leave only basic functions for SYCL CI

WA for permute(0,1,3,2) mul_mat
ggml-ci

ci/run.sh

@@ -53,6 +53,8 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
         exit 1
     fi
 
+    # Only functionality CI for SYCL now
+    GG_BUILD_LOW_PERF=True
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
 
@@ -150,7 +152,12 @@ function gg_run_ctest_release {
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
         (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
-        (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+       if [ ! -z "$GG_BUILD_SYCL" ]; then
+            # TODO(airMeng): fix iq1_xs and iq3_xs quantization in SYCL
+            (time ctest --output-on-failure -L main -E "test-quantize-fns|test-opt" ) 2>&1 | tee -a "$OUT/${ci}-ctest.log"
+       else
+            (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+       fi
     fi
 
     set +e
@@ -824,7 +831,10 @@ fi
 
 ret=0
 
-test $ret -eq 0 && gg_run ctest_debug
+if [ -z "$GG_BUILD_SYCL" ]; then
+    # to save time, remove after more machines available
+    test $ret -eq 0 && gg_run ctest_debug
+fi
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then

common/arg.cpp

@@ -128,13 +128,13 @@ static void common_params_handle_model_default(common_params & params) {
             }
             params.hf_file = params.model;
         } else if (params.model.empty()) {
-            params.model = fs_get_cache_file(string_split<std::string>(params.hf_file, '/').back());
+            params.model = fs_get_cache_file(string_split(params.hf_file, '/').back());
         }
     } else if (!params.model_url.empty()) {
         if (params.model.empty()) {
-            auto f = string_split<std::string>(params.model_url, '#').front();
-            f = string_split<std::string>(f, '?').front();
-            params.model = fs_get_cache_file(string_split<std::string>(f, '/').back());
+            auto f = string_split(params.model_url, '#').front();
+            f = string_split(f, '?').front();
+            params.model = fs_get_cache_file(string_split(f, '/').back());
         }
     } else if (params.model.empty()) {
         params.model = DEFAULT_MODEL_PATH;
@@ -251,9 +251,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         for (auto & antiprompt : params.antiprompt) {
             string_process_escapes(antiprompt);
         }
-        for (auto & seq_breaker : params.sparams.dry_sequence_breakers) {
-            string_process_escapes(seq_breaker);
-        }
     }
 
     if (!params.kv_overrides.empty()) {
@@ -882,7 +879,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--samplers"}, "SAMPLERS",
         string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()),
         [](common_params & params, const std::string & value) {
-            const auto sampler_names = string_split<std::string>(value, ';');
+            const auto sampler_names = string_split(value, ';');
             params.sparams.samplers = common_sampler_types_from_names(sampler_names, true);
         }
     ).set_sparam());
@@ -943,6 +940,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sparams.min_p = std::stof(value);
         }
     ).set_sparam());
+    add_opt(common_arg(
+        {"--tfs"}, "N",
+        string_format("tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)params.sparams.tfs_z),
+        [](common_params & params, const std::string & value) {
+            params.sparams.tfs_z = std::stof(value);
+        }
+    ).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
         string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sparams.xtc_probability),
@@ -993,64 +997,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sparams.penalty_freq = std::stof(value);
         }
     ).set_sparam());
-    add_opt(common_arg(
-        {"--dry-multiplier"}, "N",
-        string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sparams.dry_multiplier),
-        [](common_params & params, const std::string & value) {
-            params.sparams.dry_multiplier = std::stof(value);
-        }
-    ).set_sparam());
-    add_opt(common_arg(
-        {"--dry-base"}, "N",
-        string_format("set DRY sampling base value (default: %.2f)", (double)params.sparams.dry_base),
-        [](common_params & params, const std::string & value) {
-            float potential_base = std::stof(value);
-            if (potential_base >= 1.0f)
-            {
-                params.sparams.dry_base = potential_base;
-            }
-        }
-    ).set_sparam());
-    add_opt(common_arg(
-        {"--dry-allowed-length"}, "N",
-        string_format("set allowed length for DRY sampling (default: %d)", params.sparams.dry_allowed_length),
-        [](common_params & params, int value) {
-            params.sparams.dry_allowed_length = value;
-        }
-    ).set_sparam());
-    add_opt(common_arg(
-        {"--dry-penalty-last-n"}, "N",
-        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sparams.dry_penalty_last_n),
-        [](common_params & params, int value) {
-            params.sparams.dry_penalty_last_n = value;
-        }
-    ).set_sparam());
-    add_opt(common_arg(
-        {"--dry-sequence-breaker"}, "STRING",
-        string_format("add sequence breaker for DRY sampling, clearing out default breakers (%s) in the process; use \"none\" to not use any sequence breakers\n",
-            params.sparams.dry_sequence_breakers.empty() ? "none" :
-            std::accumulate(std::next(params.sparams.dry_sequence_breakers.begin()),
-                params.sparams.dry_sequence_breakers.end(),
-                std::string("'") + (params.sparams.dry_sequence_breakers[0] == "\n" ? "\\n" : params.sparams.dry_sequence_breakers[0]) + "'",
-                [](const std::string& a, const std::string& b) {
-                    std::string formatted_b = (b == "\n") ? "\\n" : b;
-                    return a + ", '" + formatted_b + "'";
-                }).c_str()),
-        [](common_params & params, const std::string & value) {
-            static bool defaults_cleared = false;
-
-            if (!defaults_cleared) {
-                params.sparams.dry_sequence_breakers.clear();
-                defaults_cleared = true;
-            }
-
-            if (value == "none") {
-                params.sparams.dry_sequence_breakers.clear();
-            } else {
-                params.sparams.dry_sequence_breakers.emplace_back(value);
-            }
-        }
-    ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-range"}, "N",
         string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sparams.dynatemp_range),
@@ -1067,7 +1013,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat"}, "N",
-        string_format("use Mirostat sampling.\nTop K, Nucleus and Locally Typical samplers are ignored if used.\n"
+        string_format("use Mirostat sampling.\nTop K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n"
         "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sparams.mirostat),
         [](common_params & params, int value) {
             params.sparams.mirostat = value;

common/common.cpp

@@ -416,6 +416,19 @@ std::string string_format(const char * fmt, ...) {
     return std::string(buf.data(), size);
 }
 
+std::vector<std::string> string_split(std::string input, char separator) {
+    std::vector<std::string> parts;
+    size_t separator_pos = input.find(separator);
+    while (separator_pos != std::string::npos) {
+        std::string part = input.substr(0, separator_pos);
+        parts.emplace_back(part);
+        input = input.substr(separator_pos + 1);
+        separator_pos = input.find(separator);
+    }
+    parts.emplace_back(input);
+    return parts;
+}
+
 std::string string_strip(const std::string & str) {
     size_t start = 0;
     size_t end = str.size();
@@ -2006,10 +2019,6 @@ void yaml_dump_non_result_info(FILE * stream, const common_params & params, cons
     fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
     fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
     fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
-    fprintf(stream, "dry_allowed_length: %d # default: 2\n", sparams.dry_allowed_length);
-    fprintf(stream, "dry_base: %.2f # default: 1.75\n", sparams.dry_base);
-    fprintf(stream, "dry_multiplier: %.1f # default: 0.0\n", sparams.dry_multiplier);
-    fprintf(stream, "dry_penalty_last_n: %d # default: -1 (0 = disable, -1 = context size)\n", sparams.dry_penalty_last_n);
     fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
     fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
     fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
@@ -2090,6 +2099,7 @@ void yaml_dump_non_result_info(FILE * stream, const common_params & params, cons
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());
     yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
 
+    fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
     fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);

common/common.h

@@ -84,15 +84,14 @@ enum llama_example {
 
 enum common_sampler_type {
     COMMON_SAMPLER_TYPE_NONE        = 0,
-    COMMON_SAMPLER_TYPE_DRY         = 1,
-    COMMON_SAMPLER_TYPE_TOP_K       = 2,
-    COMMON_SAMPLER_TYPE_TOP_P       = 3,
-    COMMON_SAMPLER_TYPE_MIN_P       = 4,
-  //COMMON_SAMPLER_TYPE_TFS_Z       = 5,
-    COMMON_SAMPLER_TYPE_TYPICAL_P   = 6,
-    COMMON_SAMPLER_TYPE_TEMPERATURE = 7,
-    COMMON_SAMPLER_TYPE_XTC         = 8,
-    COMMON_SAMPLER_TYPE_INFILL      = 9,
+    COMMON_SAMPLER_TYPE_TOP_K       = 1,
+    COMMON_SAMPLER_TYPE_TOP_P       = 2,
+    COMMON_SAMPLER_TYPE_MIN_P       = 3,
+    COMMON_SAMPLER_TYPE_TFS_Z       = 4,
+    COMMON_SAMPLER_TYPE_TYPICAL_P   = 5,
+    COMMON_SAMPLER_TYPE_TEMPERATURE = 6,
+    COMMON_SAMPLER_TYPE_XTC         = 7,
+    COMMON_SAMPLER_TYPE_INFILL      = 8,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -105,39 +104,34 @@ enum dimre_method {
 struct common_sampler_params {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
-    int32_t n_prev             = 64;    // number of previous tokens to remember
-    int32_t n_probs            = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t min_keep           = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t top_k              = 40;    // <= 0 to use vocab size
-    float   top_p              = 0.95f; // 1.0 = disabled
-    float   min_p              = 0.05f; // 0.0 = disabled
-    float   xtc_probability    = 0.00f; // 0.0 = disabled
-    float   xtc_threshold      = 0.10f; // > 0.5 disables XTC
-    float   typ_p              = 1.00f; // typical_p, 1.0 = disabled
-    float   temp               = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float   dynatemp_range     = 0.00f; // 0.0 = disabled
-    float   dynatemp_exponent  = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat     = 1.00f; // 1.0 = disabled
-    float   penalty_freq       = 0.00f; // 0.0 = disabled
-    float   penalty_present    = 0.00f; // 0.0 = disabled
-    float   dry_multiplier     = 0.0f;  // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
-    float   dry_base           = 1.75f; // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
-    int32_t dry_allowed_length = 2;     // tokens extending repetitions beyond this receive penalty
-    int32_t dry_penalty_last_n = -1;    // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
-    int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau       = 5.00f; // target entropy
-    float   mirostat_eta       = 0.10f; // learning rate
-    bool    penalize_nl        = false; // consider newlines as a repeatable token
-    bool    ignore_eos         = false;
-    bool    no_perf            = false; // disable performance metrics
-
-    std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
+    int32_t n_prev            = 64;    // number of previous tokens to remember
+    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t min_keep          = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t top_k             = 40;    // <= 0 to use vocab size
+    float   top_p             = 0.95f; // 1.0 = disabled
+    float   min_p             = 0.05f; // 0.0 = disabled
+    float   xtc_probability   = 0.00f; // 0.0 = disabled
+    float   xtc_threshold     = 0.10f; // > 0.5 disables XTC
+    float   tfs_z             = 1.00f; // 1.0 = disabled
+    float   typ_p             = 1.00f; // typical_p, 1.0 = disabled
+    float   temp              = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float   dynatemp_range    = 0.00f; // 0.0 = disabled
+    float   dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat    = 1.00f; // 1.0 = disabled
+    float   penalty_freq      = 0.00f; // 0.0 = disabled
+    float   penalty_present   = 0.00f; // 0.0 = disabled
+    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   mirostat_tau      = 5.00f; // target entropy
+    float   mirostat_eta      = 0.10f; // learning rate
+    bool    penalize_nl       = false; // consider newlines as a repeatable token
+    bool    ignore_eos        = false;
+    bool    no_perf           = false; // disable performance metrics
 
 
     std::vector<enum common_sampler_type> samplers = {
-        COMMON_SAMPLER_TYPE_DRY,
         COMMON_SAMPLER_TYPE_TOP_K,
+        COMMON_SAMPLER_TYPE_TFS_Z,
         COMMON_SAMPLER_TYPE_TYPICAL_P,
         COMMON_SAMPLER_TYPE_TOP_P,
         COMMON_SAMPLER_TYPE_MIN_P,
@@ -386,6 +380,8 @@ bool set_process_priority(enum ggml_sched_priority prio);
 LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
 std::string string_format(const char * fmt, ...);
 
+std::vector<std::string> string_split(std::string input, char separator);
+
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
@@ -393,7 +389,6 @@ void string_replace_all(std::string & s, const std::string & search, const std::
 
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
-    static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
     std::vector<T> values;
     std::istringstream str_stream(str);
     std::string token;
@@ -406,22 +401,6 @@ static std::vector<T> string_split(const std::string & str, char delim) {
     return values;
 }
 
-template<>
-std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
-{
-    std::vector<std::string> parts;
-    size_t begin_pos = 0;
-    size_t separator_pos = input.find(separator);
-    while (separator_pos != std::string::npos) {
-        std::string part = input.substr(begin_pos, separator_pos - begin_pos);
-        parts.emplace_back(part);
-        begin_pos = separator_pos + 1;
-        separator_pos = input.find(separator, begin_pos);
-    }
-    parts.emplace_back(input.substr(begin_pos, separator_pos - begin_pos));
-    return parts;
-}
-
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 

common/sampling.cpp

@@ -130,12 +130,10 @@ std::string common_sampler_params::print() const {
 
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
-            "\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
-            "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
             penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
-            dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
-            top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp,
+            top_k, tfs_z, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp,
             mirostat, mirostat_eta, mirostat_tau);
 
     return std::string(result);
@@ -176,17 +174,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
     if (params.mirostat == 0) {
         for (const auto & cnstr : params.samplers) {
             switch (cnstr) {
-                    case COMMON_SAMPLER_TYPE_DRY:
-                    {
-                        std::vector<const char*> c_breakers;
-                        c_breakers.reserve(params.dry_sequence_breakers.size());
-                        for (const auto& str : params.dry_sequence_breakers) {
-                            c_breakers.push_back(str.c_str());
-                        }
-
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
-                    }
-                        break;
                 case COMMON_SAMPLER_TYPE_TOP_K:
                     llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
                     break;
@@ -199,6 +186,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 case COMMON_SAMPLER_TYPE_XTC:
                     llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
                     break;
+                case COMMON_SAMPLER_TYPE_TFS_Z:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
+                    break;
                 case COMMON_SAMPLER_TYPE_TYPICAL_P:
                     llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
                     break;
@@ -368,8 +358,8 @@ std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx_
 
 char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
     switch (cnstr) {
-        case COMMON_SAMPLER_TYPE_DRY:         return 'd';
         case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
+        case COMMON_SAMPLER_TYPE_TFS_Z:       return 'f';
         case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
         case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
         case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
@@ -382,8 +372,8 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
 
 std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
     switch (cnstr) {
-        case COMMON_SAMPLER_TYPE_DRY:         return "dry";
         case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
+        case COMMON_SAMPLER_TYPE_TFS_Z:       return "tfs_z";
         case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
         case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
         case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
@@ -396,11 +386,11 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
 
 std::vector<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
     std::unordered_map<std::string, common_sampler_type> sampler_canonical_name_map {
-        { "dry",         COMMON_SAMPLER_TYPE_DRY },
         { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
         { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
         { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
+        { "tfs_z",       COMMON_SAMPLER_TYPE_TFS_Z },
         { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
         { "xtc",         COMMON_SAMPLER_TYPE_XTC },
         { "infill",      COMMON_SAMPLER_TYPE_INFILL },
@@ -417,6 +407,8 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         { "typ-p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "typ",         COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "min-p",       COMMON_SAMPLER_TYPE_MIN_P },
+        { "tfs-z",       COMMON_SAMPLER_TYPE_TFS_Z },
+        { "tfs",         COMMON_SAMPLER_TYPE_TFS_Z },
         { "temp",        COMMON_SAMPLER_TYPE_TEMPERATURE },
     };
 
@@ -442,8 +434,8 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
 
 std::vector<common_sampler_type> common_sampler_types_from_chars(const std::string & chars) {
     std::unordered_map<char, common_sampler_type> sampler_name_map = {
-        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_DRY),         COMMON_SAMPLER_TYPE_DRY },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TFS_Z),       COMMON_SAMPLER_TYPE_TFS_Z },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },

convert_hf_to_gguf.py

@@ -573,9 +573,6 @@ class Model:
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/BAAI/bge-small-en-v1.5
             res = "bert-bge"
-        if chkhsh == "8e62295832751ca1e8f92f2226f403dea30dc5165e448b5bfa05af5340c64ec7":
-            # ref: https://huggingface.co/BAAI/bge-large-zh-v1.5
-            res = "bert-bge-large"
         if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
             # ref: https://huggingface.co/mosaicml/mpt-7b
             res = "mpt"

convert_hf_to_gguf_update.py

@@ -72,7 +72,6 @@ models = [
     {"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
     {"name": "falcon",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
     {"name": "bert-bge",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
-    {"name": "bert-bge-large", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/BAAI/bge-large-zh-v1.5", },
     {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
     {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },

convert_lora_to_gguf.py

@@ -230,7 +230,7 @@ def get_base_tensor_name(lora_tensor_name: str) -> str:
 
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
-        description="Convert a Hugging Face PEFT LoRA adapter to a GGUF file")
+        description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
     parser.add_argument(
         "--outfile", type=Path,
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
@@ -257,11 +257,11 @@ def parse_args() -> argparse.Namespace:
     )
     parser.add_argument(
         "--base", type=Path, required=True,
-        help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required",
+        help="directory containing base model file",
     )
     parser.add_argument(
         "lora_path", type=Path,
-        help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
+        help="directory containing LoRA adapter file",
     )
 
     return parser.parse_args()

examples/llama-bench/llama-bench.cpp

@@ -21,6 +21,12 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
+#include "ggml-cuda.h"
+#include "ggml-sycl.h"
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
 
 #ifdef _WIN32
 #define WIN32_LEAN_AND_MEAN
@@ -76,27 +82,95 @@ static T stdev(const std::vector<T> & v) {
 }
 
 static std::string get_cpu_info() {
-    std::vector<std::string> cpu_list;
-    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
-        auto * dev = ggml_backend_dev_get(i);
-        auto dev_type = ggml_backend_dev_type(dev);
-        if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU || dev_type == GGML_BACKEND_DEVICE_TYPE_ACCEL) {
-            cpu_list.push_back(ggml_backend_dev_description(dev));
+    std::string id;
+#ifdef __linux__
+    FILE * f = fopen("/proc/cpuinfo", "r");
+    if (f) {
+        char buf[1024];
+        while (fgets(buf, sizeof(buf), f)) {
+            if (strncmp(buf, "model name", 10) == 0) {
+                char * p = strchr(buf, ':');
+                if (p) {
+                    p++;
+                    while (std::isspace(*p)) {
+                        p++;
+                    }
+                    while (std::isspace(p[strlen(p) - 1])) {
+                        p[strlen(p) - 1] = '\0';
+                    }
+                    id = p;
+                    break;
+                }
+            }
+        }
+        fclose(f);
+    }
+#elif defined(_WIN32)
+    HKEY hKey;
+    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                     0,
+                     KEY_READ,
+                     &hKey) != ERROR_SUCCESS) {
+        // fail to open registry key
+        return "";
+    }
+    char cpu_brand[256];
+    DWORD cpu_brand_size = sizeof(cpu_brand);
+    if (RegQueryValueExA(hKey,
+                        TEXT("ProcessorNameString"),
+                        NULL,
+                        NULL,
+                        (LPBYTE)cpu_brand,
+                        &cpu_brand_size) == ERROR_SUCCESS) {
+        id.assign(cpu_brand, cpu_brand_size);
+        if (id.find('\0') != std::string::npos) {
+            id.resize(id.find('\0'));
         }
     }
-    return join(cpu_list, ", ");
+    RegCloseKey(hKey);
+#endif
+    // TODO: other platforms
+    return id;
 }
 
 static std::string get_gpu_info() {
-    std::vector<std::string> gpu_list;
-    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
-        auto * dev = ggml_backend_dev_get(i);
-        auto dev_type = ggml_backend_dev_type(dev);
-        if (dev_type == GGML_BACKEND_DEVICE_TYPE_GPU) {
-            gpu_list.push_back(ggml_backend_dev_description(dev));
+    std::string id;
+#ifdef GGML_USE_CUDA
+    int count = ggml_backend_cuda_get_device_count();
+    for (int i = 0; i < count; i++) {
+        char buf[128];
+        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
         }
     }
-    return join(gpu_list, ", ");
+#endif
+#ifdef GGML_USE_SYCL
+    int count = ggml_backend_sycl_get_device_count();
+    for (int i = 0; i < count; i++) {
+        char buf[128];
+        ggml_backend_sycl_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
+        }
+    }
+#endif
+#ifdef GGML_USE_CANN
+    uint32_t count = ggml_backend_cann_get_device_count();
+    for (uint32_t i = 0; i < count; i++) {
+        char buf[128];
+        ggml_backend_cann_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
+        }
+    }
+#endif
+    // TODO: other backends
+    return id;
 }
 
