 2);
+        quantize &= (ggml_n_dims(tensor) == 2);
         quantize &= params->quantize_output_tensor || name != "output.weight";
         quantize &= !params->only_copy;
 

requirements.txt

@@ -2,3 +2,4 @@ numpy==1.24.4
 sentencepiece==0.1.98
 transformers>=4.34.0
 gguf>=0.1.0
+protobuf>=4.21.0

tests/test-backend-ops.cpp

@@ -54,7 +54,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
         ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
     } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16) {
         GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0);
-        std::vector<uint8_t> dataq(ggml_type_size(tensor->type)*size/ggml_blck_size(tensor->type));
+        std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
         int64_t hist[16];
         ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size, hist);
         ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
@@ -72,6 +72,8 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
 
     ggml_type_traits_t tt = ggml_internal_get_type_traits(t->type);
     size_t bs = ggml_blck_size(t->type);
+    std::vector<float> vq(ggml_blck_size(t->type));
+    bool quantized = ggml_is_quantized(t->type);
 
     // access elements by index to avoid gaps in views
     for (int64_t i3 = 0; i3 < t->ne[3]; i3++) {
@@ -85,9 +87,8 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
                         tv.push_back(*(float *) &buf[i]);
                     } else if (t->type == GGML_TYPE_I32) {
                         tv.push_back((float)*(int32_t *) &buf[i]);
-                    } else if (ggml_is_quantized(t->type)) {
-                        std::vector<float> vq(ggml_blck_size(t->type));
-                        tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type));
+                    } else if (quantized) {
+                        tt.to_float(&buf[i], vq.data(), bs);
                         tv.insert(tv.end(), vq.begin(), vq.end());
                     } else {
                         GGML_ASSERT(false);

tests/test-quantize-perf.cpp

@@ -286,7 +286,7 @@ int main(int argc, char * argv[]) {
                         qfns.from_float_reference(test_data1, test_q1, size);
                         return test_q1[0];
                     };
-                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");
@@ -300,7 +300,7 @@ int main(int argc, char * argv[]) {
                         qfns.from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
-                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");
@@ -315,7 +315,7 @@ int main(int argc, char * argv[]) {
                         qfns.to_float(test_q1, test_out, size);
                         return test_out[0];
                     };
-                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");
@@ -330,7 +330,7 @@ int main(int argc, char * argv[]) {
                         vdot.from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
-                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");
@@ -347,7 +347,7 @@ int main(int argc, char * argv[]) {
                         qfns.vec_dot(size, &result, test_q1, test_q2);
                         return result;
                     };
-                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");