vulkan: fix coopmat2 flash attention for non-contiguous inputs (#11281)

Add code similar to mul_mm_cm2 to force alignment of strides, to avoid
a performance regression.

Add noncontiguous FA tests in test-backend-ops.

Fixes #11268.

.gitignore

@@ -18,6 +18,7 @@
 *.metallib
 *.o
 *.so
+*.swp
 *.tmp
 
 # IDE / OS

README.md

@@ -204,6 +204,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 - [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
 - [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
+- [Kalavai](https://github.com/kalavai-net/kalavai-client) - Crowdsource end to end LLM deployment at any scale
 
 </details>
 

ci/run.sh

@@ -299,7 +299,7 @@ function gg_run_open_llama_7b_v2 {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -433,7 +433,7 @@ function gg_run_pythia_1_4b {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -564,7 +564,7 @@ function gg_run_pythia_2_8b {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -699,7 +699,7 @@ function gg_run_embd_bge_small {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -747,7 +747,7 @@ function gg_run_rerank_tiny {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
 
@@ -814,8 +814,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     mkdir -p ${mnt_models}
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
-    # Create a fresh python venv and enter it
-    if ! python -m venv "$MNT/venv"; then
+    # Create a fresh python3 venv and enter it
+    if ! python3 -m venv "$MNT/venv"; then
         echo "Error: Failed to create Python virtual environment at $MNT/venv."
         exit 1
     fi

common/arg.cpp

@@ -376,6 +376,30 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
     return devices;
 }
 
+static void add_rpc_devices(std::string servers) {
+    auto rpc_servers = string_split<std::string>(servers, ',');
+    if (rpc_servers.empty()) {
+        throw std::invalid_argument("no RPC servers specified");
+    }
+    ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
+    if (!rpc_reg) {
+        throw std::invalid_argument("failed to find RPC backend");
+    }
+    typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
+    ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
+    if (!ggml_backend_rpc_add_device_fn) {
+        throw std::invalid_argument("failed to find RPC device add function");
+    }
+    for (const auto & server : rpc_servers) {
+        ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
+        if (dev) {
+            ggml_backend_device_register(dev);
+        } else {
+            throw std::invalid_argument("failed to register RPC device");
+        }
+    }
+}
+
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
     auto ctx_arg = common_params_parser_init(params, ex, print_usage);
     const common_params params_org = ctx_arg.params; // the example can modify the default params
@@ -1385,7 +1409,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             {"--rpc"}, "SERVERS",
             "comma separated list of RPC servers",
             [](common_params & params, const std::string & value) {
-                params.rpc_servers = value;
+                add_rpc_devices(value);
+                GGML_UNUSED(params);
             }
         ).set_env("LLAMA_ARG_RPC"));
     }

common/common.cpp

@@ -1043,7 +1043,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
-    mparams.rpc_servers     = params.rpc_servers.c_str();
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;

common/common.h

@@ -246,7 +246,6 @@ struct common_params {
     std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
-    std::string rpc_servers          = ""; // comma separated list of RPC servers                           // NOLINT
 
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)

convert_hf_to_gguf.py

@@ -2882,6 +2882,66 @@ class InternLM2Model(Model):
             return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("InternLM3ForCausalLM")
+class InternLM3Model(Model):
+    model_arch = gguf.MODEL_ARCH.LLAMA
+
+    def set_vocab(self):
+        tokens, scores, toktypes = self._create_vocab_sentencepiece()
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+                if "added_tokens_decoder" in tokenizer_config_json:
+                    for token_id, token_data in tokenizer_config_json["added_tokens_decoder"].items():
+                        if token_data.get("special"):
+                            token_id = int(token_id)
+                            token = token_data["content"]
+                            special_vocab._set_special_token(token, token_id)
+                            # update eos token
+                            if token == '<|im_end|>' and "eos" in special_vocab.special_token_ids:
+                                special_vocab.special_token_ids["eos"] = token_id
+
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
+            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
+        self.gguf_writer.add_rope_dimension_count(rope_dim)
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear" or self.hparams["rope_scaling"].get("rope_type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        if name.endswith(("q_proj.weight", "q_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith(("k_proj.weight", "k_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @Model.register("BertModel", "BertForMaskedLM", "CamembertModel")
 class BertModel(Model):
     model_arch = gguf.MODEL_ARCH.BERT

examples/gguf-split/tests.sh

@@ -41,7 +41,7 @@ echo PASS
 echo
 
 # 2b. Test the sharded model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --n-predict 32
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --n-predict 32
 echo PASS
 echo
 
@@ -51,7 +51,7 @@ echo PASS
 echo
 
 # 3b. Test the merged model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-merge.gguf --n-predict 32
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge.gguf --n-predict 32
 echo PASS
 echo
 
@@ -61,7 +61,7 @@ echo PASS
 echo
 
 # 4b. Test the sharded model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --n-predict 32
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --n-predict 32
 echo PASS
 echo
 
@@ -71,7 +71,7 @@ echo
 #echo
 
 # 5b. Test the merged model is loading properly
-#$MAIN --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
+#$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
 #echo PASS
 #echo
 
@@ -81,7 +81,7 @@ echo PASS
 echo
 
 # 6b. Test the sharded model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --n-predict 32
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --n-predict 32
 echo PASS
 echo
 

examples/llama-bench/llama-bench.cpp

@@ -683,7 +683,7 @@ struct cmd_params_instance {
     bool               cpu_strict;
     int                poll;
     int                n_gpu_layers;
-    std::string        rpc_servers;
+    std::string        rpc_servers_str;
     llama_split_mode   split_mode;
     int                main_gpu;
     bool               no_kv_offload;
@@ -696,8 +696,37 @@ struct cmd_params_instance {
         llama_model_params mparams = llama_model_default_params();
 
         mparams.n_gpu_layers = n_gpu_layers;
-        if (!rpc_servers.empty()) {
-            mparams.rpc_servers = rpc_servers.c_str();
+        if (!rpc_servers_str.empty()) {
+            auto rpc_servers = string_split<std::string>(rpc_servers_str, ',');
+
+            // add RPC devices
+            if (!rpc_servers.empty()) {
+                ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
+                if (!rpc_reg) {
+                    fprintf(stderr, "%s: failed to find RPC backend\n", __func__);
+                    exit(1);
+                }
+
+                typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
+                ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
+                if (!ggml_backend_rpc_add_device_fn) {
+                    fprintf(stderr, "%s: failed to find RPC device add function\n", __func__);
+                    exit(1);
+                }
+                static std::vector<ggml_backend_dev_t> devices;
+                devices.clear();
+                for (const std::string & server : rpc_servers) {
+                    ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
+                    if (dev) {
+                        devices.push_back(dev);
+                    } else {
+                        fprintf(stderr, "%s: failed to add RPC device for server '%s'\n", __func__, server.c_str());
+                        exit(1);
+                    }
+                }
+                devices.push_back(nullptr);
+                mparams.devices = devices.data();
+            }
         }
         mparams.split_mode   = split_mode;
         mparams.main_gpu     = main_gpu;
@@ -708,7 +737,7 @@ struct cmd_params_instance {
     }
 
     bool equal_mparams(const cmd_params_instance & other) const {
-        return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers == other.rpc_servers &&
+        return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers_str == other.rpc_servers_str &&
                split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap &&
                tensor_split == other.tensor_split;
     }

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -347,6 +347,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
         jlong context_pointer,
         jlong batch_pointer,
         jstring jtext,
+        jboolean format_chat,
         jint n_len
     ) {
 
@@ -356,7 +357,8 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
     const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
 
-    const auto tokens_list = common_tokenize(context, text, 1);
+    bool parse_special = (format_chat == JNI_TRUE);
+    const auto tokens_list = common_tokenize(context, text, true, parse_special);
 
     auto n_ctx = llama_n_ctx(context);
     auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
@@ -368,7 +370,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     }
 
     for (auto id : tokens_list) {
-        LOGi("%s", common_token_to_piece(context, id).c_str());
+        LOGi("token: `%s`-> %d ", common_token_to_piece(context, id).c_str(), id);
     }
 
     common_batch_clear(*batch);

examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt

@@ -65,6 +65,7 @@ class LLamaAndroid {
         context: Long,
         batch: Long,
         text: String,
+        formatChat: Boolean,
         nLen: Int
     ): Int
 
@@ -115,10 +116,10 @@ class LLamaAndroid {
         }
     }
 
-    fun send(message: String): Flow<String> = flow {
+    fun send(message: String, formatChat: Boolean = false): Flow<String> = flow {
         when (val state = threadLocalState.get()) {
             is State.Loaded -> {
-                val ncur = IntVar(completion_init(state.context, state.batch, message, nlen))
+                val ncur = IntVar(completion_init(state.context, state.batch, message, formatChat, nlen))
                 while (ncur.value <= nlen) {
                     val str = completion_loop(state.context, state.batch, state.sampler, nlen, ncur)
                     if (str == null) {

examples/quantize/tests.sh

@@ -47,7 +47,7 @@ echo PASS
 echo
 
 # 3a. Test the requanted model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --n-predict 32
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --n-predict 32
 echo PASS
 echo
 
@@ -57,7 +57,7 @@ echo PASS
 echo
 
 # 4b. Test the requanted model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-requant-merge.gguf --n-predict 32
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-requant-merge.gguf --n-predict 32
 echo PASS
 echo
 

ggml/include/ggml-backend.h

@@ -203,6 +203,8 @@ extern "C" {
     // Backend registry
     //
 
+    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
+
     // Backend (reg) enumeration
     GGML_API size_t             ggml_backend_reg_count(void);
     GGML_API ggml_backend_reg_t ggml_backend_reg_get(size_t index);

ggml/include/ggml.h

@@ -1384,16 +1384,20 @@ extern "C" {
             float                 scale,
             float                 max_bias);
 
-    GGML_API struct ggml_tensor * ggml_soft_max_back(
+    GGML_API struct ggml_tensor * ggml_soft_max_ext_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            float                 scale,
+            float                 max_bias);
 
     // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
+    GGML_API struct ggml_tensor * ggml_soft_max_ext_back_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            float                 scale,
+            float                 max_bias);
 
     // rotary position embedding
     // if (mode & 1) - skip n_past elements (NOT SUPPORTED)

ggml/src/ggml-alloc.c

@@ -37,6 +37,7 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
     return true;
 }
 
+// ops that return true for this function must not use restrict pointers for their backend implementations
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
         case GGML_OP_SCALE:
@@ -52,8 +53,12 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_LOG:
         case GGML_OP_UNARY:
         case GGML_OP_ROPE:
+        case GGML_OP_ROPE_BACK:
+        case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM:
+        case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_SOFT_MAX:
+        case GGML_OP_SOFT_MAX_BACK:
             return true;
 
         default:

ggml/src/ggml-backend-impl.h

@@ -208,7 +208,6 @@ extern "C" {
 
     // Internal backend registry API
     GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
-    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
 
     // Add backend dynamic loading support to the backend
 

ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -5573,7 +5573,88 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     uint32_t utmp[4];
 
-#ifdef __ARM_NEON
+#ifdef __ARM_FEATURE_SVE
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
+
+        memcpy(utmp, x[i].scales, K_SCALE_SIZE);
+
+        uint32x2_t mins8 = { 0 };
+        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
+        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
+
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[0] &= kmask1;
+
+        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
+        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
+                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
+        sumf -= dmin * vaddvq_s32(prod);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const int vector_length = ggml_cpu_get_sve_cnt()*8;
+        const svuint8_t m4b = svdup_n_u8(0xf);
+        const svint32_t mzero = svdup_n_s32(0);
+        svint32_t sumi1 = svdup_n_s32(0);
+        svint32_t sumi1_1 = svdup_n_s32(0);
+        svint32_t sumi1_2 = svdup_n_s32(0);
+        svint32_t sumi2 = svdup_n_s32(0);
+        svint32_t sumi2_1 = svdup_n_s32(0);
+        svint32_t sumi2_2 = svdup_n_s32(0);
+        switch (vector_length) {
+            case 128:
+                {
+                    for (int j = 0; j < QK_K/64; ++j) {
+                        svint8_t q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4), m4b));
+                        svint8_t q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
+                        sumi1_1 = svmla_n_s32_x(svptrue_b32(), sumi1_1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
+                        q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4+16), m4b));
+                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
+                        sumi1_2 = svmla_n_s32_x(svptrue_b32(), sumi1_2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
+
+                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4), 4));
+                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
+                        sumi2_1 = svmla_n_s32_x(svptrue_b32(), sumi2_1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
+                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4+16), 4));
+                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
+                        sumi2_2 = svmla_n_s32_x(svptrue_b32(), sumi2_2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
+                        q4 += 32;
+                    }
+                    sumi1 = svadd_s32_x(svptrue_b32(), sumi1_1, sumi1_2);
+                    sumi2 = svadd_s32_x(svptrue_b32(), sumi2_1, sumi2_2);
+                    sumf += d * (svaddv_s32(svptrue_b32(), svadd_s32_x(svptrue_b32(), sumi1, sumi2)));
+                } break;
+            case 256:
+            case 512:
+                {
+                    for (int j = 0; j < QK_K/64; ++j) {
+                        const svuint8_t q4bits  = svld1_u8(svptrue_pat_b8(SV_VL32), q4); q4 += 32;
+                        svint8_t q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_pat_b8(SV_VL32), q4bits, m4b));
+                        svint8_t q8bytes = svld1_s8(svptrue_pat_b8(SV_VL32), q8); q8 += 32;
+                        sumi1 = svmla_n_s32_x(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
+
+                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q4bits, 4));
+                        q8bytes = svld1_s8(svptrue_pat_b8(SV_VL32), q8); q8 += 32;
+                        sumi2 = svmla_n_s32_x(svptrue_pat_b32(SV_VL8), sumi2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
+                    }
+                    sumf += d * (svaddv_s32(svptrue_pat_b32(SV_VL8), svadd_s32_x(svptrue_pat_b32(SV_VL8), sumi1, sumi2)));
+                } break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        }
+    }
+    *s = sumf;
+#elif __ARM_NEON
     const uint8x16_t m4b = vdupq_n_u8(0xf);
     const int32x4_t mzero = vdupq_n_s32(0);
 

ggml/src/ggml-cpu/ggml-cpu.c

@@ -3967,6 +3967,57 @@ static void ggml_compute_forward_dup_bytes(
     }
 }
 
+static void ggml_compute_forward_dup_q(
+        const struct ggml_compute_params * params,
+              struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const enum ggml_type type = src0->type;
+    ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
+
+    size_t qk = ggml_blck_size(type);
+    const int64_t nr = ggml_nelements(src1) / qk;
+
+    // destination must be contiguous in the first dimension
+    GGML_ASSERT(nb10 == ggml_type_size(dst->type));
+    // must either have first dimension large enough to hold a row, or fully contiguous
+    GGML_ASSERT((ne10 % qk) == 0 || ggml_is_contiguous(dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+
+        uint32_t i = ir * qk;
+
+        const int64_t i03 = i/(ne00 * ne01 * ne02);
+        const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+        const int64_t i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+        const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+        const int64_t x_offset = (i00/qk)*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+        const int64_t i13 = i/(ne10 * ne11 * ne12);
+        const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+        const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+        const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+        const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
+
+        dequantize_row_q(
+                (const void *) ((char *) src0->data + x_offset),
+                     (float *) ((char *)  dst->data + dst_offset), qk);
+    }
+}
+
 static void ggml_compute_forward_dup(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -3993,6 +4044,10 @@ static void ggml_compute_forward_dup(
             } break;
         default:
             {
+                if (ggml_is_quantized(src0->type) && dst->type == GGML_TYPE_F32) {
+                    ggml_compute_forward_dup_q(params, dst);
+                    break;
+                }
                 GGML_ABORT("fatal error");
             }
     }
@@ -6691,20 +6746,20 @@ static void ggml_compute_forward_silu_back_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * grad = dst->src[1];
+    const struct ggml_tensor * grad = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
     assert(ggml_is_contiguous_1(grad));
-    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(src1));
     assert(ggml_is_contiguous_1(dst));
-    assert(ggml_are_same_shape(src0, dst));
-    assert(ggml_are_same_shape(src0, grad));
+    assert(ggml_are_same_shape(src1, dst));
+    assert(ggml_are_same_shape(src1, grad));
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int nc = src0->ne[0];
-    const int nr = ggml_nrows(src0);
+    const int nc = src1->ne[0];
+    const int nr = ggml_nrows(src1);
 
     // rows per thread
     const int dr = (nr + nth - 1)/nth;
@@ -6716,7 +6771,7 @@ static void ggml_compute_forward_silu_back_f32(
     for (int i1 = ir0; i1 < ir1; i1++) {
         ggml_vec_silu_backward_f32(nc,
                 (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])),
+                (float *) ((char *) src1->data + i1*(src1->nb[1])),
                 (float *) ((char *) grad->data + i1*(grad->nb[1])));
 
 #ifndef NDEBUG
@@ -6895,7 +6950,7 @@ static void ggml_compute_forward_norm_f32(
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
-    GGML_ASSERT(eps > 0.0f);
+    GGML_ASSERT(eps >= 0.0f);
 
     // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
@@ -6966,7 +7021,7 @@ static void ggml_compute_forward_rms_norm_f32(
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
-    GGML_ASSERT(eps > 0.0f);
+    GGML_ASSERT(eps >= 0.0f);
 
     // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
@@ -7018,12 +7073,13 @@ static void ggml_compute_forward_rms_norm_back_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * src0 = dst->src[0]; // gradients from forward pass output
+    const struct ggml_tensor * src1 = dst->src[1]; // src1 from forward pass
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1));
 
     GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(src1->nb[0] == sizeof(float));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -7042,8 +7098,8 @@ static void ggml_compute_forward_rms_norm_back_f32(
                 const int64_t i12 = i02;
                 const int64_t i13 = i03;
 
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
-                const float * dz = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13);
+                const float * dz = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const float * x  = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13);
 
                 ggml_float sum_xx  = 0.0;
                 ggml_float sum_xdz = 0.0;
@@ -7066,9 +7122,9 @@ static void ggml_compute_forward_rms_norm_back_f32(
                 {
                     // z = rms_norm(x)
                     //
-                    // rms_norm(src0) =
+                    // rms_norm(src1) =
                     //     scale(
-                    //         src0,
+                    //         src1,
                     //         div(
                     //             1,
                     //             sqrt(
@@ -7076,13 +7132,13 @@ static void ggml_compute_forward_rms_norm_back_f32(
                     //                     scale(
                     //                         sum(
                     //                             sqr(
-                    //                                 src0)),
+                    //                                 src1)),
                     //                         (1.0/N)),
                     //                     eps))));
 
                     // postorder:
                     // ## op    args         grad
-                    // 00 param src0         grad[#00]
+                    // 00 param src1         grad[#00]
                     // 01 const 1
                     // 02 sqr   (#00)        grad[#02]
                     // 03 sum   (#02)        grad[#03]
@@ -7159,6 +7215,7 @@ static void ggml_compute_forward_rms_norm_back_f32(
                 // dx := scale(dx, rrms)
                 float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
 
+                // dx[i00] = (x*(-sum_xdz/sum_eps) + dz) / sqrtf(mean_eps)
                 ggml_vec_cpy_f32  (ne00, dx, x);
                 // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps);
                 ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps);
@@ -7750,12 +7807,13 @@ static void ggml_compute_forward_out_prod_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    GGML_ASSERT(ne0  == ne00);
-    GGML_ASSERT(ne1  == ne10);
-    GGML_ASSERT(ne2  == ne02);
-    GGML_ASSERT(ne02 == ne12);
-    GGML_ASSERT(ne3  == ne13);
-    GGML_ASSERT(ne03 == ne13);
+    GGML_ASSERT(ne0 == ne00);
+    GGML_ASSERT(ne1 == ne10);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    GGML_ASSERT(ne2 % ne02 == 0);
+    GGML_ASSERT(ne3 % ne03 == 0);
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == sizeof(float));
@@ -7797,6 +7855,10 @@ static void ggml_compute_forward_out_prod_f32(
     const int64_t blck_0 = MAX(GGML_VEC_MAD_UNROLL, 32);
     const int64_t blck_1 = 16;
 