 // command line params
@@ -864,15 +938,29 @@ struct test {
     }
 
     static std::string get_backend() {
-        std::vector<std::string> backends;
-        for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
-            auto * reg = ggml_backend_reg_get(i);
-            std::string name = ggml_backend_reg_name(reg);
-            if (name != "CPU") {
-                backends.push_back(ggml_backend_reg_name(reg));
-            }
+        if (cuda) {
+            return GGML_CUDA_NAME;
         }
-        return backends.empty() ? "CPU" : join(backends, ",");
+        if (vulkan) {
+            return "Vulkan";
+        }
+        if (kompute) {
+            return "Kompute";
+        }
+        if (metal) {
+            return "Metal";
+        }
+        if (sycl) {
+            return GGML_SYCL_NAME;
+        }
+        if (gpu_blas) {
+            return "GPU BLAS";
+        }
+        if (blas) {
+            return "BLAS";
+        }
+
+        return "CPU";
     }
 
     static const std::vector<std::string> & get_fields() {

examples/main/README.md

@@ -187,30 +187,6 @@ Use the `--no-penalize-nl` option to disable newline penalization when applying
 
 Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
 
-### DRY Repetition Penalty
-
-DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
-
-- `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
-- `--dry-base N`: Set the DRY sampling base value (default: 1.75).
-- `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
-- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
-- `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.
-
-The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
-
-The `dry-base` option sets the base value for the exponential penalty calculation in DRY sampling. Higher values lead to more aggressive penalization of repetitions.
-
-The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.
-
-The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
-
-The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.
-
-DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.
-
-Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
-
 ### Top-K Sampling
 
 -   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
@@ -235,6 +211,14 @@ The Min-P sampling method was designed as an alternative to Top-P, and aims to e
 
 Example usage: `--min-p 0.05`
 
+### Tail-Free Sampling (TFS)
+
+-   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
+
+Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS.
+
+Example usage: `--tfs 0.95`
+
 ### Locally Typical Sampling
 
 -   `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
@@ -333,15 +317,6 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
 
-## LoRA (Low-Rank Adaptation) adapters
-
--   `--lora FNAME`: Optional path to a LoRA adapter to use with scaling of 1.0. Can be mixed with `--lora-scaled` and can be repeated to use multiple adapters.
--   `--lora-scaled FNAME`: Optional path to a LoRA adapter with user-defined scaling. Can be mixed with `--lora` and can repeated to use multiple adapters.
-
-You can add LoRA adapters using `--lora` or `--lora-scaled`. For example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...` or `--lora-scaled lora_task_A.gguf 0.5 --lora-scaled lora_task_B.gguf 0.5`.
-
-LoRA adapters should be in GGUF format. To convert from Hugging Face format use the `convert-lora-to-gguf.py` script. LoRA adapters are loaded separately and applied during inference - they are not merged with the main model. This means that mmap model loading is fully supported when using LoRA adapters. The old `--lora-base` flag has been removed now that merging is no longer performed.
-
 ## Additional Options
 
 These options provide extra functionality and customization when running the LLaMA models:
@@ -350,4 +325,6 @@ These options provide extra functionality and customization when running the LLa
 -   `--verbose-prompt`: Print the prompt before generating text.
 -   `-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.
 -   `-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.
+-   `--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.
 -   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.

examples/server/README.md

@@ -99,7 +99,7 @@ The project is under active development, and we are [looking for feedback and co
 
 | Argument | Explanation |
 | -------- | ----------- |
-| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;typ_p;top_p;min_p;temperature) |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) |
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
@@ -108,19 +108,15 @@ The project is under active development, and we are [looking for feedback and co
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
 | `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
 | `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
+| `--tfs N` | tail free sampling, parameter z (default: 1.0, 1.0 = disabled) |
 | `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
 | `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dry-multiplier N` | DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
-| `--dry-base N` | DRY sampling base value (default: 1.75) |
-| `--dry-allowed-length N` | allowed length for DRY sampling (default: 2) |
-| `--dry-penalty-last-n N` | DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
-| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers (`['\n', ':', '"', '*']`) in the process; use `"none"` to not use any sequence breakers
 | `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
 | `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
-| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
+| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
 | `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
 | `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
 | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
@@ -359,6 +355,8 @@ node index.js
     `stop`: Specify a JSON array of stopping strings.
     These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`
 
+    `tfs_z`: Enable tail free sampling with parameter z. Default: `1.0`, which is disabled.
+
     `typical_p`: Enable locally typical sampling with parameter p. Default: `1.0`, which is disabled.
 
     `repeat_penalty`: Control the repetition of token sequences in the generated text. Default: `1.1`
@@ -371,16 +369,6 @@ node index.js
 
     `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
 
-    `dry_multiplier`: Set the DRY (Don't Repeat Yourself) repetition penalty multiplier. Default: `0.0`, which is disabled.
-
-    `dry_base`: Set the DRY repetition penalty base value. Default: `1.75`
-
-    `dry_allowed_length`: Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`
-
-    `dry_penalty_last_n`: How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.
-
-    `dry_sequence_breakers`: Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\n', ':', '"', '*']`
-
     `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
 
     `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
@@ -409,7 +397,7 @@ node index.js
 
     `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
 
-    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
 **Response format**
 
@@ -735,6 +723,7 @@ Example:
         "repeat_penalty": 1.100000023841858,
         "samplers": [
             "top_k",
+            "tfs_z",
             "typical_p",
             "top_p",
             "min_p",
@@ -748,6 +737,7 @@ Example:
         "stream": false,
         "task_id": 0,
         "temperature": 0.0,
+        "tfs_z": 1.0,
         "top_k": 40,
         "top_p": 0.949999988079071,
         "typical_p": 1.0

examples/server/public/index-new.html

@@ -40,15 +40,12 @@
       repeat_last_n: 0, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.0, // 1.0 = disabled
       penalize_nl: false, // true only useful for infinite completion
-      dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
-      dry_base: 1.75,     // 0.0 = disabled
-      dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
-      dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
       top_k: 0, // <= 0 to use vocab size
       top_p: 1.0, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled; recommended for non-english: ~ 0.4
       xtc_probability: 0.0, // 0 = disabled;
       xtc_threshold: 0.1, // > 0.5 disables XTC;
+      tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -836,16 +833,13 @@ return html`
         <fieldset class="params">
           ${IntField({ label: "Top-K", title: "Limits the selection of the next token to the K most probable tokens. 1 means no randomness = greedy sampling. If set to 0, it means the entire vocabulary size is considered.", max: 100, min: 0, step: 1, name: "top_k", value: params.value.top_k })}
           ${IntField({ label: "Penalize Last N", title: "The last n tokens that are taken into account to penalise repetitions. A value of 0 means that this function is deactivated and -1 means that the entire size of the context is taken into account.", max: 2048, min: 0, step: 16, name: "repeat_last_n", value: params.value.repeat_last_n })}
-          ${FloatField({ label: "Presence Penalty", title: "A penalty that is applied if certain tokens appear repeatedly in the generated text. A higher value leads to fewer repetitions.", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
-          ${FloatField({ label: "Frequency Penalty", title: "A penalty that is applied based on the frequency with which certain tokens occur in the training data set. A higher value results in rare tokens being favoured.", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
           ${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
+          ${FloatField({ label: "Presence Penalty", title: "A penalty that is applied if certain tokens appear repeatedly in the generated text. A higher value leads to fewer repetitions.", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
+          ${FloatField({ label: "TFS-Z", title: "Activates tail-free sampling, a method used to limit the prediction of tokens that are too frequent. The parameter z controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
+          ${FloatField({ label: "Frequency Penalty", title: "A penalty that is applied based on the frequency with which certain tokens occur in the training data set. A higher value results in rare tokens being favoured.", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
           ${FloatField({ label: "Typical-P", title: "Activates local typical sampling, a method used to limit the prediction of tokens that are atypical in the current context. The parameter p controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
           ${FloatField({ label: "XTC probability", title: "Sets the chance for token removal (checked once on sampler start)", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
           ${FloatField({ label: "XTC threshold", title: "Sets a minimum probability threshold for tokens to be removed", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
-          ${FloatField({ label: "DRY Penalty Multiplier", title: "Set the DRY repetition penalty multiplier. Default is 0.0, which disables DRY.", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
-          ${FloatField({ label: "DRY Base", title: "Set the DRY repetition penalty base value. Default is 1.75", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
-          ${IntField({ label: "DRY Allowed Length", title: "Tokens that extend repetition beyond this receive exponentially increasing penalty. Default is 2", max: 10, min: 1, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
-          ${IntField({ label: "DRY Penalty Last N", title: "How many tokens to scan for repetitions. Default is -1, where 0 is disabled and -1 is context size", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
           ${IntField({ label: "Min Keep", title: "If greater than 0, samplers are forced to return N possible tokens at minimum. Default is 0", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
         </fieldset>
 
@@ -1145,12 +1139,11 @@ document.addEventListener('DOMContentLoaded', (event) => {
     xtc_probability: { snapValue: 0.0, snapRangeMultiplier: 4 },
     xtc_threshold: { snapValue: 0.5, snapRangeMultiplier: 4 },
     top_p: { snapValue: 1.0, snapRangeMultiplier: 4 },
+    tfs_z: { snapValue: 1.0, snapRangeMultiplier: 4 },
     typical_p: { snapValue: 1.0, snapRangeMultiplier: 4 },
     repeat_penalty: { snapValue: 1.0, snapRangeMultiplier: 4 },
     presence_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
     frequency_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
-    dry_multiplier: { snapValue: 0.0, snapRangeMultiplier: 4 },
-    dry_base: { snapValue: 1.75, snapRangeMultiplier: 4 },
   };
   // add an event listener for each slider
   Object.keys(snapSettings).forEach(sliderName => {

examples/server/public/index.html

@@ -304,15 +304,12 @@
       repeat_last_n: 256, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.18, // 1.0 = disabled
       penalize_nl: false,
-      dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
-      dry_base: 1.75,     // 0.0 = disabled
-      dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
-      dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
       xtc_probability: 0.0, // 0 = disabled;
       xtc_threshold: 0.1, // > 0.5 disables XTC;
+      tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -1014,13 +1011,10 @@
           <details>
             <summary>More options</summary>
             <fieldset class="two">
+              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
               ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
               ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
               ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
-              ${FloatField({ label: "DRY Penalty Multiplier", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
-              ${FloatField({ label: "DRY Base", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
-              ${IntField({ label: "DRY Allowed Length", max: 10, min: 2, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
-              ${IntField({ label: "DRY Penalty Last N", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
               ${FloatField({ label: "XTC probability", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
               ${FloatField({ label: "XTC threshold", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
             </fieldset>

examples/server/server.cpp

@@ -800,57 +800,35 @@ struct server_context {
             slot.oaicompat_model = "";
         }
 
-        slot.params.stream              = json_value(data, "stream",             false);
-        slot.params.cache_prompt        = json_value(data, "cache_prompt",       false);
-        slot.params.n_predict           = json_value(data, "n_predict",          json_value(data, "max_tokens", default_params.n_predict));
-        slot.params.n_indent            = json_value(data, "n_indent",           default_params.n_indent);
-        slot.sparams.top_k              = json_value(data, "top_k",              default_sparams.top_k);
-        slot.sparams.top_p              = json_value(data, "top_p",              default_sparams.top_p);
-        slot.sparams.min_p              = json_value(data, "min_p",              default_sparams.min_p);
-        slot.sparams.xtc_probability    = json_value(data, "xtc_probability",    default_sparams.xtc_probability);
-        slot.sparams.xtc_threshold      = json_value(data, "xtc_threshold",      default_sparams.xtc_threshold);
-        slot.sparams.typ_p              = json_value(data, "typical_p",          default_sparams.typ_p);
-        slot.sparams.temp               = json_value(data, "temperature",        default_sparams.temp);
-        slot.sparams.dynatemp_range     = json_value(data, "dynatemp_range",     default_sparams.dynatemp_range);
-        slot.sparams.dynatemp_exponent  = json_value(data, "dynatemp_exponent",  default_sparams.dynatemp_exponent);
-        slot.sparams.penalty_last_n     = json_value(data, "repeat_last_n",      default_sparams.penalty_last_n);
-        slot.sparams.penalty_repeat     = json_value(data, "repeat_penalty",     default_sparams.penalty_repeat);
-        slot.sparams.penalty_freq       = json_value(data, "frequency_penalty",  default_sparams.penalty_freq);
-        slot.sparams.penalty_present    = json_value(data, "presence_penalty",   default_sparams.penalty_present);
-        slot.sparams.dry_multiplier     = json_value(data, "dry_multiplier",     default_sparams.dry_multiplier);
-        slot.sparams.dry_base           = json_value(data, "dry_base",           default_sparams.dry_base);
-        slot.sparams.dry_allowed_length = json_value(data, "dry_allowed_length", default_sparams.dry_allowed_length);
-        slot.sparams.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", default_sparams.dry_penalty_last_n);
-        slot.sparams.mirostat           = json_value(data, "mirostat",           default_sparams.mirostat);
-        slot.sparams.mirostat_tau       = json_value(data, "mirostat_tau",       default_sparams.mirostat_tau);
-        slot.sparams.mirostat_eta       = json_value(data, "mirostat_eta",       default_sparams.mirostat_eta);
-        slot.sparams.penalize_nl        = json_value(data, "penalize_nl",        default_sparams.penalize_nl);
-        slot.params.n_keep              = json_value(data, "n_keep",             default_params.n_keep);
-        slot.params.n_discard           = json_value(data, "n_discard",          default_params.n_discard);
-        slot.sparams.seed               = json_value(data, "seed",               default_sparams.seed);
-        slot.sparams.n_probs            = json_value(data, "n_probs",            default_sparams.n_probs);
-        slot.sparams.min_keep           = json_value(data, "min_keep",           default_sparams.min_keep);
-      //slot.params.t_max_prompt_ms     = json_value(data, "t_max_prompt_ms",    default_params.t_max_prompt_ms); // TODO: implement
-        slot.params.t_max_predict_ms    = json_value(data, "t_max_predict_ms",   default_params.t_max_predict_ms);
-
-        if (slot.sparams.dry_base < 1.0f)
-        {
-           slot.sparams.dry_base = default_sparams.dry_base;
-        }
-
-        // sequence breakers for DRY
-        {
-            // Currently, this is not compatible with TextGen WebUI, Koboldcpp and SillyTavern format
-            // Ref: https://github.com/oobabooga/text-generation-webui/blob/d1af7a41ade7bd3c3a463bfa640725edb818ebaf/extensions/openai/typing.py#L39
-
-            if (data.contains("dry_sequence_breakers")) {
-                slot.sparams.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector<std::string>());
-                if (slot.sparams.dry_sequence_breakers.empty()) {
-                    send_error(task, "Error: dry_sequence_breakers must be a non-empty array of strings", ERROR_TYPE_INVALID_REQUEST);
-                    return false;
-                }
-            }
-        }
+        slot.params.stream             = json_value(data, "stream",            false);
+        slot.params.cache_prompt       = json_value(data, "cache_prompt",      false);
+        slot.params.n_predict          = json_value(data, "n_predict",         json_value(data, "max_tokens", default_params.n_predict));
+        slot.params.n_indent           = json_value(data, "n_indent",          default_params.n_indent);
+        slot.sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
+        slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
+        slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
+        slot.sparams.xtc_probability   = json_value(data, "xtc_probability",   default_sparams.xtc_probability);
+        slot.sparams.xtc_threshold     = json_value(data, "xtc_threshold",     default_sparams.xtc_threshold);
+        slot.sparams.tfs_z             = json_value(data, "tfs_z",             default_sparams.tfs_z);
+        slot.sparams.typ_p             = json_value(data, "typical_p",         default_sparams.typ_p);
+        slot.sparams.temp              = json_value(data, "temperature",       default_sparams.temp);
+        slot.sparams.dynatemp_range    = json_value(data, "dynatemp_range",    default_sparams.dynatemp_range);
+        slot.sparams.dynatemp_exponent = json_value(data, "dynatemp_exponent", default_sparams.dynatemp_exponent);
+        slot.sparams.penalty_last_n    = json_value(data, "repeat_last_n",     default_sparams.penalty_last_n);
+        slot.sparams.penalty_repeat    = json_value(data, "repeat_penalty",    default_sparams.penalty_repeat);
+        slot.sparams.penalty_freq      = json_value(data, "frequency_penalty", default_sparams.penalty_freq);
+        slot.sparams.penalty_present   = json_value(data, "presence_penalty",  default_sparams.penalty_present);
+        slot.sparams.mirostat          = json_value(data, "mirostat",          default_sparams.mirostat);
+        slot.sparams.mirostat_tau      = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
+        slot.sparams.mirostat_eta      = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
+        slot.sparams.penalize_nl       = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
+        slot.params.n_keep             = json_value(data, "n_keep",            default_params.n_keep);
+        slot.params.n_discard          = json_value(data, "n_discard",         default_params.n_discard);
+        slot.sparams.seed              = json_value(data, "seed",              default_sparams.seed);
+        slot.sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
+        slot.sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
+      //slot.params.t_max_prompt_ms    = json_value(data, "t_max_prompt_ms",   default_params.t_max_prompt_ms); // TODO: implement
+        slot.params.t_max_predict_ms   = json_value(data, "t_max_predict_ms",  default_params.t_max_predict_ms);
 
         // process "json_schema" and "grammar"
         if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) {
@@ -1148,16 +1126,12 @@ struct server_context {
             {"min_p",                     slot.sparams.min_p},
             {"xtc_probability",           slot.sparams.xtc_probability},
             {"xtc_threshold",             slot.sparams.xtc_threshold},
+            {"tfs_z",                     slot.sparams.tfs_z},
             {"typical_p",                 slot.sparams.typ_p},
             {"repeat_last_n",             slot.sparams.penalty_last_n},
             {"repeat_penalty",            slot.sparams.penalty_repeat},
             {"presence_penalty",          slot.sparams.penalty_present},
             {"frequency_penalty",         slot.sparams.penalty_freq},
-            {"dry_multiplier",            slot.sparams.dry_multiplier},
-            {"dry_base",                  slot.sparams.dry_base},
-            {"dry_allowed_length",        slot.sparams.dry_allowed_length},
-            {"dry_penalty_last_n",        slot.sparams.dry_penalty_last_n},
-            {"dry_sequence_breakers",     slot.sparams.dry_sequence_breakers},
             {"mirostat",                  slot.sparams.mirostat},
             {"mirostat_tau",              slot.sparams.mirostat_tau},
             {"mirostat_eta",              slot.sparams.mirostat_eta},
@@ -1878,7 +1852,6 @@ struct server_context {
                     if (slot.state == SLOT_STATE_STARTED) {
                         slot.t_start_process_prompt = ggml_time_us();
                         slot.t_start_generation = 0;
-
                         slot.n_past = 0;
                         slot.n_prompt_tokens = prompt_tokens.size();
                         slot.state = SLOT_STATE_PROCESSING_PROMPT;
@@ -1909,17 +1882,12 @@ struct server_context {
                         }
 
                         if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) {
+                            // this prompt is too large to process - discard it
                             if (slot.n_prompt_tokens > n_ubatch) {
                                 slot.release();
                                 send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER);
                                 continue;
                             }
-
-                            if (slot.n_prompt_tokens > slot.n_ctx) {
-                                slot.release();
-                                send_error(slot, "input is larger than the max context size. skipping", ERROR_TYPE_SERVER);
-                                continue;
-                            }
                         } else {
                             if (!params.ctx_shift) {
                                 // if context shift is disabled, we make sure prompt size is smaller than KV size
@@ -1997,6 +1965,7 @@ struct server_context {
 
                                             for (size_t i = 0; i < n_match; i++) {
                                                 slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
+
                                                 slot.n_past++;
                                             }
 
@@ -2406,7 +2375,7 @@ int main(int argc, char ** argv) {
     auto middleware_server_state = [&res_error, &state](const httplib::Request & req, httplib::Response & res) {
         server_state current_state = state.load();
         if (current_state == SERVER_STATE_LOADING_MODEL) {
-            auto tmp = string_split<std::string>(req.path, '.');
+            auto tmp = string_split(req.path, '.');
             if (req.path == "/" || tmp.back() == "html") {
                 res.set_content(reinterpret_cast<const char*>(loading_html), loading_html_len, "text/html; charset=utf-8");
                 res.status = 503;
@@ -3259,7 +3228,7 @@ int main(int argc, char ** argv) {
         ctx_server.queue_tasks.terminate();
     };
 
-    LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port);
+    LOG_INF("%s: server is listening on %s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port);
 
     ctx_server.queue_tasks.start_loop();
 

examples/server/themes/buttons-top/index.html

@@ -226,6 +226,7 @@
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
+      tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -787,6 +788,7 @@
           <details>
             <summary>More options</summary>
             <fieldset class="two">
+              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
               ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
               ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
               ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}

examples/server/themes/wild/index.html

@@ -229,6 +229,7 @@
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
+      tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -790,6 +791,7 @@
           <details>
             <summary>More options</summary>
             <fieldset class="two">
+              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
               ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
               ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
               ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}

examples/server/utils.hpp

@@ -266,10 +266,8 @@ static llama_tokens format_infill(
     }
 
     // for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
-    const int n_prefix_take = std::min<int>(tokens_prefix.size(),                3*(n_batch/4));
-    const int n_suffix_take = std::min<int>(tokens_suffix.size(), std::max<int>(0, (n_batch/4) - (2 + tokens_prompt.size())));
-
-    SRV_DBG("n_prefix_take = %d, n_suffix_take = %d, total = %d\n", n_prefix_take, n_suffix_take, (n_prefix_take + n_suffix_take));
+    const int n_suffix_take = std::min<int>(tokens_suffix.size(),   (n_batch/4));
+    const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4) - 3);
 
