gguf-py : Numpy dequantization for grid-based i-quants

* gguf-py : use classes for quants

* convert_hf : simplify internal quantization type selection

* gguf-py : fix flake8 lint

* gguf-py : fix BF16 numpy view type

* gguf-py : remove LlamaFileTypeMap

Too specific to 'llama.cpp', and would be a maintenance burden
to keep up to date.

* gguf-py : add generic quantize and dequantize functions

The quant classes no longer need to be known,
only the target or the source type,
for 'quantize' and 'dequantize', respectively.

.devops/llama-server.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev curl
+    apt-get install -y build-essential git libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -16,7 +16,7 @@ RUN make -j$(nproc) llama-server
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
 COPY --from=build /app/llama-server /llama-server
 

CONTRIBUTING.md

@@ -5,6 +5,7 @@
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
 - Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
   - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
+- Consider allowing write access to your branch for faster review
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
 # Pull requests (for collaborators)

Makefile

@@ -888,15 +888,16 @@ ggml/src/ggml-metal-embed.o: \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
 	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
+	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
+	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
+	@rmdir ${TEMP_ASSEMBLY}
 endif
 endif # GGML_METAL
 
@@ -1205,6 +1206,7 @@ clean:
 	rm -rvf ggml/*.dll
 	rm -rvf ggml/*.so
 	rm -vrf ggml/src/*.o
+	rm -rvf ggml/src/llamafile/*.o
 	rm -rvf common/build-info.cpp
 	rm -vrf ggml/src/ggml-metal-embed.metal
 	rm -vrf ggml/src/ggml-cuda/*.o

README.md

@@ -95,8 +95,16 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
+- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
+- [x] [Smaug](https://huggingface.co/models?search=Smaug)
+- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
+- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
+- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
+- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
+- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -145,6 +153,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
 - [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
+- [ramalama](https://github.com/containers/ramalama) (MIT)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)

common/common.cpp

@@ -684,14 +684,24 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     }
     if (arg == "--lora") {
         CHECK_ARG
-        params.lora_adapter.emplace_back(argv[i], 1.0f);
+        params.lora_adapters.push_back({
+            std::string(argv[i]),
+            1.0,
+        });
         return true;
     }
     if (arg == "--lora-scaled") {
         CHECK_ARG
-        const char* lora_adapter = argv[i];
+        std::string lora_adapter = argv[i];
         CHECK_ARG
-        params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
+        params.lora_adapters.push_back({
+            lora_adapter,
+            std::stof(argv[i]),
+        });
+        return true;
+    }
+    if (arg == "--lora-init-without-apply") {
+        params.lora_init_without_apply = true;
         return true;
     }
     if (arg == "--control-vector") {
@@ -1654,6 +1664,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
                                                                         "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
     options.push_back({ "server",      "-sps,  --slot-prompt-similarity SIMILARITY",
                                                                         "how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity });
+    options.push_back({ "server",      "       --lora-init-without-apply",     "load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"});
 
 #ifndef LOG_DISABLE_LOGS
     options.push_back({ "logging" });
@@ -2039,8 +2050,8 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
-
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
+struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
+    llama_init_result iparams;
     auto mparams = llama_model_params_from_gpt_params(params);
 
     llama_model * model = nullptr;
@@ -2055,7 +2066,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     auto cparams = llama_context_params_from_gpt_params(params);
@@ -2064,7 +2075,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     if (!params.control_vectors.empty()) {
@@ -2075,7 +2086,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
 
         int err = llama_control_vector_apply(lctx,
@@ -2087,21 +2098,26 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
     }
 
-    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
-        const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
-        float lora_scale = std::get<1>(params.lora_adapter[i]);
-        auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
-        if (adapter == nullptr) {
-            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+    // load and optionally apply lora adapters
+    for (auto & la : params.lora_adapters) {
+        llama_lora_adapter_container loaded_la;
+        loaded_la.path = la.path;
+        loaded_la.scale = la.scale;
+        loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
+        if (loaded_la.adapter == nullptr) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
-        llama_lora_adapter_set(lctx, adapter, lora_scale);
+        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
+    }
+    if (!params.lora_init_without_apply) {
+        llama_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
     if (params.ignore_eos) {
@@ -2135,7 +2151,18 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         llama_reset_timings(lctx);
     }
 