+    // dps == dst per src0, used for group query attention
+    const int64_t dps2 = ne2 / ne02;
+    const int64_t dps3 = ne3 / ne03;
+
     for (int64_t bir = ir0; bir < ir1; bir += blck_1) {
         const int64_t bir1 = MIN(bir + blck_1, ir1);
         for (int64_t bi01 = 0; bi01 < ne01; bi01 += blck_0) {
@@ -7807,8 +7869,8 @@ static void ggml_compute_forward_out_prod_f32(
                 const int64_t i2 = (ir - i3*ne2*ne1)/ne1;
                 const int64_t i1 = (ir - i3*ne2*ne1 - i2*ne1);
 
-                const int64_t i02 = i2;
-                const int64_t i03 = i3;
+                const int64_t i02 = i2 / dps2;
+                const int64_t i03 = i3 / dps3;
 
                 //const int64_t i10 = i1;
                 const int64_t i12 = i2;
@@ -8906,9 +8968,9 @@ static void ggml_compute_forward_soft_max(
 }
 
 
-// ggml_compute_forward_soft_max_back
+// ggml_compute_forward_soft_max_ext_back
 
-static void ggml_compute_forward_soft_max_back_f32(
+static void ggml_compute_forward_soft_max_ext_back_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
@@ -8921,6 +8983,14 @@ static void ggml_compute_forward_soft_max_back_f32(
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_are_same_shape(src1, dst));
 
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (const float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float));
+
+    GGML_ASSERT(max_bias == 0.0f);
+
     // TODO: handle transposed/permuted matrices
 
     const int ith = params->ith;
@@ -8969,10 +9039,11 @@ static void ggml_compute_forward_soft_max_back_f32(
 
         // linear runtime, no additional memory
         float dot_y_dy = 0;
-        ggml_vec_dot_f32 (nc, &dot_y_dy, 0, y, 0, dy, 0, 1);
-        ggml_vec_cpy_f32 (nc, dx, dy);
-        ggml_vec_acc1_f32(nc, dx, -dot_y_dy);
-        ggml_vec_mul_f32 (nc, dx, dx, y);
+        ggml_vec_dot_f32  (nc, &dot_y_dy, 0, y, 0, dy, 0, 1);
+        ggml_vec_cpy_f32  (nc, dx, dy);
+        ggml_vec_acc1_f32 (nc, dx, -dot_y_dy);
+        ggml_vec_mul_f32  (nc, dx, dx, y);
+        ggml_vec_scale_f32(nc, dx, scale);
 
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
@@ -8983,7 +9054,7 @@ static void ggml_compute_forward_soft_max_back_f32(
     }
 }
 
-static void ggml_compute_forward_soft_max_back(
+static void ggml_compute_forward_soft_max_ext_back(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
@@ -8992,7 +9063,7 @@ static void ggml_compute_forward_soft_max_back(
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_soft_max_back_f32(params, dst);
+                ggml_compute_forward_soft_max_ext_back_f32(params, dst);
             } break;
         default:
             {
@@ -9985,9 +10056,10 @@ static void ggml_compute_forward_im2col_back_f32(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * src0 = dst->src[0]; // gradients of forward pass output
+    const struct ggml_tensor * src1 = dst->src[1]; // convolution kernel
 
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -10009,11 +10081,11 @@ static void ggml_compute_forward_im2col_back_f32(
     const int64_t IH = is_2D ? ne1 : 1;
     const int64_t IW = ne0;
 
-    const int64_t KH = is_2D ? ne01 : 1;
-    const int64_t KW = ne00;
+    const int64_t KH = is_2D ? ne11 : 1;
+    const int64_t KW = ne10;
 
-    const int64_t OH = is_2D ? ne12 : 1;
-    const int64_t OW = ne11;
+    const int64_t OH = is_2D ? ne02 : 1;
+    const int64_t OW = ne01;
 
     int ofs0 = is_2D ? nb3 : nb2;
     int ofs1 = is_2D ? nb2 : nb1;
@@ -10059,9 +10131,9 @@ static void ggml_compute_forward_im2col_back_f32(
                                     continue;
                                 }
 
-                                const float * const src_data = (const float *) src1->data
+                                const float * const grad_in = (const float *) src0->data
                                     + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
-                                grad += src_data[iic*(KH*KW) + ikh*KW + ikw];
+                                grad += grad_in[iic*(KH*KW) + ikh*KW + ikw];
                             }
                         }
                         float * dst_data = (float *)((char *) wdata + (in*ofs0 + iic*ofs1)); // [IH, IW]
@@ -12484,22 +12556,22 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-    const struct ggml_tensor * opt0 = dst->src[2];
+    const struct ggml_tensor * grad  = dst->src[0]; // gradient of forward pass output
+    const struct ggml_tensor * src0f = dst->src[1]; // src0 of forward pass
+    const struct ggml_tensor * src1f = dst->src[2]; // src1 of forward pass
 
     GGML_ASSERT(ggml_is_contiguous(dst));
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(src1));
-    GGML_ASSERT(ggml_is_contiguous(opt0));
-    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_is_contiguous(src0f));
+    GGML_ASSERT(ggml_is_contiguous(src1f));
+    GGML_ASSERT(ggml_is_contiguous(grad));
+    GGML_ASSERT(ggml_are_same_shape(src0f, src1f) && ggml_are_same_shape(src0f, dst));
 
     const int64_t ith = params->ith;
     const int64_t nth = params->nth;
 
     // TODO: handle transposed/permuted matrices
-    const int64_t nc = src0->ne[0];
-    const int64_t nr = ggml_nrows(src0);
+    const int64_t nc = src0f->ne[0];
+    const int64_t nr = ggml_nrows(src0f);
 
     // rows per thread
     const int64_t dr = (nr + nth - 1)/nth;
@@ -12508,12 +12580,12 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
     const int64_t ir0 = dr*ith;
     const int64_t ir1 = MIN(ir0 + dr, nr);
 
-    const float d_by_nr = ((const float *) opt0->data)[0] / (float) nr;
+    const float d_by_nr = ((const float *) grad->data)[0] / (float) nr;
 
     for (int64_t i1 = ir0; i1 < ir1; i1++) {
-        float * ds0 = (float *)((char *) dst->data  + i1*dst->nb[1]);
-        float * s0  = (float *)((char *) src0->data + i1*src0->nb[1]);
-        float * s1  = (float *)((char *) src1->data + i1*src1->nb[1]);
+        float       * ds0 = (float       *)((char       *) dst->data   + i1*dst->nb[1]);
+        const float * s0  = (const float *)((const char *) src0f->data + i1*src0f->nb[1]);
+        const float * s1  = (const float *)((const char *) src1f->data + i1*src1f->nb[1]);
 
 #ifndef NDEBUG
         for (int64_t i = 0; i < nc; ++i) {
@@ -12526,11 +12598,11 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
         // soft_max
         float max = -INFINITY;
         ggml_vec_max_f32(nc, &max, s0);
-        ggml_float sum = ggml_vec_soft_max_f32(nc, ds0, s0, max);
+        const ggml_float sum = ggml_vec_soft_max_f32(nc, ds0, s0, max);
         assert(sum > 0.0);
         ggml_vec_scale_f32(nc, ds0, 1.0/sum);
 
-        // grad(src0) = (softmax(src0) - src1) * grad(cross_entropy_loss(src0, src1)) / nr
+        // grad(src0f) = (softmax(src0f) - src1f) * grad(cross_entropy_loss(src0f, src1f)) / nr
         ggml_vec_sub_f32(nc, ds0, ds0, s1);
         ggml_vec_scale_f32(nc, ds0, d_by_nr);
 
@@ -12827,7 +12899,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_SOFT_MAX_BACK:
             {
-                ggml_compute_forward_soft_max_back(params, tensor);
+                ggml_compute_forward_soft_max_ext_back(params, tensor);
             } break;
         case GGML_OP_ROPE:
             {

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -403,6 +403,16 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
                 op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
+        case GGML_OP_SOFT_MAX_BACK: {
+            if (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type != GGML_TYPE_F32) {
+                return false;
+            }
+            float max_bias = 0.0f;
+
+            memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
+
+            return max_bias == 0.0f;
+        }
         case GGML_OP_IM2COL_BACK:
             return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
         case GGML_OP_OUT_PROD:

ggml/src/ggml-cuda/cross-entropy-loss.cu

@@ -5,95 +5,89 @@
 #include <cmath>
 #include <cstdint>
 
-static __global__ void cross_entropy_loss_f32(const float * logits, const float * labels, float * dst, const int nclasses, const int k) {
-    const int warp_id = threadIdx.x / WARP_SIZE;
-    const int lane_id = threadIdx.x % WARP_SIZE;
-    const int i0 = blockDim.x*blockIdx.x + warp_id*WARP_SIZE;
+template <bool use_shared>
+static __global__ void cross_entropy_loss_f32(
+        const float * __restrict__ logits, const float * __restrict__ labels, float * __restrict__ dst, const int nclasses, const int k) {
+    extern __shared__ float tmp[];
 
-    const int ne_tmp = WARP_SIZE*nclasses;
-
-    extern __shared__ float tmp_all[];
-    float * tmp_logits = tmp_all + (2*warp_id + 0)*ne_tmp;
-    float * tmp_labels = tmp_all + (2*warp_id + 1)*ne_tmp;
-
-    // Each warp first loads ne_tmp logits/labels into shared memory:
-    for (int i = lane_id; i < ne_tmp; i += WARP_SIZE) {
-        const int ig = i0*nclasses + i; // ig == i global
-
-        tmp_logits[i] = ig < k*nclasses ? logits[ig] : 0.0f;
-        tmp_labels[i] = ig < k*nclasses ? labels[ig] : 0.0f;
-    }
-
-    // Each thread in the warp then calculates the cross entropy loss for a single row.
-    // TODO: pad in order to avoid shared memory bank conflicts.
+    logits += int64_t(blockIdx.x)*nclasses;
+    labels += int64_t(blockIdx.x)*nclasses;
 
     // Find maximum for softmax:
-    float max = -INFINITY;
-    for (int i = 0; i < nclasses; ++i) {
-        max = fmaxf(max, tmp_logits[lane_id*nclasses + i]);
+    float max_logit = -INFINITY;
+    for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
+        const float val = logits[i];
+        max_logit = fmaxf(max_logit, val);
+
+        if (use_shared) {
+            tmp[i] = val;
+        }
     }
+    max_logit = warp_reduce_max(max_logit);
 
     // Calculate log(softmax(logits)) which is just logits - max:
     float sum = 0.0f;
-    for (int i = 0; i < nclasses; ++i) {
-        float val = tmp_logits[lane_id*nclasses + i] - max;
-        sum += expf(val);
-        tmp_logits[lane_id*nclasses + i] = val;
+    for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
+        const float logit_i = use_shared ? tmp[i] : logits[i];
+        sum += expf(logit_i - max_logit);
     }
+    sum = warp_reduce_sum(sum);
     sum = logf(sum);
 
     // log(exp(logits - max) / sum) = (logits - max) - log(sum)
     float loss = 0.0f;
-    for (int i = 0; i < nclasses; ++i) {
-        loss += (tmp_logits[lane_id*nclasses + i] - sum) * tmp_labels[lane_id*nclasses + i];
+    for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
+        const float logit_i = use_shared ? tmp[i] : logits[i];
+        loss += (logit_i - max_logit - sum) * labels[i];
     }
     loss = -warp_reduce_sum(loss) / (float)k;
 
-    __syncthreads();
-
-    if (lane_id == 0) {
-        tmp_all[warp_id] = loss;
-    }
-
-    __syncthreads();
-
-    if (warp_id != 0) {
-        return;
-    }
-
-    loss = lane_id < CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE/WARP_SIZE ? tmp_all[lane_id] : 0.0f;
-    loss = warp_reduce_sum(loss);
-
-    if (lane_id != 0) {
+    if (threadIdx.x != 0) {
         return;
     }
 
     dst[blockIdx.x] = loss;
 }
 
-static __global__ void cross_entropy_loss_back_f32(const float * logits, const float * labels, const float * loss, float * dst, const int nclasses) {
+template <bool use_shared>
+static __global__ void cross_entropy_loss_back_f32(
+        const float * __restrict__ grad, const float * __restrict__ logits, const float * __restrict__ labels,
+        float * __restrict__ dst, const int nclasses) {
     extern __shared__ float tmp[];
 
+    logits += int64_t(blockIdx.x)*nclasses;
+    labels += int64_t(blockIdx.x)*nclasses;
+    dst    += int64_t(blockIdx.x)*nclasses;
+
     float maxval = -INFINITY;
     for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
-        const float val = logits[blockIdx.x*nclasses + i];
+        const float val = logits[i];
         maxval = fmaxf(maxval, val);
-        tmp[i] = val;
+
+        if (use_shared) {
+            tmp[i] = val;
+        }
     }
     maxval = warp_reduce_max(maxval);
 
     float sum = 0.0f;
     for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
-        const float val = expf(tmp[i] - maxval);
+        const float val = expf((use_shared ? tmp[i] : logits[i]) - maxval);
         sum += val;
-        tmp[i] = val;
+
+        if (use_shared) {
+            tmp[i] = val;
+        } else {
+            dst[i] = val;
+        }
     }
     sum = warp_reduce_sum(sum);
     const float sm_scale = 1.0f/sum;
 
-    const float d_by_nrows = *loss/gridDim.x;
+    const float d_by_nrows = *grad/gridDim.x;
     for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
-        dst[blockIdx.x*nclasses + i] = (tmp[i]*sm_scale - labels[blockIdx.x*nclasses + i])*d_by_nrows;
+        const float val = use_shared ? tmp[i] : dst[i];
+        dst[i] = (val*sm_scale - labels[i])*d_by_nrows;
     }
 }
 
@@ -119,48 +113,77 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
     ggml_cuda_pool & pool = ctx.pool();
     cudaStream_t stream = ctx.stream();
 
-    const dim3 blocks_dim(CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE, 1, 1);
-    const dim3 blocks_num((nrows + CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE - 1) / CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE, 1, 1);
-    const int shmem = 2*CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE*ne00*sizeof(float);
+    const dim3 blocks_dim(WARP_SIZE, 1, 1);
+    const dim3 blocks_num(nrows, 1, 1);
+    const size_t nbytes_shared = ne00*sizeof(float);
+
+    const int    id    = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
     ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);
 
-    cross_entropy_loss_f32<<<blocks_num, blocks_dim, shmem, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
+    if (nbytes_shared <= smpbo) {
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+        if (!shared_memory_limit_raised[id]) {
+            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
+            shared_memory_limit_raised[id] = true;
+        }
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+        cross_entropy_loss_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
+    } else {
+        cross_entropy_loss_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
+    }
+    CUDA_CHECK(cudaGetLastError());
 
     // Combine results from individual blocks:
     sum_f32_cuda(pool, dst_tmp.ptr, dst_d, blocks_num.x, stream);
 }
 
 void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-    const ggml_tensor * opt0 = dst->src[2];
+    const ggml_tensor * grad  = dst->src[0];
+    const ggml_tensor * src0f = dst->src[1];
+    const ggml_tensor * src1f = dst->src[2];
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(opt0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0f->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1f->type == GGML_TYPE_F32);
+    GGML_ASSERT( grad->type == GGML_TYPE_F32);
+    GGML_ASSERT(  dst->type == GGML_TYPE_F32);
 
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(src1));
-    GGML_ASSERT(ggml_is_contiguous(opt0));
+    GGML_ASSERT(ggml_is_scalar(grad));
+    GGML_ASSERT(ggml_is_contiguous(src0f));
+    GGML_ASSERT(ggml_is_contiguous(src1f));
     GGML_ASSERT(ggml_is_contiguous(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, src1));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_are_same_shape(src0f, src1f));
+    GGML_ASSERT(ggml_are_same_shape(src0f, dst));
 
-    const int64_t ne00  = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
+    const int64_t ne00  = src0f->ne[0];
+    const int64_t nrows = ggml_nrows(src0f);
 
-    const float * src0_d = (const float *) src0->data;
-    const float * src1_d = (const float *) src1->data;
-    const float * opt0_d = (const float *) opt0->data;
-    float       * dst_d  = (float       *) dst->data;
+    const float * grad_d  = (const float *) grad->data;
+    const float * src0f_d = (const float *) src0f->data;
+    const float * src1f_d = (const float *) src1f->data;
+    float       * dst_d   = (float       *) dst->data;
 
     cudaStream_t stream = ctx.stream();
 
     const dim3 blocks_dim(WARP_SIZE, 1, 1);
     const dim3 blocks_num(nrows, 1, 1);
-    const int shmem = ne00*sizeof(float);
+    const size_t nbytes_shared = ne00*sizeof(float);
 
-    cross_entropy_loss_back_f32<<<blocks_num, blocks_dim, shmem, stream>>>(src0_d, src1_d, opt0_d, dst_d, ne00);
+    const int    id    = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+
+    if (nbytes_shared <= smpbo) {
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+        if (!shared_memory_limit_raised[id]) {
+            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
+            shared_memory_limit_raised[id] = true;
+        }
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+        cross_entropy_loss_back_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
+    } else {
+        cross_entropy_loss_back_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
+    }
 }

ggml/src/ggml-cuda/getrows.cu

@@ -3,15 +3,15 @@
 
 template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void k_get_rows(
-            const void * src0, const int32_t * src1, dst_t * dst,
-            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
-            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
-            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
-            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
-            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+        const void * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
+        /*const int64_t ne10, const int64_t ne11,*/ const int64_t ne12, /*const int64_t ne13,*/
+        /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
+        /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
     const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
-    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i10 =  blockDim.y*blockIdx.y + threadIdx.y;
     const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
     const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
 
@@ -22,10 +22,10 @@ static __global__ void k_get_rows(
     const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
 
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
+    const void * src0_row = (const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03;
 
-    const int ib = i00/qk; // block index
-    const int iqs = (i00%qk)/qr; // quant index
+    const int ib   =  i00/qk;      // block index
+    const int iqs  = (i00%qk)/qr;  // quant index
     const int iybs = i00 - i00%qk; // dst block start index
     const int y_offset = qr == 1 ? 1 : qk/2;
 
@@ -39,15 +39,15 @@ static __global__ void k_get_rows(
 
 template<typename src0_t, typename dst_t>
 static __global__ void k_get_rows_float(
-            const src0_t * src0, const int32_t * src1, dst_t * dst,
-            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
-            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
-            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
-            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
-            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+        const src0_t * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
+        /*const int64_t ne10, const int64_t ne11,*/ const int64_t ne12, /*const int64_t ne13,*/
+        /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
+        /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    const int i00 = blockIdx.x*blockDim.x + threadIdx.x;
-    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i00 =  blockIdx.x*blockDim.x + threadIdx.x;
+    const int i10 =  blockDim.y*blockIdx.y + threadIdx.y;
     const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
     const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
 
@@ -58,14 +58,38 @@ static __global__ void k_get_rows_float(
     const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
 
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
+    const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
 
     dst_row[i00] = src0_row[i00];
 }
 
+template<typename grad_t, typename dst_t>
+static __global__ void k_get_rows_back_float(
+        const grad_t * __restrict__ grad, const int32_t * __restrict__ rows, dst_t * __restrict__ dst, const int64_t ncols, const int64_t nrows_grad) {
+    const int col = blockIdx.x*blockDim.x + threadIdx.x;
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int dst_row = blockIdx.y*blockDim.y + threadIdx.y;
+
+    float sum = 0.0f;
+
+    for (int64_t i = 0; i < nrows_grad; ++i) {
+        if (rows[i] != dst_row) {
+            continue;
+        }
+        sum += grad[i*ncols + col];
+    }
+
+    dst[dst_row*ncols + col] = sum;
+}
+
 template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-                            const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+static void get_rows_cuda(
+        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+        const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -87,22 +111,25 @@ static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, gg
     GGML_ASSERT(ne00 % 2 == 0);
 
     k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
-            src0_dd, src1_dd, dst_dd,
-            ne00, /*ne01, ne02, ne03,*/
-            /*ne10, ne11,*/ ne12, /*ne13,*/
-            /* s0,*/ s1, s2, s3,
-            /* nb00,*/ nb01, nb02, nb03,
-            s10, s11, s12/*, s13*/);
+        src0_dd, src1_dd, dst_dd,
+        ne00, /*ne01, ne02, ne03,*/
+        /*ne10, ne11,*/ ne12, /*ne13,*/
+        /* s0,*/ s1, s2, s3,
+        /* nb00,*/ nb01, nb02, nb03,
+        s10, s11, s12/*, s13*/);
 