     // fill the rest of the context with extra chunks
     const int n_extra_take = std::min<int>(std::max<int>(0, n_ctx - (n_batch) - 2*n_predict), extra_tokens.size());
@@ -607,7 +605,7 @@ static json oaicompat_completion_params_parse(
     }
 
     // Copy remaining properties to llama_params
-    // This allows user to use llama.cpp-specific params like "mirostat", ... via OAI endpoint.
+    // This allows user to use llama.cpp-specific params like "mirostat", "tfs_z",... via OAI endpoint.
     // See "launch_slot_with_task()" for a complete list of params supported by llama.cpp
     for (const auto & item : body.items()) {
         // Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens"

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1729665710,
-        "narHash": "sha256-AlcmCXJZPIlO5dmFzV3V2XF6x/OpNWUV8Y/FMPGd8Z4=",
+        "lastModified": 1729256560,
+        "narHash": "sha256-/uilDXvCIEs3C9l73JTACm4quuHUsIHcns1c+cHUJwA=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "2768c7d042a37de65bb1b5b3268fc987e534c49d",
+        "rev": "4c2fcb090b1f3e5b47eaa7bd33913b574a11e0a0",
         "type": "github"
       },
       "original": {

ggml/include/ggml-backend.h

@@ -114,12 +114,11 @@ extern "C" {
     //
 
     enum ggml_backend_dev_type {
-        // CPU device using system memory
         GGML_BACKEND_DEVICE_TYPE_CPU,
-        // GPU device using dedicated memory
         GGML_BACKEND_DEVICE_TYPE_GPU,
-        // accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
-        GGML_BACKEND_DEVICE_TYPE_ACCEL
+        // devices with full capabilities (excludes backends such as BLAS that only support matrix multiplication)
+        GGML_BACKEND_DEVICE_TYPE_CPU_FULL,
+        GGML_BACKEND_DEVICE_TYPE_GPU_FULL
     };
 
     // functionality supported by the device
@@ -168,14 +167,10 @@ extern "C" {
     GGML_API ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index);
     GGML_API void *             ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name);
 
-    // Common functions that may be obtained using ggml_backend_reg_get_proc_address
 
-    // Split buffer type for tensor parallelism
-    typedef ggml_backend_buffer_type_t   (*ggml_backend_split_buffer_type_t)(int main_device, const float * tensor_split);
-    // Set the number of threads for the backend
-    typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
-    // Get additional buffer types provided by the device (returns a NULL-terminated array)
-    typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device);
+    // Functions that may be obtained using ggml_backend_reg_get_proc_address
+    typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(const float *);
+    typedef void (*ggml_backend_set_n_threads_t)(ggml_backend_t, int);
 
     //
     // Backend registry
@@ -197,7 +192,7 @@ extern "C" {
     GGML_API ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params);
     // = ggml_backend_dev_init(ggml_backend_dev_by_type(type), params)
     GGML_API ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params);
-    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU) OR ggml_backend_dev_by_type(CPU), NULL)
+    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU_FULL) OR ggml_backend_dev_by_type(CPU_FULL), NULL)
     GGML_API ggml_backend_t ggml_backend_init_best(void);
 
     //

ggml/include/ggml-cuda.h

@@ -28,7 +28,7 @@ GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);

ggml/include/ggml-kompute.h

@@ -11,8 +11,6 @@
 extern "C" {
 #endif
 
-#define GGML_KOMPUTE_MAX_DEVICES 16
-
 struct ggml_vk_device {
     int index;
     int type; // same as VkPhysicalDeviceType
@@ -43,8 +41,6 @@ GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
 
 GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
 
-GGML_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
-
 #ifdef __cplusplus
 }
 #endif

ggml/src/CMakeLists.txt

@@ -800,7 +800,6 @@ if (GGML_KOMPUTE)
             kompute-shaders/op_mul_mat_q8_0.comp
             kompute-shaders/op_mul_mat_q4_0.comp
             kompute-shaders/op_mul_mat_q4_1.comp
-            kompute-shaders/op_mul_mat_q4_k.comp
             kompute-shaders/op_mul_mat_q6_k.comp
             kompute-shaders/op_getrows_f32.comp
             kompute-shaders/op_getrows_f16.comp
@@ -834,7 +833,6 @@ if (GGML_KOMPUTE)
             shaderop_mul_mat_q8_0.h
             shaderop_mul_mat_q4_0.h
             shaderop_mul_mat_q4_1.h
-            shaderop_mul_mat_q4_k.h
             shaderop_mul_mat_q6_k.h
             shaderop_getrows_f32.h
             shaderop_getrows_f16.h

ggml/src/ggml-aarch64.c

@@ -991,73 +991,6 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         }
     }
     return;
-#elif defined(__riscv_v_intrinsic)
-    if (__riscv_vlenb() >= QK4_0) {
-        const size_t vl = QK4_0;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-            for (int l = 0; l < nb; l++) {
-                const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
-                const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
-                const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
-                const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
-                __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
-                const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
-                const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
-                const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
-
-                const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
-                const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
-                const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
-                const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
-                const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
-                const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
-                const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
-
-                const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
-                const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                // vector version needs Zvfhmin extension
-                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
-                const float b_scales[8] = {
-                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
-                };
-                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
-                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
-                sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
-            }
-            __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
-        }
-        return;
-    }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
     {
         float sumf[8];
@@ -3238,207 +3171,6 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
                 }
             }
         }
-        return;
-    }
-#elif defined(__riscv_v_intrinsic)
-    if (__riscv_vlenb() >= QK4_0) {
-        const size_t vl = QK4_0;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-                vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                for (int l = 0; l < nb; l++) {
-                    const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
-                    const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
-                    const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
-                    const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
-                    const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
-                    const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
-                    const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
-
-                    // vector version needs Zvfhmin extension
-                    const float a_scales[4] = {
-                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
-                    };
-                    const float b_scales[8] = {
-                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
-                    };
-                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
-
-                    const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
-                    const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
-                    const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
-                    const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l0;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l0 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
-                        sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
-                    const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
-                    const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
-                    const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l1;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l1 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
-                        sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
-                    const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
-                    const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
-                    const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l2;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l2 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
-                        sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
-                    const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
-                    const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
-                    const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l3;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l3 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
-                        sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
-                    }
-                }
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
-            }
-        }
-
         return;
     }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)

ggml/src/ggml-amx.cpp

@@ -16,6 +16,12 @@
 #if defined(__AMX_INT8__)
 
 // AMX buffer interface
+static const char * ggml_backend_amx_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "AMX";
+
+    GGML_UNUSED(buffer);
+}
+
 static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     free(buffer->context);
 }
@@ -66,6 +72,7 @@ static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
 }
 
 static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
+    /* .get_name        = */ ggml_backend_amx_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_amx_buffer_get_base,
     /* .init_tensor     = */ NULL, // no initialization required
@@ -114,14 +121,14 @@ static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft
 ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
     static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
         /* .iface = */ {
-            /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
-            /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
+        /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
+        /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
+        /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
+        /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+        /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
+        /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
         },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
+        /* .device  = */ NULL,
         /* .context = */ NULL,
     };
 
@@ -142,6 +149,12 @@ static void ggml_backend_amx_free(ggml_backend_t backend) {
     delete backend;
 }
 
+static ggml_backend_buffer_type_t ggml_backend_amx_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_amx_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
 static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
 
@@ -174,6 +187,7 @@ static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, s
 static struct ggml_backend_i ggml_backend_amx_i = {
     /* .get_name                = */ ggml_backend_amx_name,
     /* .free                    = */ ggml_backend_amx_free,
+    /* .get_default_buffer_type = */ ggml_backend_amx_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
@@ -183,6 +197,9 @@ static struct ggml_backend_i ggml_backend_amx_i = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_amx_graph_compute,
+    /* .supports_op             = */ NULL,
+    /* .supports_buft           = */ NULL,
+    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -262,7 +279,7 @@ static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t *
 }
 
 static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+    return GGML_BACKEND_DEVICE_TYPE_CPU;
 
     GGML_UNUSED(dev);
 }

ggml/src/ggml-backend-impl.h

@@ -22,7 +22,7 @@ extern "C" {
         size_t                (*get_max_size)  (ggml_backend_buffer_type_t buft);
         // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
         size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
-        // (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
+        // (optional) check if tensor data is in host memory (defaults to false)
         bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
     };
 
@@ -37,6 +37,7 @@ extern "C" {
     //
 
     struct ggml_backend_buffer_i {
+        const char * (*get_name)     (ggml_backend_buffer_t buffer);
         // (optional) free the buffer
         void         (*free_buffer)  (ggml_backend_buffer_t buffer);
         // base address of the buffer
@@ -87,16 +88,19 @@ extern "C" {
 
         void (*free)(ggml_backend_t backend);
 
+        // Will be moved to the device interface
+        // buffer allocation
+        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
+
         // (optional) asynchronous tensor data access
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
         bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
-        // (optional) complete all pending operations (required if the backend supports async operations)
+        // (optional) complete all pending operations
         void (*synchronize)(ggml_backend_t backend);
 
-        // (optional) graph plans (not used currently)
-        // compute graph with a plan
+        // (optional) compute graph with a plan (not used currently)
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
@@ -107,6 +111,13 @@ extern "C" {
         // compute graph (always async if supported by the backend)
         enum ggml_status          (*graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
+        // IMPORTANT: these functions have been moved to the device interface and will be removed from the backend interface
+        //            new backends should implement the device interface instead
+        // These functions are being moved to the device interface
+        bool (*supports_op)  (ggml_backend_t backend, const struct ggml_tensor * op);
+        bool (*supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
+        bool (*offload_op)   (ggml_backend_t backend, const struct ggml_tensor * op);
+
         // (optional) event synchronization
         // record an event on this stream
         void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);

ggml/src/ggml-backend.cpp

@@ -34,11 +34,6 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
 }
 
 ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    if (size == 0) {
-        // return a dummy buffer for zero-sized allocations
-        return ggml_backend_buffer_init(buft, {}, NULL, 0);
-    }
-
     return buft->iface.alloc_buffer(buft, size);
 }
 
@@ -94,7 +89,7 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
 }
 
 const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_name(ggml_backend_buffer_get_type(buffer));
+    return buffer->iface.get_name(buffer);
 }
 
 void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
@@ -113,11 +108,6 @@ size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
 }
 
 void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
-    // get_base is optional if the buffer is zero-sized
-    if (buffer->size == 0) {
-        return NULL;
-    }
-
     void * base = buffer->iface.get_base(buffer);
 
     GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
@@ -132,15 +122,6 @@ void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_t
     }
 }
 
-void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    // clear is optional if the buffer is zero-sized
-    if (buffer->size == 0) {
-        return;
-    }
-
-    buffer->iface.clear(buffer, value);
-}
-
 size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
     return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
 }
@@ -153,6 +134,10 @@ size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct g
     return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
 }
 
+void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    buffer->iface.clear(buffer, value);
+}
+
 bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
     return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer));
 }
@@ -213,7 +198,7 @@ void ggml_backend_free(ggml_backend_t backend) {
 }
 
 ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
-    return ggml_backend_dev_buffer_type(backend->device);
+    return backend->iface.get_default_buffer_type(backend);
 }
 
 ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
@@ -253,42 +238,43 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
 void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    if (size == 0) {
-        return;
-    }
-
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
+    if (!size) {
+        return;
+    }
+
     buf->iface.set_tensor(buf, tensor, data, offset, size);
 }
 
 void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    if (size == 0) {
-        return;
-    }
-
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
+    if (!size) {
+        return;
+    }
+
     buf->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
 GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    if (size == 0) {
-        return;
-    }
-
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not implemented by backend buffer");
+
+    if (!size) {
+        return;
+    }
+
+    GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not supported by backend buffer");
 
     buf->iface.memset_tensor(buf, tensor, value, offset, size);
 }
@@ -330,15 +316,32 @@ enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct
 }
 
 bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return ggml_backend_dev_supports_op(backend->device, op);
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_supports_op(backend->device, op);
+    }
+
+    return backend->iface.supports_op(backend, op);
 }
 
 bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_dev_supports_buft(backend->device, buft);
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_supports_buft(backend->device, buft);
+    }
+    return backend->iface.supports_buft(backend, buft);
 }
 
 bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return ggml_backend_dev_offload_op(backend->device, op);
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_offload_op(backend->device, op);
+    }
+
+    if (backend->iface.offload_op != NULL) {
+        return backend->iface.offload_op(backend, op);
+    }
+    return false;
 }
 
 ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
@@ -562,10 +565,6 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
 #include "ggml-cann.h"
 #endif
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 struct ggml_backend_registry {
     std::vector<ggml_backend_reg_t> backends;
     std::vector<ggml_backend_dev_t> devices;
@@ -583,9 +582,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_VULKAN
         register_backend(ggml_backend_vk_reg());
 #endif
-#ifdef GGML_USE_CANN
-        register_backend(ggml_backend_cann_reg());
-#endif
 #ifdef GGML_USE_BLAS
         register_backend(ggml_backend_blas_reg());
 #endif
@@ -595,10 +591,12 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_AMX
         register_backend(ggml_backend_amx_reg());
 #endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
+#ifdef GGML_USE_CANN
+        register_backend(ggml_backend_cann_reg());
 #endif
 
+        // TODO: kompute
+
         register_backend(ggml_backend_cpu_reg());
     }
 
@@ -703,9 +701,9 @@ ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const
 }
 
 ggml_backend_t ggml_backend_init_best(void) {
-    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL);
     if (!dev) {
-        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU_FULL);
     }
     if (!dev) {
         return NULL;
@@ -713,7 +711,13 @@ ggml_backend_t ggml_backend_init_best(void) {
     return ggml_backend_dev_init(dev, NULL);
 }
 
-// CPU backend - buffer
+// backend CPU
+
+static const char * ggml_backend_cpu_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "CPU";
+
+    GGML_UNUSED(buffer);
+}
 
 static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
     uintptr_t data = (uintptr_t)buffer->context;
@@ -763,6 +767,7 @@ static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
 }
 
 static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+    /* .get_name        = */ ggml_backend_cpu_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
     /* .init_tensor     = */ NULL, // no initialization required
@@ -775,6 +780,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
 };
 
 static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
+    /* .get_name        = */ ggml_backend_cpu_buffer_get_name,
     /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
     /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
     /* .init_tensor     = */ NULL, // no initialization required
@@ -786,8 +792,6 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
     /* .reset           = */ NULL,
 };
 
-// CPU backend - buffer type
-
 static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     return "CPU";
 
@@ -795,14 +799,19 @@ static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_ty
 }
 
 static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * data = ggml_aligned_malloc(size);
+    auto alloc_size = size;
+    if (alloc_size == 0) {
+        alloc_size = 1;
+    }
+
+    void * data = ggml_aligned_malloc(alloc_size);
 
     if (data == NULL) {
-        GGML_LOG_ERROR("%s: failed to allocate buffer of size %zu\n", __func__, size);
+        GGML_LOG_ERROR("%s: failed to allocate buffer of size %zu\n", __func__, alloc_size);
         return NULL;
     }
 
-    return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, size);
+    return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, alloc_size);
 }
 
 static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@@ -834,29 +843,6 @@ ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
     return &ggml_backend_cpu_buffer_type;
 }
 
-static const char * ggml_backend_cpu_buffer_from_ptr_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "CPU_Mapped";
-
-    GGML_UNUSED(buft);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_cpu_buffer_from_ptr_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
-        /* .iface   = */ {
-            /* .get_name         = */ ggml_backend_cpu_buffer_from_ptr_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
-        },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_cpu_buffer_type;
-}
-
 #ifdef GGML_USE_CPU_HBM
 
 // buffer type HBM
@@ -869,11 +855,18 @@ static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffe
     GGML_UNUSED(buft);
 }
 
+static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buf);
+}
+
 static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     hbw_free(buffer->context);
 }
 
 static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    //void * ptr = hbw_malloc(size);
     void * ptr;
     int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
     if (result != 0) {
@@ -883,6 +876,7 @@ static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_
 
     ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
     buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_cpu_hbm_buffer_get_name;
     buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
 
     return buffer;
@@ -905,21 +899,6 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
 }
 #endif
 
-static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
-    static ggml_backend_buffer_type_t bufts[] = {
-#ifdef GGML_USE_CPU_HBM
-        ggml_backend_cpu_hbm_buffer_type(),
-#endif
-        NULL
-    };
-
-    return bufts;
-
-    GGML_UNUSED(device);
-}
-
-// CPU backend - backend (stream)
-
 struct ggml_backend_cpu_context {
     int                 n_threads;
     ggml_threadpool_t   threadpool;
@@ -944,6 +923,12 @@ static void ggml_backend_cpu_free(ggml_backend_t backend) {
     delete backend;
 }
 
+static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
 struct ggml_backend_plan_cpu {
     struct ggml_cplan cplan;
     struct ggml_cgraph cgraph;
@@ -1013,6 +998,7 @@ static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, s
 static const struct ggml_backend_i ggml_backend_cpu_i = {
     /* .get_name                = */ ggml_backend_cpu_get_name,
     /* .free                    = */ ggml_backend_cpu_free,
+    /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
@@ -1022,6 +1008,9 @@ static const struct ggml_backend_i ggml_backend_cpu_i = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
     /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op             = */ NULL,
+    /* .supports_buft           = */ NULL,
+    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -1092,10 +1081,10 @@ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_
 
 ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
     GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
-    return ggml_backend_buffer_init(ggml_backend_cpu_buffer_from_ptr_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
+    return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
 }
 
-// CPU backend - device
+////////////////////////
 
 struct ggml_backend_cpu_device_context {
     std::string description = "CPU";
@@ -1182,7 +1171,7 @@ static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t *
 }
 
 static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_CPU;
+    return GGML_BACKEND_DEVICE_TYPE_CPU_FULL;
 
     GGML_UNUSED(dev);
 }
@@ -1200,7 +1189,7 @@ static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggm
     };
 }
 
-static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+static ggml_backend_t ggml_backend_cpu_device_init(ggml_backend_dev_t dev, const char * params) {
     return ggml_backend_cpu_init();
 
     GGML_UNUSED(dev);
@@ -1213,7 +1202,7 @@ static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_b
     GGML_UNUSED(dev);
 }
 
-static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
     return ggml_backend_cpu_buffer_from_ptr(ptr, size);
 
     GGML_UNUSED(dev);
@@ -1255,10 +1244,10 @@ static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
     /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
     /* .get_type             = */ ggml_backend_cpu_device_get_type,
     /* .get_props            = */ ggml_backend_cpu_device_get_props,
-    /* .init_backend         = */ ggml_backend_cpu_device_init_backend,
+    /* .init_backend         = */ ggml_backend_cpu_device_init,
     /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
     /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
+    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_ptr,
     /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
     /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
     /* .offload_op           = */ NULL,
@@ -1267,7 +1256,7 @@ static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
     /* .event_synchronize    = */ NULL,
 };
 
-// CPU backend - backend (reg)
+////////////////////////
 
 static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
     return "CPU";
@@ -1298,10 +1287,6 @@ static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const ch
     if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
         return (void *)ggml_backend_cpu_set_n_threads;
     }
-    if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
-        return (void *)ggml_backend_cpu_get_extra_bufts;
-    }
-
     return NULL;
 
     GGML_UNUSED(reg);
@@ -1330,6 +1315,12 @@ struct ggml_backend_multi_buffer_context {
     size_t n_buffers;
 };
 
+static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+
+    return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
+}
+
 static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
     for (size_t i = 0; i < ctx->n_buffers; i++) {
@@ -1348,6 +1339,7 @@ static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_
 }
 
 static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = {
+    /* .get_name        = */ ggml_backend_multi_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
     /* .get_base        = */ NULL,
     /* .init_tensor     = */ NULL,
@@ -1376,7 +1368,7 @@ ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer
 }
 
 bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
-    return buffer->iface.free_buffer == ggml_backend_multi_buffer_free_buffer;
+    return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
 }
 
 void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
@@ -1468,7 +1460,7 @@ struct ggml_backend_sched {
     char * context_buffer;
     size_t context_buffer_size;
 
-    int debug;
+    bool debug;
 };
 
 #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
@@ -1508,7 +1500,7 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
     return -1;
 }
 
-#if 1
+#if 0
 #define GGML_SCHED_MAX_SPLITS_DEBUG 4096
 static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
 #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
@@ -1556,9 +1548,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         if (src == NULL) {
             continue;
         }
-        // skip ROPE since the rope freqs tensor is too small to choose a backend based on it
-        // not an ideal solution
-        if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+        if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
             if (src_backend_id == sched->n_backends - 1) {
@@ -1605,21 +1595,19 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         if (ggml_is_view_op(node->op)) {
             continue;
         }
-        if (sched->debug > 1) {
-            ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
-                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * src = node->src[j];
-                if (src == NULL) {
-                    continue;
-                }
-                ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
-                GGML_LOG_DEBUG(" %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
-                    fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
+        GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
+            fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
             }
-            GGML_LOG_DEBUG("\n");
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
+            GGML_LOG_DEBUG(" %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
+                fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
         }
+        GGML_LOG_DEBUG("\n");
     }
 }
 