-    return std::make_tuple(model, lctx);
+    iparams.model   = model;
+    iparams.context = lctx;
+    return iparams;
+}
+
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters) {
+    llama_lora_adapter_clear(ctx);
+    for (auto & la : lora_adapters) {
+        if (la.scale != 0.0f) {
+            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+        }
+    }
 }
 
 struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
@@ -3160,19 +3187,18 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     }
 
     fprintf(stream, "lora:\n");
-    for (std::tuple<std::string, float> la : params.lora_adapter) {
-        if (std::get<1>(la) != 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale == 1.0f) {
+            fprintf(stream, "  - %s\n", la.path.c_str());
         }
-        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
     }
     fprintf(stream, "lora_scaled:\n");
-    for (std::tuple<std::string, float> la : params.lora_adapter) {
-        if (std::get<1>(la) == 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale != 1.0f) {
+            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
         }
-        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
     }
+    fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);

common/common.h

@@ -33,6 +33,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct llama_lora_adapter_info {
+    std::string path;
+    float scale;
+};
+
+struct llama_lora_adapter_container : llama_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
+};
+
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -126,8 +135,8 @@ struct gpt_params {
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
 
-    // TODO: avoid tuple, use struct
-    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
     std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -308,8 +317,13 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-// TODO: avoid tuplue, use struct
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_init_result {
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
+};
+
+struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
 
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
@@ -317,6 +331,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 
+// clear LoRA adapters from context, then apply new list of adapters
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
+
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);

convert_hf_to_gguf.py

@@ -251,12 +251,7 @@ class Model:
 
         return [(self.map_tensor_name(name), data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del name, new_name, bid, n_dims  # unused
-
-        return False
-
-    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid, n_dims  # unused
 
         return False
@@ -285,55 +280,47 @@ class Model:
             for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
                 data: np.ndarray  # type hint
                 n_dims = len(data.shape)
-                data_dtype = data.dtype
-                data_qtype: gguf.GGMLQuantizationType | None = None
-
-                # when both are True, f32 should win
-                extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
-                extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims)
+                data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
                 # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
-                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
-                extra_f32 = any(cond for cond in (
-                    extra_f32,
-                    n_dims == 1,
-                    new_name.endswith("_norm.weight"),
-                ))
-
-                # Some tensor types are always in float32
-                extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
-                    gguf.MODEL_TENSOR.FFN_GATE_INP,
-                    gguf.MODEL_TENSOR.POS_EMBD,
-                    gguf.MODEL_TENSOR.TOKEN_TYPES,
-                ))
-
-                # if f16 desired, convert any float32 2-dim weight tensors to float16
-                extra_f16 = any(cond for cond in (
-                    extra_f16,
-                    (name.endswith(".weight") and n_dims >= 2),
-                ))
-
-                if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
-                    if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                        data = gguf.quantize_bf16(data)
-                        assert data.dtype == np.uint16
-                        data_qtype = gguf.GGMLQuantizationType.BF16
-
-                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
-                        data = gguf.quantize_q8_0(data)
-                        assert data.dtype == np.uint8
-                        data_qtype = gguf.GGMLQuantizationType.Q8_0
-
-                    else:  # default to float16 for quantized tensors
-                        if data_dtype != np.float16:
-                            data = data.astype(np.float16)
-                        data_qtype = gguf.GGMLQuantizationType.F16
-
-                if data_qtype is None:  # by default, convert to float32
-                    if data_dtype != np.float32:
-                        data = data.astype(np.float32)
+                if n_dims <= 1 or new_name.endswith("_norm.weight"):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
+                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
+                # Some tensor types are always in float32
+                if data_qtype is False and (
+                    any(
+                        self.match_model_tensor_name(new_name, key, bid)
+                        for key in (
+                            gguf.MODEL_TENSOR.FFN_GATE_INP,
+                            gguf.MODEL_TENSOR.POS_EMBD,
+                            gguf.MODEL_TENSOR.TOKEN_TYPES,
+                        )
+                    )
+                    or not name.endswith(".weight")
+                ):
+                    data_qtype = gguf.GGMLQuantizationType.F32
+
+                # No override (data_qtype is False), or wants to be quantized (data_qtype is True)
+                if isinstance(data_qtype, bool):
+                    if self.ftype == gguf.LlamaFileType.ALL_F32:
+                        data_qtype = gguf.GGMLQuantizationType.F32
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                        data_qtype = gguf.GGMLQuantizationType.F16
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
+                        data_qtype = gguf.GGMLQuantizationType.Q8_0
+                    else:
+                        raise ValueError(f"Unknown file type: {self.ftype.name}")
+
+                try:
+                    data = gguf.quants.quantize(data, data_qtype)
+                except gguf.QuantError as e:
+                    logger.warning("%s, %s", e, "falling back to F16")
+                    data_qtype = gguf.GGMLQuantizationType.F16
+                    data = gguf.quants.quantize(data, data_qtype)
+
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 
                 # reverse shape to make it similar to the internal ggml dimension order
@@ -1765,7 +1752,7 @@ class DbrxModel(Model):
 