     GGML_UNUSED(dst);
 }
 
 template<typename src0_t>
-static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-                                const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+static void get_rows_cuda_float(
+        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+        const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
+    GGML_ASSERT(ne13 == 1);
+
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
@@ -119,12 +146,12 @@ static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * sr
     //const size_t s13 = nb13 / ggml_element_size(src1);
 
     k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
-            src0_dd, src1_dd, dst_dd,
-            ne00, /*ne01, ne02, ne03,*/
-            /*ne10, ne11,*/ ne12, /*ne13,*/
-            /* s0,*/ s1, s2, s3,
-            /* nb00,*/ nb01, nb02, nb03,
-            s10, s11, s12/*, s13*/);
+        src0_dd, src1_dd, dst_dd,
+        ne00, /*ne01, ne02, ne03,*/
+        /*ne10, ne11,*/ ne12, /*ne13,*/
+        /* s0,*/ s1, s2, s3,
+        /* nb00,*/ nb01, nb02, nb03,
+        s10, s11, s12/*, s13*/);
 
     GGML_UNUSED(dst);
 }
@@ -132,42 +159,41 @@ static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * sr
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
-    const float * src0_d = (const float *)src0->data;
-    const float * src1_d = (const float *)src1->data;
-    float * dst_d = (float *)dst->data;
+
+    const void    * src0_d = (const void    *) src0->data;
+    const int32_t * src1_d = (const int32_t *) src1->data;
+    float         * dst_d  = (float         *) dst->data;
+
     cudaStream_t stream = ctx.stream();
 
-
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
 
     GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
     GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
-    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
-
-    const int32_t * src1_i32 = (const int32_t *) src1_d;
+    GGML_ASSERT(dst->nb[0]  == ggml_type_size(dst->type));
 
     switch (src0->type) {
         case GGML_TYPE_F16:
-            get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream);
+            get_rows_cuda_float(src0, src1, dst, (const half *) src0_d, src1_d, dst_d, stream);
             break;
         case GGML_TYPE_F32:
-            get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_cuda_float(src0, src1, dst, (const float *) src0_d, src1_d, dst_d, stream);
             break;
         case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
             break;
         case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
             break;
         case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
             break;
         case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
             break;
         case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
             break;
         default:
             // TODO: k-quants
@@ -175,3 +201,34 @@ void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
             break;
     }
 }
+
+void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0]; // gradients of forward pass output
+    const ggml_tensor * src1 = dst->src[1]; // src1 in forward pass
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const float   * src0_d = (const float   *) src0->data;
+    const int32_t * src1_d = (const int32_t *) src1->data;
+    float         * dst_d  = (float         *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    GGML_ASSERT(ne02*ne03 == 1);
+    GGML_ASSERT(ne12*ne13 == 1);
+    GGML_ASSERT(ne2*ne3 == 1);
+
+    const dim3 block_dims(CUDA_GET_ROWS_BACK_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + CUDA_GET_ROWS_BACK_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BACK_BLOCK_SIZE;
+    const dim3 block_nums(block_num_x, ne1, 1);
+
+    k_get_rows_back_float<<<block_nums, block_dims, 0, stream>>>(src0_d, src1_d, dst_d, ne00, ne10);
+}

ggml/src/ggml-cuda/getrows.cuh

@@ -1,5 +1,8 @@
 #include "common.cuh"
 
 #define CUDA_GET_ROWS_BLOCK_SIZE 256
+#define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256
 
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2003,6 +2003,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_GET_ROWS:
             ggml_cuda_op_get_rows(ctx, dst);
             break;
+        case GGML_OP_GET_ROWS_BACK:
+            ggml_cuda_op_get_rows_back(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2091,9 +2094,15 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_LEAKY_RELU:
             ggml_cuda_op_leaky_relu(ctx, dst);
             break;
+        case GGML_OP_SILU_BACK:
+            ggml_cuda_op_silu_back(ctx, dst);
+            break;
         case GGML_OP_RMS_NORM:
             ggml_cuda_op_rms_norm(ctx, dst);
             break;
+        case GGML_OP_RMS_NORM_BACK:
+            ggml_cuda_op_rms_norm_back(ctx, dst);
+            break;
         case GGML_OP_MUL_MAT:
             if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) {
                 GGML_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
@@ -2138,6 +2147,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SOFT_MAX:
             ggml_cuda_op_soft_max(ctx, dst);
             break;
+        case GGML_OP_SOFT_MAX_BACK:
+            ggml_cuda_op_soft_max_back(ctx, dst);
+            break;
         case GGML_OP_ROPE:
             ggml_cuda_op_rope(ctx, dst);
             break;
@@ -2912,7 +2924,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 }
             } break;
         case GGML_OP_OUT_PROD:
-            return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
+            return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
         case GGML_OP_GET_ROWS:
             {
                 switch (op->src[0]->type) {
@@ -2928,6 +2940,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                         return false;
                 }
             } break;
+        case GGML_OP_GET_ROWS_BACK:
+            {
+                return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -3001,8 +3017,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 }
                 return false;
             } break;
+        case GGML_OP_SILU_BACK:
+            return ggml_is_contiguous(op->src[0]);
+            break;
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
+        case GGML_OP_RMS_NORM_BACK:
             return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0;
             break;
         case GGML_OP_NONE:
@@ -3027,6 +3047,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
             return true;
+        case GGML_OP_SOFT_MAX_BACK: {
+            float max_bias = 0.0f;
+            memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
+            return max_bias == 0.0f;
+        }
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK: {
             const size_t ts = ggml_type_size(op->src[0]->type);

ggml/src/ggml-cuda/norm.cu

@@ -5,20 +5,24 @@ static __global__ void norm_f32(const float * x, float * dst, const int ncols, c
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
-    float2 mean_var = make_float2(0.f, 0.f);
+    x   += int64_t(row)*ncols;
+    dst += int64_t(row)*ncols;
+
+    float2 mean_var = make_float2(0.0f, 0.0f);
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row*ncols + col];
+        const float xi = x[col];
         mean_var.x += xi;
         mean_var.y += xi * xi;
     }
 
     // sum up partial sums
     mean_var = warp_reduce_sum(mean_var);
-    if (block_size > WARP_SIZE) {
+    if constexpr (block_size > WARP_SIZE) {
+        static_assert(block_size == 1024, "unexpected block_size");
         __shared__ float2 s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
+        const int warp_id = threadIdx.x / WARP_SIZE;
+        const int lane_id = threadIdx.x % WARP_SIZE;
         if (lane_id == 0) {
             s_sum[warp_id] = mean_var;
         }
@@ -32,7 +36,7 @@ static __global__ void norm_f32(const float * x, float * dst, const int ncols, c
     const float inv_std = rsqrtf(var + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
+        dst[col] = (x[col] - mean) * inv_std;
     }
 }
 
@@ -40,14 +44,8 @@ template <int block_size>
 static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
     // blockIdx.x: num_groups idx
     // threadIdx.x: block_size idx
-    int start = blockIdx.x * group_size;
-    int end = start + group_size;
-
-    start += threadIdx.x;
-
-    if (end >= ne_elements) {
-        end = ne_elements;
-    }
+    const int start =     blockIdx.x*group_size + threadIdx.x;
+    const int end   = min(blockIdx.x*group_size + group_size,  ne_elements);
 
     float tmp = 0.0f; // partial sum for thread in warp
 
@@ -56,10 +54,11 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
     }
 
     tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
+    if constexpr (block_size > WARP_SIZE) {
+        static_assert(block_size == 1024, "unexpected block_size");
         __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
+        const int warp_id = threadIdx.x / WARP_SIZE;
+        const int lane_id = threadIdx.x % WARP_SIZE;
         if (lane_id == 0) {
             s_sum[warp_id] = tmp;
         }
@@ -68,11 +67,11 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
         tmp = warp_reduce_sum(tmp);
     }
 
-    float mean = tmp / group_size;
+    const float mean = tmp / group_size;
     tmp = 0.0f;
 
     for (int j = start; j < end; j += block_size) {
-        float xi = x[j] - mean;
+        const float xi = x[j] - mean;
         dst[j] = xi;
         tmp += xi * xi;
     }
@@ -80,8 +79,8 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
     tmp = warp_reduce_sum(tmp);
     if (block_size > WARP_SIZE) {
         __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
+        const int warp_id = threadIdx.x / WARP_SIZE;
+        const int lane_id = threadIdx.x % WARP_SIZE;
         if (lane_id == 0) {
             s_sum[warp_id] = tmp;
         }
@@ -90,8 +89,8 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
         tmp = warp_reduce_sum(tmp);
     }
 
-    float variance = tmp / group_size;
-    float scale = rsqrtf(variance + eps);
+    const float variance = tmp / group_size;
+    const float scale = rsqrtf(variance + eps);
     for (int j = start; j < end; j += block_size) {
         dst[j] *= scale;
     }
@@ -102,19 +101,23 @@ static __global__ void rms_norm_f32(const float * x, float * dst, const int ncol
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
+    x   += int64_t(row)*ncols;
+    dst += int64_t(row)*ncols;
+
     float tmp = 0.0f; // partial sum for thread in warp
 
     for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row*ncols + col];
+        const float xi = x[col];
         tmp += xi * xi;
     }
 
     // sum up partial sums
     tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
+    if constexpr (block_size > WARP_SIZE) {
+        static_assert(block_size == 1024, "unexpected block_size");
         __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
+        const int warp_id = threadIdx.x / WARP_SIZE;
+        const int lane_id = threadIdx.x % WARP_SIZE;
         if (lane_id == 0) {
             s_sum[warp_id] = tmp;
         }
@@ -127,12 +130,63 @@ static __global__ void rms_norm_f32(const float * x, float * dst, const int ncol
     const float scale = rsqrtf(mean + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[row*ncols + col] = scale * x[row*ncols + col];
+        dst[col] = scale * x[col];
+    }
+}
+
+template <int block_size>
+static __global__ void rms_norm_back_f32(
+        const float * grad, const float * xf, float * dst, const int ncols, const float eps) {
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int tid = threadIdx.x;
+
+    grad += int64_t(row)*ncols;
+    xf   += int64_t(row)*ncols;
+    dst  += int64_t(row)*ncols;
+
+    float sum_xx = 0.0f; // sum for squares of x, equivalent to forward pass
+    float sum_xg = 0.0f; // sum for x * gradient, needed because RMS norm mixes inputs
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xfi = xf[col];
+        sum_xx += xfi * xfi;
+        sum_xg += xfi * grad[col];
+    }
+
+    // sum up partial sums
+    sum_xx = warp_reduce_sum(sum_xx);
+    sum_xg = warp_reduce_sum(sum_xg);
+    if constexpr (block_size > WARP_SIZE) {
+        static_assert(block_size == 1024, "unexpected block_size");
+        __shared__ float s_sum_xx[32];
+        __shared__ float s_sum_xg[32];
+        const int warp_id = threadIdx.x / WARP_SIZE;
+        const int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum_xx[warp_id] = sum_xx;
+            s_sum_xg[warp_id] = sum_xg;
+        }
+        __syncthreads();
+
+        sum_xx = s_sum_xx[lane_id];
+        sum_xx = warp_reduce_sum(sum_xx);
+
+        sum_xg = s_sum_xg[lane_id];
+        sum_xg = warp_reduce_sum(sum_xg);
+    }
+
+    const float mean_eps = sum_xx / ncols + eps;
+    const float sum_eps  = sum_xx + ncols*eps;
+
+    const float scale_grad = rsqrtf(mean_eps);
+    const float scale_x    = -scale_grad * sum_xg/sum_eps;
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[col] = scale_grad*grad[col] + scale_x*xf[col];
     }
 }
 
 static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
         norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
@@ -142,7 +196,8 @@ static void norm_f32_cuda(const float * x, float * dst, const int ncols, const i
     }
 }
 
-static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
+static void group_norm_f32_cuda(
+        const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
     if (group_size < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
         group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
@@ -153,7 +208,6 @@ static void group_norm_f32_cuda(const float * x, float * dst, const int num_grou
 }
 
 static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
         rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
@@ -163,6 +217,16 @@ static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, con
     }
 }
 
+static void rms_norm_back_f32_cuda(const float * grad, const float * xf, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        rms_norm_back_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(grad, xf, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        rms_norm_back_f32<1024><<<nrows, block_dims, 0, stream>>>(grad, xf, dst, ncols, eps);
+    }
+}
+
 void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;
@@ -179,6 +243,7 @@ void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
 
     norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
 }
@@ -198,6 +263,7 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 
     float eps;
     memcpy(&eps, dst->op_params + 1, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
 
     int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
     group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
@@ -219,6 +285,33 @@ void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
 
     rms_norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
 }
+
+void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * grad  = dst->src[0]; // gradients
+    const ggml_tensor * src0f = dst->src[1]; // src0 from forward pass
+
+    const float * grad_d  = (const float *) grad->data;
+    const float * src0f_d = (const float *) src0f->data;
+    float       * dst_d   = (float       *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(grad));
+
+    GGML_ASSERT( grad->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0f->type == GGML_TYPE_F32);
+    GGML_ASSERT(  dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0f->ne[0];
+    const int64_t nrows = ggml_nrows(src0f);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
+
+    rms_norm_back_f32_cuda(grad_d, src0f_d, dst_d, ne00, nrows, eps, stream);
+}

ggml/src/ggml-cuda/norm.cuh

@@ -5,3 +5,5 @@ void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/out-prod.cu

@@ -11,16 +11,15 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type  == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(dst));
 
     GGML_ASSERT(ne01 == ne11);
     GGML_ASSERT(ne0 == ne00);
     GGML_ASSERT(ne1 == ne10);
 
-    GGML_ASSERT(ne2 == src0->ne[2]);
+    GGML_ASSERT(ne2 % src0->ne[2] == 0);
+    GGML_ASSERT(ne3 % src0->ne[3] == 0);
+
     GGML_ASSERT(ne2 == src1->ne[2]);
-    GGML_ASSERT(ne3 == src0->ne[3]);
     GGML_ASSERT(ne3 == src1->ne[3]);
 
     const float * src0_d = (const float *) src0->data;
@@ -33,8 +32,6 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const float alpha = 1.0f;
     const float beta = 0.0f;
 
-    GGML_ASSERT(ne2 == 1);
-    GGML_ASSERT(ne3 == 1);
     CUBLAS_CHECK(cublasSetStream(handle, stream));
 
     const bool src1_T = ggml_is_transposed(src1);
@@ -42,10 +39,27 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t           ldb            = (src1_T ?        nb10 :        nb11) /  sizeof(float);
     GGML_ASSERT(                             (src1_T ?        nb11 :        nb10) == sizeof(float));
 
-    CUBLAS_CHECK(
-        cublasSgemm(handle, CUBLAS_OP_N, src1_cublas_op,
-                ne0, ne1, ne01,
-                &alpha, src0_d, ne00,
-                        src1_d, ldb,
-                &beta,  dst_d,  ne0));
+    // data strides in dimensions 2/3
+    const size_t s02 = nb02 / sizeof(float);
+    const size_t s03 = nb03 / sizeof(float);
+    const size_t s12 = nb12 / sizeof(float);
+    const size_t s13 = nb13 / sizeof(float);
+    const size_t s2  = nb2  / sizeof(float);
+    const size_t s3  = nb3  / sizeof(float);
+
+    // dps == dst per src0, used for group query attention
+    const int64_t dps2 = ne2 / ne02;
+    const int64_t dps3 = ne3 / ne03;
+
+    // TODO batched matrix multiplication
+    for (int64_t i3 = 0; i3 < ne3; ++i3) {
+        for (int64_t i2 = 0; i2 < ne2; ++i2) {
+            CUBLAS_CHECK(
+                cublasSgemm(handle, CUBLAS_OP_N, src1_cublas_op,
+                        ne0, ne1, ne01,
+                        &alpha, src0_d + (i3/dps3)*s03 + (i2/dps2)*s02, ne00,
+                                src1_d +  i3      *s13 +  i2      *s12, ldb,
+                        &beta,  dst_d  +  i3      *s3  +  i2      *s2,  ne0));
+        }
+    }
 }

ggml/src/ggml-cuda/rope.cu

@@ -39,9 +39,9 @@ static __device__ void rope_yarn(
 
 template<bool forward, bool has_ff, typename T>
 static __global__ void rope_norm(
-        const T * __restrict__ x, T * __restrict__ dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
-        const int32_t * __restrict__ pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * __restrict__ freq_factors) {
+        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
+        const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -83,9 +83,9 @@ static __global__ void rope_norm(
 
 template<bool forward, bool has_ff, typename T>
 static __global__ void rope_neox(
-        const T * __restrict__ x, T * __restrict__ dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
-        const int32_t * __restrict__ pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * __restrict__ freq_factors) {
+        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
+        const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -127,9 +127,9 @@ static __global__ void rope_neox(
 
 template<bool forward, bool has_ff, typename T>
 static __global__ void rope_multi(
-        const T * __restrict__ x, T * __restrict__ dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2,
-        const int n_dims, const int32_t * __restrict__ pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * __restrict__ freq_factors, const mrope_sections sections) {
+        const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2,
+        const int n_dims, const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -187,9 +187,9 @@ static __global__ void rope_multi(
 
 template<bool forward, bool has_ff, typename T>
 static __global__ void rope_vision(
-        const T * __restrict__ x, T * __restrict__ dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims,
-        const int32_t * __restrict__ pos, const float freq_scale, const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
-        const float theta_scale, const float * __restrict__ freq_factors, const mrope_sections sections) {
+        const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims,
+        const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
+        const float theta_scale, const float * freq_factors, const mrope_sections sections) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -234,9 +234,9 @@ static __global__ void rope_vision(
 
 template<bool forward, typename T>
 static void rope_norm_cuda(
-        const T * __restrict__ x, T * __restrict__ dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * __restrict__ pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * __restrict__ freq_factors, cudaStream_t stream) {
+        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims, const int nr,
+        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -257,9 +257,9 @@ static void rope_norm_cuda(
 
 template<bool forward, typename T>
 static void rope_neox_cuda(
-        const T * __restrict__ x, T * __restrict__ dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * __restrict__ pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * __restrict__ freq_factors, cudaStream_t stream) {
+        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims, const int nr,
+        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -280,9 +280,9 @@ static void rope_neox_cuda(
 
 template<bool forward, typename T>
 static void rope_multi_cuda(
-        const T * __restrict__ x, T * __restrict__ dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * __restrict__ pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * __restrict__ freq_factors, const mrope_sections sections, cudaStream_t stream) {
+        const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
+        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, cudaStream_t stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -303,9 +303,9 @@ static void rope_multi_cuda(
 
 template<bool forward, typename T>
 static void rope_vision_cuda(
-        const T * __restrict__ x, T * __restrict__ dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * __restrict__ pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * __restrict__ freq_factors, const mrope_sections sections, cudaStream_t stream) {
+        const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
+        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
+        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, cudaStream_t stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);

ggml/src/ggml-cuda/softmax.cu

@@ -1,5 +1,7 @@
 #include "common.cuh"
+#include "ggml.h"
 #include "softmax.cuh"
+#include <cstdint>
 
 template <typename T>
 static __device__ __forceinline__ float t2f32(T val) {
@@ -11,14 +13,20 @@ __device__ float __forceinline__ t2f32<half>(half val) {
     return __half2float(val);
 }
 
-template <bool vals_smem, int ncols_template, int block_size_template, typename T>
-static __global__ void soft_max_f32(const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
+template <bool use_shared, int ncols_template, int block_size_template, typename T>
+static __global__ void soft_max_f32(
+        const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y,
+        const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
     const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
 
     const int tid  = threadIdx.x;
     const int rowx = blockIdx.x;
     const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
 