@@ -1911,11 +1899,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     if (src == NULL) {
                         continue;
                     }
-                    // check if a weight is on a different and incompatible backend
+                    // check if a weight is on a different backend
                     // by starting a new split, the memory of the previously offloaded weights can be reused
                     if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
                         int src_backend_id = tensor_backend_id(src);
-                        if (src_backend_id != cur_backend_id && !ggml_backend_sched_buffer_supported(sched, src, cur_backend_id)) {
+                        if (src_backend_id != cur_backend_id) {
                             need_new_split = true;
                             break;
                         }
@@ -1927,6 +1915,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                         int src_backend_id = sched->hv_tensor_backend_ids[id];
                         bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
                         if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
+                            //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
                             need_new_split = true;
                             break;
                         }
@@ -2251,8 +2240,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
 
     struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
 
-    const char * GGML_SCHED_DEBUG = getenv("GGML_SCHED_DEBUG");
-    sched->debug = GGML_SCHED_DEBUG ? atoi(GGML_SCHED_DEBUG) : 0;
+    sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
     sched->n_backends = n_backends;
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 

ggml/src/ggml-blas.cpp

@@ -224,6 +224,12 @@ static void ggml_backend_blas_free(ggml_backend_t backend) {
     delete backend;
 }
 
+static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
 static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
 
@@ -259,6 +265,7 @@ static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend,
 static struct ggml_backend_i blas_backend_i = {
     /* .get_name                = */ ggml_backend_blas_get_name,
     /* .free                    = */ ggml_backend_blas_free,
+    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
@@ -268,6 +275,9 @@ static struct ggml_backend_i blas_backend_i = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_blas_graph_compute,
+    /* .supports_op             = */ NULL,
+    /* .supports_buft           = */ NULL,
+    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -346,7 +356,7 @@ static void ggml_backend_blas_device_get_memory(ggml_backend_dev_t dev, size_t *
 }
 
 static enum ggml_backend_dev_type ggml_backend_blas_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+    return GGML_BACKEND_DEVICE_TYPE_CPU;
 
     GGML_UNUSED(dev);
 }
@@ -364,7 +374,7 @@ static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct gg
     };
 }
 
-static ggml_backend_t ggml_backend_blas_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+static ggml_backend_t ggml_backend_blas_device_init(ggml_backend_dev_t dev, const char * params) {
     return ggml_backend_blas_init();
 
     GGML_UNUSED(dev);
@@ -377,7 +387,7 @@ static ggml_backend_buffer_type_t ggml_backend_blas_device_get_buffer_type(ggml_
     GGML_UNUSED(dev);
 }
 
-static ggml_backend_buffer_t ggml_backend_blas_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+static ggml_backend_buffer_t ggml_backend_blas_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
     return ggml_backend_cpu_buffer_from_ptr(ptr, size);
 
     GGML_UNUSED(dev);
@@ -446,10 +456,10 @@ static const struct ggml_backend_device_i ggml_backend_blas_device_i = {
     /* .get_memory           = */ ggml_backend_blas_device_get_memory,
     /* .get_type             = */ ggml_backend_blas_device_get_type,
     /* .get_props            = */ ggml_backend_blas_device_get_props,
-    /* .init_backend         = */ ggml_backend_blas_device_init_backend,
+    /* .init_backend         = */ ggml_backend_blas_device_init,
     /* .get_buffer_type      = */ ggml_backend_blas_device_get_buffer_type,
     /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ ggml_backend_blas_device_buffer_from_host_ptr,
+    /* .buffer_from_host_ptr = */ ggml_backend_blas_device_buffer_from_ptr,
     /* .supports_op          = */ ggml_backend_blas_device_supports_op,
     /* .supports_buft        = */ ggml_backend_blas_device_supports_buft,
     /* .offload_op           = */ NULL,

ggml/src/ggml-cann.cpp

@@ -489,6 +489,23 @@ struct ggml_backend_cann_buffer_context {
     ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
 };
 
+/**
+ * @brief Retrieve the name associated with a CANN buffer.
+ *
+ * This function returns the name of a CANN buffer, which is stored in the
+ * context of the buffer.
+ *
+ * @param buffer The CANN buffer whose name is to be retrieved.
+ * @return A pointer to a C-string containing the name of the buffer.
+ */
+
+static const char* ggml_backend_cann_buffer_get_name(
+    ggml_backend_buffer_t buffer) {
+    return "CANN";
+
+    GGML_UNUSED(buffer);
+}
+
 /**
  * @brief Check if a buffer is a CANN buffer.
  *
@@ -498,10 +515,9 @@ struct ggml_backend_cann_buffer_context {
  * @param buffer The buffer to check.
  * @return true if the buffer is a CANN buffer, false otherwise.
  */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
 static bool ggml_backend_buffer_is_cann(
     ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_is_cann(buffer->buft);
+    return buffer->iface.get_name == ggml_backend_cann_buffer_get_name;
 }
 
 /**
@@ -949,6 +965,7 @@ static void ggml_backend_cann_buffer_clear(
  * on a CANN buffer within the backend.
  */
 static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
+    /* .get_name        = */ ggml_backend_cann_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cann_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
@@ -982,10 +999,9 @@ struct ggml_backend_cann_buffer_type_context {
  */
 static const char* ggml_backend_cann_buffer_type_name(
     ggml_backend_buffer_type_t buft) {
-    ggml_backend_cann_buffer_type_context* buft_ctx =
-        (ggml_backend_cann_buffer_type_context*)buft->context;
+    return "CANN";
 
-    return buft_ctx->name.c_str();
+    GGML_UNUSED(buft);
 }
 
 /**
@@ -1449,6 +1465,24 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
     delete backend;
 }
 
+/**
+ * @brief Retrieves the default buffer type associated with the CANN backend.
+ *
+ * This function returns the buffer type specific to the device associated
+ * with the CANN backend. It is used to allocate buffers for computations
+ * performed by the backend.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @return Pointer to the buffer type structure for the CANN backend.
+ */
+static ggml_backend_buffer_type_t
+ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    return ggml_backend_cann_buffer_type(cann_ctx->device);
+}
+
 /**
  * @brief Sets tensor data asynchronously in the CANN backend.
  *
@@ -1829,6 +1863,7 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
 static const ggml_backend_i ggml_backend_cann_interface = {
     /* .get_name                = */ ggml_backend_cann_name,
     /* .free                    = */ ggml_backend_cann_free,
+    /* .get_default_buffer_type = */ ggml_backend_cann_get_default_buffer_type,
     /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
     /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
@@ -1838,6 +1873,9 @@ static const ggml_backend_i ggml_backend_cann_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_cann_graph_compute,
+    /* .supports_op             = */ NULL, // moved to device
+    /* .supports_buft           = */ NULL, // moved to device
+    /* .offload_op              = */ NULL, // moved to device
     /* .event_record            = */ ggml_backend_cann_event_record,
     /* .event_wait              = */ ggml_backend_cann_event_wait,
 };
@@ -1880,7 +1918,7 @@ static void ggml_backend_cann_device_get_memory(ggml_backend_dev_t dev, size_t *
 
 static enum ggml_backend_dev_type ggml_backend_cann_device_get_type(ggml_backend_dev_t dev) {
     GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
 }
 
 static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {

ggml/src/ggml-cuda.cu

@@ -421,13 +421,18 @@ struct ggml_backend_cuda_buffer_context {
     }
 };
 
-static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
-    delete ctx;
+    return ctx->name.c_str();
 }
 
 static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
-    return buffer->iface.free_buffer == ggml_backend_cuda_buffer_free_buffer;
+    return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name;
+}
+
+static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+    delete ctx;
 }
 
 static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -510,6 +515,7 @@ static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
 }
 
 static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
+    /* .get_name        = */ ggml_backend_cuda_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cuda_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cuda_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_cuda_buffer_init_tensor,
@@ -542,6 +548,8 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
 
     ggml_cuda_set_device(buft_ctx->device);
 
+    size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
+
     void * dev_ptr;
     cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device);
     if (err != cudaSuccess) {
@@ -649,9 +657,7 @@ static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_spl
 }
 
 struct ggml_backend_cuda_split_buffer_type_context {
-    int main_device;
     std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
-    std::string name;
 };
 
 struct ggml_backend_cuda_split_buffer_context {
@@ -674,6 +680,16 @@ struct ggml_backend_cuda_split_buffer_context {
     std::vector<ggml_tensor_extra_gpu *> tensor_extras;
 };
 
+static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return GGML_CUDA_NAME "_Split";
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
+    GGML_UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds
+}
 
 static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
@@ -817,6 +833,7 @@ static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, u
 }
 
 static const ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
+    /* .get_name        = */ ggml_backend_cuda_split_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cuda_split_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cuda_split_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_cuda_split_buffer_init_tensor,
@@ -831,9 +848,9 @@ static const ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
 // cuda split buffer type
 
 static const char * ggml_backend_cuda_split_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;
+    return GGML_CUDA_NAME "_Split";
 
-    return ctx->name.c_str();
+    GGML_UNUSED(buft);
 }
 
 static bool ggml_backend_buft_is_cuda_split(ggml_backend_buffer_type_t buft) {
@@ -898,11 +915,11 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_inte
     /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
 };
 
-ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split) {
+ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
 
-    static std::map<std::pair<int, std::array<float, GGML_CUDA_MAX_DEVICES>>, struct ggml_backend_buffer_type> buft_map;
+    static std::map<std::array<float, GGML_CUDA_MAX_DEVICES>, struct ggml_backend_buffer_type> buft_map;
 
     std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split_arr = {};
 
@@ -920,23 +937,18 @@ ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device,
         }
     }
 
-    auto it = buft_map.find({main_device, tensor_split_arr});
+    auto it = buft_map.find(tensor_split_arr);
     if (it != buft_map.end()) {
         return &it->second;
     }
-    auto * ctx = new ggml_backend_cuda_split_buffer_type_context{
-        main_device,
-        tensor_split_arr,
-        GGML_CUDA_NAME + std::to_string(main_device) + "_Split",
-    };
 
     struct ggml_backend_buffer_type buft {
         /* .iface   = */ ggml_backend_cuda_split_buffer_type_interface,
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), main_device),
-        /* .context = */ ctx,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), 0),
+        /* .context = */ new ggml_backend_cuda_split_buffer_type_context{tensor_split_arr},
     };
 
-    auto result = buft_map.emplace(std::make_pair(main_device, tensor_split_arr), buft);
+    auto result = buft_map.emplace(tensor_split_arr, buft);
     return &result.first->second;
 }
 
@@ -948,6 +960,12 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
     GGML_UNUSED(buft);
 }
 
+static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return GGML_CUDA_NAME "_Host";
+
+    GGML_UNUSED(buffer);
+}
+
 static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     CUDA_CHECK(cudaFreeHost(buffer->context));
 }
@@ -980,6 +998,7 @@ static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggm
 
     ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
     buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_cuda_host_buffer_name;
     buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;
 
     return buffer;
@@ -1381,7 +1400,7 @@ static void ggml_cuda_op_mul_mat(
 
     const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
 
-    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
+    const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
     GGML_ASSERT(!(split && ne02 > 1));
     GGML_ASSERT(!(split && ne03 > 1));
     GGML_ASSERT(!(split && ne02 < ne12));
@@ -1465,19 +1484,14 @@ static void ggml_cuda_op_mul_mat(
             const size_t nbytes_data    = ggml_nbytes(src0);
             const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
-        // TODO: remove this for MUSA once the Guilty Lockup issue is resolved
-#ifndef GGML_USE_MUSA
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream));
-#else // GGML_USE_MUSA
-            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream));
-#endif // !GGML_USE_MUSA
         }
 
         // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
         if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
             const size_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
             const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
-            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream));
+            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
         }
 
         if (src1_on_device && src1_is_contiguous) {
@@ -1871,7 +1885,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 }
 
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
+    const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
 
     bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
@@ -1998,7 +2012,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+    GGML_ASSERT(!ggml_backend_buffer_is_cuda_split(src0->buffer) && "mul_mat_id does not support split buffers");
 
     cudaStream_t stream = ctx.stream();
 
@@ -2131,7 +2145,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 
 static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {
     // why is this here instead of mul_mat?
-    if (dst->src[0] != nullptr && ggml_backend_buft_is_cuda_split(dst->src[0]->buffer->buft)) {
+    if (dst->src[0] != nullptr && ggml_backend_buffer_is_cuda_split(dst->src[0]->buffer)) {
         ggml_cuda_set_peer_access(dst->src[1]->ne[1], ctx.device);
     }
 
@@ -2352,6 +2366,12 @@ static void ggml_backend_cuda_free(ggml_backend_t backend) {
     delete backend;
 }
 
+static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    return ggml_backend_cuda_buffer_type(cuda_ctx->device);
+}
+
 static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -2557,7 +2577,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
                 continue;
             }
 
-            if (node->src[0] && node->src[0]->buffer && ggml_backend_buft_is_cuda_split(node->src[0]->buffer->buft)) {
+            if (node->src[0] && node->src[0]->buffer && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
                 use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
 #ifndef NDEBUG
                 GGML_LOG_DEBUG("%s: disabling CUDA graphs due to split buffer\n", __func__);
@@ -2644,8 +2664,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
                 for (int j = 0; j < GGML_MAX_SRC; j++) {
                     if (node->src[j] != nullptr) {
                         assert(node->src[j]->buffer);
-                        assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) ||
-                               ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft));
+                        assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
                     }
                 }
 #endif
@@ -2738,7 +2757,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
         if (stat == cudaErrorGraphExecUpdateFailure) {
 #ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: CUDA graph update failed\n", __func__);
+            GGML_LOG_ERROR("%s: CUDA graph update failed\n", __func__);
 #endif
             // The pre-existing graph exec cannot be updated due to violated constraints
             // so instead clear error and re-instantiate
@@ -2787,6 +2806,7 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
 static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .get_name                = */ ggml_backend_cuda_get_name,
     /* .free                    = */ ggml_backend_cuda_free,
+    /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
     /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
     /* .cpy_tensor_async        = */ ggml_backend_cuda_cpy_tensor_async,
@@ -2796,6 +2816,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
+    /* .supports_op             = */ NULL, // moved to device
+    /* .supports_buft           = */ NULL, // moved to device
+    /* .offload_op              = */ NULL, // moved to device
     /* .event_record            = */ ggml_backend_cuda_event_record,
     /* .event_wait              = */ ggml_backend_cuda_event_wait,
 };
@@ -2885,7 +2908,7 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
 
 static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
     GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
 }
 
 static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
@@ -2909,7 +2932,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
     };
 }
 
-static ggml_backend_t ggml_backend_cuda_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+static ggml_backend_t ggml_backend_cuda_device_init(ggml_backend_dev_t dev, const char * params) {
     GGML_UNUSED(params);
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
     return ggml_backend_cuda_init(ctx->device);
@@ -2925,29 +2948,18 @@ static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_host_buffer_type(
     return ggml_backend_cuda_host_buffer_type();
 }
 
+static ggml_backend_buffer_t ggml_backend_cuda_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    GGML_UNUSED(dev);
+    GGML_UNUSED(ptr);
+    GGML_UNUSED(size);
+    GGML_UNUSED(max_tensor_size);
+    return nullptr;
+}
+
 // TODO: move these functions here
 static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
     ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
 
-    // split buffers can only be used with GGML_OP_MUL_MAT
-    if (op->op != GGML_OP_MUL_MAT) {
-        for (int i = 0; i < GGML_MAX_SRC; i++) {
-            if (op->src[i] && op->src[i]->buffer && ggml_backend_buft_is_cuda_split(op->src[i]->buffer->buft)) {
-                return false;
-            }
-        }
-    }
-
-    // check if all the sources are allocated on this device
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (op->src[i] && op->src[i]->buffer && ggml_backend_buft_is_cuda(op->src[i]->buffer->buft)) {
-            ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)op->src[i]->buffer->buft->context;
-            if (buft_ctx->device != dev_ctx->device) {
-                return false;
-            }
-        }
-    }
-
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -3173,27 +3185,24 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 }
 
 static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return (ggml_backend_buft_is_cuda(buft) || ggml_backend_buft_is_cuda_split(buft)) && buft->device == dev;
-}
-
-static int64_t get_op_batch_size(const ggml_tensor * op) {
-    switch (op->op) {
-        case GGML_OP_GET_ROWS:
-            return 0;
-        case GGML_OP_MUL_MAT:
-            return op->ne[1];
-        case GGML_OP_MUL_MAT_ID:
-        case GGML_OP_ROPE:
-            return op->ne[2];
-        default:
-            return ggml_nrows(op);
+    if (ggml_backend_buft_is_cuda_split(buft)) {
+        return true;
     }
+
+    if (ggml_backend_buft_is_cuda(buft)) {
+        ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context;
+        ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
+        return buft_ctx->device == dev_ctx->device;
+    }
+
+    return false;
 }
 
 static bool ggml_backend_cuda_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
     const int min_batch_size = 32;
 
-    return get_op_batch_size(op) >= min_batch_size;
+    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
+           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
 
     GGML_UNUSED(dev);
 }
@@ -3234,10 +3243,10 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
     /* .get_memory              = */ ggml_backend_cuda_device_get_memory,
     /* .get_type                = */ ggml_backend_cuda_device_get_type,
     /* .get_props               = */ ggml_backend_cuda_device_get_props,
-    /* .init_backend            = */ ggml_backend_cuda_device_init_backend,
+    /* .init_backend            = */ ggml_backend_cuda_device_init,
     /* .get_buffer_type         = */ ggml_backend_cuda_device_get_buffer_type,
     /* .get_host_buffer_type    = */ ggml_backend_cuda_device_get_host_buffer_type,
-    /* .buffer_from_host_ptr    = */ NULL,
+    /* .buffer_from_host_ptr    = */ ggml_backend_cuda_device_buffer_from_host_ptr,
     /* .supports_op             = */ ggml_backend_cuda_device_supports_op,
     /* .supports_buft           = */ ggml_backend_cuda_device_supports_buft,
     /* .offload_op              = */ ggml_backend_cuda_device_offload_op,

ggml/src/ggml-cuda/cpy.cuh

@@ -1,6 +1,6 @@
 #include "common.cuh"
 
-#define CUDA_CPY_BLOCK_SIZE 64
+#define CUDA_CPY_BLOCK_SIZE 32
 
 void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
 

ggml/src/ggml-kompute.cpp

@@ -20,7 +20,6 @@
 #include "shaderop_mul_mat_q8_0.h"
 #include "shaderop_mul_mat_q4_0.h"
 #include "shaderop_mul_mat_q4_1.h"
-#include "shaderop_mul_mat_q4_k.h"
 #include "shaderop_mul_mat_q6_k.h"
 #include "shaderop_mul_mat_mat_f32.h"
 #include "shaderop_getrows_f32.h"
@@ -43,7 +42,6 @@
 #include <cstring>
 #include <iostream>
 #include <memory>
-#include <mutex>
 #include <stdexcept>
 #include <string>
 #include <unordered_map>
@@ -275,9 +273,18 @@ static std::vector<ggml_vk_device> ggml_vk_available_devices_internal(size_t mem
     return results;
 }
 
-static std::vector<ggml_vk_device>& ggml_vk_available_devices() {
-    static std::vector<ggml_vk_device> devices = ggml_vk_available_devices_internal(0);
-    return devices;
+// public API returns a C-style array
+ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count) {
+    auto devices = ggml_vk_available_devices_internal(memoryRequired);
+    *count = devices.size();
+    if (devices.empty()) {
+        return nullptr;
+    }
+
+    size_t nbytes = sizeof (ggml_vk_device) * (devices.size());
+    auto * arr = static_cast<ggml_vk_device *>(malloc(nbytes));
+    memcpy(arr, devices.data(), nbytes);
+    return arr;
 }
 
 static void ggml_vk_filterByVendor(std::vector<ggml_vk_device>& devices, const std::string& targetVendor) {
@@ -334,7 +341,7 @@ ggml_vk_device ggml_vk_current_device() {
     if (!komputeManager()->hasDevice())
         return ggml_vk_device();
 
-    auto devices = ggml_vk_available_devices();
+    auto devices = ggml_vk_available_devices_internal(0);
     ggml_vk_filterByName(devices, komputeManager()->physicalDevice()->getProperties().deviceName.data());
     GGML_ASSERT(!devices.empty());
     return devices.front();
@@ -1068,40 +1075,6 @@ static void ggml_vk_mul_mat_q8_0(Args&&... args) {
     ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
 }
 
-static void ggml_vk_mul_mat_q4_k(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne10,
-    int32_t ne11, int32_t ne12, int32_t ne13, int32_t ne0,
-    int32_t ne1, int32_t r2, int32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv,
-        kp::shader_data::op_mul_mat_q4_k_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3;
-    } pushConsts {
-        0, 0, 0,
-        ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
 static void ggml_vk_mul_mat_q6_k(
     kp::Sequence& seq,
     const std::shared_ptr<kp::Tensor>& inA,
@@ -1350,7 +1323,17 @@ static void ggml_vk_cpy_f16_f32(Args&&... args) {
     ggml_vk_cpy(spirv, 2, 4, std::forward<Args>(args)...);
 }
 