         return [(new_name, data_torch)]
 
-    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid  # unused
 
         return n_dims > 1
@@ -2506,6 +2493,112 @@ class NomicBertModel(BertModel):
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
 
 
+@Model.register("XLMRobertaModel")
+class XLMRobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'sentencepiece.bpe.model'
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+        assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        # realign tokens (see HF tokenizer code)
+        tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
+        scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
+        toktypes = [
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.UNKNOWN,
+        ] + toktypes[3:-1]
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        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)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_token_type_count(1)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @Model.register("GemmaForCausalLM")
 class GemmaModel(Model):
     model_arch = gguf.MODEL_ARCH.GEMMA
@@ -2680,18 +2773,22 @@ class MambaModel(Model):
 
         return [(new_name, data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del n_dims  # unused
-
-        return bid is not None and new_name in (
-            self.format_tensor_name(n, bid, ".weight" if name.endswith(".weight") else "") for n in [
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
+        if bid is not None and new_name in (
+            self.format_tensor_name(
+                n, bid, ".weight" if name.endswith(".weight") else ""
+            )
+            for n in [
                 gguf.MODEL_TENSOR.SSM_CONV1D,
                 gguf.MODEL_TENSOR.SSM_X,
                 gguf.MODEL_TENSOR.SSM_DT,
                 gguf.MODEL_TENSOR.SSM_A,
                 gguf.MODEL_TENSOR.SSM_D,
             ]
-        )
+        ):
+            return gguf.GGMLQuantizationType.F32
+
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
 
 @Model.register("CohereForCausalLM")

examples/cvector-generator/cvector-generator.cpp

@@ -414,9 +414,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // int n_ctx = llama_n_ctx(ctx);
     int n_layers = llama_n_layer(model);

examples/embedding/embedding.cpp

@@ -79,11 +79,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;

examples/eval-callback/eval-callback.cpp

@@ -163,9 +163,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;

examples/export-lora/export-lora.cpp

@@ -135,7 +135,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<std::tuple<std::string, float>> & lora_files,
+            std::vector<llama_lora_adapter_info> & lora_files,
             std::string & outfile,
             int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
         fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@@ -144,9 +144,9 @@ struct lora_merge_ctx {
             throw std::runtime_error("split model is not yet supported");
         }
 