+    x    += int64_t(rowx)*ncols;
+    mask += int64_t(rowy)*ncols * (mask != nullptr);
+    dst  += int64_t(rowx)*ncols;
+
     const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
 
     const int warp_id = threadIdx.x / WARP_SIZE;
@@ -29,7 +37,7 @@ static __global__ void soft_max_f32(const float * x, const T * mask, float * dst
     extern __shared__ float data_soft_max_f32[];
     float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
     // shared memory buffer to cache values between iterations:
-    float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + (int64_t)rowx*ncols;
+    float * vals = use_shared ? buf_iw + WARP_SIZE : dst;
 
     float max_val = -INFINITY;
 
@@ -41,10 +49,7 @@ static __global__ void soft_max_f32(const float * x, const T * mask, float * dst
             break;
         }
 
-        const int64_t ix = (int64_t)rowx*ncols + col;
-        const int64_t iy = (int64_t)rowy*ncols + col;
-
-        const float val = x[ix]*scale + (mask ? slope*t2f32(mask[iy]) : 0.0f);
+        const float val = x[col]*scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
 
         vals[col] = val;
         max_val = max(max_val, val);
@@ -110,8 +115,29 @@ static __global__ void soft_max_f32(const float * x, const T * mask, float * dst
             return;
         }
 
-        const int64_t idst = (int64_t)rowx*ncols + col;
-        dst[idst] = vals[col] * inv_sum;
+        dst[col] = vals[col] * inv_sum;
+    }
+}
+
+static __global__ void soft_max_back_f32(
+        const float * grad, const float * dstf, float * dst, const int ncols, const float scale) {
+    const int tid  = threadIdx.x;
+    const int rowx = blockIdx.x;
+
+    grad += int64_t(rowx)*ncols;
+    dstf += int64_t(rowx)*ncols;
+    dst  += int64_t(rowx)*ncols;
+
+    float dgf_dot = 0.0f; // dot product of dst from forward pass and gradients
+
+    for (int col = tid; col < ncols; col += WARP_SIZE) {
+        dgf_dot += dstf[col]*grad[col];
+    }
+
+    dgf_dot = warp_reduce_sum(dgf_dot);
+
+    for (int col = tid; col < ncols; col += WARP_SIZE) {
+        dst[col] = scale * (grad[col] - dgf_dot) * dstf[col];
     }
 }
 
@@ -121,7 +147,7 @@ static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, cons
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
     const dim3 block_nums(nrows_x, 1, 1);
-    const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
+    const size_t nbytes_shared = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
     const uint32_t n_head      = nrows_x/nrows_y;
@@ -131,50 +157,68 @@ static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, cons
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
     // FIXME: this limit could be raised by ~2-4x on Ampere or newer
-    if (shmem < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
+    if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
         switch (ncols_x) {
             case 32:
-                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true,   32,   32><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 64:
-                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true,   64,   64><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 128:
-                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true,  128,  128><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 256:
-                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true,  256,  256><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 512:
-                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true,  512,  512><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             default:
-                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true,    0,    0><<<block_nums, block_dims, nbytes_shared, stream>>>
+                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
         }
     } else {
-        const size_t shmem_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+        const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
     }
 }
 
+static void soft_max_back_f32_cuda(
+        const float * grad, const float * dstf, float * dst,
+        const int ncols, const int nrows, const float scale, cudaStream_t stream) {
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows,     1, 1);
+
+    soft_max_back_f32<<<block_nums, block_dims, 0, stream>>>(grad, dstf, dst, ncols, scale);
+}
+
 void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    const float * src0_d = (const float *)src0->data;
-    const void  * src1_d = src1 ? (const void *)src1->data : nullptr;
+    const float * src0_d = (const float *) src0->data;
+    const void  * src1_d = src1 ? (const void *) src1->data : nullptr;
+    float       *  dst_d = (float *) dst->data;
 
-    float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
@@ -189,18 +233,42 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float scale    = 1.0f;
     float max_bias = 0.0f;
 
-    memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
+    memcpy(&scale,    (const float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float));
 
     const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
     if (use_f16) {
-        const half * src1_dd = (const half *)src1_d;
-
-        soft_max_f32_cuda(src0_d, src1_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
     } else {
-        const float * src1_dd = (const float *)src1_d;
-
-        soft_max_f32_cuda(src0_d, src1_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
     }
 }
+
+void ggml_cuda_op_soft_max_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0]; // grad
+    const ggml_tensor * src1 = dst->src[1]; // forward pass output
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (const float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float));
+
+    GGML_ASSERT(max_bias == 0.0f);
+
+    soft_max_back_f32_cuda(src0_d, src1_d, dst_d, ncols, nrows, scale, stream);
+}

ggml/src/ggml-cuda/softmax.cuh

@@ -3,3 +3,5 @@
 #define CUDA_SOFT_MAX_BLOCK_SIZE 1024
 
 void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_soft_max_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/unary.cu

@@ -51,6 +51,19 @@ static __global__ void silu_f32(const float * x, float * dst, const int k) {
     dst[i] = x[i] / (1.0f + expf(-x[i]));
 }
 
+static __global__ void silu_back_f32(
+        const float * grad, const float * xf, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    const float xfi = xf[i];
+    const float s = 1.0f / (1.0f + expf(-xfi));
+    dst[i] = grad[i] * s * (1.0f + xfi * (1.0f - s));
+}
+
 static __global__ void tanh_f32(const float * x, float * dst, int k) {
     const int i  = blockDim.x*blockIdx.x + threadIdx.x;
     if (i >= k) {
@@ -173,6 +186,11 @@ static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_
     silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void silu_back_f32_cuda(const float * grad, const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SILU_BACK_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
+    silu_back_f32<<<num_blocks, CUDA_SILU_BACK_BLOCK_SIZE, 0, stream>>>(grad, x, dst, k);
+}
+
 static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
     tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -284,6 +302,24 @@ void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     silu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
 }
 
+void ggml_cuda_op_silu_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0]; // input from forward pass
+    const ggml_tensor * src1 = dst->src[1]; // grads of forward pass output
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    silu_back_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(src0), stream);
+}
+
 void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;

ggml/src/ggml-cuda/unary.cuh

@@ -4,6 +4,7 @@
 #define CUDA_STEP_BLOCK_SIZE 256
 #define CUDA_GELU_BLOCK_SIZE 256
 #define CUDA_SILU_BLOCK_SIZE 256
+#define CUDA_SILU_BACK_BLOCK_SIZE 256
 #define CUDA_TANH_BLOCK_SIZE 256
 #define CUDA_RELU_BLOCK_SIZE 256
 #define CUDA_SIGMOID_BLOCK_SIZE 256
@@ -23,6 +24,8 @@ void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_silu_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -228,6 +228,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_repeat_f32;
     vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
     vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16;
+    vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_norm_f32;
     vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
@@ -384,10 +386,13 @@ struct vk_flash_attn_push_constants {
     uint32_t nev3;
     uint32_t nem1;
 
+    uint32_t nb01;
     uint32_t nb02;
     uint32_t nb03;
+    uint32_t nb11;
     uint32_t nb12;
     uint32_t nb13;
+    uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
     uint32_t nb31;
@@ -1965,6 +1970,20 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_1], "cpy_q4_1_f32", cpy_q4_1_f32_len, cpy_q4_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q5_0], "cpy_q5_0_f32", cpy_q5_0_f32_len, cpy_q5_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q5_1], "cpy_q5_1_f32", cpy_q5_1_f32_len, cpy_q5_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q8_0], "cpy_q8_0_f32", cpy_q8_0_f32_len, cpy_q8_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_IQ4_NL], "cpy_iq4_nl_f32", cpy_iq4_nl_f32_len, cpy_iq4_nl_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_add_f32_norepeat, "add_f32_norepeat", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
@@ -3689,6 +3708,33 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_cpy_f16_f16;
         }
     }
+    if (src->type == GGML_TYPE_F32) {
+        switch (to) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_IQ4_NL:
+            return ctx->device->pipeline_cpy_f32_quant[to];
+        default:
+            break;
+        }
+    }
+
+    if (to == GGML_TYPE_F32) {
+        switch (src->type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_IQ4_NL:
+            return ctx->device->pipeline_cpy_quant_f32[src->type];
+        default:
+            break;
+        }
+    }
 
     std::cerr << "Missing CPY op for types: " << ggml_type_name(src->type) << " " << ggml_type_name(to) << std::endl;
     GGML_ABORT("fatal error");
@@ -4766,7 +4812,14 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
     assert(pipelines);
 
-    bool aligned = (KV % pipelines[1]->align) == 0;
+    const uint32_t q_stride = (uint32_t)(nbq1 / ggml_type_size(q->type));
+    const uint32_t k_stride = (uint32_t)(nbk1 / ggml_type_size(k->type));
+    const uint32_t v_stride = (uint32_t)(nbv1 / ggml_type_size(v->type));
+
+    bool aligned = (KV % pipelines[1]->align) == 0 &&
+                   // the "aligned" shader variant will forcibly align strides, for performance
+                   (q_stride & 7) == 0 && (k_stride & 7) == 0 && (v_stride & 7) == 0;
+
     vk_pipeline pipeline = pipelines[aligned];
     assert(pipeline);
 
@@ -4802,15 +4855,15 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     if (ctx->device->uma) {
         ggml_vk_host_get(ctx->device, q->data, d_Q, q_buf_offset);
-        ggml_vk_host_get(ctx->device, k->data, d_K, q_buf_offset);
-        ggml_vk_host_get(ctx->device, v->data, d_V, q_buf_offset);
-        ggml_vk_host_get(ctx->device, dst->data, d_D, q_buf_offset);
+        ggml_vk_host_get(ctx->device, k->data, d_K, k_buf_offset);
+        ggml_vk_host_get(ctx->device, v->data, d_V, v_buf_offset);
+        ggml_vk_host_get(ctx->device, dst->data, d_D, d_buf_offset);
         Q_uma = d_Q != nullptr;
         K_uma = d_K != nullptr;
         V_uma = d_V != nullptr;
         D_uma = d_D != nullptr;
         if (mask) {
-            ggml_vk_host_get(ctx->device, mask->data, d_M, q_buf_offset);
+            ggml_vk_host_get(ctx->device, mask->data, d_M, m_buf_offset);
             M_uma = d_M != nullptr;
         }
     }
@@ -4848,7 +4901,18 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         }
     }
 
-    const vk_flash_attn_push_constants pc = { N, KV, (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3, (uint32_t)neq2, (uint32_t)neq3, (uint32_t)nek2, (uint32_t)nek3, (uint32_t)nev2, (uint32_t)nev3, nem1, (uint32_t)nbq2, (uint32_t)nbq3, (uint32_t)nbk2, (uint32_t)nbk3, (uint32_t)nbv2, (uint32_t)nbv3, nbm1, scale, max_bias, logit_softcap, mask != nullptr, n_head_log2, m0, m1 };
+    const vk_flash_attn_push_constants pc = { N, KV,
+                                              (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3,
+                                              (uint32_t)neq2, (uint32_t)neq3,
+                                              (uint32_t)nek2, (uint32_t)nek3,
+                                              (uint32_t)nev2, (uint32_t)nev3,
+                                              nem1,
+                                              q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
+                                              k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
+                                              v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
+                                              nbm1,
+                                              scale, max_bias, logit_softcap,
+                                              mask != nullptr, n_head_log2, m0, m1 };
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                 {
                                     vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE},
@@ -5160,7 +5224,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     }
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << ggml_op_name(op) << ", " << (dryrun ? "dryrun" : "") << ")");
-    GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
+    GGML_ASSERT(op == GGML_OP_GET_ROWS || op == GGML_OP_CPY || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
     GGML_ASSERT(ggml_vk_op_supports_incontiguous(op) || ggml_vk_dim01_contiguous(src0));  // NOLINT
     GGML_ASSERT(dst->buffer != nullptr);
     const uint64_t ne00 = src0->ne[0];
@@ -7905,12 +7969,36 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1] != nullptr ? op->src[1]->type : src0_type;
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
-                    return true;
+
+                if (src0_type == GGML_TYPE_F32) {
+                    switch (src1_type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_IQ4_NL:
+                        return true;
+                    default:
+                        break;
+                    }
                 }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
-                    return true;
+                if (src1_type == GGML_TYPE_F32) {
+                    switch (src0_type) {
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_IQ4_NL:
+                        return true;
+                    default:
+                        break;
+                    }
                 }
+
                 if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
                     return true;
                 }
@@ -8601,6 +8689,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
     ggml_tensor * src0 = tensor->src[0];
     ggml_tensor * src1 = tensor->src[1];
     ggml_tensor * src2 = tensor->src[2];
+    ggml_tensor * src3 = tensor->src[3];
 
     void * tensor_data = tensor->data;
 
@@ -8663,6 +8752,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
                         if (src2 != nullptr) {
                             std::cerr << "src2=" << src2 << " src2->name=" << src2->name << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
                         }
+                        if (src3 != nullptr) {
+                            std::cerr << "src3=" << src3 << " src3->name=" << src3->name << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl;
+                        }
                         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
                         std::cerr << std::endl << "Result:" << std::endl;
                         ggml_vk_print_tensor_area(tensor, tensor_data, i0, i1, i2, i3);
@@ -8707,6 +8799,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
         if (src2 != nullptr) {
             std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
         }
+        if (src3 != nullptr) {
+            std::cerr << "src3=" << src3 << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl;
+        }
         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
         std::cerr << std::endl << "Result:" << std::endl;
         ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
@@ -8729,6 +8824,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
         if (src2 != nullptr) {
             std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
         }
+        if (src3 != nullptr) {
+            std::cerr << "src3=" << src3 << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl;
+        }
         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
         std::cerr << std::endl << "Result:" << std::endl;
         ggml_vk_print_tensor_area(tensor, tensor_data, first_error[0], first_error[1], first_error[2], first_error[3]);

ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp

@@ -0,0 +1,51 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+#include "dequant_funcs.comp"
+
+#if defined(DATA_A_IQ4_NL)
+// 16 invocations needed for init_iq4nl_shmem
+layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
+#else
+layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+#endif
+
+void main() {
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+    if (gl_LocalInvocationIndex.x != 0) {
+        return;
+    }
+#endif
+
+    const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    uint dst_idx = get_doffset() + dst_idx(idx);
+    uint src_idx = src0_idx_quant(idx, QUANT_K);
+
+    const uint a_offset = 0;
+    const uint ib = src_idx;
+    const vec2 dm = get_dm(ib, a_offset);
+
+    [[unroll]] for (int j = 0; j < QUANT_K; j += 4) {
+        vec4 v = dequantize4(ib, j / QUANT_R, a_offset);
+        v = v * dm.x + vec4(dm.y);
+
+#if QUANT_R == 2
+        data_d[dst_idx + j/2 +             0] = v[0];
+        data_d[dst_idx + j/2 + QUANT_K/2 + 0] = v[1];
+        data_d[dst_idx + j/2 +             1] = v[2];
+        data_d[dst_idx + j/2 + QUANT_K/2 + 1] = v[3];
+#else
+        data_d[dst_idx + j + 0] = v[0];
+        data_d[dst_idx + j + 1] = v[1];
+        data_d[dst_idx + j + 2] = v[2];
+        data_d[dst_idx + j + 3] = v[3];
+#endif
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp

@@ -0,0 +1,237 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+#if defined(DATA_A_IQ4_NL)
+// 16 invocations needed for init_iq4nl_shmem
+layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
+#else
+layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+#endif
+
+layout (binding = 0) readonly buffer S {float data_s[];};
+layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];};
+
+#if defined(DATA_A_Q4_0)
+void quantize(uint dst_idx, uint src_idx)
+{
+    float amax = 0.0;
+    float vmax = 0.0;
+
+    [[unroll]] for (int j = 0; j < QUANT_K_Q4_0; ++j) {
+        const float v = data_s[src_idx + j];
+        if (amax < abs(v)) {
+            amax = abs(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = (d != 0.0) ? 1.0/d : 0.0;
+
+    data_q[dst_idx].d = float16_t(d);
+
+    [[unroll]] for (int j = 0; j < QUANT_K_Q4_0/2; ++j) {
+        const float x0 = data_s[src_idx + 0              + j]*id;
+        const float x1 = data_s[src_idx + QUANT_K_Q4_0/2 + j]*id;
+
+        const uint xi0 = min(15, int(x0 + 8.5));
+        const uint xi1 = min(15, int(x1 + 8.5));
+
+        data_q[dst_idx].qs[j]  = uint8_t(xi0 | (xi1 << 4));
+    }
+}
+#endif
+
+#if defined(DATA_A_Q4_1)
+void quantize(uint dst_idx, uint src_idx)
+{
+    float vmin = 1.0/0.0;
+    float vmax = -vmin;
+
+    [[unroll]] for (int j = 0; j < QUANT_K_Q4_1; ++j) {
+        const float v = data_s[src_idx + j];
+
+        if (v < vmin) vmin = v;
+        if (v > vmax) vmax = v;
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = (d != 0.0) ? 1.0/d : 0.0;
+
+    data_q[dst_idx].d = float16_t(d);
+    data_q[dst_idx].m = float16_t(vmin);
+
+    [[unroll]] for (int j = 0; j < QUANT_K_Q4_1/2; ++j) {
+        const float x0 = (data_s[src_idx + 0              + j] - vmin)*id;
+        const float x1 = (data_s[src_idx + QUANT_K_Q4_1/2 + j] - vmin)*id;
+
+        const uint xi0 = min(15, int(x0 + 0.5));
+        const uint xi1 = min(15, int(x1 + 0.5));
+
+        data_q[dst_idx].qs[j]  = uint8_t(xi0 | (xi1 << 4));
+    }
+}
+#endif
+
+#if defined(DATA_A_Q5_0)
+void quantize(uint dst_idx, uint src_idx)
+{
+    float amax = 0.0;
+    float vmax = 0.0;
+
+    [[unroll]] for (int j = 0; j < QUANT_K_Q5_0; ++j) {
+        const float v = data_s[src_idx + j];
+        if (amax < abs(v)) {
+            amax = abs(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -16;
+    const float id = (d != 0.0) ? 1.0/d : 0.0;
+
+    data_q[dst_idx].d = float16_t(d);
+
+    uint32_t qh = 0;
+    [[unroll]] for (int j = 0; j < QUANT_K_Q5_0/2; ++j) {
+        const float x0 = data_s[src_idx + 0              + j]*id;
+        const float x1 = data_s[src_idx + QUANT_K_Q5_0/2 + j]*id;
+
+        const uint xi0 = min(31, int(x0 + 16.5));
+        const uint xi1 = min(31, int(x1 + 16.5));
+
+        data_q[dst_idx].qs[j]  = uint8_t((xi0 & 0xf) | ((xi1 & 0xf) << 4));
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QUANT_K_Q5_0/2);
+    }
+    data_q[dst_idx].qh[0] = uint16_t(qh & 0xFFFF);
+    data_q[dst_idx].qh[1] = uint16_t(qh >> 16);
+}
+#endif
+
+#if defined(DATA_A_Q5_1)
+void quantize(uint dst_idx, uint src_idx)
+{
+    float min = data_s[src_idx + 0];
+    float max = min;
+
+    [[unroll]] for (int j = 1; j < QUANT_K_Q5_1; ++j) {
+        const float v = data_s[src_idx + j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d  = (max - min) / 31;
+    const float id = (d != 0) ? 1.0/d : 0.0;
+
+    data_q[dst_idx].d = float16_t(d);
+    data_q[dst_idx].m = float16_t(min);
+
+    uint32_t qh = 0;
+    [[unroll]] for (int j = 0; j < QUANT_K_Q5_1/2; ++j) {
+        const float x0 = (data_s[src_idx + 0              + j] - min)*id;
+        const float x1 = (data_s[src_idx + QUANT_K_Q5_1/2 + j] - min)*id;
+
+        const uint xi0 = uint(x0 + 0.5);
+        const uint xi1 = uint(x1 + 0.5);
+
+        data_q[dst_idx].qs[j]  = uint8_t((xi0 & 0xf) | ((xi1 & 0xf) << 4));
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QUANT_K_Q5_1/2);
+    }
+    data_q[dst_idx].qh = qh;
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+void quantize(uint dst_idx, uint src_idx)
+{
+    float amax = 0.0; // absolute max
+
+    [[unroll]] for (int j = 0; j < QUANT_K_Q8_0; j++) {
+        const float v = data_s[src_idx + j];
+        amax = max(amax, abs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = (d != 0.0) ? 1.0/d : 0.0;
+
+    data_q[dst_idx].d = float16_t(d);
+
+    [[unroll]] for (int j = 0; j < QUANT_K_Q8_0; ++j) {
+        const float x0 = data_s[src_idx + j]*id;
+
+        data_q[dst_idx].qs[j] = int8_t(round(x0));
+    }
+}
+#endif
+
+#if defined(DATA_A_IQ4_NL)
+uint best_index(float x) {
+    if (x <= kvalues_iq4nl[0]) return 0;
+    if (x >= kvalues_iq4nl[15]) return 15;
+    int ml = 0, mu = 15;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < kvalues_iq4nl[mav]) mu = mav; else ml = mav;
+    }
+    return x - kvalues_iq4nl[mu-1] < kvalues_iq4nl[mu] - x ? mu-1 : mu;
+}
+
+void quantize(uint dst_idx, uint src_idx)
+{
+    float amax = 0.0;
+    float vmax = 0.0;
+
+    [[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL; ++j) {
+        const float v = data_s[src_idx + j];
+        if (amax < abs(v)) {
+            amax = abs(v);
+            vmax = v;
+        }
+    }
+
+    float d = vmax / kvalues_iq4nl[0];
+    const float id = (d != 0.0) ? 1.0/d : 0.0;
+
+    float sumqx = 0, sumq2 = 0;
+    [[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL/2; ++j) {
+        const float x0 = data_s[src_idx + 0                + j]*id;
+        const float x1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*id;
+        const uint xi0 = best_index(x0);
+        const uint xi1 = best_index(x1);
+        data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4));
+        const float v0 = kvalues_iq4nl[xi0];
+        const float v1 = kvalues_iq4nl[xi1];
+        const float w0 = data_s[src_idx + 0                + j]*data_s[src_idx + 0                + j];
+        const float w1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*data_s[src_idx + QUANT_K_IQ4_NL/2 + j];
+        sumqx += w0*v0*data_s[src_idx + j] + w1*v1*data_s[src_idx + QUANT_K_IQ4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+    }
+
+    data_q[dst_idx].d = float16_t(sumq2 > 0 ? sumqx/sumq2 : d);
+
+}
+#endif
+
+void main() {
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+    if (gl_LocalInvocationIndex.x != 0) {
+        return;
+    }
+#endif
+
+    const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    uint dst_idx = dst_idx_quant(idx, QUANT_K);
+    uint src_idx = get_aoffset() + src0_idx(idx);
+
+    quantize(dst_idx, src_idx);
+}