-static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+static bool ggml_vk_supports_op(const struct ggml_tensor * op) {
+    switch (op->type) {
+        case GGML_TYPE_F16:
+        case GGML_TYPE_F32:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+            break;
+        default:
+            return false;
+    }
+
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -1419,7 +1402,6 @@ static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, cons
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
                 case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q4_K:
                     return true;
                 default:
                     ;
@@ -1428,8 +1410,6 @@ static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, cons
             ;
     }
     return false;
-
-    GGML_UNUSED(dev);
 }
 
 static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) {
@@ -1478,6 +1458,11 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
 
             any_commands_recorded = true;
 
+            if (!ggml_vk_supports_op(dst)) {
+                 fprintf(stderr, "%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
+                 GGML_ABORT("unsupported op");
+            }
+
             const int32_t ne00 = src0 ? src0->ne[0] : 0;
             const int32_t ne01 = src0 ? src0->ne[1] : 0;
             const int32_t ne02 = src0 ? src0->ne[2] : 0;
@@ -1671,12 +1656,6 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                                     ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
                                 );
                                 break;
-                            case GGML_TYPE_Q4_K:
-                                ggml_vk_mul_mat_q4_k(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, ne12/ne02, ne13/ne03
-                                );
-                                break;
                             case GGML_TYPE_Q6_K:
                                 ggml_vk_mul_mat_q6_k(
                                     seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
@@ -1841,6 +1820,11 @@ static void ggml_backend_kompute_device_unref(ggml_backend_buffer_type_t buft) {
     }
 }
 
+static const char * ggml_backend_kompute_buffer_get_name(ggml_backend_buffer_t buffer) {
+    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buffer->buft->context);
+    return ctx->name.c_str();
+}
+
 static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     auto * memory = (ggml_vk_memory *)buffer->context;
     if (ggml_vk_has_device()) {
@@ -1884,6 +1868,7 @@ static void ggml_backend_kompute_buffer_clear(ggml_backend_buffer_t buffer, uint
 }
 
 static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = {
+    /* .get_name        = */ ggml_backend_kompute_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_kompute_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_kompute_buffer_get_base,
     /* .init_tensor     = */ NULL,
@@ -1928,31 +1913,25 @@ static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = {
 };
 
 ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) {
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
+    static std::vector<ggml_backend_buffer_type> bufts = []() {
+        std::vector<ggml_backend_buffer_type> vec;
+        auto devices = ggml_vk_available_devices_internal(0);
+        vec.reserve(devices.size());
 
-    auto devices = ggml_vk_available_devices();
-    int32_t device_count = (int32_t) devices.size();
-    GGML_ASSERT(device < device_count);
-    GGML_ASSERT(devices.size() <= GGML_KOMPUTE_MAX_DEVICES);
-
-    static ggml_backend_buffer_type
-        ggml_backend_kompute_buffer_types[GGML_KOMPUTE_MAX_DEVICES];
-
-    static bool ggml_backend_kompute_buffer_type_initialized = false;
-
-    if (!ggml_backend_kompute_buffer_type_initialized) {
-        for (int32_t i = 0; i < device_count; i++) {
-            ggml_backend_kompute_buffer_types[i] = {
-                /* .iface    = */ ggml_backend_kompute_buffer_type_interface,
-                /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_kompute_reg(), i),
-                /* .context  = */ new ggml_backend_kompute_buffer_type_context{ i, devices[i].bufferAlignment, devices[i].maxAlloc },
-            };
+        for (const auto & dev : devices) {
+            vec.push_back({
+                /* .iface   = */ ggml_backend_kompute_buffer_type_interface,
+                /* .device  = */ nullptr,
+                /* .context = */ new ggml_backend_kompute_buffer_type_context(dev.index, dev.bufferAlignment, dev.maxAlloc)
+            });
         }
-        ggml_backend_kompute_buffer_type_initialized = true;
-    }
+        return vec;
+    }();
 
-    return &ggml_backend_kompute_buffer_types[device];
+    auto it = std::find_if(bufts.begin(), bufts.end(), [device](const ggml_backend_buffer_type & t) {
+        return device == static_cast<ggml_backend_kompute_buffer_type_context *>(t.context)->device;
+    });
+    return it < bufts.end() ? &*it : nullptr;
 }
 
 // backend
@@ -1974,15 +1953,31 @@ static void ggml_backend_kompute_free(ggml_backend_t backend) {
     delete backend;
 }
 
+static ggml_backend_buffer_type_t ggml_backend_kompute_get_default_buffer_type(ggml_backend_t backend) {
+    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
+    return ggml_backend_kompute_buffer_type(ctx->device);
+}
+
 static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
     ggml_vk_graph_compute(ctx, cgraph);
     return GGML_STATUS_SUCCESS;
 }
 
+static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    GGML_UNUSED(backend);
+    return ggml_vk_supports_op(op);
+}
+
+static bool ggml_backend_kompute_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    GGML_UNUSED(backend);
+    return buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name;
+}
+
 static struct ggml_backend_i kompute_backend_i = {
     /* .get_name                = */ ggml_backend_kompute_name,
     /* .free                    = */ ggml_backend_kompute_free,
+    /* .get_default_buffer_type = */ ggml_backend_kompute_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
@@ -1992,6 +1987,9 @@ static struct ggml_backend_i kompute_backend_i = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_kompute_graph_compute,
+    /* .supports_op             = */ ggml_backend_kompute_supports_op,
+    /* .supports_buft           = */ ggml_backend_kompute_supports_buft,
+    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -2008,7 +2006,7 @@ ggml_backend_t ggml_backend_kompute_init(int device) {
     ggml_backend_t kompute_backend = new ggml_backend {
         /* .guid      = */ ggml_backend_kompute_guid(),
         /* .interface = */ kompute_backend_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_kompute_reg(), device),
+        /* .device    = */ nullptr,
         /* .context   = */ s_kompute_context,
     };
 
@@ -2018,167 +2016,3 @@ ggml_backend_t ggml_backend_kompute_init(int device) {
 bool ggml_backend_is_kompute(ggml_backend_t backend) {
     return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
 }
-
-static size_t ggml_backend_kompute_get_device_count() {
-    auto devices = ggml_vk_available_devices();
-    return devices.size();
-}
-
-static void ggml_backend_kompute_get_device_description(int device, char * description, size_t description_size) {
-    auto devices = ggml_vk_available_devices();
-    GGML_ASSERT((size_t) device < devices.size());
-    snprintf(description, description_size, "%s", devices[device].name);
-}
-
-static void ggml_backend_kompute_get_device_memory(int device, size_t * free, size_t * total) {
-    auto devices = ggml_vk_available_devices();
-    GGML_ASSERT((size_t) device < devices.size());
-    *total = devices[device].heapSize;
-    *free = devices[device].heapSize;
-}
-
-//////////////////////////
-
-struct ggml_backend_kompute_device_context {
-    int device;
-    std::string name;
-    std::string description;
-};
-
-static const char * ggml_backend_kompute_device_get_name(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ctx->name.c_str();
-}
-
-static const char * ggml_backend_kompute_device_get_description(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ctx->description.c_str();
-}
-
-static void ggml_backend_kompute_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    ggml_backend_kompute_get_device_memory(ctx->device, free, total);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_kompute_device_get_buffer_type(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ggml_backend_kompute_buffer_type(ctx->device);
-}
-
-static bool ggml_backend_kompute_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    if (buft->iface.get_name != ggml_backend_kompute_buffer_type_get_name) {
-        return false;
-    }
-
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    ggml_backend_kompute_buffer_type_context * buft_ctx = (ggml_backend_kompute_buffer_type_context *)buft->context;
-
-    return buft_ctx->device == ctx->device;
-}
-
-static enum ggml_backend_dev_type ggml_backend_kompute_device_get_type(ggml_backend_dev_t dev) {
-    GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
-}
-
-static void ggml_backend_kompute_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_kompute_device_get_name(dev);
-    props->description = ggml_backend_kompute_device_get_description(dev);
-    props->type        = ggml_backend_kompute_device_get_type(dev);
-    ggml_backend_kompute_device_get_memory(dev, &props->memory_free, &props->memory_total);
-    props->caps = {
-        /* async                  = */ false,
-        /* host_buffer            = */ false,
-        /* .buffer_from_host_ptr  = */ false,
-        /* events                 = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_kompute_device_init(ggml_backend_dev_t dev, const char * params) {
-    GGML_UNUSED(params);
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ggml_backend_kompute_init(ctx->device);
-}
-
-static bool ggml_backend_kompute_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-
-    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
-           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_kompute_device_i = {
-    /* .get_name             = */ ggml_backend_kompute_device_get_name,
-    /* .get_description      = */ ggml_backend_kompute_device_get_description,
-    /* .get_memory           = */ ggml_backend_kompute_device_get_memory,
-    /* .get_type             = */ ggml_backend_kompute_device_get_type,
-    /* .get_props            = */ ggml_backend_kompute_device_get_props,
-    /* .init_backend         = */ ggml_backend_kompute_device_init,
-    /* .get_buffer_type      = */ ggml_backend_kompute_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ NULL,
-    /* .supports_op          = */ ggml_backend_kompute_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_kompute_device_supports_buft,
-    /* .offload_op           = */ ggml_backend_kompute_device_offload_op,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-static const char * ggml_backend_kompute_reg_get_name(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return "Kompute";
-}
-
-static size_t ggml_backend_kompute_reg_get_device_count(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return ggml_backend_kompute_get_device_count();
-}
-
-static ggml_backend_dev_t ggml_backend_kompute_reg_get_device(ggml_backend_reg_t reg, size_t device) {
-    static std::vector<ggml_backend_dev_t> devices;
-
-    static bool initialized = false;
-
-    {
-        static std::mutex mutex;
-        std::lock_guard<std::mutex> lock(mutex);
-        if (!initialized) {
-            for (size_t i = 0; i < ggml_backend_kompute_get_device_count(); i++) {
-                ggml_backend_kompute_device_context * ctx = new ggml_backend_kompute_device_context;
-                char desc[256];
-                ggml_backend_kompute_get_device_description(i, desc, sizeof(desc));
-                ctx->device = i;
-                ctx->name = "Kompute" + std::to_string(i);
-                ctx->description = desc;
-                devices.push_back(new ggml_backend_device {
-                    /* .iface   = */ ggml_backend_kompute_device_i,
-                    /* .reg     = */ reg,
-                    /* .context = */ ctx,
-                });
-            }
-            initialized = true;
-        }
-    }
-
-    GGML_ASSERT(device < devices.size());
-    return devices[device];
-}
-
-static const struct ggml_backend_reg_i ggml_backend_kompute_reg_i = {
-    /* .get_name         = */ ggml_backend_kompute_reg_get_name,
-    /* .get_device_count = */ ggml_backend_kompute_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_kompute_reg_get_device,
-    /* .get_proc_address = */ NULL,
-};
-
-ggml_backend_reg_t ggml_backend_kompute_reg() {
-    static ggml_backend_reg reg = {
-        /* .iface   = */ ggml_backend_kompute_reg_i,
-        /* .context = */ nullptr,
-    };
-
-    return &reg;
-}

ggml/src/ggml-metal.m

@@ -1015,21 +1015,19 @@ static void ggml_metal_encode_node(
     id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
     id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(dst,  &offs_dst)  : nil;
 
-#if 0
-    GGML_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
-    if (src0) {
-        GGML_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03,
-                ggml_is_contiguous(src0), src0->name);
-    }
-    if (src1) {
-        GGML_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
-                ggml_is_contiguous(src1), src1->name);
-    }
-    if (dst) {
-        GGML_LOG_INFO("%s: dst  - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2, ne3, nb0, nb1, nb2, nb3,
-                dst->name);
-    }
-#endif
+    //GGML_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+    //if (src0) {
+    //    GGML_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+    //            ggml_is_contiguous(src0), src0->name);
+    //}
+    //if (src1) {
+    //    GGML_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+    //            ggml_is_contiguous(src1), src1->name);
+    //}
+    //if (dst) {
+    //    GGML_LOG_INFO("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+    //            dst->name);
+    //}
 
     id<MTLDevice> device = ctx_dev->mtl_device;
 
@@ -1812,16 +1810,14 @@ static void ggml_metal_encode_node(
                             [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
                             [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
                             [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
-                            [encoder setBytes:&nb03    length:sizeof(nb03) atIndex:7];
-                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:8];
-                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:9];
-                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:10];
-                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:11];
-                            [encoder setBytes:&nb13    length:sizeof(nb13) atIndex:12];
-                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
-                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
-                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:15];
-                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:16];
+                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
+                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:8];
+                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:9];
+                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:10];
+                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:11];
+                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
+                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:13];
+                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:14];
                             [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                         } else {
@@ -1990,22 +1986,20 @@ static void ggml_metal_encode_node(
                             [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
                             [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
                             [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:13];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:14];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:15];
-                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:16];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:17];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:18];
-                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:19];
-                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:20];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
+                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:17];
+                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:18];
 
                             if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
-                                src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
-                                src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
+                                    src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
+                                    src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@@ -2054,9 +2048,6 @@ static void ggml_metal_encode_node(
 
                 GGML_ASSERT(src1t == GGML_TYPE_F32);
 
-                GGML_ASSERT(ne03 == 1);
-                GGML_ASSERT(ne13 == 1);
-
                 // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                 // to the matrix-vector kernel
                 // ne20 = n_used_experts
@@ -3247,6 +3238,12 @@ static enum ggml_status ggml_metal_graph_compute(
 
 // backend interface
 
+static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "Metal";
+
+    UNUSED(buffer);
+}
+
 static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
@@ -3301,6 +3298,7 @@ static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_
 }
 
 static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
+    /* .get_name        = */ ggml_backend_metal_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_metal_buffer_get_base,
     /* .init_tensor     = */ NULL,
@@ -3425,29 +3423,6 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
     return &ggml_backend_buffer_type_metal;
 }
 
-static const char * ggml_backend_metal_buffer_from_ptr_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "Metal_Mapped";
-
-    UNUSED(buft);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_metal_buffer_from_ptr_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_buffer_from_ptr_type_metal = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_metal_buffer_from_ptr_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
-            /* .get_max_size     = */ ggml_backend_metal_buffer_type_get_max_size,
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
-        },
-        /* .device  = */ &g_ggml_backend_metal_device,
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_buffer_from_ptr_type_metal;
-}
-
 // TODO: obsoleted by ggml_backend_metal_device_buffer_from_ptr
 ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
     struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context));
@@ -3524,7 +3499,7 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
         }
     }
 
-    return ggml_backend_buffer_init(ggml_backend_metal_buffer_from_ptr_type(), ggml_backend_metal_buffer_i, ctx, size);
+    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
 }
 
 // backend
@@ -3545,6 +3520,12 @@ static void ggml_backend_metal_free(ggml_backend_t backend) {
     free(backend);
 }
 
+static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_metal_buffer_type();
+
+    UNUSED(backend);
+}
+
 static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     return ggml_metal_graph_compute(backend, cgraph);
 }
@@ -3611,6 +3592,7 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
 static struct ggml_backend_i ggml_backend_metal_i = {
     /* .get_name                = */ ggml_backend_metal_name,
     /* .free                    = */ ggml_backend_metal_free,
+    /* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
@@ -3620,6 +3602,9 @@ static struct ggml_backend_i ggml_backend_metal_i = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_metal_graph_compute,
+    /* .supports_op             = */ NULL,
+    /* .supports_buft           = */ NULL,
+    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -3714,7 +3699,7 @@ static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t
 }
 
 static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
 
     GGML_UNUSED(dev);
 }

ggml/src/ggml-rpc.cpp

@@ -178,6 +178,7 @@ struct ggml_backend_rpc_buffer_context {
     std::shared_ptr<socket_t> sock;
     std::unordered_map<ggml_backend_buffer_t, void *> base_cache;
     uint64_t remote_ptr;
+    std::string name;
 };
 
 // RPC helper functions
@@ -408,6 +409,11 @@ static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
     return sock;
 }
 
+static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+    return ctx->name.c_str();
+}
+
 static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_msg_free_buffer_req request = {ctx->remote_ptr};
@@ -518,6 +524,7 @@ static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
 }
 
 static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
+    /* .get_name        = */ ggml_backend_rpc_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_rpc_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_rpc_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_rpc_buffer_init_tensor,
@@ -544,7 +551,7 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
     if (response.remote_ptr != 0) {
         ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
             ggml_backend_rpc_buffer_interface,
-            new ggml_backend_rpc_buffer_context{sock, {}, response.remote_ptr},
+            new ggml_backend_rpc_buffer_context{sock, {}, response.remote_ptr, "RPC[" + std::string(buft_ctx->endpoint) + "]"},
             response.remote_size);
         return buffer;
     } else {
@@ -602,6 +609,11 @@ static void ggml_backend_rpc_free(ggml_backend_t backend) {
     delete backend;
 }
 
+static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;
+    return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str());
+}
+
 static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
     UNUSED(backend);
     // this is no-op because we don't have any async operations
@@ -658,6 +670,7 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
 static ggml_backend_i ggml_backend_rpc_interface = {
     /* .get_name                = */ ggml_backend_rpc_name,
     /* .free                    = */ ggml_backend_rpc_free,
+    /* .get_default_buffer_type = */ ggml_backend_rpc_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
@@ -667,6 +680,9 @@ static ggml_backend_i ggml_backend_rpc_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_rpc_graph_compute,
+    /* .supports_op             = */ NULL,
+    /* .supports_buft           = */ NULL,
+    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -1262,7 +1278,7 @@ static void ggml_backend_rpc_device_get_memory(ggml_backend_dev_t dev, size_t *
 
 static enum ggml_backend_dev_type ggml_backend_rpc_device_get_type(ggml_backend_dev_t dev) {
     // TODO: obtain value from the server
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
 
     UNUSED(dev);
 }

ggml/src/ggml-sycl.cpp

@@ -249,10 +249,13 @@ struct ggml_backend_sycl_buffer_context {
     }
 };
 
-static const char * ggml_backend_sycl_buffer_type_get_name(ggml_backend_buffer_type_t buft);
+static const char * ggml_backend_sycl_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
+    return ctx->name.c_str();
+}
 
 static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) {
-    return buffer->buft->iface.get_name == ggml_backend_sycl_buffer_type_get_name;
+    return buffer->iface.get_name == ggml_backend_sycl_buffer_get_name;
 }
 
 static void
@@ -437,6 +440,7 @@ catch (sycl::exception const &exc) {
 }
 
 static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = {
+    /* .get_name        = */ ggml_backend_sycl_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_sycl_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_sycl_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_sycl_buffer_init_tensor,
@@ -694,6 +698,16 @@ struct ggml_backend_sycl_split_buffer_context {
     std::vector<queue_ptr> streams;
 };
 
+static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return GGML_SYCL_NAME "_Split";
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) {
+   return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name;
+}
+
 static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
     delete ctx;
@@ -901,6 +915,7 @@ static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, u
 }
 
 static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = {
+    /* .get_name        = */ ggml_backend_sycl_split_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_sycl_split_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_sycl_split_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_sycl_split_buffer_init_tensor,
@@ -920,10 +935,6 @@ static const char * ggml_backend_sycl_split_buffer_type_get_name(ggml_backend_bu
     GGML_UNUSED(buft);
 }
 
-static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) {
-   return buffer->buft->iface.get_name == ggml_backend_sycl_split_buffer_type_get_name;
-}
-
 static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
     // instead, we allocate them for each tensor separately in init_tensor
@@ -1029,6 +1040,12 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
     GGML_UNUSED(buft);
 }
 
+static const char * ggml_backend_sycl_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return GGML_SYCL_NAME "_Host";
+
+    GGML_UNUSED(buffer);
+}
+
 static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_sycl_host_free(buffer->context);
 }
@@ -1044,6 +1061,7 @@ static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggm
     // FIXME: this is a hack to avoid having to implement a new buffer type
     ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
     buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_sycl_host_buffer_name;
     buffer->iface.free_buffer = ggml_backend_sycl_host_buffer_free_buffer;
 
     return buffer;
@@ -4871,6 +4889,12 @@ static void ggml_backend_sycl_free(ggml_backend_t backend) {
     delete backend;
 }
 
+
+static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    return ggml_backend_sycl_buffer_type(sycl_ctx->device);
+}
+
 static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                ggml_tensor *tensor,
                                                const void *data, size_t offset,
@@ -5007,6 +5031,7 @@ static void ggml_backend_sycl_event_wait(ggml_backend_t backend, ggml_backend_ev
 static ggml_backend_i ggml_backend_sycl_interface = {
     /* .get_name                = */ ggml_backend_sycl_get_name,
     /* .free                    = */ ggml_backend_sycl_free,
+    /* .get_default_buffer_type = */ ggml_backend_sycl_get_default_buffer_type,
     /* .set_tensor_async        = */ ggml_backend_sycl_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_sycl_get_tensor_async,
     /* .cpy_tensor_async        = */ NULL, // ggml_backend_sycl_cpy_tensor_async,
@@ -5018,6 +5043,9 @@ static ggml_backend_i ggml_backend_sycl_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_sycl_graph_compute,
+    /* .supports_op             = */ NULL, // moved to device
+    /* .supports_buft           = */ NULL, // moved to device
+    /* .offload_op              = */ NULL, // moved to device
     /* .event_record            = */ ggml_backend_sycl_event_record,
     /* .event_wait              = */ ggml_backend_sycl_event_wait,
 };
@@ -5064,7 +5092,7 @@ static void ggml_backend_sycl_device_get_memory(ggml_backend_dev_t dev, size_t *
 
 static enum ggml_backend_dev_type ggml_backend_sycl_device_get_type(ggml_backend_dev_t dev) {
     GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
 }
 
 static void ggml_backend_sycl_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
@@ -5145,6 +5173,10 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (op->op == GGML_OP_MUL_MAT) {
                     a = op->src[0];
                     b = op->src[1];
+                    if (ggml_is_permuted(a) || ggml_is_permuted(b)) {
+                        // TODO: fix like https://github.com/ggerganov/llama.cpp/pull/10021
+                        return false;
+                    }
                 } else {
                     a = op->src[2];
                     b = op->src[1];
@@ -5360,14 +5392,12 @@ static ggml_backend_dev_t ggml_backend_sycl_reg_get_device(ggml_backend_reg_t re
     return ctx->devices[index];
 }
 