-        for (auto lora_inp : lora_files) {
-            auto fname = std::get<0>(lora_inp);
-            auto scale = std::get<1>(lora_inp);
+        for (auto & lora_inp : lora_files) {
+            auto fname = lora_inp.path;
+            auto scale = lora_inp.scale;
             std::unique_ptr<file_input> adapter(new file_input(fname, scale));
             check_metadata_lora(adapter.get());
             adapters.push_back(std::move(adapter));
@@ -407,7 +407,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity == 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapter, params.lora_outfile, params.n_threads);
+        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());

examples/imatrix/imatrix.cpp

@@ -611,10 +611,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;

examples/infill/infill.cpp

@@ -179,7 +179,10 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
 
     if (model == NULL) {
         LOG_TEE("%s: error: unable to load model\n", __func__);

examples/llama-bench/llama-bench.cpp

@@ -27,6 +27,14 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 // utils
 static uint64_t get_time_ns() {
     using clock = std::chrono::high_resolution_clock;
@@ -96,6 +104,27 @@ static std::string get_cpu_info() {
         }
         fclose(f);
     }
+#elif defined(_WIN32)
+    HKEY hKey;
+    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                     0,
+                     KEY_READ,
+                     &hKey) != ERROR_SUCCESS) {
+        // fail to open registry key
+        return "";
+    }
+    char cpu_brand[256];
+    DWORD cpu_brand_size = sizeof(cpu_brand);
+    if (RegQueryValueExA(hKey,
+                        TEXT("ProcessorNameString"),
+                        NULL,
+                        NULL,
+                        (LPBYTE)cpu_brand,
+                        &cpu_brand_size) == ERROR_SUCCESS) {
+        id.assign(cpu_brand, cpu_brand_size);
+    }
+    RegCloseKey(hKey);
 #endif
     // TODO: other platforms
     return id;

examples/lookahead/lookahead.cpp

@@ -58,11 +58,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // Tokenize the prompt
     std::vector<llama_token> inp;

examples/lookup/lookup-create.cpp

@@ -22,11 +22,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt

examples/lookup/lookup-stats.cpp

@@ -26,11 +26,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;

examples/lookup/lookup.cpp

@@ -34,11 +34,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;

examples/main/main.cpp

@@ -207,7 +207,10 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
     if (sparams.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
         ctx_guidance = llama_new_context_with_model(model, lparams);

examples/parallel/parallel.cpp

@@ -129,11 +129,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {

examples/perplexity/perplexity.cpp

@@ -2018,11 +2018,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model and apply lora adapter, if any
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;

examples/quantize/quantize.cpp

@@ -91,7 +91,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 }
 
 // usage:
-//  ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
 [[noreturn]]
 static void usage(const char * executable) {

examples/retrieval/retrieval.cpp

@@ -148,11 +148,12 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;

examples/save-load-state/save-load-state.cpp

@@ -28,10 +28,11 @@ int main(int argc, char ** argv) {
     std::string result2;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;

examples/server/README.md

@@ -207,41 +207,6 @@ model:
   -hff,  --hf-file FILE           Hugging Face model file (default: unused)
   -hft,  --hf-token TOKEN         Hugging Face access token (default: value from HF_TOKEN environment variable)
 
-retrieval:
-
-         --context-file FNAME     file to load context from (repeat to specify multiple files)
-         --chunk-size N           minimum length of embedded text chunks (default: 64)
-         --chunk-separator STRING
-                                  separator between chunks (default: '
-                                  ')
-
-passkey:
-
-         --junk N                 number of times to repeat the junk text (default: 250)
-         --pos N                  position of the passkey in the junk text (default: -1)
-
-imatrix:
-
-  -o,    --output FNAME           output file (default: 'imatrix.dat')
-         --output-frequency N     output the imatrix every N iterations (default: 10)
-         --save-frequency N       save an imatrix copy every N iterations (default: 0)
-         --process-output         collect data for the output tensor (default: false)
-         --no-ppl                 do not compute perplexity (default: true)
-         --chunk N                start processing the input from chunk N (default: 0)
-
-bench:
-
-  -pps                            is the prompt shared across parallel sequences (default: false)
-  -npp n0,n1,...                  number of prompt tokens
-  -ntg n0,n1,...                  number of text generation tokens
-  -npl n0,n1,...                  number of parallel prompts
-
-embedding:
-
-         --embd-normalize         normalisation for embendings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
-         --embd-output-format     empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-         --embd-separator         separator of embendings (default \n) for example "<#sep#>"
-
 server:
 