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -101,19 +101,25 @@ layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ2_
    block_q2_K block;
 };
 
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2_K_packed16 {
+   block_q2_K_packed16 block;
+};
+
 float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
+    decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl);
     const f16vec2 d = bl.block.d;
     const uint idx = coordInBlock[1];
-    const uint iqs = idx;
 
-    const uint qsi = (iqs / 128) * 32 + (iqs % 32);     // 0..31
-    const uint scalesi = iqs / 16;                      // 0..15
-    const uint qsshift = ((iqs % 128) / 32) * 2;        // 0,2,4,6
+    const uint scalesi = (idx & 0xF0) >> 4;             // 0..15
+    const uint qsshift = (idx & 0x60) >> 4;             // 0,2,4,6
+
+    uint qs = uint32_t(bl16.block.qs[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
+    qs = (qs >> qsshift) & 0x0303;
+    qs = unpack8(qs)[idx & 1];
 
-    uint32_t qs = bl.block.qs[qsi];
     const uint scales = bl.block.scales[scalesi];
-    float16_t ret = d.x * float16_t(scales & 0xF) * float16_t((qs >> qsshift) & 3) - d.y * float16_t(scales >> 4);
+    float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4);
     return ret;
 }
 
@@ -157,39 +163,47 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4
    block_q4_K_packed16 block;
 };
 
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed128 {
+   block_q4_K_packed128 block;
+};
+
 float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
     decodeBufQ4_K_packed16 bl16 = decodeBufQ4_K_packed16(bl);
+    decodeBufQ4_K_packed128 bl128 = decodeBufQ4_K_packed128(bl);
     const uint idx = coordInBlock[1];
 
     const uint b = (idx & 0x20) >> 5;            // 0,1
     const uint is = (idx & 0xE0) >> 5;         // 0..7
 
-    const f16vec2 loadd = bl.block.d;
+    uvec4 v = bl128.block.q4k[0];
+
+    const f16vec2 loadd = unpackFloat2x16(v.x);
 
     uint32_t sc;
     uint32_t mbyte;
 
-    uint32_t scidx0 = (is < 4) ? is : (is + 4);
-    uint32_t scidx1 = (is < 4) ? is : (is - 4);
-    uint32_t scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-    uint32_t scidxshift1 = (is < 4) ? 0 : 2;
-    uint32_t mbidx0 = is + 4;
-    uint32_t mbidx1 = (is < 4) ? is + 4 : is;
-    uint32_t mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-    uint32_t mbidxshift0 = (is < 4) ? 0 : 4;
-    uint32_t mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-    uint32_t mbidxshift1 = (is < 4) ? 0 : 2;
+    uint32_t scale0 = v.y;
+    uint32_t scale4 = v.z;
+    uint32_t scale8 = v.w;
 
-    sc    = uint8_t((bl.block.scales[scidx0] & 0xF)                         | ((bl.block.scales[scidx1] & scidxmask1) >> scidxshift1));
-    mbyte = uint8_t(((bl.block.scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((bl.block.scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+    uint32_t sc_lo = scale0;
+    uint32_t mb_lo = scale4;
+    uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2);
+    uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2);
+
+    sc = is < 4 ? sc_lo : sc_hi;
+    mbyte = is < 4 ? mb_lo : mb_hi;
+    sc = sc >> (8 * (is & 3));
+    mbyte = mbyte >> (8 * (is & 3));
+    sc &= 0x3F;
+    mbyte &= 0x3F;
 
     const float16_t d = loadd.x * float16_t(sc);
     const float16_t m = loadd.y * float16_t(mbyte);
 
     uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
-    qs = (qs >> (b * 4)) & 0x0F0F;
-    qs = unpack8(qs)[idx & 1];
+    qs = (qs >> (b * 4 + 8 * (idx & 1))) & 0xF;
 
     float16_t ret = d * float16_t(qs) - m;
 
@@ -204,47 +218,53 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5
    block_q5_K_packed16 block;
 };
 
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed128 {
+   block_q5_K_packed128 block;
+};
+
 float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
     decodeBufQ5_K_packed16 bl16 = decodeBufQ5_K_packed16(bl);
+    decodeBufQ5_K_packed128 bl128 = decodeBufQ5_K_packed128(bl);
     const uint idx = coordInBlock[1];
 
     const uint b = (idx & 0x20) >> 5;          // 0,1
     const uint is = (idx & 0xE0) >> 5;         // 0..7
 
-    const uint32_t hm = 0x0101 << is;
+    uvec4 v = bl128.block.q5k[0];
 
-    const f16vec2 loadd = bl.block.d;
+    const f16vec2 loadd = unpackFloat2x16(v.x);
 
     uint32_t sc;
     uint32_t mbyte;
 
-    uint32_t scidx0 = (is < 4) ? is : (is + 4);
-    uint32_t scidx1 = (is < 4) ? is : (is - 4);
-    uint32_t scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-    uint32_t scidxshift1 = (is < 4) ? 0 : 2;
-    uint32_t mbidx0 = is + 4;
-    uint32_t mbidx1 = (is < 4) ? is + 4 : is;
-    uint32_t mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-    uint32_t mbidxshift0 = (is < 4) ? 0 : 4;
-    uint32_t mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-    uint32_t mbidxshift1 = (is < 4) ? 0 : 2;
+    uint32_t scale0 = v.y;
+    uint32_t scale4 = v.z;
+    uint32_t scale8 = v.w;
 
-    sc    = uint8_t((bl.block.scales[scidx0] & 0xF)                         | ((bl.block.scales[scidx1] & scidxmask1) >> scidxshift1));
-    mbyte = uint8_t(((bl.block.scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((bl.block.scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+    uint32_t sc_lo = scale0;
+    uint32_t mb_lo = scale4;
+    uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2);
+    uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2);
+
+    sc = is < 4 ? sc_lo : sc_hi;
+    mbyte = is < 4 ? mb_lo : mb_hi;
+    sc = sc >> (8 * (is & 3));
+    mbyte = mbyte >> (8 * (is & 3));
+    sc &= 0x3F;
+    mbyte &= 0x3F;
 
     const float16_t d = loadd.x * float16_t(sc);
     const float16_t m = loadd.y * float16_t(mbyte);
 
     uint qh = uint32_t(bl16.block.qh[(idx & 0x1E) >> 1]);
-    qh = qh & hm;
-    qh = unpack8(qh)[idx & 1];
+    qh = ((qh >> is) & 0x101) << 4;
 
     uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
     qs = (qs >> (b * 4)) & 0x0F0F;
-    qs = unpack8(qs)[idx & 1];
+    qs = unpack8(qs | qh)[idx & 1];
 
-    float16_t ret = d * (float16_t(qs) + (qh != 0 ? float16_t(16) : float16_t(0))) - m;
+    float16_t ret = d * (float16_t(qs)) - m;
 
     return ret;
 }

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -42,10 +42,13 @@ layout (push_constant) uniform parameter {
     uint32_t nev3;
     uint32_t nem1;
 
+    uint32_t nb01;
     uint32_t nb02;
     uint32_t nb03;
+    uint32_t nb11;
     uint32_t nb12;
     uint32_t nb13;
+    uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
     uint32_t nb31;
@@ -146,6 +149,23 @@ void main() {
     tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
     tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
 
+    // nb?1 are already divided by the type size and are in units of elements
+    uint32_t q_stride = p.nb01;
+    uint32_t k_stride = p.nb11;
+    uint32_t v_stride = p.nb21;
+    // hint to the compiler that strides are aligned for the aligned variant of the shader
+    if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
+    {
+        q_stride &= ~7;
+#if !defined(BLOCK_SIZE)
+        k_stride &= ~7;
+        v_stride &= ~7;
+#endif
+    }
+    tensorLayoutQ = setTensorLayoutStrideNV(tensorLayoutQ, q_stride, 1);
+    tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
+    tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
+
     coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Q;
     coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16;
 

ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp

@@ -54,3 +54,23 @@ uint dst_idx(uint idx) {
     const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
     return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
 }
+
+uint src0_idx_quant(uint idx, uint qk) {
+    const uint i03 = fastdiv(idx, p.ne0_012mp, p.ne0_012L);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, p.ne0_01L);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, p.ne0_0L);
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + (i00/qk)*p.nb00;
+}
+
+uint dst_idx_quant(uint idx, uint qk) {
+    const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L);
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+    return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + (i10/qk)*p.nb10;
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -5,6 +5,80 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
+shared FLOAT_TYPE sccache1[BLOCK_SIZE/16][16];
+shared FLOAT_TYPE sccache2[BLOCK_SIZE/16][16];
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
+    const uint y_idx = i * QUANT_K + y_offset;
+
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+
+        barrier();
+        if (!all_threads) { // when we don't have enough blocks to use all threads
+            if (i < num_blocks_per_row) {
+                const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]);
+                sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF);
+                sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF);
+            }
+            barrier();
+
+            if (i >= num_blocks_per_row)
+                continue;
+        } else {
+            const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]);
+            sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF);
+            sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF);
+            barrier();
+        }
+
+        const uint32_t qs_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
+        const vec4 qs_u32_0 = vec4(unpack8(qs_u32 & 0x03030303));
+        const vec4 qs_u32_2 = vec4(unpack8((qs_u32 >> 2) & 0x03030303));
+        const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
+        const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
+
+        vec2 d = vec2(data_a[ib0 + i].d);
+        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);
+            vec2 b16 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  8]);
+            vec2 b32 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
+            vec2 b48 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
+            vec2 b64 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
+            vec2 b80 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
+            vec2 b96 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
+            vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);
+
+            FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
+            FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
+            [[unroll]] for (int l = 0; l < 2; ++l) {
+                sum1 = fma(FLOAT_TYPE(b0[l]),   sccache1[ix][    8*v_im] * qs_u32_0[l  ],
+                       fma(FLOAT_TYPE(b16[l]),  sccache1[ix][1 + 8*v_im] * qs_u32_0[l+2],
+                       fma(FLOAT_TYPE(b32[l]),  sccache1[ix][2 + 8*v_im] * qs_u32_2[l  ],
+                       fma(FLOAT_TYPE(b48[l]),  sccache1[ix][3 + 8*v_im] * qs_u32_2[l+2],
+                       fma(FLOAT_TYPE(b64[l]),  sccache1[ix][4 + 8*v_im] * qs_u32_4[l  ],
+                       fma(FLOAT_TYPE(b80[l]),  sccache1[ix][5 + 8*v_im] * qs_u32_4[l+2],
+                       fma(FLOAT_TYPE(b96[l]),  sccache1[ix][6 + 8*v_im] * qs_u32_6[l  ],
+                       fma(FLOAT_TYPE(b112[l]), sccache1[ix][7 + 8*v_im] * qs_u32_6[l+2], sum1))))))));
+                sum2 = fma(FLOAT_TYPE(b0[l]),   sccache2[ix][    8*v_im],
+                       fma(FLOAT_TYPE(b16[l]),  sccache2[ix][1 + 8*v_im],
+                       fma(FLOAT_TYPE(b32[l]),  sccache2[ix][2 + 8*v_im],
+                       fma(FLOAT_TYPE(b48[l]),  sccache2[ix][3 + 8*v_im],
+                       fma(FLOAT_TYPE(b64[l]),  sccache2[ix][4 + 8*v_im],
+                       fma(FLOAT_TYPE(b80[l]),  sccache2[ix][5 + 8*v_im],
+                       fma(FLOAT_TYPE(b96[l]),  sccache2[ix][6 + 8*v_im],
+                       fma(FLOAT_TYPE(b112[l]), sccache2[ix][7 + 8*v_im], sum2))))))));
+            }
+            temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
+        }
+    }
+}
+
 void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
@@ -14,88 +88,28 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
     const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint itid = tid%16;  // 0...15
+    const uint ix = tid/16;
 
-    const uint step = 8;
-
-    const uint v_im = itid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = itid - step*v_im;                      // 0...15 or 0...7
+    const uint v_im = itid/8;                                // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = itid - 8*v_im;                         // 0...7
 
     const uint l0 = 2*v_in;                                  // 0...15
     const uint q_offset = 32*v_im + l0;
-    const uint s_offset = 8*v_im;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
-
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
             temp[j][i] = FLOAT_TYPE(0);
         }
     }
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
-        const uint y_idx = i * QUANT_K + y_offset;
-
-        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-            vec2 d = vec2(data_a[ib0 + i].d);
-            const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-            const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-
-            uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
-            uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
-
-            uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
-            uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
-            uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
-            uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
-
-            uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
-            uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
-            uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
-            uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
-
-            uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
-            uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
-            uvec2 qs0 =  uvec2(unpack8(qs0_u16));
-            uvec2 qs16 = uvec2(unpack8(qs16_u16));
-
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);
-                vec2 b16 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  8]);
-                vec2 b32 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
-                vec2 b48 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
-                vec2 b64 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
-                vec2 b80 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
-                vec2 b96 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
-                vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);
-
-                FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
-                FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
-                [[unroll]] for (int l = 0; l < 2; ++l) {
-                    sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
-                           fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
-                           fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
-                           fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
-                           fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 4) & 3),
-                           fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
-                           fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 6) & 3),
-                           fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
-                    sum2 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_hi4[0]),
-                           fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_hi4[1]),
-                           fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_hi4[2]),
-                           fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_hi4[3]),
-                           fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_hi4[0]),
-                           fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_hi4[1]),
-                           fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
-                           fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
-                }
-                temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
-            }
-        }
-    }
+    const uint nbr_par_th = num_blocks_per_row%it_size;
+    const uint nbr_all_th = num_blocks_per_row - nbr_par_th;
+    uint i0 = 0;
+    [[unroll]] for (; i0 < nbr_all_th; i0 += it_size)
+        calc_superblock(a_offset, b_offset, itid, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, true);
+    calc_superblock(a_offset, b_offset, itid, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, false);
 
     reduce_result(temp, d_offset, first_row, num_rows, tid);
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -5,6 +5,74 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
+shared FLOAT_TYPE sccache[BLOCK_SIZE/16][2][8];
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, const uint itid8, const uint v_im, const uint v_im4, const uint v_in, const uint32_t hm_m[4], const uint q_offset, const uint y_offset, const uint s_shift, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
+    const uint y_idx = i * QUANT_K + y_offset;
+
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+
+        if (!all_threads) { // when we don't have enough blocks to use all threads
+            barrier();
+            if (i < num_blocks_per_row)
+                sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
+            barrier();
+
+            if (i >= num_blocks_per_row)
+                continue;
+        }
+
+        const uint32_t hmk = ~(uint32_t(data_a_packed16[ib0 + i].hmask[v_in]) | (uint32_t(data_a_packed16[ib0 + i].hmask[v_in + 8]) << 16));
+        const vec4 hmk_0 = vec4(unpack8(((hmk & hm_m[0]) >> (    v_im4)) << 2));
+        const vec4 hmk_1 = vec4(unpack8(((hmk & hm_m[1]) >> (1 + v_im4)) << 2));
+        const vec4 hmk_2 = vec4(unpack8(((hmk & hm_m[2]) >> (2 + v_im4)) << 2));
+        const vec4 hmk_3 = vec4(unpack8(((hmk & hm_m[3]) >> (3 + v_im4)) << 2));
+
+        // 0, 1, 16, 17
+        uint32_t qs_u32 = uint32_t(data_a[ib0 + i].qs[q_offset]) | (uint32_t(data_a[ib0 + i].qs[q_offset + 1]) << 8);
+        qs_u32 |= (uint32_t(data_a[ib0 + i].qs[q_offset + 16]) | (uint32_t(data_a[ib0 + i].qs[q_offset + 17]) << 8)) << 16;
+        const vec4 qs_u32_0 = vec4(unpack8(qs_u32 & 0x03030303));
+        const vec4 qs_u32_2 = vec4(unpack8((qs_u32 >> 2) & 0x03030303));
+        const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
+        const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
+
+        if (all_threads) {
+            barrier();
+            sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
+            barrier();
+        }
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);
+            vec2 b16 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  8]);
+            vec2 b32 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
+            vec2 b48 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
+            vec2 b64 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
+            vec2 b80 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
+            vec2 b96 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
+            vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);
+
+            FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+            [[unroll]] for (int l = 0; l < 2; ++l) {
+                sum = fma(FLOAT_TYPE(  b0[l]) * sccache[ix][v_im][0], qs_u32_0[l  ] - hmk_0[l  ],
+                      fma(FLOAT_TYPE( b16[l]) * sccache[ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2],
+                      fma(FLOAT_TYPE( b32[l]) * sccache[ix][v_im][2], qs_u32_2[l  ] - hmk_1[l  ],
+                      fma(FLOAT_TYPE( b48[l]) * sccache[ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2],
+                      fma(FLOAT_TYPE( b64[l]) * sccache[ix][v_im][4], qs_u32_4[l  ] - hmk_2[l  ],
+                      fma(FLOAT_TYPE( b80[l]) * sccache[ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2],
+                      fma(FLOAT_TYPE( b96[l]) * sccache[ix][v_im][6], qs_u32_6[l  ] - hmk_3[l  ],
+                      fma(FLOAT_TYPE(b112[l]) * sccache[ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum))))))));
+            }
+            temp[j][n] = fma(d, sum, temp[j][n]);
+        }
+    }
+}
+
 void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
@@ -14,76 +82,37 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
     const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint itid = tid%16;  // 0...15
+    const uint ix = tid/16;
+    const uint itid8 = itid%8;
 