-static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, const char *name) {
+static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, const char *name)
+{
     GGML_UNUSED(reg);
-
-    // TODO: update to the current function signature
-    //if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
-    //    return (void *)ggml_backend_sycl_split_buffer_type;
-    //}
-
+    if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
+        return (void *)ggml_backend_sycl_split_buffer_type;
+    }
     // SYCL doesn't support registering host memory, left here for reference
     // "ggml_backend_register_host_buffer"
     // "ggml_backend_unregister_host_buffer"

ggml/src/ggml-sycl/norm.cpp

@@ -8,7 +8,6 @@ static void norm_f32(const float* x, float* dst, const int ncols, const float ep
 
     const int nthreads = item_ct1.get_local_range(2);
     const int nwarps = nthreads / WARP_SIZE;
-    assert(nwarps % WARP_SIZE == 0);
     sycl::float2 mean_var = sycl::float2(0.f, 0.f);
 
     for (int col = tid; col < ncols; col += block_size) {
@@ -55,7 +54,6 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
     int end = start + group_size;
     const int nthreads = item_ct1.get_local_range(2);
     const int nwarps = nthreads / WARP_SIZE;
-    assert(nwarps % WARP_SIZE == 0);
     start += item_ct1.get_local_id(2);
     int nreduce = nwarps / WARP_SIZE;
 
@@ -144,7 +142,6 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const floa
     const int tid = item_ct1.get_local_id(2);
     const int nthreads = item_ct1.get_local_range(2);
     const int nwarps = nthreads / WARP_SIZE;
-    assert(nwarps % WARP_SIZE == 0);
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
@@ -202,6 +199,7 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols,
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
+        assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
         const sycl::range<3> block_dims(1, 1, work_group_size);
         /*
         DPCT1049:17: The work-group size passed to the SYCL kernel may exceed
@@ -244,6 +242,7 @@ static void group_norm_f32_sycl(const float* x, float* dst,
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
+        assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
         const sycl::range<3> block_dims(1, 1, work_group_size);
         /*
         DPCT1049:18: The work-group size passed to the SYCL kernel may exceed
@@ -290,6 +289,7 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols,
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
+        assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
         const sycl::range<3> block_dims(1, 1, work_group_size);
         /*
         DPCT1049:19: The work-group size passed to the SYCL kernel may exceed

ggml/src/ggml-vulkan.cpp

@@ -213,7 +213,6 @@ struct vk_device_struct {
     vk_pipeline pipeline_sum_rows_f32;
     vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
     vk_pipeline pipeline_timestep_embedding_f32;
-    vk_pipeline pipeline_pool2d_f32;
 
     std::unordered_map<std::string, vk_pipeline_ref> pipelines;
     std::unordered_map<std::string, uint64_t> pipeline_descriptor_set_requirements;
@@ -404,17 +403,6 @@ struct vk_op_timestep_embedding_push_constants {
     uint32_t max_period;
 };
 
-struct vk_op_pool2d_push_constants {
-    uint32_t IW; uint32_t IH;
-    uint32_t OW; uint32_t OH;
-    uint32_t OC;
-    uint32_t pelements;
-    uint32_t op;
-    int32_t k0; int32_t k1;
-    int32_t s0; int32_t s1;
-    int32_t p0; int32_t p1;
-};
-
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -1815,8 +1803,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
-
     for (auto &c : compiles) {
         c.wait();
     }
@@ -4248,11 +4234,6 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_timestep_embedding_f32;
         }
         return nullptr;
-    case GGML_OP_POOL_2D:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_pool2d_f32;
-        }
-        return nullptr;
     case GGML_OP_LEAKY_RELU:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_leaky_relu_f32;
@@ -4483,14 +4464,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             uint32_t half_ceil = (dim + 1) / 2;
             elements = { half_ceil, (uint32_t)src0->ne[0], 1 };
         } break;
-    case GGML_OP_POOL_2D:
-        {
-            const uint32_t N = dst->ne[3];
-            const uint32_t OC = dst->ne[2];
-            const uint32_t OH = dst->ne[1];
-            const uint32_t OW = dst->ne[0];
-            elements = { N * OC * OH * OW, 1, 1};
-        } break;
     case GGML_OP_ADD:
     case GGML_OP_DIV:
     case GGML_OP_MUL:
@@ -4941,34 +4914,6 @@ static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context
     }, dryrun);
 }
 
-static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
-    const int32_t k1 = dst->op_params[1];
-    const int32_t k0 = dst->op_params[2];
-    const int32_t s1 = dst->op_params[3];
-    const int32_t s0 = dst->op_params[4];
-    const int32_t p1 = dst->op_params[5];
-    const int32_t p0 = dst->op_params[6];
-
-    const uint32_t IH = src0->ne[1];
-    const uint32_t IW = src0->ne[0];
-
-    const uint32_t N = dst->ne[3];
-
-    const uint32_t OC = dst->ne[2];
-    const uint32_t OH = dst->ne[1];
-    const uint32_t OW = dst->ne[0];
-
-    const uint32_t parallel_elements = N * OC * OH * OW;
-
-    ggml_vk_op_f32<vk_op_pool2d_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_POOL_2D, {
-        IW, IH, OW, OH, OC,
-        parallel_elements,
-        op,
-        k0, k1, s0, s1, p0, p1,
-    }, dryrun);
-}
-
 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const float * op_params = (const float *)dst->op_params;
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
@@ -5847,7 +5792,6 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_SUM_ROWS:
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
-    case GGML_OP_POOL_2D:
     case GGML_OP_LEAKY_RELU:
         break;
     default:
@@ -5983,10 +5927,6 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_TIMESTEP_EMBEDDING:
         ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun);
 
-        break;
-    case GGML_OP_POOL_2D:
-        ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
-
         break;
     case GGML_OP_LEAKY_RELU:
         ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun);
@@ -6078,7 +6018,6 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_SUM_ROWS:
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
-    case GGML_OP_POOL_2D:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_REPEAT:
         buf = tensor->buffer;
@@ -6247,8 +6186,13 @@ static void ggml_vk_get_device_description(int device, char * description, size_
 
 // device backend
 
+static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
+    return ctx->name.c_str();
+}
+
 static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) {
-    return buffer->buft->iface.get_name == ggml_backend_vk_buffer_type_name;
+    return buffer->iface.get_name == ggml_backend_vk_buffer_get_name;
 }
 
 static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
@@ -6312,6 +6256,7 @@ static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t v
 }
 
 static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = {
+    /* .get_name        = */ ggml_backend_vk_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_vk_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_vk_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_vk_buffer_init_tensor,
@@ -6407,6 +6352,7 @@ static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_
 
     ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
     buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_vk_host_buffer_name;
     buffer->iface.free_buffer = ggml_backend_vk_host_buffer_free_buffer;
 
     return buffer;
@@ -6639,6 +6585,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 static ggml_backend_i ggml_backend_vk_interface = {
     /* .get_name                = */ ggml_backend_vk_name,
     /* .free                    = */ ggml_backend_vk_free,
+    /* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,  // ggml_backend_vk_set_tensor_async,
     /* .get_tensor_async        = */ NULL,  // ggml_backend_vk_get_tensor_async,
     /* .cpy_tensor_async        = */ NULL,  // ggml_backend_vk_cpy_tensor_async,
@@ -6648,6 +6595,9 @@ static ggml_backend_i ggml_backend_vk_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_vk_graph_compute,
+    /* .supports_op             = */ NULL,
+    /* .supports_buft           = */ NULL,
+    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -6706,7 +6656,7 @@ void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total
 //////////////////////////
 
 struct ggml_backend_vk_device_context {
-    size_t device;
+    int device;
     std::string name;
     std::string description;
 };
@@ -6738,7 +6688,7 @@ static ggml_backend_buffer_type_t ggml_backend_vk_device_get_host_buffer_type(gg
 
 static enum ggml_backend_dev_type ggml_backend_vk_device_get_type(ggml_backend_dev_t dev) {
     UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
 }
 
 static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
@@ -6747,10 +6697,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
     props->type        = ggml_backend_vk_device_get_type(dev);
     ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
     props->caps = {
-        /* .async                 = */ false,
-        /* .host_buffer           = */ true,
-        /* .buffer_from_host_ptr  = */ false,
-        /* .events                = */ false,
+        /* async       */ false,
+        /* host_buffer */ true,
+        /* events      */ false,
     };
 }
 
@@ -6872,7 +6821,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SUM_ROWS:
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
-        case GGML_OP_POOL_2D:
         case GGML_OP_LEAKY_RELU:
             return true;
         default:
@@ -6939,7 +6887,7 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
         static std::mutex mutex;
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
-            for (int i = 0; i < ggml_backend_vk_get_device_count(); i++) {
+            for (size_t i = 0; i < ggml_backend_vk_get_device_count(); i++) {
                 ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
                 char desc[256];
                 ggml_backend_vk_get_device_description(i, desc, sizeof(desc));
@@ -7386,16 +7334,6 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         const int32_t dim = tensor->op_params[0];
         const int32_t max_period = tensor->op_params[1];
         tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period);
-    } else if (tensor->op == GGML_OP_POOL_2D) {
-        enum ggml_op_pool op = static_cast<ggml_op_pool>(dst->op_params[0]);
-        const int32_t k0 = tensor->op_params[1];
-        const int32_t k1 = tensor->op_params[2];
-        const int32_t s0 = tensor->op_params[3];
-        const int32_t s1 = tensor->op_params[4];
-        const int32_t p0 = tensor->op_params[5];
-        const int32_t p1 = tensor->op_params[6];
-
-        tensor_clone = ggml_pool_2d(ggml_ctx, src0_clone, op, k0, k1, s0, s1, p0, p1);
     } else if (tensor->op == GGML_OP_LEAKY_RELU) {
         const float * op_params = (const float *)tensor->op_params;
         tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false);

ggml/src/ggml.c

@@ -4028,9 +4028,7 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
     if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
         GGML_LOG_WARN("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
                 __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
-#ifndef NDEBUG
-        GGML_ABORT("not enough space in the context's memory pool");
-#endif
+        assert(false);
         return NULL;
     }
 
@@ -22102,46 +22100,18 @@ static size_t gguf_type_size(enum gguf_type type) {
     return GGUF_TYPE_SIZE[type];
 }
 
-static bool gguf_tensor_info_sanitize(struct gguf_tensor_info * info) {
-    if (info->n_dims > GGML_MAX_DIMS) {
-        fprintf(stderr, "%s: invalid number of dimensions (%" PRIu32 ")\n", __func__, info->n_dims);
-        return false;
-    }
-
-    if (info->type < 0 || info->type >= GGML_TYPE_COUNT) {
-        fprintf(stderr, "%s: invalid type (%d)\n", __func__, info->type);
-        return false;
-    }
-
-    if (strlen(info->name.data) >= GGML_MAX_NAME) {
-        fprintf(stderr, "%s: tensor '%s' name is too long\n", __func__, info->name.data);
-        return false;
-    }
+static void gguf_tensor_info_sanitize(struct gguf_tensor_info * info) {
+    GGML_ASSERT(info->n_dims <= GGML_MAX_DIMS);
+    GGML_ASSERT(0 <= info->type && info->type < GGML_TYPE_COUNT);
 
     for (uint32_t i = 0; i < info->n_dims; ++i) {
-        if (info->ne[i] <= 0) {
-            fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[i]);
-            return false;
-        }
+        GGML_ASSERT(info->ne[i] > 0);
     }
 
     // prevent overflow for total number of elements
-    if (INT64_MAX/info->ne[1] <= info->ne[0]) {
-        fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[1]);
-        return false;
-    }
-
-    if (INT64_MAX/info->ne[2] <= info->ne[0]*info->ne[1]) {
-        fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[2]);
-        return false;
-    }
-
-    if (INT64_MAX/info->ne[3] <= info->ne[0]*info->ne[1]*info->ne[2]) {
-        fprintf(stderr, "%s: invalid number of elements (%" PRIu64 ")\n", __func__, info->ne[3]);
-        return false;
-    }
-
-    return true;
+    GGML_ASSERT(INT64_MAX/info->ne[1] > info->ne[0]);
+    GGML_ASSERT(INT64_MAX/info->ne[2] > info->ne[0]*info->ne[1]);
+    GGML_ASSERT(INT64_MAX/info->ne[3] > info->ne[0]*info->ne[1]*info->ne[2]);
 }
 
 static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset) {
@@ -22164,11 +22134,7 @@ static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) {
         return false;
     }
 
-    p->data = calloc(p->n + 1, 1);
-    if (!p->data) {
-        fprintf(stderr, "%s: failed to allocate memory for string of length %" PRIu64 "\n", __func__, p->n);
-        return false;
-    }
+    p->data = GGML_CALLOC(p->n + 1, 1);
 
     ok = ok && gguf_fread_el(file,  p->data, p->n, offset);
 
@@ -22202,11 +22168,7 @@ static void gguf_free_kv(struct gguf_kv * kv) {
 }
 
 struct gguf_context * gguf_init_empty(void) {
-    struct gguf_context * ctx = calloc(1, sizeof(struct gguf_context));
-    if (!ctx) {
-        fprintf(stderr, "%s: failed to allocate memory for context\n", __func__);
-        return NULL;
-    }
+    struct gguf_context * ctx = GGML_CALLOC(1, sizeof(struct gguf_context));
 
     memcpy(ctx->header.magic, GGUF_MAGIC, sizeof(ctx->header.magic));
     ctx->header.version   = GGUF_VERSION;
@@ -22252,12 +22214,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
     bool ok = true;
 
-    struct gguf_context * ctx = calloc(1, sizeof(struct gguf_context));
-    if (!ctx) {
-        fprintf(stderr, "%s: failed to allocate memory for context\n", __func__);
-        fclose(file);
-        return NULL;
-    }
+    struct gguf_context * ctx = GGML_CALLOC(1, sizeof(struct gguf_context));
 
     // read the header
     {
@@ -22296,13 +22253,9 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     {
         const uint64_t n_kv = ctx->header.n_kv;
 
-        ctx->kv = calloc(n_kv, sizeof(struct gguf_kv));
-        if (!ctx->kv) {
-            fprintf(stderr, "%s: failed to allocate memory for kv pairs\n", __func__);
-            fclose(file);
-            gguf_free(ctx);
-            return NULL;
-        }
+        // header.n_kv will hold the actual value of pairs that were successfully read in the loop below
+        ctx->header.n_kv = 0;
+        ctx->kv = GGML_CALLOC(n_kv, sizeof(struct gguf_kv));
 
         for (uint64_t i = 0; i < n_kv; ++i) {
             struct gguf_kv * kv = &ctx->kv[i];
@@ -22353,13 +22306,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                                         return NULL;
                                     }
 
-                                    kv->value.arr.data = calloc(kv->value.arr.n, gguf_type_size(kv->value.arr.type));
-                                    if (!kv->value.arr.data) {
-                                        fprintf(stderr, "%s: failed to allocate memory for array\n", __func__);
-                                        fclose(file);
-                                        gguf_free(ctx);
-                                        return NULL;
-                                    }
+                                    kv->value.arr.data = GGML_CALLOC(kv->value.arr.n, gguf_type_size(kv->value.arr.type));
 
                                     ok = ok && gguf_fread_el(file, kv->value.arr.data, kv->value.arr.n * gguf_type_size(kv->value.arr.type), &offset);
                                 } break;
@@ -22373,36 +22320,24 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                                         return NULL;
                                     }
 
-                                    kv->value.arr.data = calloc(kv->value.arr.n, sizeof(struct gguf_str));
-                                    if (!kv->value.arr.data) {
-                                        fprintf(stderr, "%s: failed to allocate memory for array\n", __func__);
-                                        fclose(file);
-                                        gguf_free(ctx);
-                                        return NULL;
-                                    }
+                                    kv->value.arr.data = GGML_CALLOC(kv->value.arr.n, sizeof(struct gguf_str));
 
                                     for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
                                         ok = ok && gguf_fread_str(file, &((struct gguf_str *) kv->value.arr.data)[j], &offset);
                                     }
                                 } break;
                             case GGUF_TYPE_ARRAY:
-                            default:
-                                {
-                                    fprintf(stderr, "%s: invalid array type %d\n", __func__, kv->value.arr.type);
-                                    ok = false;
-                                } break;
+                            default: GGML_ABORT("invalid type");
                         }
                     } break;
-                default:
-                    {
-                        fprintf(stderr, "%s: invalid type %d\n", __func__, kv->type);
-                        ok = false;
-                    } break;
+                default: GGML_ABORT("invalid type");
             }
 
             if (!ok) {
                 break;
             }
+
+            ctx->header.n_kv++;
         }
 
         if (!ok) {
@@ -22415,13 +22350,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
     // read the tensor infos
     if (ctx->header.n_tensors > 0) {
-        ctx->infos = calloc(ctx->header.n_tensors, sizeof(struct gguf_tensor_info));
-        if (!ctx->infos) {
-            fprintf(stderr, "%s: failed to allocate memory for tensor infos\n", __func__);
-            fclose(file);
-            gguf_free(ctx);
-            return NULL;
-        }
+        ctx->infos = GGML_CALLOC(ctx->header.n_tensors, sizeof(struct gguf_tensor_info));
 
         for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
             struct gguf_tensor_info * info = &ctx->infos[i];
@@ -22442,7 +22371,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
             ok = ok && gguf_fread_el (file, &info->type,   sizeof(info->type),    &offset);
             ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset),  &offset);
 
-            ok = ok && gguf_tensor_info_sanitize(info);
+            // TODO: return an error instead of crashing with GGML_ASSERT
+            gguf_tensor_info_sanitize(info);
 
             // make sure there is no duplicated tensor names
             for (uint64_t j = 0; j < i && ok; ++j) {

ggml/src/kompute-shaders/common.comp

@@ -15,7 +15,6 @@
 #define TWOPI_F 6.283185307179586f
 
 #define QK_K 256
-#define K_SCALE_SIZE 12
 
 #define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
 #define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
@@ -65,14 +64,6 @@ mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
     return reg;
 }
 
-#define sizeof_block_q4_k 144
-struct block_q4_k {
-    float16_t d;
-    float16_t dmin;
-    uint8_t scales[K_SCALE_SIZE];
-    uint8_t qs[QK_K/2];
-};
-
 #define sizeof_block_q6_k 210
 struct block_q6_k {
     uint8_t ql[QK_K/2];      // quants, lower 4 bits

ggml/src/kompute-shaders/op_mul_mat_q4_k.comp

@@ -1,133 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define N_DST 4
-#define SIZE_OF_BLOCK sizeof_block_q4_k
-
-layout(local_size_x = 4) in;
-layout(local_size_y = 8) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { block_q4_k inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne10;
-    int ne0;
-    int ne1;
-    int ne01;
-    int ne02;
-    int ne12;
-    int r2;
-    int r3;
-} pcs;
-
-void main() {
-    const uint16_t kmask1 = uint16_t(0x3f3f);
-    const uint16_t kmask2 = uint16_t(0x0f0f);
-    const uint16_t kmask3 = uint16_t(0xc0c0);
-
-    const uint ix = gl_SubgroupInvocationID/8;  // 0...3
-    const uint it = gl_SubgroupInvocationID%8;  // 0...7
-    const uint iq = it/4;     // 0 or 1
-    const uint ir = it%4;     // 0...3
-
-    const uint nb = pcs.ne00/QK_K;
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = r0 * N_DST;
-    const uint ib_row = first_row * nb;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
-
-    const uint xblk = ib_row + offset0 + pcs.inAOff;
-    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff;
-
-    float yl[16];
-    float yh[16];
-    float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
-    float all_sum = 0.f;
-
-    uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
-
-    for (uint ib = ix; ib < nb; ib += 4) {
-        const uint blk_idx = ib + xblk;
-
-        float sumy[4] = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+0] = inB[y4+i+  0]; sumy[0] += yl[i+0];
-            yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
-            yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
-            yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
-        }
-
-        for (int row = 0; row < N_DST; row++) {
-            uint row_idx = row * nb;
-
-            uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
-            uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
-            uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
-            uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
-            uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
-
-            uint16_t sc16[4];
-            sc16[0] = sc_0 & kmask1;
-            sc16[1] = sc_2 & kmask1;
-            sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
-            sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
-
-            float acc1[4] = {0.f, 0.f, 0.f, 0.f};
-            float acc2[4] = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
-                uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
-                acc1[0] += yl[i+0] * (q1 & 0x000F);
-                acc1[1] += yl[i+1] * (q1 & 0x0F00);
-                acc1[2] += yl[i+8] * (q1 & 0x00F0);
-                acc1[3] += yl[i+9] * (q1 & 0xF000);
-                acc2[0] += yh[i+0] * (q2 & 0x000F);
-                acc2[1] += yh[i+1] * (q2 & 0x0F00);
-                acc2[2] += yh[i+8] * (q2 & 0x00F0);
-                acc2[3] += yh[i+9] * (q2 & 0xF000);
-            }
-
-            uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
-            uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
-            uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
-            uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
-            uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
-            uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
-            uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
-            uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
-
-            float dall = float(inA[blk_idx + row_idx].d);
-            float dmin = float(inA[blk_idx + row_idx].dmin);
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 +
-                               (acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f +
-                               (acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 +
-                               (acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
-                dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
-        }
-
-        y4 += 4 * QK_K;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = subgroupAdd(sumf[row]);
-        if (subgroupElect()) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
-        }
-    }
-}

ggml/src/llamafile/sgemm.cpp

@@ -942,36 +942,6 @@ class tinyBLAS_Q0_AVX {
         return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
     }
 