          --host HOST              ip address to listen (default: 127.0.0.1)
@@ -267,7 +232,8 @@ server:
                                   https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
   -sps,  --slot-prompt-similarity SIMILARITY
                                   how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
-
+         --lora-init-without-apply
+                                  load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)
 
 logging:
 
@@ -279,15 +245,6 @@ logging:
          --log-file FNAME         Specify a log filename (without extension)
          --log-new                Create a separate new log file on start. Each log file will have unique name: "<name>.<ID>.log"
          --log-append             Don't truncate the old log file.
-
-cvector:
-
-  -o,    --output FNAME           output file (default: 'control_vector.gguf')
-         --positive-file FNAME    positive prompts file, one prompt per line (default: 'examples/cvector-generator/positive.txt')
-         --negative-file FNAME    negative prompts file, one prompt per line (default: 'examples/cvector-generator/negative.txt')
-         --pca-batch N            batch size used for PCA. Larger batch runs faster, but uses more memory (default: 100)
-         --pca-iter N             number of iterations used for PCA (default: 1000)
-         --method {pca,mean}      dimensionality reduction method to be used (default: pca)
 ```
 
 
@@ -411,7 +368,8 @@ node index.js
 
 ## API Endpoints
 
-- **GET** `/health`: Returns the current state of the server:
+### GET `/health`: Returns the current state of the server
+
   - 503 -> `{"status": "loading model"}` if the model is still being loaded.
   - 500 -> `{"status": "error"}` if the model failed to load.
   - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
@@ -420,7 +378,7 @@ node index.js
 
   If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
 
-- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
+### POST `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
@@ -498,7 +456,7 @@ node index.js
 
     `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
-### Result JSON
+**Response format**
 
 - Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
 
@@ -537,7 +495,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
 - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
-- **POST** `/tokenize`: Tokenize a given text.
+### POST `/tokenize`: Tokenize a given text
 
     *Options:*
 
@@ -545,13 +503,15 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
 
-- **POST** `/detokenize`: Convert tokens to text.
+### POST `/detokenize`: Convert tokens to text
 
     *Options:*
 
     `tokens`: Set the tokens to detokenize.
 
-- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
+### POST `/embedding`: Generate embedding of a given text
+
+The same as [the embedding example](../embedding) does.
 
     *Options:*
 
@@ -559,7 +519,9 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
-- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+### POST `/infill`: For code infilling.
+
+Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 
@@ -571,7 +533,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - **GET** `/props`: Return current server settings.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -589,7 +551,9 @@ Notice that each `probs` is an array of length `n_probs`.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `chat_template` - the model's original Jinja2 prompt template
 
-- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
+### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
+
+Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
     *Options:*
 
@@ -641,7 +605,7 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API.
+### POST `/v1/embeddings`: OpenAI-compatible embeddings API
 
     *Options:*
 
@@ -675,9 +639,9 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+### GET `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
 
-### Result JSON
+**Response format**
 
 ```json
 [
@@ -738,7 +702,7 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
-- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
+### GET `/metrics`: Prometheus compatible metrics exporter endpoint if `--metrics` is enabled:
 
 Available metrics:
 - `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
@@ -750,13 +714,13 @@ Available metrics:
 - `llamacpp:requests_processing`: Number of requests processing.
 - `llamacpp:requests_deferred`: Number of requests deferred.
 