-    const uint step = 8;
+    const uint v_im = itid/8;                               // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_im4 = v_im*4;
+    const uint v_in = itid - 8*v_im;                        // 0...7
 
-    const uint v_im = itid/step;                            // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = itid - step*v_im;                     // 0...15 or 0...7
-
-    const uint8_t m = uint8_t(1 << (4 * v_im));
+    const uint32_t m = 0x01010101 << (4 * v_im);
+    uint32_t hm_m[4];
+    [[unroll]] for (uint j = 0; j < 4; ++j)
+        hm_m[j] = m << j;
 
     const uint l0 = 2*v_in;                                 // 0...15
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
-
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
             temp[j][i] = FLOAT_TYPE(0);
         }
     }
 
-    const uint s_shift = 4 * v_im;
+    const uint s_shift = v_im4 + 2*(itid8/4);
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
-        const uint y_idx = i * QUANT_K + y_offset;
-
-        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-            const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
-
-            uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
-            uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
-            uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
-            uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
-            uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
-            uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
-            u8vec2 s0 = unpack8(s0_16);
-            u8vec2 s2 = unpack8(s2_16);
-            u8vec2 s4 = unpack8(s4_16);
-            u8vec2 s6 = unpack8(s6_16);
-            u8vec2 s8 = unpack8(s8_16);
-            u8vec2 s10 = unpack8(s10_16);
-
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-
-                vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);
-                vec2 b16 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  8]);
-                vec2 b32 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
-                vec2 b48 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
-                vec2 b64 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
-                vec2 b80 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
-                vec2 b96 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
-                vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);
-
-                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-                [[unroll]] for (int l = 0; l < 2; ++l) {
-                    sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
-                          fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
-                          fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
-                          fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
-                          fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
-                          fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
-                          fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
-                          fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
-                }
-                temp[j][n] = fma(d, sum, temp[j][n]);
-            }
-        }
-    }
+    const uint nbr_par_th = num_blocks_per_row%it_size;
+    const uint nbr_all_th = num_blocks_per_row - nbr_par_th;
+    uint i0 = 0;
+    [[unroll]] for (; i0 < nbr_all_th; i0 += it_size)
+        calc_superblock(a_offset, b_offset, ix, itid8, v_im, v_im4, v_in, hm_m, q_offset, y_offset, s_shift, i0 + ix, num_blocks_per_row, first_row, num_rows, true);
+    calc_superblock(a_offset, b_offset, ix, itid8, v_im, v_im4, v_in, hm_m, q_offset, y_offset, s_shift, i0 + ix, num_blocks_per_row, first_row, num_rows, false);
 
     reduce_result(temp, d_offset, first_row, num_rows, tid);
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -6,6 +6,86 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y1_idx = i * QUANT_K + y_offset;
+    const uint y2_idx = y1_idx + 128;
+
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        vec2 d = vec2(data_a[ib0 + i].d);
+        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+        const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+        const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+        const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+
+        const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32;
+        const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2;
+        const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F));
+        const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h));
+
+        const FLOAT_TYPE sc0 = scale_0_4_l_f.x;
+        const FLOAT_TYPE sc1 = scale_0_4_l_f.y;
+        const FLOAT_TYPE sc2 = scale_0_4_l_f.z;
+        const FLOAT_TYPE sc3 = scale_0_4_l_f.w;
+        const FLOAT_TYPE sc4 = scale8_f.x;
+        const FLOAT_TYPE sc5 = scale8_f.y;
+        const FLOAT_TYPE sc6 = scale8_f.z;
+        const FLOAT_TYPE sc7 = scale8_f.w;
+
+        const uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
+        const uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
+
+        const uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
+        const uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
+        const uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
+        const uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
+
+        const vec4 qs0_lo4 = vec4(unpack8(qs0_u32_lo4));
+        const vec4 qs64_lo4 = vec4(unpack8(qs64_u32_lo4));
+        const vec4 qs0_hi4 = vec4(unpack8(qs0_u32_hi4));
+        const vec4 qs64_hi4 = vec4(unpack8(qs64_u32_hi4));
+
+        const FLOAT_TYPE q4_0  = qs0_lo4.x;
+        const FLOAT_TYPE q4_1  = qs0_lo4.y;
+        const FLOAT_TYPE q4_2  = qs0_lo4.z;
+        const FLOAT_TYPE q4_3  = qs0_lo4.w;
+        const FLOAT_TYPE q4_4  = qs0_hi4.x;
+        const FLOAT_TYPE q4_5  = qs0_hi4.y;
+        const FLOAT_TYPE q4_6  = qs0_hi4.z;
+        const FLOAT_TYPE q4_7  = qs0_hi4.w;
+        const FLOAT_TYPE q4_8  = qs64_lo4.x;
+        const FLOAT_TYPE q4_9  = qs64_lo4.y;
+        const FLOAT_TYPE q4_10 = qs64_lo4.z;
+        const FLOAT_TYPE q4_11 = qs64_lo4.w;
+        const FLOAT_TYPE q4_12 = qs64_hi4.x;
+        const FLOAT_TYPE q4_13 = qs64_hi4.y;
+        const FLOAT_TYPE q4_14 = qs64_hi4.z;
+        const FLOAT_TYPE q4_15 = qs64_hi4.w;
+
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            vec4 by10 =  vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4    ]);
+            vec4 by132 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8]);
+            vec4 by20 =  vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4    ]);
+            vec4 by232 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8]);
+
+            const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
+            const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
+            const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
+            const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
+            const FLOAT_TYPE smin =
+                fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
+                fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
+                fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
+                fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
+            temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+        }
+    }
+}
+
 void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
@@ -15,13 +95,11 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
     const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint itid = tid%16;  // 0...15
+    const uint ix = tid/16;
 
-    const uint step = 4;
-
-    const uint il = itid/step;                      // 0...3
-    const uint ir = itid - step*il;                 // 0...7 or 0...3
+    const uint il = itid/4;                         // 0...3
+    const uint ir = itid - 4*il;                    // 0...3
     const uint n =  4;
 
     const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
@@ -31,89 +109,14 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*v_im + l0;
 
-    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
-
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
             temp[j][i] = FLOAT_TYPE(0);
         }
     }
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
-        const uint y1_idx = i * QUANT_K + y_offset;
-        const uint y2_idx = y1_idx + 128;
-
-        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-            vec2 d = vec2(data_a[ib0 + i].d);
-            const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-            const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-
-            uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
-            uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
-            uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
-            uvec4 scale0 = uvec4(unpack8(scale0_u32));
-            uvec4 scale4 = uvec4(unpack8(scale4_u32));
-            uvec4 scale8 = uvec4(unpack8(scale8_u32));
-
-            const uint32_t sc0 = (  scale0.x       & 0x3f);
-            const uint32_t sc1 = (  scale0.y       & 0x3f);
-            const uint32_t sc2 = (  scale4.x       & 0x3f);
-            const uint32_t sc3 = (  scale4.y       & 0x3f);
-            const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
-            const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
-            const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
-            const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
-
-            uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
-            uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
-
-            uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
-            uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
-            uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
-            uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
-
-            uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
-            uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
-            uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
-            uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
-
-            const uint32_t q4_0  = qs0_lo4.x;
-            const uint32_t q4_1  = qs0_lo4.y;
-            const uint32_t q4_2  = qs0_lo4.z;
-            const uint32_t q4_3  = qs0_lo4.w;
-            const uint32_t q4_4  = qs0_hi4.x;
-            const uint32_t q4_5  = qs0_hi4.y;
-            const uint32_t q4_6  = qs0_hi4.z;
-            const uint32_t q4_7  = qs0_hi4.w;
-            const uint32_t q4_8  = qs64_lo4.x;
-            const uint32_t q4_9  = qs64_lo4.y;
-            const uint32_t q4_10 = qs64_lo4.z;
-            const uint32_t q4_11 = qs64_lo4.w;
-            const uint32_t q4_12 = qs64_hi4.x;
-            const uint32_t q4_13 = qs64_hi4.y;
-            const uint32_t q4_14 = qs64_hi4.z;
-            const uint32_t q4_15 = qs64_hi4.w;
-
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec4 by10 =  vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4    ]);
-                vec4 by132 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8]);
-                vec4 by20 =  vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4    ]);
-                vec4 by232 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8]);
-
-                const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
-                const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
-                const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
-                const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
-                const FLOAT_TYPE smin =
-                    fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
-                    fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
-                    fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
-                    fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
-                temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
-            }
-        }
-    }
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size)
+        calc_superblock(a_offset, b_offset, v_im, q_offset, y_offset, i, num_blocks_per_row, first_row, num_rows);
 
     reduce_result(temp, d_offset, first_row, num_rows, tid);
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -6,6 +6,118 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im, const uint l0, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y1_idx = i * QUANT_K + y_offset;
+    const uint y2_idx = y1_idx + 128;
+
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        vec2 d = vec2(data_a[ib0 + i].d);
+        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+        const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+        const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+        const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+
+        const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32;
+        const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2;
+        const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F));
+        const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h));
+
+        const FLOAT_TYPE sc0 = scale_0_4_l_f.x;
+        const FLOAT_TYPE sc1 = scale_0_4_l_f.y;
+        const FLOAT_TYPE sc2 = scale_0_4_l_f.z;
+        const FLOAT_TYPE sc3 = scale_0_4_l_f.w;
+        const FLOAT_TYPE sc4 = scale8_f.x;
+        const FLOAT_TYPE sc5 = scale8_f.y;
+        const FLOAT_TYPE sc6 = scale8_f.z;
+        const FLOAT_TYPE sc7 = scale8_f.w;
+
+        const uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
+        const uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
+
+        uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
+        uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
+        uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
+
+        const uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
+
+        const uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
+        const uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
+        const uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010);
+        const uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
+
+        qs0_16_u32_lo4 += qs0_16_lo4_offset16;
+        qs0_16_u32_hi4 += qs0_16_hi4_offset16;
+        qs64_80_u32_lo4 += qs64_80_lo4_offset16;
+        qs64_80_u32_hi4 += qs64_80_hi4_offset16;
+
+        const vec4 qs0_16_lo4 = vec4(unpack8(qs0_16_u32_lo4));
+        const vec4 qs64_80_lo4 = vec4(unpack8(qs64_80_u32_lo4));
+        const vec4 qs0_16_hi4 = vec4(unpack8(qs0_16_u32_hi4));
+        const vec4 qs64_80_hi4 = vec4(unpack8(qs64_80_u32_hi4));
+
+        const FLOAT_TYPE q4_0  = qs0_16_lo4.x;
+        const FLOAT_TYPE q4_1  = qs0_16_lo4.y;
+        const FLOAT_TYPE q4_2  = qs0_16_lo4.z;
+        const FLOAT_TYPE q4_3  = qs0_16_lo4.w;
+        const FLOAT_TYPE q4_4  = qs0_16_hi4.x;
+        const FLOAT_TYPE q4_5  = qs0_16_hi4.y;
+        const FLOAT_TYPE q4_6  = qs0_16_hi4.z;
+        const FLOAT_TYPE q4_7  = qs0_16_hi4.w;
+        const FLOAT_TYPE q4_8  = qs64_80_lo4.x;
+        const FLOAT_TYPE q4_9  = qs64_80_lo4.y;
+        const FLOAT_TYPE q4_10 = qs64_80_lo4.z;
+        const FLOAT_TYPE q4_11 = qs64_80_lo4.w;
+        const FLOAT_TYPE q4_12 = qs64_80_hi4.x;
+        const FLOAT_TYPE q4_13 = qs64_80_hi4.y;
+        const FLOAT_TYPE q4_14 = qs64_80_hi4.z;
+        const FLOAT_TYPE q4_15 = qs64_80_hi4.w;
+
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            vec2 by10 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2     ]);
+            vec2 by116 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 +  8]);
+            vec2 by132 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16]);
+            vec2 by148 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24]);
+            vec2 by20 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2     ]);
+            vec2 by216 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 +  8]);
+            vec2 by232 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16]);
+            vec2 by248 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24]);
+
+            const FLOAT_TYPE sx =
+              fma(FLOAT_TYPE(by10.x), q4_0,
+              fma(FLOAT_TYPE(by10.y), q4_1,
+              fma(FLOAT_TYPE(by116.x), q4_2,
+                 FLOAT_TYPE(by116.y) * q4_3)));
+            const FLOAT_TYPE sy =
+              fma(FLOAT_TYPE(by132.x), q4_4,
+              fma(FLOAT_TYPE(by132.y), q4_5,
+              fma(FLOAT_TYPE(by148.x), q4_6,
+                 FLOAT_TYPE(by148.y) * q4_7)));
+            const FLOAT_TYPE sz =
+              fma(FLOAT_TYPE(by20.x), q4_8,
+              fma(FLOAT_TYPE(by20.y), q4_9,
+              fma(FLOAT_TYPE(by216.x), q4_10,
+                 FLOAT_TYPE(by216.y) * q4_11)));
+            const FLOAT_TYPE sw =
+              fma(FLOAT_TYPE(by232.x), q4_12,
+              fma(FLOAT_TYPE(by232.y), q4_13,
+              fma(FLOAT_TYPE(by248.x), q4_14,
+                 FLOAT_TYPE(by248.y) * q4_15)));
+            const FLOAT_TYPE smin =
+              fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
+              fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
+              fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
+                  (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
+            temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+        }
+    }
+}
+
 void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
@@ -15,11 +127,11 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
     const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint itid = tid%16;  // 0...15
+    const uint ix = tid/16;
 
     const uint il = itid/4;                          // 0...3
-    const uint ir = itid - 4*il;                     // 0...7 or 0...3
+    const uint ir = itid - 4*il;                     // 0...3
 
     const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
     const uint v_in = il % 2;
@@ -28,121 +140,14 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*v_im + l0;
 
-    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
-
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
             temp[j][i] = FLOAT_TYPE(0);
         }
     }
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
-        const uint y1_idx = i * QUANT_K + y_offset;
-        const uint y2_idx = y1_idx + 128;
-
-        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-            vec2 d = vec2(data_a[ib0 + i].d);
-            const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-            const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-
-            uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
-            uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
-            uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
-            uvec4 scale0 = uvec4(unpack8(scale0_u32));
-            uvec4 scale4 = uvec4(unpack8(scale4_u32));
-            uvec4 scale8 = uvec4(unpack8(scale8_u32));
-
-            const uint32_t sc0 = (  scale0.x       & 0x3f);
-            const uint32_t sc1 = (  scale0.y       & 0x3f);
-            const uint32_t sc2 = (  scale4.x       & 0x3f);
-            const uint32_t sc3 = (  scale4.y       & 0x3f);
-            const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
-            const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
-            const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
-            const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
-
-            uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
-            uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
-
-            uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
-            uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
-            uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
-            uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
-
-            uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
-
-            uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
-            uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
-            uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
-            uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
-
-            qs0_16_u32_lo4 += qs0_16_lo4_offset16;
-            qs0_16_u32_hi4 += qs0_16_hi4_offset16;
-            qs64_80_u32_lo4 += qs64_80_lo4_offset16;
-            qs64_80_u32_hi4 += qs64_80_hi4_offset16;
-
-            uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
-            uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
-            uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
-            uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
-
-            const uint32_t q4_0  = qs0_16_lo4.x;
-            const uint32_t q4_1  = qs0_16_lo4.y;
-            const uint32_t q4_2  = qs0_16_lo4.z;
-            const uint32_t q4_3  = qs0_16_lo4.w;
-            const uint32_t q4_4  = qs0_16_hi4.x;
-            const uint32_t q4_5  = qs0_16_hi4.y;
-            const uint32_t q4_6  = qs0_16_hi4.z;
-            const uint32_t q4_7  = qs0_16_hi4.w;
-            const uint32_t q4_8  = qs64_80_lo4.x;
-            const uint32_t q4_9  = qs64_80_lo4.y;
-            const uint32_t q4_10 = qs64_80_lo4.z;
-            const uint32_t q4_11 = qs64_80_lo4.w;
-            const uint32_t q4_12 = qs64_80_hi4.x;
-            const uint32_t q4_13 = qs64_80_hi4.y;
-            const uint32_t q4_14 = qs64_80_hi4.z;
-            const uint32_t q4_15 = qs64_80_hi4.w;
-
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec2 by10 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2     ]);
-                vec2 by116 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 +  8]);
-                vec2 by132 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16]);
-                vec2 by148 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24]);
-                vec2 by20 =  vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2     ]);
-                vec2 by216 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 +  8]);
-                vec2 by232 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16]);
-                vec2 by248 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24]);
-
-                const FLOAT_TYPE sx =
-                  fma(FLOAT_TYPE(by10.x), q4_0,
-                  fma(FLOAT_TYPE(by10.y), q4_1,
-                  fma(FLOAT_TYPE(by116.x), q4_2,
-                     FLOAT_TYPE(by116.y) * q4_3)));
-                const FLOAT_TYPE sy =
-                  fma(FLOAT_TYPE(by132.x), q4_4,
-                  fma(FLOAT_TYPE(by132.y), q4_5,
-                  fma(FLOAT_TYPE(by148.x), q4_6,
-                     FLOAT_TYPE(by148.y) * q4_7)));
-                const FLOAT_TYPE sz =
-                  fma(FLOAT_TYPE(by20.x), q4_8,
-                  fma(FLOAT_TYPE(by20.y), q4_9,
-                  fma(FLOAT_TYPE(by216.x), q4_10,
-                     FLOAT_TYPE(by216.y) * q4_11)));
-                const FLOAT_TYPE sw =
-                  fma(FLOAT_TYPE(by232.x), q4_12,
-                  fma(FLOAT_TYPE(by232.y), q4_13,
-                  fma(FLOAT_TYPE(by248.x), q4_14,
-                     FLOAT_TYPE(by248.y) * q4_15)));
-                const FLOAT_TYPE smin =
-                  fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
-                  fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
-                  fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
-                      (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
-                temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
-            }
-        }
-    }
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size)
+        calc_superblock(a_offset, b_offset, v_im, l0, q_offset, y_offset, i, num_blocks_per_row, first_row, num_rows);
 
     reduce_result(temp, d_offset, first_row, num_rows, tid);
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -6,7 +6,77 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+shared FLOAT_TYPE sccache[BLOCK_SIZE/16][16];
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint ix, const uint ql_offset, const uint qh_offset, const uint s_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
+    const uint y_idx = i * QUANT_K + y_offset;
+
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+
+        if (!all_threads) { // when we don't have enough blocks to use all threads
+            barrier();
+            if (i < num_blocks_per_row)
+                sccache[ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]);
+            barrier();
+
+            if (i >= num_blocks_per_row)
+                continue;
+        }
+
+        const uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
+        const uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
+
+        const uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
+        const uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
+        const uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
+        const uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
+
+        const uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
+        const uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
+        const uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
+        const uint32_t qh4_u32 = (qh_u32 & 0x30303030);
+        const uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
+
+        const uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
+        const uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
+        const uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
+        const uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
+
+        const vec4 q0 = vec4(unpack8(q0_u32)) - 32;
+        const vec4 q1 = vec4(unpack8(q1_u32)) - 32;
+        const vec4 q2 = vec4(unpack8(q2_u32)) - 32;
+        const vec4 q3 = vec4(unpack8(q3_u32)) - 32;
+
+        if (all_threads) {
+            barrier();
+            sccache[ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]);
+            barrier();
+        }
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            vec4 by0  = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4     ]);
+            vec4 by32 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 +  8]);
+            vec4 by64 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 16]);
+            vec4 by96 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 24]);
+
+            FLOAT_TYPE sum[4] = {0, 0, 0, 0};
+            [[unroll]] for (uint l = 0; l < 4; ++l) {
+                sum[0] = fma(FLOAT_TYPE(by0[l]), q0[l], sum[0]);
+                sum[1] = fma(FLOAT_TYPE(by32[l]), q1[l], sum[1]);
+                sum[2] = fma(FLOAT_TYPE(by64[l]), q2[l], sum[2]);
+                sum[3] = fma(FLOAT_TYPE(by96[l]), q3[l], sum[3]);
+            }
+            temp[j][n] = fma(fma(sum[0], sccache[ix][s_offset], fma(sum[1], sccache[ix][s_offset + 2], fma(sum[2], sccache[ix][s_offset + 4], sum[3] * sccache[ix][s_offset + 6]))), d, temp[j][n]);
+        }
+    }
+}
+
+void compute_outputs(const uint first_row, const uint num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -15,13 +85,11 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
     const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid%16;  // 0...16
-    const uint ix  = tid/16;
+    const uint itid = tid%16;  // 0...15
+    const uint ix = tid/16;
 