-    inline __m256i load(const block_q5_0 *b) {
-        return _mm256_or_si256(denibble(b->qs), bittobyte(b->qh));
-    }
-
-    inline __m128i load0(const block_q5_0* b) {
-        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
-        uint32_t x32;
-        memcpy(&x32, b->qh, sizeof(uint32_t));
-        __m128i qxl = _mm_and_si128(_mm_set1_epi8(15), x);
-        __m128i bytesl = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
-                                        _mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
-                                                     _mm_shuffle_epi8(_mm_set1_epi32(x32),
-                                                                      _mm_set_epi64x(0x0101010101010101, 0x0000000000000000))));
-        bytesl = _mm_andnot_si128(bytesl, _mm_set1_epi8((char)0xF0));
-        return _mm_or_si128(qxl, bytesl);
-    }
-
-    inline __m128i load1(const block_q5_0* b) {
-        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
-        uint32_t x32;
-        memcpy(&x32, b->qh, sizeof(uint32_t));
-        __m128i qxh = _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4));
-        __m128i bytesh = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
-                                        _mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
-                                                     _mm_shuffle_epi8(_mm_set1_epi32(x32),
-                                                                      _mm_set_epi64x(0x0303030303030303, 0x0202020202020202))));
-        bytesh = _mm_andnot_si128(bytesh, _mm_set1_epi8((char)0xF0));
-        return _mm_or_si128(qxh, bytesh);
-    }
-
     inline __m256i load(const block_iq4_nl *b) {
         return MM256_SET_M128I(load1(b), load0(b));
     }
@@ -1003,17 +973,6 @@ class tinyBLAS_Q0_AVX {
                                                         _mm_srli_epi16(x, 4), 1));
     }
 
-    static inline __m256i bittobyte(const uint8_t *p) {
-        uint32_t x32;
-        memcpy(&x32, p, sizeof(uint32_t));
-        __m256i bytes = _mm256_cmpeq_epi8(_mm256_set1_epi64x(-1),
-                                          _mm256_or_si256(_mm256_set1_epi64x(0x7fbfdfeff7fbfdfe),
-                                                          _mm256_shuffle_epi8(_mm256_set1_epi32(x32),
-                                                                              _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202,
-                                                                                                0x0101010101010101, 0x0000000000000000))));
-        return _mm256_andnot_si256(bytes, _mm256_set1_epi8((char)0xF0));
-    }
-
     const TA *const A;
     const TB *const B;
     TC *const C;
@@ -1223,22 +1182,6 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
 #endif
     }
 
-    case GGML_TYPE_Q5_0: {
-        if (Btype != GGML_TYPE_Q8_0)
-            return false;
-#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
-        tinyBLAS_Q0_AVX<block_q5_0, block_q8_0, float> tb{
-            k, (const block_q5_0 *)A, lda,
-            (const block_q8_0 *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#else
-        return false;
-#endif
-    }
-
     case GGML_TYPE_IQ4_NL: {
         if (Btype != GGML_TYPE_Q8_0)
             return false;

ggml/src/vulkan-shaders/pool2d.comp

@@ -1,74 +0,0 @@
-#version 450
-
-#include "types.comp"
-
-#extension GL_EXT_shader_16bit_storage : require
-
-layout(push_constant) uniform parameter {
-    uint IW; uint IH;
-    uint OW; uint OH;
-    uint OC;
-    uint pelements;
-    uint op;
-    int k0; int k1;
-    int s0; int s1;
-    int p0; int p1;
-} p;
-
-#define BLOCK_SIZE 512
-#define FLT_MAX 3.402823466e+38F
-#define OP_POOL_MAX 0u
-#define OP_POOL_AVG 1u
-
-layout (local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
-
-layout(binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout(binding = 1) writeonly buffer D {D_TYPE data_d[];};
-
-void main() {
-    const uint idx = gl_GlobalInvocationID.x;
-    if (idx >= p.pelements) {
-        return;
-    }
-
-    const uint O_HW = p.OW * p.OH;
-
-    const uint nc = idx / O_HW;
-    const uint cur_oh = (idx % O_HW) / p.OW;
-    const uint cur_ow = (idx % O_HW) % p.OW;
-
-    const int start_h = int(cur_oh) * p.s0 - p.p0;
-    const uint bh = max(start_h, 0);
-    const uint eh = min(start_h + p.k0, p.IH);
-
-    const int start_w = int(cur_ow) * p.s1 - p.p1;
-    const uint bw = max(start_w, 0);
-    const uint ew = min(start_w + p.k1, p.IW);
-
-    const float scale = 1.0 / float(p.k0 * p.k1);
-    float res;
-
-    if (p.op == OP_POOL_AVG) {
-        res = 0.0;
-    } else if (p.op == OP_POOL_MAX) {
-        res = -FLT_MAX;
-    } else {
-        return;
-    }
-
-    #pragma unroll
-    for (uint i = bh; i < eh; i++) {
-        #pragma unroll
-        for (uint j = bw; j < ew; j++) {
-            const float cur = D_TYPE(data_a[nc * p.IH * p.IW + i * p.IW + j]);
-
-            if (p.op == OP_POOL_AVG) {
-                res += cur * scale;
-            } else if (p.op == OP_POOL_MAX) {
-                res = max(res, cur);
-            }
-        }
-    }
-
-    data_d[nc * O_HW + cur_oh * p.OW + cur_ow] = res;
-}

ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -493,10 +493,6 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
-
-    tasks.push_back(std::async(std::launch::async, [=] {
-        string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
-    }));
 }
 
 void write_output_files() {

include/llama.h

@@ -205,7 +205,7 @@ extern "C" {
     enum llama_split_mode {
         LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
         LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW   = 2, // split layers and KV across GPUs, use tensor parallelism if supported
+        LLAMA_SPLIT_MODE_ROW   = 2, // split rows across GPUs
     };
 
     // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
@@ -274,7 +274,10 @@ extern "C" {
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
-        // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE
+        // main_gpu interpretation depends on split_mode:
+        // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model
+        // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results
+        // LLAMA_SPLIT_MODE_LAYER: ignored
         int32_t main_gpu;
 
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
@@ -1084,6 +1087,9 @@ extern "C" {
     /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
     LLAMA_API struct llama_sampler * llama_sampler_init_min_p      (float   p, size_t min_keep);
 
+    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
+    LLAMA_API struct llama_sampler * llama_sampler_init_tail_free  (float   z, size_t min_keep);
+
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
     LLAMA_API struct llama_sampler * llama_sampler_init_typical    (float   p, size_t min_keep);
 
@@ -1135,16 +1141,6 @@ extern "C" {
                                 bool   penalize_nl,     // consider newlines as a repeatable token
                                 bool   ignore_eos);     // ignore the end-of-sequence token
 
-    ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
-    LLAMA_API struct llama_sampler *    llama_sampler_init_dry(
-            const struct llama_model *  model,
-                               float    dry_multiplier,
-                               float    dry_base,
-                             int32_t    dry_allowed_length,
-                             int32_t    dry_penalty_last_n,
-                          const char ** seq_breakers,
-                              size_t    num_breakers);
-
     LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
                              int32_t   n_vocab,
                              int32_t   n_logit_bias,

scripts/compare-llama-bench.py

@@ -20,7 +20,7 @@ logger = logging.getLogger("compare-llama-bench")
 # Properties by which to differentiate results per commit:
 KEY_PROPERTIES = [
     "cpu_info", "gpu_info", "n_gpu_layers", "cuda", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas",
-    "blas", "model_filename", "model_type", "n_batch", "n_ubatch", "embeddings", "n_threads",
+    "blas", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_ubatch", "embeddings", "n_threads",
     "type_k", "type_v", "use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen"
 ]
 

scripts/run-with-preset.py

@@ -20,7 +20,7 @@ CLI_ARGS_LLAMA_CLI_PERPLEXITY = [
     "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt",
     "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "repeat-last-n",
     "repeat-penalty", "reverse-prompt", "rope-freq-base", "rope-freq-scale", "rope-scale", "seed",
-    "simple-io", "tensor-split", "threads", "temp", "top-k", "top-p", "typical",
+    "simple-io", "tensor-split", "threads", "temp", "tfs", "top-k", "top-p", "typical",
     "verbose-prompt"
 ]
 

scripts/sync-ggml-am.sh

@@ -76,7 +76,6 @@ while read c; do
         src/ggml*.m \
         src/ggml*.metal \
         src/ggml*.cu \
-        src/ggml-amx/* \
         src/ggml-cann/* \
         src/ggml-cuda/* \
         src/ggml-sycl/* \
@@ -122,8 +121,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-aarch64.c      -> ggml/src/ggml-aarch64.c
     # src/ggml-aarch64.h      -> ggml/src/ggml-aarch64.h
     # src/ggml-alloc.c        -> ggml/src/ggml-alloc.c
-    # src/ggml-amx/*          -> ggml/src/ggml-amx/
-    # src/ggml-amx.cpp        -> ggml/src/ggml-amx.cpp
     # src/ggml-backend-impl.h -> ggml/src/ggml-backend-impl.h
     # src/ggml-backend.cpp    -> ggml/src/ggml-backend.cpp
     # src/ggml-cann/*         -> ggml/src/ggml-cann/
@@ -144,7 +141,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     #
     # include/ggml.h         -> ggml/include/ggml.h
     # include/ggml-alloc.h   -> ggml/include/ggml-alloc.h
-    # include/ggml-amx.h     -> ggml/include/ggml-amx.h
     # include/ggml-backend.h -> ggml/include/ggml-backend.h
     # include/ggml-blas.h    -> ggml/include/ggml-blas.h
     # include/ggml-cann.h    -> ggml/include/ggml-cann.h
@@ -172,8 +168,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.c/\1ggml\/src\/ggml-aarch64.c/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.h/\1ggml\/src\/ggml-aarch64.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-alloc\.c/\1ggml\/src\/ggml-alloc.c/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-amx\//\1ggml\/src\/ggml-amx\//g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml-amx\.cpp/\1ggml\/src\/ggml-amx.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-backend-impl\.h/\1ggml\/src\/ggml-backend-impl.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-backend\.cpp/\1ggml\/src\/ggml-backend.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
@@ -193,7 +187,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)src\/vulkan-shaders\//\1ggml\/src\/vulkan-shaders\//g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml\.h/\1ggml\/include\/ggml.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-alloc\.h/\1ggml\/include\/ggml-alloc.h/g' \
-        -e 's/([[:space:]]|[ab]\/)include\/ggml-amx\.h/\1ggml\/include\/ggml-amx.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-backend\.h/\1ggml\/include\/ggml-backend.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-blas\.h/\1ggml\/include\/ggml-blas.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-cann\.h/\1ggml\/include\/ggml-cann.h/g' \

scripts/sync-ggml.last

@@ -1 +1 @@
-162e232411ee98ceb0cccfa84886118d917d2123
+6dccc647264f5429df2624f36138f601e7ce23e5

scripts/sync-ggml.sh

@@ -8,8 +8,6 @@ cp -rpv ../ggml/src/ggml.c              ./ggml/src/ggml.c
 cp -rpv ../ggml/src/ggml-aarch64.c      ./ggml/src/ggml-aarch64.c
 cp -rpv ../ggml/src/ggml-aarch64.h      ./ggml/src/ggml-aarch64.h
 cp -rpv ../ggml/src/ggml-alloc.c        ./ggml/src/ggml-alloc.c
-cp -rpv ../ggml/src/ggml-amx/*          ./ggml/src/ggml-amx/
-cp -rpv ../ggml/src/ggml-amx.cpp        ./ggml/src/ggml-amx.cpp
 cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml/src/ggml-backend-impl.h
 cp -rpv ../ggml/src/ggml-backend.cpp    ./ggml/src/ggml-backend.cpp
 cp -rpv ../ggml/src/ggml-cann/*         ./ggml/src/ggml-cann/
@@ -31,7 +29,6 @@ cp -rpv ../ggml/src/vulkan-shaders/*    ./ggml/src/vulkan-shaders/
 
 cp -rpv ../ggml/include/ggml.h         ./ggml/include/ggml.h
 cp -rpv ../ggml/include/ggml-alloc.h   ./ggml/include/ggml-alloc.h
-cp -rpv ../ggml/include/ggml-amx.h     ./ggml/include/ggml-amx.h
 cp -rpv ../ggml/include/ggml-backend.h ./ggml/include/ggml-backend.h
 cp -rpv ../ggml/include/ggml-blas.h    ./ggml/include/ggml-blas.h
 cp -rpv ../ggml/include/ggml-cann.h    ./ggml/include/ggml-cann.h

src/llama-sampling.cpp

@@ -113,7 +113,7 @@ static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
 }
 
 static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) {
-    // TODO: move bucket sort to separate function so that top_p/typical/softmax first is equally fast
+    // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast
     // if (k >= (int32_t)cur_p->size) {
     //     return;
     // }
@@ -733,6 +733,101 @@ struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) {
     };
 }
 
+// tail-free
+
+struct llama_sampler_tail_free {
+    const float  z;
+    const size_t min_keep;
+};
+
+static const char * llama_sampler_tail_free_name(const struct llama_sampler * /*smpl*/) {
+    return "tail-free";
+}
+
+static void llama_sampler_tail_free_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_tail_free *) smpl->ctx;
+
+    if (ctx->z >= 1.0f || cur_p->size <= 2) {
+        return;
+    }
+
+    llama_sampler_softmax_impl(cur_p);
+
+    // Compute the first and second derivatives
+    std::vector<float> first_derivatives(cur_p->size - 1);
+    std::vector<float> second_derivatives(cur_p->size - 2);
+
+    for (size_t i = 0; i < first_derivatives.size(); ++i) {
+        first_derivatives[i] = cur_p->data[i].p - cur_p->data[i + 1].p;
+    }
+    for (size_t i = 0; i < second_derivatives.size(); ++i) {
+        second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1];
+    }
+
+    // Calculate absolute value of second derivatives
+    for (size_t i = 0; i < second_derivatives.size(); ++i) {
+        second_derivatives[i] = std::abs(second_derivatives[i]);
+    }
+
+    // Normalize the second derivatives
+    {
+        const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f);
+
+        if (second_derivatives_sum > 1e-6f) {
+            for (float & value : second_derivatives) {
+                value /= second_derivatives_sum;
+            }
+        } else {
+            for (float & value : second_derivatives) {
+                value = 1.0f / second_derivatives.size();
+            }
+        }
+    }
+
+    float cum_sum = 0.0f;
+    size_t last_idx = cur_p->size;
+    for (size_t i = 0; i < second_derivatives.size(); ++i) {
+        cum_sum += second_derivatives[i];
+
+        // Check if the running sum is greater than z or if we have kept at least min_keep tokens
+        if (cum_sum > ctx->z && i >= ctx->min_keep) {
+            last_idx = i;
+            break;
+        }
+    }
+
+    // Resize the output vector to keep only the tokens above the tail location
+    cur_p->size = last_idx;
+}
+
+static struct llama_sampler * llama_sampler_tail_free_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_tail_free *) smpl->ctx;
+    return llama_sampler_init_tail_free(ctx->z, ctx->min_keep);
+}
+
+static void llama_sampler_tail_free_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_tail_free *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_tail_free_i = {
+    /* .name   = */ llama_sampler_tail_free_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_tail_free_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_tail_free_clone,
+    /* .free   = */ llama_sampler_tail_free_free,
+};
+
+struct llama_sampler * llama_sampler_init_tail_free(float z, size_t min_keep) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_tail_free_i,
+        /* .ctx   = */ new llama_sampler_tail_free {
+            /* .z        = */ z,
+            /*. min_keep = */ min_keep,
+        },
+    };
+}
+
 // typical
 
 struct llama_sampler_typical {
@@ -1588,397 +1683,6 @@ struct llama_sampler * llama_sampler_init_penalties(
     };
 }
 