-    const uint step = 8;
-
-    const uint v_im = itid/step;                            // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = itid - step*v_im;                     // 0...15 or 0...7
+    const uint v_im = itid/8;                               // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = itid - 8*v_im;                        // 0...7
 
     const uint l0 = 4 * v_in;                               // 0, 4, 8, ..., 28
     const uint is = v_in / 4;
@@ -31,68 +99,18 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint s_offset  =  8*v_im + is;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
-
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
             temp[j][i] = FLOAT_TYPE(0);
         }
     }
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
-        const uint y_idx = i * QUANT_K + y_offset;
-
-        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-            const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
-
-            FLOAT_TYPE scales[4];
-            scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
-            scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
-            scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
-            scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
-
-            uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
-            uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
-
-            uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
-            uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
-            uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
-            uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
-
-            uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
-            uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
-            uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
-            uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
-            uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
-
-            uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
-            uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
-            uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
-            uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
-
-            uvec4 q0 = uvec4(unpack8(q0_u32));
-            uvec4 q1 = uvec4(unpack8(q1_u32));
-            uvec4 q2 = uvec4(unpack8(q2_u32));
-            uvec4 q3 = uvec4(unpack8(q3_u32));
-
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec4 by0  = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4     ]);
-                vec4 by32 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 +  8]);
-                vec4 by64 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 16]);
-                vec4 by96 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 24]);
-
-                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-                [[unroll]] for (int l = 0; l < 4; ++l) {
-                    sum = fma(FLOAT_TYPE(by0[l])  * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
-                          fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
-                          fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
-                          fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
-                }
-                temp[j][n] += sum * d;
-            }
-        }
-    }
+    const uint nbr_par_th = num_blocks_per_row%it_size;
+    const uint nbr_all_th = num_blocks_per_row - nbr_par_th;
+    uint i0 = 0;
+    [[unroll]] for (; i0 < nbr_all_th; i0 += it_size)
+        calc_superblock(a_offset, b_offset, itid, ix, ql_offset, qh_offset, s_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, true);
+    calc_superblock(a_offset, b_offset, itid, ix, ql_offset, qh_offset, s_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, false);
 
     reduce_result(temp, d_offset, first_row, num_rows, tid);
 }

ggml/src/ggml-vulkan/vulkan-shaders/types.comp

@@ -227,6 +227,11 @@ struct block_q4_K_packed32
     uint32_t qs[QUANT_K_Q4_K/2/4];
 };
 
+struct block_q4_K_packed128
+{
+    uvec4 q4k[9];
+};
+
 #if defined(DATA_A_Q4_K)
 #define QUANT_K QUANT_K_Q4_K
 #define A_TYPE block_q4_K
@@ -252,6 +257,11 @@ struct block_q5_K_packed16
     uint16_t qs[QUANT_K_Q5_K/2/2];
 };
 
+struct block_q5_K_packed128
+{
+    uvec4 q5k[11];
+};
+
 #if defined(DATA_A_Q5_K)
 #define QUANT_K QUANT_K_Q5_K
 #define A_TYPE block_q5_K

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

@@ -417,6 +417,11 @@ void process_shaders() {
     string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
     string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
 
+    for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
+        string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+        string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }
+
     string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
 

ggml/src/ggml.c

@@ -3450,12 +3450,14 @@ struct ggml_tensor * ggml_soft_max_ext(
     return ggml_soft_max_impl(ctx, a, mask, scale, max_bias, false);
 }
 
-// ggml_soft_max_back
+// ggml_soft_max_ext_back
 
-static struct ggml_tensor * ggml_soft_max_back_impl(
+static struct ggml_tensor * ggml_soft_max_ext_back_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
+        float                 scale,
+        float                 max_bias,
         bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
@@ -3463,21 +3465,28 @@ static struct ggml_tensor * ggml_soft_max_back_impl(
     result->src[0] = a;
     result->src[1] = b;
 
+    memcpy((float *) result->op_params + 0, &scale,    sizeof(float));
+    memcpy((float *) result->op_params + 1, &max_bias, sizeof(float));
+
     return result;
 }
 
-struct ggml_tensor * ggml_soft_max_back(
+struct ggml_tensor * ggml_soft_max_ext_back(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        struct ggml_tensor  * b) {
-    return ggml_soft_max_back_impl(ctx, a, b, false);
+        struct ggml_tensor  * b,
+        float                 scale,
+        float                 max_bias) {
+    return ggml_soft_max_ext_back_impl(ctx, a, b, scale, max_bias, false);
 }
 
-struct ggml_tensor * ggml_soft_max_back_inplace(
+struct ggml_tensor * ggml_soft_max_ext_back_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        struct ggml_tensor  * b) {
-    return ggml_soft_max_back_impl(ctx, a, b, true);
+        struct ggml_tensor  * b,
+        float                 scale,
+        float                 max_bias) {
+    return ggml_soft_max_ext_back_impl(ctx, a, b, scale, max_bias, true);
 }
 
 // ggml_rope
@@ -5076,10 +5085,10 @@ struct ggml_tensor * ggml_cross_entropy_loss_back(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         struct ggml_tensor  * c) {
-    GGML_ASSERT(ggml_are_same_shape(a, b));
-    GGML_ASSERT(ggml_is_scalar(c));
+    GGML_ASSERT(ggml_is_scalar(a));
+    GGML_ASSERT(ggml_are_same_shape(b, c));
 
-    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, b);
 
     result->op     = GGML_OP_CROSS_ENTROPY_LOSS_BACK;
     result->src[0] = a;
@@ -5258,7 +5267,7 @@ static void ggml_sub_or_set(
 }
 
 static void ggml_compute_backward(
-        struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i, bool * grads_needed) {
+        struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i, const bool * grads_needed) {
     struct ggml_tensor * tensor = cgraph->nodes[i];
     struct ggml_tensor * grad   = ggml_graph_get_grad(cgraph, tensor);
 
@@ -5402,7 +5411,7 @@ static void ggml_compute_backward(
             if (src0_needs_grads) {
                 float eps;
                 memcpy(&eps, tensor->op_params, sizeof(float));
-                ggml_add_or_set(ctx, cgraph, isrc0, ggml_rms_norm_back(ctx, src0, grad, eps));
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_rms_norm_back(ctx, grad, src0, eps));
             }
         } break;
         case GGML_OP_MUL_MAT: {
@@ -5585,7 +5594,13 @@ static void ggml_compute_backward(
         } break;
         case GGML_OP_SOFT_MAX: {
             if (src0_needs_grads) {
-                ggml_add_or_set(ctx, cgraph, isrc0, ggml_soft_max_back(ctx, grad, tensor));
+                float scale    = 1.0f;
+                float max_bias = 0.0f;
+
+                memcpy(&scale,    (const float *) tensor->op_params + 0, sizeof(float));
+                memcpy(&max_bias, (const float *) tensor->op_params + 1, sizeof(float));
+
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_soft_max_ext_back(ctx, grad, tensor, scale, max_bias));
             }
             GGML_ASSERT((!src1 || !src1_needs_grads) && "backward pass for softmax mask not implemented");
         } break;
@@ -5626,7 +5641,7 @@ static void ggml_compute_backward(
                 const int32_t d1    = ggml_get_op_params_i32(tensor, 5);
                 const bool    is_2D = ggml_get_op_params_i32(tensor, 6) == 1;
 
-                ggml_add_or_set(ctx, cgraph, isrc1, ggml_im2col_back(ctx, src0, grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D));
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_im2col_back(ctx, grad, src0, src1->ne, s0, s1, p0, p1, d0, d1, is_2D));
             }
         } break;
         case GGML_OP_POOL_2D: {
@@ -5669,7 +5684,7 @@ static void ggml_compute_backward(
                 } break;
                 case GGML_UNARY_OP_SILU: {
                     if (src0_needs_grads) {
-                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, src0, grad));
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, grad, src0));
                     }
                 } break;
                 case GGML_UNARY_OP_EXP: {
@@ -5686,7 +5701,7 @@ static void ggml_compute_backward(
         } break;
         case GGML_OP_CROSS_ENTROPY_LOSS: {
             if (src0_needs_grads) {
-                ggml_add_or_set(ctx, cgraph, isrc0, ggml_cross_entropy_loss_back(ctx, src0, src1, grad));
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_cross_entropy_loss_back(ctx, grad, src0, src1));
             }
             GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
         } break;

include/llama.h

@@ -288,9 +288,6 @@ extern "C" {
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
         const float * tensor_split;
 
-        // comma separated list of RPC servers to use for offloading
-        const char * rpc_servers;
-
         // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
         // If the provided progress_callback returns true, model loading continues.
         // If it returns false, model loading is immediately aborted.
@@ -418,10 +415,20 @@ extern "C" {
               struct llama_model_params   params),
             "use llama_model_load_from_file instead");
 
+    // Load the model from a file
+    // If the file is split into multiple parts, the file name must follow this pattern: <name>-%05d-of-%05d.gguf
+    // If the split file name does not follow this pattern, use llama_model_load_from_splits
     LLAMA_API struct llama_model * llama_model_load_from_file(
                              const char * path_model,
               struct llama_model_params   params);
 
+    // Load the model from multiple splits (support custom naming scheme)
+    // The paths must be in the correct order
+    LLAMA_API struct llama_model * llama_model_load_from_splits(
+                             const char ** paths,
+                                 size_t    n_paths,
+              struct llama_model_params    params);
+
     DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model),
             "use llama_model_free instead");
 
@@ -951,7 +958,7 @@ extern "C" {
     LLAMA_API llama_token llama_vocab_fim_rep(const struct llama_vocab * vocab);
     LLAMA_API llama_token llama_vocab_fim_sep(const struct llama_vocab * vocab);
 
-    DEPRECATED(LLAMA_API const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token), "use llama_vocabable_get_text instead");
+    DEPRECATED(LLAMA_API const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_text instead");
     DEPRECATED(LLAMA_API float llama_token_get_score(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_score instead");
     DEPRECATED(LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_attr instead");
     DEPRECATED(LLAMA_API bool llama_token_is_eog(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_eog instead");

scripts/hf.sh

@@ -1,112 +0,0 @@
-#!/bin/bash
-#
-# Shortcut for downloading HF models
-#
-# Usage:
-#   ./llama-cli -m $(./scripts/hf.sh https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
-#   ./llama-cli -m $(./scripts/hf.sh --url https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/blob/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
-#   ./llama-cli -m $(./scripts/hf.sh --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf)
-#
-
-# all logs go to stderr
-function log {
-    echo "$@" 1>&2
-}
-
-function usage {
-    log "Usage: $0 [[--url] <url>] [--repo <repo>] [--file <file>] [--outdir <dir> [-h|--help]"
-    exit 1
-}
-
-# check for curl or wget
-function has_cmd {
-    if ! [ -x "$(command -v $1)" ]; then
-        return 1
-    fi
-}
-
-if has_cmd wget; then
-    cmd="wget -q -c -O %s/%s %s"
-elif has_cmd curl; then
-    cmd="curl -C - -f --output-dir %s -o %s -L %s"
-else
-    log "[E] curl or wget not found"
-    exit 1
-fi
-
-url=""
-repo=""
-file=""
-outdir="."
-
-# parse args
-while [[ $# -gt 0 ]]; do
-    case "$1" in
-        --url)
-            url="$2"
-            shift 2
-            ;;
-        --repo)
-            repo="$2"
-            shift 2
-            ;;
-        --file)
-            file="$2"
-            shift 2
-            ;;
-        --outdir)
-            outdir="$2"
-            shift 2
-            ;;
-        -h|--help)
-            usage
-            ;;
-        *)
-            url="$1"
-            shift
-            ;;
-    esac
-done
-
-if [ -n "$repo" ] && [ -n "$file" ]; then
-    url="https://huggingface.co/$repo/resolve/main/$file"
-fi
-
-if [ -z "$url" ]; then
-    log "[E] missing --url"
-    usage
-fi
-
-# check if the URL is a HuggingFace model, and if so, try to download it
-is_url=false
-
-if [[ ${#url} -gt 22 ]]; then
-    if [[ ${url:0:22} == "https://huggingface.co" ]]; then
-        is_url=true
-    fi
-fi
-
-if [ "$is_url" = false ]; then
-    log "[E] invalid URL, must start with https://huggingface.co"
-    exit 0
-fi
-
-# replace "blob/main" with "resolve/main"
-url=${url/blob\/main/resolve\/main}
-
-basename=$(basename $url)
-
-log "[+] attempting to download $basename"
-
-if [ -n "$cmd" ]; then
-    cmd=$(printf "$cmd" "$outdir" "$basename" "$url")
-    log "[+] $cmd"
-    if $cmd; then
-        echo $outdir/$basename
-        exit 0
-    fi
-fi
-
-log "[-] failed to download"
-
-exit 1

src/llama-model-loader.cpp

@@ -64,6 +64,33 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
     }
 }
 
+// return a list of splits for a given path
+// for example, given "<name>-00002-of-00004.gguf", returns list of all 4 splits
+static std::vector<std::string> llama_get_list_splits(const std::string & path, const int idx, const int n_split) {
+    std::vector<std::string> paths;
+    std::string split_prefix;
+    std::vector<char> buf(llama_path_max(), 0);
+
+    {
+        int ret = llama_split_prefix(buf.data(), buf.size(), path.c_str(), idx, n_split);
+        if (!ret) {
+            throw std::runtime_error(format("invalid split file name: %s", path.c_str()));
+        }
+        split_prefix = std::string(buf.data(), ret);
+    }
+
+    if (split_prefix.empty()) {
+        throw std::runtime_error(format("invalid split file: %s", path.c_str()));
+    }
+
+    for (int idx = 0; idx < n_split; ++idx) {
+        int ret = llama_split_path(buf.data(), buf.size(), split_prefix.c_str(), idx, n_split);
+        paths.push_back(std::string(buf.data(), ret));
+    }
+
+    return paths;
+}
+
 namespace GGUFMeta {
     template <typename T, gguf_type gt_, T (*gfun)(const gguf_context *, const int64_t)>
     struct GKV_Base_Type {
@@ -413,7 +440,12 @@ namespace GGUFMeta {
     template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
     template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
 
-llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p) {
+llama_model_loader::llama_model_loader(
+        const std::string & fname,
+        std::vector<std::string> & splits,
+        bool use_mmap,
+        bool check_tensors,
+        const struct llama_model_kv_override * param_overrides_p) {
     int trace = 0;
     if (getenv("LLAMA_TRACE")) {
         trace = atoi(getenv("LLAMA_TRACE"));
@@ -425,6 +457,7 @@ llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap,
         }
     }
 
+    // Load the main GGUF
     struct ggml_context * ctx = NULL;
     struct gguf_init_params params = {
         /*.no_alloc = */ true,
@@ -460,35 +493,54 @@ llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap,
 
     // Load additional GGML contexts
     if (n_split > 1) {
+        // make sure the main file is loaded first
         uint16_t idx = 0;
-        get_key(llm_kv(LLM_KV_SPLIT_NO), idx);
+        const std::string kv_split_no = llm_kv(LLM_KV_SPLIT_NO);
+        get_key(kv_split_no, idx);
         if (idx != 0) {
-            throw std::runtime_error(format("illegal split file: %d, model must be loaded with the first split", idx));
+            throw std::runtime_error(format("illegal split file idx: %d (file: %s), model must be loaded with the first split", idx, fname.c_str()));
         }
 
-        std::vector<char> split_prefix(llama_path_max(), 0);
-        if (!llama_split_prefix(split_prefix.data(), split_prefix.size(), fname.c_str(), idx, n_split)) {
-            throw std::runtime_error(format("invalid split file: %s", fname.c_str()));
+        // generate list of splits if needed
+        if (splits.empty()) {
+            splits = llama_get_list_splits(fname, idx, n_split);
+        }
+
+        // in case user give a custom list of splits, check if it matches the expected number
+        if (n_split != (uint16_t)splits.size()) {
+            throw std::runtime_error(format("invalid split count, given: %zu splits, but expected %d", splits.size(), n_split));
         }
 
         if (trace > 0) {
             LLAMA_LOG_INFO("%s: loading additional %d GGUFs\n", __func__, n_split);
         }
 
-        std::vector<char> split_path(llama_path_max(), 0);
+        // load other splits
         for (idx = 1; idx < n_split; idx++) {
-            llama_split_path(split_path.data(), split_path.size(), split_prefix.data(), idx, n_split);
+            const char * fname_split = splits[idx].c_str();
 
             struct gguf_init_params split_params = {
                 /*.no_alloc = */ true,
                 /*.ctx      = */ &ctx,
             };
-            gguf_context_ptr ctx_gguf { gguf_init_from_file(split_path.data(), split_params) };
+            gguf_context_ptr ctx_gguf { gguf_init_from_file(fname_split, split_params) };
             if (!ctx_gguf) {
-                throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, split_path.data()));
+                throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, fname_split));
             }
 
-            files.emplace_back(new llama_file(split_path.data(), "rb"));
+            // check idx
+            {
+                const int kid = gguf_find_key(ctx_gguf.get(), kv_split_no.c_str());
+                if (kid < 0) {
+                    throw std::runtime_error(format("missing key %s in GGUF split %s", kv_split_no.c_str(), fname_split));
+                }
+                int idx_gguf = gguf_get_val_u16(ctx_gguf.get(), kid);
+                if (idx_gguf != idx) {
+                    throw std::runtime_error(format("invalid split file idx: %d (file: %s), expected %d", idx_gguf, fname_split, idx));
+                }
+            }
+
+            files.emplace_back(new llama_file(fname_split, "rb"));
             contexts.emplace_back(ctx);
 
             // Save tensors data offset info of the shard.

src/llama-model-loader.h

@@ -90,7 +90,12 @@ struct llama_model_loader {
     size_t size_data = 0;
     std::vector<std::pair<size_t, size_t>> mmaps_used;
 
-    llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p);
+    llama_model_loader(
+        const std::string & fname,
+        std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
+        bool use_mmap,
+        bool check_tensors,
+        const struct llama_model_kv_override * param_overrides_p);
 
     template<typename T>
     typename std::enable_if<std::is_integral<T>::value, bool>::type

src/llama-model.cpp

@@ -3717,7 +3717,6 @@ struct llama_model_params llama_model_default_params() {
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,
-        /*.rpc_servers                 =*/ nullptr,
         /*.progress_callback           =*/ nullptr,
         /*.progress_callback_user_data =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,

src/llama-model.h

@@ -323,8 +323,6 @@ struct llama_model {
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;
 
-    std::vector<std::string> rpc_servers;
-
     // list of devices used in this model
     std::vector<ggml_backend_dev_t> devices;
 

src/llama-quant.cpp

@@ -526,7 +526,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         kv_overrides = v->data();
     }
 