-// DRY
-
-struct llama_sampler_dry {
-    int32_t total_context_size;
-
-    const float   dry_multiplier;
-    const float   dry_base;
-    const int32_t dry_allowed_length;
-    const int32_t dry_penalty_last_n;
-
-    std::unordered_multimap<llama_token, std::vector<llama_token>> dry_processed_breakers;
-    std::vector<int> dry_repeat_count;
-    std::unordered_map<llama_token, int> dry_max_token_repeat;
-    ring_buffer<llama_token> last_tokens;
-};
-
-// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
-static void get_overlapping_token_sequences(const llama_vocab & vocab, const std::string& str, std::unordered_multimap<llama_token, std::vector<llama_token>>& token_sequences, int max_tail_len = -1) {
-    for (llama_token token_id = 0; token_id < (llama_token)vocab.n_vocab; token_id++) {
-        std::string word = llama_detokenize(vocab, {token_id}, true);
-        if (word.find(str) != std::string::npos) {
-            token_sequences.emplace(token_id, std::vector<llama_token>());
-        } else {
-            size_t word_len = word.size(), str_len = str.size();
-            size_t pos = -1;
-            while ((pos = word.find(str[0], pos + 1)) != std::string::npos) {
-                bool match = true;
-                size_t i;
-                for (i = 1; i < str_len && i + pos < word_len; ++i) {
-                    if (word[pos + i] != str[i]) {
-                        match = false;
-                        break;
-                    }
-                }
-                if (match) {
-                    std::vector<llama_token> tokenization = llama_tokenize_internal(vocab, str.substr(i), false, false);
-                    if (max_tail_len >= 0 && tokenization.size() > (size_t)max_tail_len) {
-                        tokenization.resize(max_tail_len);
-                    }
-
-                    // Ensure we don't already have a duplicate matching tokenization
-                    auto its = token_sequences.equal_range(token_id);
-                    bool found = false;
-                    for (auto it = its.first; it != its.second; ++it) {
-                        if (tokenization == it->second) {
-                            found = true;
-                            break;
-                        }
-                    }
-                    if (!found) {
-                        token_sequences.emplace(token_id, tokenization);
-                    }
-                }
-            }
-        }
-    }
-}
-
-static const char * llama_sampler_dry_name(const struct llama_sampler * /*smpl*/) {
-    return "dry";
-}
-
-static void llama_sampler_dry_accept(struct llama_sampler * smpl, llama_token token) {
-    auto * ctx = (llama_sampler_dry *) smpl->ctx;
-    if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
-        return;
-    }
-
-    ctx->last_tokens.push_back(token);
-}
-
-// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
-static void llama_sampler_dry_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    auto * ctx = (llama_sampler_dry *) smpl->ctx;
-
-    if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
-        return;
-    }
-
-    int32_t effective_dry_penalty_last_n = (ctx->dry_penalty_last_n == -1) ? ctx->total_context_size : std::max(ctx->dry_penalty_last_n, 0);
-    int last_n_repeat = std::min(std::min((int)ctx->last_tokens.size(), effective_dry_penalty_last_n), ctx->total_context_size);
-
-    if (last_n_repeat <= ctx->dry_allowed_length) {
-        return;
-    }
-
-    ctx->dry_repeat_count.assign(last_n_repeat, 0);
-    ctx->dry_max_token_repeat.clear();
-
-    // Step 1: Look for restart sequences to limit the maximum repetition length.
-    // Work backwards through the context looking for any token that begins a restart sequence.
-    //
-    // The collection `restart_sequences` is a mapping from a "head" token to all "tail"
-    // sequences that together comprise a restart sequence. This allows us to quickly check
-    // whether each token is the head of a complete sequence. Most restart sequences are actually
-    // a single token, and for these the "tail" is an empty vector.
-    //
-    // If the token is a "head", test all restart sequences that begin with this token
-    // (there will often only be one sequence for each token, but if sequences like 'aaaq1' and
-    // 'aaa1' are used as restart strings, both could start with 'aaa' when tokenized). The
-    // longest matching sequence (if any) is used to limit the maximum repetition length.
-    //
-    // Note that in the case case of a short sequence contained in a longer one, this might fail to
-    // find the smallest value for `rep_limit`. For example, if 'amniotic' and 'ni' are both used as
-    // restart sequences, 'ni' will be found first, and since it's shorter it will fail to suppress
-    // 'otic'. This is a minor issue since fully contained restart sequences are likely to be rare.
-    //
-    // This is theoretically worst-case O(N^2) for arbitrary restart sequences, which is why we
-    // have already clamped the maximum tail sequence length when generating `restart_sequences`.
-    // With clamping, this scan is O(N) in the context length.
-
-    int rep_limit = last_n_repeat;
-    for (int i = 0; i < last_n_repeat; ++i) {
-        llama_token token = ctx->last_tokens.rat(i);
-        auto its = ctx->dry_processed_breakers.equal_range(token);
-        if (its.first == ctx->dry_processed_breakers.end()) {
-            continue;
-        }
-        int longest_match = -1;
-        for (auto it = its.first; it != its.second; ++it) {
-            // Note that (*it) does not contain the head character, so seq_len will be
-            // the restart sequence length minus 1.
-            // In the common case of a single-token restart sequence, (*it) will be empty
-            // and we will trivially match.
-            int seq_len = (int)it->second.size();
-            if (seq_len > longest_match && seq_len <= (int)i) {
-                bool match = true;
-                for (int offset = 0; offset < seq_len; ++offset) {
-                    // The -1 when indexing `last_tokens` is because we already matched the head.
-                    if (it->second[offset] != ctx->last_tokens.rat(i - offset - 1)) {
-                        match = false;
-                        break;
-                    }
-                }
-                if (match) {
-                    longest_match = seq_len;
-                }
-            }
-        }
-        if (longest_match >= 0) {
-            // We found a restart sequence starting `i` tokens from the end and continuing for
-            // `longest_match` tokens.
-            rep_limit = i - longest_match;
-            break;
-        }
-    }
-    if (rep_limit < ctx->dry_allowed_length) {
-        return;
-    }
-
-    // Step 2: Iterate in reverse over the last N tokens of the context, using the "Z-algorithm" (in
-    // the reverse direction) to efficiently compute the positions and lengths of suffixes appearing
-    // elsewhere in the context. We limit the suffix length to `rep_limit` to respect restart sequences.
-    //
-    // This algorithm is not currently documented on Wikipedia, but there is a clear description here:
-    // https://ivanyu.me/blog/2014/10/15/z-algorithm/
-    //
-    // The code below is adapted from the public domain implementation by the same author here:
-    // https://github.com/ivanyu/string-algorithms/blob/master/z_algorithm.py
-    //
-    // Example:
-    // Last N tokens: a b c c b c y a b c
-    // Repeat counts: 0 0 3 1 0 2 0 0 0 0
-    //                    ^
-    //   This `3` means that the last three tokens of the context (a b c) also appear here.
-    //
-    // This step is worst case O(N) since the Z-algorithm is linear, despite the appearance of nested
-    // for/while loops. This can be seen by observing that the `lt` and `rt` bounds are set after each
-    // repeated suffix is detected (i.e. after each while loop when n > 0). These bound variables
-    // ensure that the inner while loops only examine each token in the context once as the outer
-    // for loop iterates over the context.
-
-    {
-        const int last = last_n_repeat - 1;
-        int rt = 0, lt = 0;
-
-        for (int k = 1; k < last_n_repeat; ++k) {
-            if (k > rt) {
-                // If k is outside the current Z-box, do naive computation.
-                int n = 0;
-                while (n + k < last_n_repeat && ctx->last_tokens.rat(n) == ctx->last_tokens.rat(n+k)) {
-                    ++n;
-                }
-                ctx->dry_repeat_count[last - k] = std::min(n, rep_limit);
-                if (n > 0) {
-                    lt = k;
-                    rt = k+n-1;
-                }
-            } else {
-                // If k is inside the current Z-box, consider two cases.
-
-                int p = k - lt; // Pair index.
-                int right_part_len = rt - k + 1;
-
-                if (ctx->dry_repeat_count[last - p] < right_part_len) {
-                    int n = std::min(ctx->dry_repeat_count[last - p], rep_limit);
-                    ctx->dry_repeat_count[last - k] = n;
-                } else {
-                    int i = rt + 1;
-                    while (i < last_n_repeat && ctx->last_tokens.rat(i) == ctx->last_tokens.rat(i - k)) {
-                        i += 1;
-                    }
-
-                    int n = std::min(i - k, rep_limit);
-                    ctx->dry_repeat_count[last - k] = n;
-                    lt = k;
-                    rt = i - 1;
-                }
-            }
-        }
-    }
-
-    // Step 3: Iterate over dry_repeat_count and last_tokens, examining the maximum repeat length
-    // that would be generated by emitting each new token that would extend a sequence.
-    //
-    // Following the same example as above:
-    // Last N tokens: a b c c b c y a b c
-    // Repeat counts: 0 0 3 1 0 2 0 0 0 0
-    //
-    // For each non-zero, look ahead one token. This token, if emitted, would extend the repetition.
-    // c: 3 -> 4 (from `a b c` to `a b c c`)
-    // b: 1 -> 2 (from `c` to `c b`)
-    // y: 2 -> 3 (from `b c` to `b c y`)
-
-    for (int i = 0; i < last_n_repeat - 1; ++i) {
-        int repeat_len = ctx->dry_repeat_count[i];
-        if (repeat_len >= ctx->dry_allowed_length) {
-            // This token ends a repeat, so the next token would continue one.
-            // By convention, the value of `repeat_len` only includes the tokens currently
-            // in the context, not the new token that would be added.
-            llama_token token = ctx->last_tokens.rat(last_n_repeat - 2 - i);
-            // Track the maximum sequence ending in this token.
-            const auto& it = ctx->dry_max_token_repeat.find(token);
-            if (it == ctx->dry_max_token_repeat.end() || it->second < repeat_len) {
-                ctx->dry_max_token_repeat[token] = repeat_len;
-            }
-        }
-    }
-
-    // Step 4: Apply logit penalties based on the maximum repeat length for relevant tokens.
-
-    // Prevent floating point overflow in `pow(penalty_base, exponent)` by clamping to `max_exponent`.
-    // Compute it from `penalty_base` and the approximate log of `std::numeric_limits<float>::max()`
-    const float FLOAT_MAX_LOG = 88.7228391f;
-    int max_exponent = 0;
-    if (ctx->dry_base > 1.000001f) {
-        max_exponent = FLOAT_MAX_LOG / std::log(ctx->dry_base);
-    }
-
-    for (size_t i = 0; i < cur_p->size; ++i) {
-        const auto& af_kvp = ctx->dry_max_token_repeat.find(cur_p->data[i].id);
-        if (af_kvp != ctx->dry_max_token_repeat.end()) {
-            // Check all sequence breakers starting with this token
-            auto range = ctx->dry_processed_breakers.equal_range(cur_p->data[i].id);
-            bool is_single_token_breaker = false;
-
-            for (auto it = range.first; it != range.second; ++it) {
-                if (it->second.empty()) {
-                    is_single_token_breaker = true;
-                    break;
-                }
-            }
-
-            // Apply penalty only if it's not a single-token sequence breaker
-            if (!is_single_token_breaker) {
-                int repeat_exp = af_kvp->second - ctx->dry_allowed_length;
-                if (max_exponent > 0 && repeat_exp > max_exponent) {
-                    repeat_exp = max_exponent;
-                }
-                float penalty = ctx->dry_multiplier * std::pow(ctx->dry_base, repeat_exp);
-                cur_p->data[i].logit -= penalty;
-            }
-        }
-    }
-
-    cur_p->sorted = false;
-}
-
-static void llama_sampler_dry_reset(struct llama_sampler * smpl) {
-    auto * ctx = (llama_sampler_dry *) smpl->ctx;
-    ctx->last_tokens.clear();
-    ctx->dry_repeat_count.clear();
-    ctx->dry_max_token_repeat.clear();
-}
-
-static struct llama_sampler * llama_sampler_dry_clone(const struct llama_sampler * smpl) {
-    const auto * ctx = (llama_sampler_dry *) smpl->ctx;
-
-    // nullptr is passed as vocab because it is only needed for raw sequence breaker processing, which we have already done and will be copying
-    auto * result = llama_sampler_init_dry(nullptr, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);
-    // Copy the state, including the processed breakers
-    {
-        auto * result_ctx = (llama_sampler_dry *) result->ctx;
-        result_ctx->dry_processed_breakers = ctx->dry_processed_breakers;
-        result_ctx->dry_repeat_count = ctx->dry_repeat_count;
-        result_ctx->dry_max_token_repeat = ctx->dry_max_token_repeat;
-        result_ctx->last_tokens = ctx->last_tokens;
-    }
-
-    return result;
-}
-
-static void llama_sampler_dry_free(struct llama_sampler * smpl) {
-    delete (llama_sampler_dry *) smpl->ctx;
-}
-
-static struct llama_sampler_i llama_sampler_dry_i = {
-    /* .name   = */ llama_sampler_dry_name,
-    /* .accept = */ llama_sampler_dry_accept,
-    /* .apply  = */ llama_sampler_dry_apply,
-    /* .reset  = */ llama_sampler_dry_reset,
-    /* .clone  = */ llama_sampler_dry_clone,
-    /* .free   = */ llama_sampler_dry_free,
-};
-
-struct llama_sampler * llama_sampler_init_dry_impl(const struct llama_vocab & vocab, int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
-    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? context_size : std::max(dry_penalty_last_n, 0);
-    std::unordered_multimap<llama_token, std::vector<llama_token>> processed_breakers;
-    const int MAX_CHAR_LEN = 40;
-    const int MAX_SEQ_LEN = 20;
-
-    const bool dry_enabled = (dry_multiplier != 0.0f && dry_base >= 1.0f && dry_penalty_last_n != 0);
-
-    if (dry_enabled && seq_breakers != nullptr && num_breakers > 0) {
-        // Process sequence breakers
-        for (size_t i = 0; i < num_breakers; ++i) {
-            if (seq_breakers[i] == nullptr || std::strlen(seq_breakers[i]) == 0) {
-                LLAMA_LOG_WARN("skipping null or empty DRY sequence breaker at index %zu\n", i);
-                continue;
-            }
-
-            std::string sequence_break(seq_breakers[i]);
-            if (sequence_break.empty()) {
-                LLAMA_LOG_WARN("skipping empty DRY sequence breaker\n");
-                continue;
-            }
-
-            if (sequence_break.size() > MAX_CHAR_LEN) {
-                LLAMA_LOG_WARN("truncating DRY sequence breaker to %d characters\n", MAX_CHAR_LEN);
-                sequence_break.resize(MAX_CHAR_LEN);
-            }
-
-            get_overlapping_token_sequences(vocab, sequence_break, processed_breakers, MAX_SEQ_LEN);
-        }
-    }
-
-    return new llama_sampler {
-        /* .iface = */ &llama_sampler_dry_i,
-        /* .ctx   = */ new llama_sampler_dry {
-            /* .total_context_size     = */ context_size,
-            /* .dry_multiplier         = */ dry_multiplier,
-            /* .dry_base               = */ dry_base,
-            /* .dry_allowed_length     = */ dry_allowed_length,
-            /* .dry_penalty_last_n     = */ dry_penalty_last_n,
-            /* .dry_processed_breakers = */ std::move(processed_breakers),
-            /* .dry_repeat_count       = */ dry_enabled ? std::vector<int>(effective_dry_penalty_last_n, 0) : std::vector<int>{},
-            /* .dry_max_token_repeat   = */ {},
-            /* .last_tokens            = */ dry_enabled ? ring_buffer<llama_token>(effective_dry_penalty_last_n) : ring_buffer<llama_token>(0),
-        },
-    };
-}
-
-// wrapper for test-sampling.cpp
-struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers) {
-    llama_vocab dummy_vocab;
-    auto * result = llama_sampler_init_dry_impl(dummy_vocab, context_size, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, NULL, 0);
-    auto * ctx = (llama_sampler_dry *) result->ctx;
-
-    // Process the token-based sequence breakers
-    ctx->dry_processed_breakers.clear();
-    if (seq_breakers.empty()) {
-        LLAMA_LOG_WARN("empty DRY sequence breakers list in llama_sampler_init_dry_testing\n");
-    } else {
-        for (const auto& breaker : seq_breakers) {
-            if (breaker.empty()) {
-                LLAMA_LOG_WARN("skipping DRY empty sequence breaker\n");
-                continue;
-            }
-            llama_token head_token = breaker[0];
-            std::vector<llama_token> tail_tokens(breaker.begin() + 1, breaker.end());
-            ctx->dry_processed_breakers.emplace(head_token, std::move(tail_tokens));
-        }
-
-        if (ctx->dry_processed_breakers.empty()) {
-            LLAMA_LOG_WARN("no valid DRY sequence breakers processed in llama_sampler_init_dry_testing\n");
-        }
-    }
-
-    return result;
-}
-
 // logit-bias
 
 struct llama_sampler_logit_bias {

src/llama-sampling.h

@@ -28,21 +28,3 @@ struct llama_sampler * llama_sampler_init_grammar_impl(
 
 struct llama_sampler * llama_sampler_init_infill_impl(
         const struct llama_vocab & vocab);
-
-struct llama_sampler * llama_sampler_init_dry_impl(
-        const struct llama_vocab &  vocab,
-                         int32_t    context_size,
-                           float    dry_multiplier,
-                           float    dry_base,
-                         int32_t    dry_allowed_length,
-                         int32_t    dry_penalty_last_n,
-                      const char ** seq_breakers,
-                          size_t    num_breakers);
-
-struct llama_sampler * llama_sampler_init_dry_testing(
-                         int32_t   context_size,
-                           float   dry_multiplier,
-                           float   dry_base,
-                         int32_t   dry_allowed_length,
-                         int32_t   dry_penalty_last_n,
-  const std::vector<std::vector<llama_token>>& seq_breakers);

src/llama-vocab.cpp

@@ -1966,19 +1966,3 @@ int32_t llama_detokenize_impl(
 
     return total <= text_len_max ? total : -total;
 }
-
-std::string llama_detokenize(const struct llama_vocab & vocab, const std::vector<llama_token> & tokens, bool special) {
-    std::string text;
-    text.resize(std::max(text.capacity(), tokens.size()));
-    int32_t n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
-    if (n_chars < 0) {
-        text.resize(-n_chars);
-        n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
-        GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
-    }
-
-    text.resize(n_chars);
-
-    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
-    return text;
-}

src/llama-vocab.h

@@ -163,8 +163,3 @@ int32_t llama_detokenize_impl(
                          int32_t   text_len_max,
                             bool   remove_special,
                             bool   unparse_special);
-
-std::string llama_detokenize(
-        const struct llama_vocab & vocab,
-  const std::vector<llama_token> & tokens,
-                            bool   special);

tests/test-chat-template.cpp

@@ -65,8 +65,6 @@ int main(void) {
         u8"{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}",
         // DeepSeek-V2
         "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{{ bos_token }}{% for message in messages %}{% if message['role'] == 'user' %}{{ 'User: ' + message['content'] + '\n\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: ' + message['content'] + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'] + '\n\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}",
-        // ibm-granite/granite-3.0-8b-instruct
-        "{%- if tools %}\n    {{- '<|start_of_role|>available_tools<|end_of_role|>\n' }}\n    {%- for tool in tools %}\n    {{- tool | tojson(indent=4) }}\n    {%- if not loop.last %}\n        {{- '\n\n' }}\n    {%- endif %}\n    {%- endfor %}\n    {{- '<|end_of_text|>\n' }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'system' %}\n    {{- '<|start_of_role|>system<|end_of_role|>' + message['content'] + '<|end_of_text|>\n' }}\n    {%- elif message['role'] == 'user' %}\n    {{- '<|start_of_role|>user<|end_of_role|>' + message['content'] + '<|end_of_text|>\n' }}\n    {%- elif message['role'] == 'assistant' %}\n    {{- '<|start_of_role|>assistant<|end_of_role|>'  + message['content'] + '<|end_of_text|>\n' }}\n    {%- elif message['role'] == 'assistant_tool_call' %}\n    {{- '<|start_of_role|>assistant<|end_of_role|><|tool_call|>' + message['content'] + '<|end_of_text|>\n' }}\n    {%- elif message['role'] == 'tool_response' %}\n    {{- '<|start_of_role|>tool_response<|end_of_role|>' + message['content'] + '<|end_of_text|>\n' }}\n    {%- endif %}\n    {%- if loop.last and add_generation_prompt %}\n    {{- '<|start_of_role|>assistant<|end_of_role|>' }}\n    {%- endif %}\n{%- endfor %}",
     };
     std::vector<std::string> expected_output = {
         // teknium/OpenHermes-2.5-Mistral-7B
@@ -111,8 +109,6 @@ int main(void) {
         u8"You are a helpful assistant<用户>Hello<AI>Hi there<用户>Who are you<AI>I am an assistant<用户>Another question<AI>",
         // DeepSeek-V2
         u8"You are a helpful assistant\n\nUser: Hello\n\nAssistant: Hi there<｜end▁of▁sentence｜>User: Who are you\n\nAssistant:    I am an assistant   <｜end▁of▁sentence｜>User: Another question\n\nAssistant:",
-        // ibm-granite/granite-3.0-8b-instruct
-        "<|start_of_role|>system<|end_of_role|>You are a helpful assistant<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Hello<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>Hi there<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Who are you<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>   I am an assistant   <|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Another question<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>\n",
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;

tests/test-sampling.cpp

@@ -10,8 +10,6 @@
 #include <string>
 #include <vector>
 
-extern struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers);
-
 static void dump(const llama_token_data_array * cur_p) {
     for (size_t i = 0; i < cur_p->size; i++) {
         printf("%d: %f (%f)\n", cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
@@ -105,6 +103,16 @@ static void test_top_p(const std::vector<float> & probs, const std::vector<float
     tester.check();
 }
 
+static void test_tfs(const std::vector<float> & probs, const std::vector<float> & probs_expected, float z) {
+    sampler_tester tester(probs, probs_expected);
+
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_tail_free(z, 1));
+    DUMP(&tester.cur_p);
+
+    tester.check();
+}
+
 static void test_min_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
     sampler_tester tester(probs, probs_expected);
 
@@ -159,29 +167,6 @@ static void test_penalties(
     tester.check();
 }
 
-static void test_dry(
-    const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
-    const std::vector<float> & expected_probs, float dry_multiplier, float dry_base,
-    int dry_allowed_length, int dry_penalty_last_n,
-    const std::vector<std::vector<llama_token>> & seq_breakers
-) {
-    GGML_ASSERT(probs.size() == expected_probs.size());
-
-    sampler_tester tester(probs, expected_probs);
-
-    auto * sampler = llama_sampler_init_dry_testing(1024, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, seq_breakers);
-
-    for (size_t i = 0; i < last_tokens.size(); i++) {
-        llama_sampler_accept(sampler, last_tokens[i]);
-    }
-
-    DUMP(&tester.cur_p);
-    tester.apply(sampler);
-    tester.apply(llama_sampler_init_dist(0));
-    DUMP(&tester.cur_p);
-    tester.check();
-}
-
 static void test_sampler_queue(const size_t n_vocab, const std::string & samplers_sequence, const int top_k, const float top_p, const float min_p
 ) {
     sampler_tester tester(n_vocab);
@@ -192,6 +177,7 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler
     for (auto s : samplers_sequence) {
         switch (s){
             case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break;
+            case 'f': GGML_ABORT("tail_free test not implemented");
             case 'y': GGML_ABORT("typical test not implemented");
             case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break;
             case 'm': tester.apply(llama_sampler_init_min_p(min_p, 1)); break;
@@ -288,11 +274,12 @@ static void test_perf() {
         data.emplace_back(llama_token_data{i, logit, 0.0f});
     }
 
-    BENCH(llama_sampler_init_top_k  (40),                     data, 32);
-    BENCH(llama_sampler_init_top_p  (0.8f, 1),                data, 32);
-    BENCH(llama_sampler_init_min_p  (0.2f, 1),                data, 32);
-    BENCH(llama_sampler_init_typical(0.5f, 1),                data, 32);
-    BENCH(llama_sampler_init_xtc    (1.0f, 0.1f, 1, 1),       data, 32);
+    BENCH(llama_sampler_init_top_k    (40),                     data, 32);
+    BENCH(llama_sampler_init_top_p    (0.8f, 1),                data, 32);
+    BENCH(llama_sampler_init_min_p    (0.2f, 1),                data, 32);
+    BENCH(llama_sampler_init_tail_free(0.5f, 1),                data, 32);
+    BENCH(llama_sampler_init_typical  (0.5f, 1),                data, 32);
+    BENCH(llama_sampler_init_xtc      (1.0f, 0.1f, 1, 1),       data, 32);
 }
 
 int main(void) {
@@ -331,6 +318,10 @@ int main(void) {
     printf("XTC should not:\n");
     test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.4f, 0.3f, 0.2f, 0.1f},              0.99f, 0.39f);
 
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f);
+
     test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f);
     test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);
 
@@ -342,13 +333,6 @@ int main(void) {
     test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
     test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f);
 
-
-    test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1}, {0.25f, 0.25f, 0.25f, 0.25f}, 1.0f, 1.1f, 2, 4, {});
-    test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1, 2, 0, 1}, {0.296923f, 0.296923f, 0.296923f, 0.109232f}, 1.0f, 1.1f, 2, 5, {});
-    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 2, 6, {{3}});
-    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.241818f, 0.241818f, 0.032727f}, 2.0f, 1.1f, 2, 5, {});
-    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 4, 7, {});
-
     test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
     test_sampler_queue(10000, "k",     1, 1.0f, 1.0f);
     test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f);