-    llama_model_loader ml(fname_inp, use_mmap, /*check_tensors*/ true, kv_overrides);
+    std::vector<std::string> splits = {};
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());

src/llama-vocab.cpp

@@ -439,7 +439,7 @@ struct llm_tokenizer_bpe_session {
                 "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
                 "Are you sure this is what you want?\n", __FUNCTION__);
         }
-        if (vocab.get_add_bos() && output.size() >= 2 && *(output.end()-2) == vocab.token_eos()) {
+        if (vocab.get_add_eos() && output.size() >= 2 && *(output.end()-2) == vocab.token_eos()) {
             LLAMA_LOG_WARN(
                 "%s: Added a EOS token to the prompt as specified by the model but the prompt "
                 "also ends with a EOS token. So now the final prompt ends with 2 EOS tokens. "

src/llama.cpp

@@ -31,7 +31,7 @@
 #endif
 
 // Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
-static int llama_model_load(const std::string & fname, llama_model & model, llama_model_params & params) {
+static int llama_model_load(const std::string & fname, std::vector<std::string> & splits, llama_model & model, llama_model_params & params) {
     // loading time will be recalculated after the first eval, so
     // we take page faults deferred by mmap() into consideration
     model.t_load_us = 0;
@@ -40,7 +40,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
     model.t_start_us = tm.t_start_us;
 
     try {
-        llama_model_loader ml(fname, params.use_mmap, params.check_tensors, params.kv_overrides);
+        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides);
 
         ml.print_info();
 
@@ -9374,14 +9374,9 @@ int64_t llama_time_us(void) {
     return ggml_time_us();
 }
 
-struct llama_model * llama_load_model_from_file(
-        const char * path_model,
-        struct llama_model_params params) {
-    return llama_model_load_from_file(path_model, params);
-}
-
-struct llama_model * llama_model_load_from_file(
-        const char * path_model,
+static struct llama_model * llama_model_load_from_file_impl(
+        const std::string & path_model,
+        std::vector<std::string> & splits,
         struct llama_model_params params) {
     ggml_time_init();
 
@@ -9404,47 +9399,6 @@ struct llama_model * llama_model_load_from_file(
         };
     }
 
-    if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') {
-        // split the servers set them into model->rpc_servers
-        std::string servers(params.rpc_servers);
-        size_t pos = 0;
-        while ((pos = servers.find(',')) != std::string::npos) {
-            std::string server = servers.substr(0, pos);
-            model->rpc_servers.push_back(server);
-            servers.erase(0, pos + 1);
-        }
-        model->rpc_servers.push_back(servers);
-    }
-
-    // add RPC devices
-    if (!model->rpc_servers.empty()) {
-        ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
-        if (!rpc_reg) {
-            LLAMA_LOG_ERROR("%s: failed to find RPC backend\n", __func__);
-            llama_model_free(model);
-            return nullptr;
-        }
-
-        typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
-        ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
-        if (!ggml_backend_rpc_add_device_fn) {
-            LLAMA_LOG_ERROR("%s: failed to find RPC device add function\n", __func__);
-            llama_model_free(model);
-            return nullptr;
-        }
-
-        for (const std::string & server : model->rpc_servers) {
-            ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
-            if (dev) {
-                model->devices.push_back(dev);
-            } else {
-                LLAMA_LOG_ERROR("%s: failed to add RPC device for server '%s'\n", __func__, server.c_str());
-                llama_model_free(model);
-                return nullptr;
-            }
-        }
-    }
-
     // create list of devices to use with this model
     if (params.devices) {
         for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
@@ -9485,7 +9439,7 @@ struct llama_model * llama_model_load_from_file(
         LLAMA_LOG_INFO("%s: using device %s (%s) - %zu MiB free\n", __func__, ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), free/1024/1024);
     }
 
-    const int status = llama_model_load(path_model, *model, params);
+    const int status = llama_model_load(path_model, splits, *model, params);
     GGML_ASSERT(status <= 0);
     if (status < 0) {
         if (status == -1) {
@@ -9501,6 +9455,35 @@ struct llama_model * llama_model_load_from_file(
     return model;
 }
 
+// deprecated
+struct llama_model * llama_load_model_from_file(
+        const char * path_model,
+        struct llama_model_params params) {
+    return llama_model_load_from_file(path_model, params);
+}
+
+struct llama_model * llama_model_load_from_file(
+        const char * path_model,
+        struct llama_model_params params) {
+    std::vector<std::string> splits = {};
+    return llama_model_load_from_file_impl(path_model, splits, params);
+}
+
+struct llama_model * llama_model_load_from_splits(
+        const char ** paths,
+        size_t n_paths,
+        struct llama_model_params params) {
+    std::vector<std::string> splits;
+    if (n_paths == 0) {
+        LLAMA_LOG_ERROR("%s: list of splits is empty\n", __func__);
+        return nullptr;
+    }
+    for (size_t i = 0; i < n_paths; ++i) {
+        splits.push_back(paths[i]);
+    }
+    return llama_model_load_from_file_impl(splits.front(), splits, params);
+}
+
 struct llama_context * llama_init_from_model(
                  struct llama_model * model,
         struct llama_context_params   params) {

tests/test-backend-ops.cpp

@@ -780,7 +780,7 @@ struct test_case {
             }
         }
         if (!any_params) {
-            printf("not supported [%s] \n", op_name);
+            printf("not supported [%s] \n", op_desc(out).c_str());
             supported = false;
         }
         if (!supported) {
@@ -1130,6 +1130,59 @@ struct test_get_rows : public test_case {
     }
 };
 
+// GGML_OP_GET_ROWS_BACK
+struct test_get_rows_back : public test_case {
+    const ggml_type type;
+    const int n; // cols
+    const int m; // rows
+    const int r; // rows to get
+    const int b; // batch size
+    const bool v; // view (non-contiguous src1)
+
+    std::string vars() override {
+        return VARS_TO_STR6(type, n, m, r, b, v);
+    }
+
+    test_get_rows_back(ggml_type type = GGML_TYPE_F32, int n = 10, int m = 5, int r = 3, int b = 1, bool v = false)
+        : type(type), n(n), m(m), r(r), b(b), v(v) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * in_forward = ggml_new_tensor_3d(ctx, type, n, m, b);
+        ggml_set_name(in_forward, "in_forward");
+
+        ggml_tensor * rows = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, r, b);
+        ggml_set_name(rows, "rows");
+        if (v) {
+            rows = ggml_view_2d(ctx, rows, r/2, b, rows->nb[1], 0);
+            ggml_set_name(rows, "view_of_rows");
+        }
+
+        ggml_tensor * grad = ggml_new_tensor_3d(ctx, type, n, r, b);
+        ggml_set_name(grad, "grad");
+
+        ggml_tensor * out = ggml_get_rows_back(ctx, grad, rows, in_forward);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_I32) {
+                if (ggml_is_view_op(t->op)) { continue; }
+                // rows
+                std::vector<int> data(r*b);
+                for (int i = 0; i < r*b; i++) {
+                    data[i] = rand() % m;
+                }
+                ggml_backend_tensor_set(t, data.data(), 0, r * b * sizeof(int));
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
+};
+
 // GGML_OP_ARGMAX
 struct test_argmax : public test_case {
     const ggml_type type;
@@ -1531,6 +1584,39 @@ struct test_scale : public test_case {
     }
 };
 
+// GGML_OP_SILU_BACK
+struct test_silu_back : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    float eps;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, eps);
+    }
+
+    test_silu_back(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {64, 5, 4, 3},
+            float eps = 1e-6f)
+        : type(type), ne(ne), eps(eps) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(a, "a");
+
+        ggml_tensor * grad = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(grad, "grad");
+
+        ggml_tensor * out = ggml_silu_back(ctx, a, grad);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    bool grad_precise() override {
+        return true;
+    }
+};
+
 // GGML_OP_NORM
 struct test_norm : public test_case {
     const ggml_type type;
@@ -1583,11 +1669,56 @@ struct test_rms_norm : public test_case {
         return out;
     }
 
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t, -10.f, 10.f);
+        }
+    }
+
+    float grad_eps() override {
+        return 1.0f;
+    }
+
     bool grad_precise() override {
         return true;
     }
 };
 
+// GGML_OP_RMS_NORM_BACK
+struct test_rms_norm_back : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    float eps;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, eps);
+    }
+
+    test_rms_norm_back(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {64, 5, 4, 3},
+            float eps = 1e-6f)
+        : type(type), ne(ne), eps(eps) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(a, "a");
+
+        ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(b, "b");
+
+        ggml_tensor * out = ggml_rms_norm_back(ctx, a, b, eps);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t, -10.f, 10.f);
+        }
+    }
+};
+
 // GGML_OP_SSM_CONV
 struct test_ssm_conv : public test_case {
     const ggml_type type;
@@ -1855,10 +1986,11 @@ struct test_out_prod : public test_case {
     const int64_t n;
     const int64_t k;
     const std::array<int64_t, 2> bs; // dims 3 and 4
+    const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
     const bool trans_b;
 
     std::string vars() override {
-        return VARS_TO_STR7(type_a, type_b, m, n, k, bs, trans_b);
+        return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, trans_b);
     }
 
     double max_nmse_err() override {
@@ -1868,8 +2000,9 @@ struct test_out_prod : public test_case {
     test_out_prod(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
             int64_t m = 32, int64_t n = 32, int64_t k = 32,
             std::array<int64_t, 2> bs = {10, 10},
+            std::array<int64_t, 2> nr = {2, 2},
             bool trans_b = false)
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), trans_b(trans_b) {}
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), trans_b(trans_b) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type_a, m, k, bs[0], bs[1]);
@@ -1877,10 +2010,10 @@ struct test_out_prod : public test_case {
 
         ggml_tensor * b;
         if (trans_b) {
-            b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0], bs[1]);
+            b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
             b = ggml_transpose(ctx, b);
         } else {
-            b = ggml_new_tensor_4d(ctx, type_b, n, k, bs[0], bs[1]);
+            b = ggml_new_tensor_4d(ctx, type_b, n, k, bs[0]*nr[0], bs[1]*nr[1]);
         }
         ggml_set_name(b, "b");
 
@@ -2191,6 +2324,36 @@ struct test_soft_max : public test_case {
     }
 };
 
+// GGML_OP_SOFT_MAX_BACK
+struct test_soft_max_back : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const float scale;
+    const float max_bias;
+
+    std::string vars() override {
+        return VARS_TO_STR4(type, ne, scale, max_bias);
+    }
+
+    test_soft_max_back(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 5, 4, 3},
+            float scale = 1.0f,
+            float max_bias = 0.0f)
+        : type(type), ne(ne), scale(scale), max_bias(max_bias) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(a, "a");
+
+        ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(a, "a");
+
+        ggml_tensor * out = ggml_soft_max_ext_back(ctx, a, b, scale, max_bias);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+};
 
 // GGML_OP_ROPE + GGML_OP_ROPE_BACK
 struct test_rope : public test_case {
@@ -2883,9 +3046,10 @@ struct test_flash_attn_ext : public test_case {
     const float logit_softcap; // Gemma 2
 
     const ggml_type type_KV;
+    std::array<int32_t, 4> permute;
 
     std::string vars() override {
-        return VARS_TO_STR8(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV);
+        return VARS_TO_STR9(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV, permute);
     }
 
     double max_nmse_err() override {
@@ -2900,19 +3064,33 @@ struct test_flash_attn_ext : public test_case {
     }
 
     test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8,
-                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16)
-        : hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV) {}
+                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16,
+                        std::array<int32_t, 4> permute = {0, 1, 2, 3})
+        : hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV), permute(permute) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
 
-        ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs_padded, nb, nh, 1);
+        auto const &create_permuted = [&](ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) -> ggml_tensor * {
+            int64_t ne[4] = {ne0, ne1, ne2, ne3};
+            int64_t ne_perm[4];
+            for (int i = 0; i < 4; ++i) {
+                ne_perm[permute[i]] = ne[i];
+            }
+            ggml_tensor * t = ggml_new_tensor_4d(ctx, type, ne_perm[0], ne_perm[1], ne_perm[2], ne_perm[3]);
+            if (permute != std::array<int32_t, 4>{0, 1, 2, 3}) {
+                t = ggml_permute(ctx, t, permute[0], permute[1], permute[2], permute[3]);
+            }
+            return t;
+        };
+
+        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hs_padded, nb, nh, 1);
         ggml_set_name(q, "q");
 
-        ggml_tensor * k = ggml_new_tensor_4d(ctx, type_KV,       hs_padded, kv, nh, 1);
+        ggml_tensor * k = create_permuted(type_KV,       hs_padded, kv, nh, 1);
         ggml_set_name(k, "k");
 
-        ggml_tensor * v = ggml_new_tensor_4d(ctx, type_KV,       hs_padded, kv, nh, 1);
+        ggml_tensor * v = create_permuted(type_KV,       hs_padded, kv, nh, 1);
         ggml_set_name(v, "v");
 
         ggml_tensor * m = nullptr;
@@ -2980,6 +3158,40 @@ struct test_cross_entropy_loss : public test_case {
     }
 };
 
+// GGML_OP_CROSS_ENTROPY_LOSS_BACK
+struct test_cross_entropy_loss_back : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_cross_entropy_loss_back(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 5, 4, 3})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * grad = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+        ggml_set_name(grad, "grad");
+
+        ggml_tensor * logits = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(logits, "logits");
+
+        ggml_tensor * labels = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(labels, "labels");
+
+        // Ensure labels add up to 1:
+        labels = ggml_soft_max(ctx, labels);
+        ggml_set_name(labels, "labels_normalized");
+
+        ggml_tensor * out = ggml_cross_entropy_loss_back(ctx, grad, logits, labels);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+};
+
 // GGML_OP_OPT_STEP_ADAMW
 struct test_opt_step_adamw : public test_case {
     const ggml_type type;
@@ -3479,6 +3691,16 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
+    test_cases.emplace_back(new test_get_rows_back(GGML_TYPE_F32, 1, 8, 2, 1, false));
+    for (ggml_type type : all_types) {
+        for (bool v : {false, true}) {
+            test_cases.emplace_back(new test_get_rows_back(type, 256, 5, 4, 1, v));
+        }
+    }
+    for (bool v : {false, true}) {
+        test_cases.emplace_back(new test_get_rows_back(GGML_TYPE_I32, 256, 5, 4, 1, v));
+    }
+
     for (ggml_type type_input : {GGML_TYPE_F32}) {
         for (ggml_op_pool pool_type : {GGML_OP_POOL_AVG, GGML_OP_POOL_MAX}) {
             for (int k0 : {1, 3}) {
@@ -3601,6 +3823,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
     }
+    for (ggml_type type_dst : {GGML_TYPE_F32}) {
+        for (ggml_type type_src : all_types) {
+            test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
+            test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
+        }
+    }
     for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) {
         for (ggml_type type_dst : {GGML_TYPE_F16, GGML_TYPE_F32}) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {1, 0, 2, 3})); // cpy not-contiguous
@@ -3657,10 +3885,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_add1());
     test_cases.emplace_back(new test_scale());
+    test_cases.emplace_back(new test_silu_back());
 
-    for (float eps : {1e-6f, 1e-5f, 1e-3f, 1e-1f}) {
-        test_cases.emplace_back(new test_norm(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
-        test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
+    for (float eps : {0.0f, 1e-7f, 1e-4f, 1e-1f}) {
+        test_cases.emplace_back(new test_norm         (GGML_TYPE_F32, {64, 5, 4, 3}, eps));
+        test_cases.emplace_back(new test_rms_norm     (GGML_TYPE_F32, {64, 5, 4, 3}, eps));
+        test_cases.emplace_back(new test_rms_norm_back(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
     }
 
     test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 1, 1}, {4, 1536, 1, 1}));
@@ -3800,22 +4030,19 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     for (ggml_type type_a : base_types) {
         for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 1, 16, { 1,  1}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 1, 16, {10,  1}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 1, 16, {10,  1}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 1, 16, {10, 10}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 1, 16, {10, 10}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 1, 16, {10, 10}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 1, 16, {10, 10}));
-
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 16, 16, { 1,  1}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 16, 16, { 1,  1}, true));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 16, 16, {10,  1}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 16, 16, {10,  1}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 16, 16, {10, 10}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 16, 16, {10, 10}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 16, 16, {10, 10}));
-            test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, 16, 16, {10, 10}));
+            for (int n : {1, 16}) {
+                for (int k : {1, 16}) {
+                    for (int bs2 : {1, 3}) {
+                        for (int bs3 : {1, 3}) {
+                            for (int nr2 : {1, 2}) {
+                                for (int nr3 : {1, 2}) {
+                                    test_cases.emplace_back(new test_out_prod(type_a, type_b, 256, n, k, {bs2, bs3}, {nr2, nr3}));
+                                }
+                            }
+                        }
+                    }
+                }
+            }
         }
     }
 
@@ -3858,11 +4085,22 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             }
         }
     }
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, 0.1f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true,  0.1f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 8.0f));
 
+    for (float max_bias : {0.0f, 8.0f}) {
+        for (float scale : {1.0f, 0.1f}) {
+            for (int64_t ne0 : {16, 1024}) {
+                for (int64_t ne1 : {16, 1024}) {
+                    test_cases.emplace_back(new test_soft_max_back(GGML_TYPE_F32, {ne0,   ne1,   1, 1}, scale, max_bias));
+                    test_cases.emplace_back(new test_soft_max_back(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, scale, max_bias));
+                }
+            }
+        }
+    }
+
     for (bool fw : {true, false}) { // fw == forward
         bool all = true;
 
@@ -3944,6 +4182,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                             for (int nb : { 1, 3, 32, 35, }) {
                                 for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
                                     test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV));
+                                    // run fewer test cases permuted
+                                    if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+                                        test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV, {0, 2, 1, 3}));
+                                    }
                                 }
                             }
                         }
@@ -3953,7 +4195,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
-    test_cases.emplace_back(new test_cross_entropy_loss());
+    test_cases.emplace_back(new test_cross_entropy_loss     (GGML_TYPE_F32, {   10, 5, 4, 3}));
+    test_cases.emplace_back(new test_cross_entropy_loss     (GGML_TYPE_F32, {30000, 1, 1, 1}));
+    test_cases.emplace_back(new test_cross_entropy_loss_back(GGML_TYPE_F32, {   10, 5, 4, 3}));
+    test_cases.emplace_back(new test_cross_entropy_loss_back(GGML_TYPE_F32, {30000, 1, 1, 1}));
+
     test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}));
 
     // these tests are disabled to save execution time, but they can be handy for debugging

tests/test-lora-conversion-inference.sh

@@ -80,18 +80,18 @@ run_conversion_and_inference_lora() {
     # Run inference
     echo -e "\n\n---------------------------\n\n"
     echo "Running llama-cli without lora for $model_name with hidden_size $hidden_size..."
-    OUTPUT_BASE=$(./llama-cli -m $MODELS_REPO/$model_name/hidden_size=$hidden_size/base/Base-F32.gguf \
+    OUTPUT_BASE=$(./llama-cli -no-cnv -m $MODELS_REPO/$model_name/hidden_size=$hidden_size/base/Base-F32.gguf \
         -p "$EXPECTED_BASE_FIRST_WORD" -n 50 --seed 42 --temp 0)
 
     echo -e "\n\n---------------------------\n\n"
     echo "Running llama-cli with hot lora for $model_name with hidden_size $hidden_size..."
-    OUTPUT_LORA_HOT=$(./llama-cli -m $MODELS_REPO/$model_name/hidden_size=$hidden_size/base/Base-F32.gguf \
+    OUTPUT_LORA_HOT=$(./llama-cli -no-cnv -m $MODELS_REPO/$model_name/hidden_size=$hidden_size/base/Base-F32.gguf \
         --lora $MODELS_REPO/$model_name/hidden_size=$hidden_size/lora/Lora-F32-LoRA.gguf \
         -p "$EXPECTED_LORA_FIRST_WORD" -n 50 --seed 42 --temp 0)
 
     echo -e "\n\n---------------------------\n\n"
     echo "Running llama-cli with merged lora for $model_name with hidden_size $hidden_size..."
-    OUTPUT_LORA_MERGED=$(./llama-cli -m $MODELS_REPO/$model_name/hidden_size=$hidden_size/base/Base-F32-lora-merged.gguf \
+    OUTPUT_LORA_MERGED=$(./llama-cli -no-cnv -m $MODELS_REPO/$model_name/hidden_size=$hidden_size/base/Base-F32-lora-merged.gguf \
         -p "$EXPECTED_LORA_FIRST_WORD" -n 50 --seed 42 --temp 0)
 
     # Remove any initial white space