-- **POST** `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
+### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
 
     *Options:*
 
     `filename`: Name of the file to save the slot's prompt cache. The file will be saved in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -770,13 +734,13 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
+### POST `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
 
     *Options:*
 
     `filename`: Name of the file to restore the slot's prompt cache from. The file should be located in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -790,9 +754,9 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
+### POST `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -801,6 +765,42 @@ Available metrics:
 }
 ```
 
+### GET `/lora-adapters`: Get list of all LoRA adapters
+
+If an adapter is disabled, the scale will be set to 0.
+
+**Response format**
+
+```json
+[
+    {
+        "id": 0,
+        "path": "my_adapter_1.gguf",
+        "scale": 0.0
+    },
+    {
+        "id": 1,
+        "path": "my_adapter_2.gguf",
+        "scale": 0.0
+    }
+]
+```
+
+### POST `/lora-adapters`: Set list of LoRA adapters
+
+To disable an adapter, either remove it from the list below, or set scale to 0.
+
+**Request format**
+
+To know the `id` of the adapter, use GET `/lora-adapters`
+
+```json
+[
+  {"id": 0, "scale": 0.2},
+  {"id": 1, "scale": 0.8}
+]
+```
+
 ## More examples
 
 ### Change system prompt on runtime

examples/server/server.cpp

@@ -78,6 +78,7 @@ enum server_task_type {
     SERVER_TASK_TYPE_SLOT_SAVE,
     SERVER_TASK_TYPE_SLOT_RESTORE,
     SERVER_TASK_TYPE_SLOT_ERASE,
+    SERVER_TASK_TYPE_SET_LORA,
 };
 
 struct server_task {
@@ -622,6 +623,7 @@ struct server_response {
 struct server_context {
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
 
     gpt_params params;
 
@@ -677,7 +679,11 @@ struct server_context {
         // dedicate one sequence to the system prompt
         params.n_parallel += 1;
 
-        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+        model = llama_init.model;
+        ctx = llama_init.context;
+        lora_adapters = llama_init.lora_adapters;
         params.n_parallel -= 1; // but be sneaky about it
         if (model == nullptr) {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -900,7 +906,7 @@ struct server_context {
 
         slot.params.stream             = json_value(data, "stream",            false);
         slot.params.cache_prompt       = json_value(data, "cache_prompt",      false);
-        slot.params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
+        slot.params.n_predict          = json_value(data, "n_predict",         json_value(data, "max_tokens", default_params.n_predict));
         slot.sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
         slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
         slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
@@ -1847,6 +1853,14 @@ struct server_context {
                     };
                     queue_results.send(result);
                 } break;
+            case SERVER_TASK_TYPE_SET_LORA:
+                {
+                    llama_lora_adapters_apply(ctx, lora_adapters);
+                    server_task_result result;
+                    result.id = task.id;
+                    result.data = json{{ "success", true }};
+                    queue_results.send(result);
+                } break;
         }
     }
 
@@ -3325,6 +3339,55 @@ int main(int argc, char ** argv) {
         return res.set_content(root.dump(), "application/json; charset=utf-8");
     };
 
+    const auto handle_lora_adapters_list = [&](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        json result = json::array();
+        for (size_t i = 0; i < ctx_server.lora_adapters.size(); ++i) {
+            auto & la = ctx_server.lora_adapters[i];
+            result.push_back({
+                {"id", i},
+                {"path", la.path},
+                {"scale", la.scale},
+            });
+        }
+        res.set_content(result.dump(), "application/json");
+        res.status = 200; // HTTP OK
+    };
+
+    const auto handle_lora_adapters_apply = [&](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+
+        const std::vector<json> body = json::parse(req.body);
+        int max_idx = ctx_server.lora_adapters.size();
+
+        // clear existing value
+        for (auto & la : ctx_server.lora_adapters) {
+            la.scale = 0.0f;
+        }
+
+        // set value
+        for (auto entry : body) {
+            int id      = entry.at("id");
+            float scale = entry.at("scale");
+            if (0 <= id && id < max_idx) {
+                ctx_server.lora_adapters[id].scale = scale;
+            } else {
+                throw std::runtime_error("invalid adapter id");
+            }
+        }
+
+        server_task task;
+        task.type = SERVER_TASK_TYPE_SET_LORA;
+        const int id_task = ctx_server.queue_tasks.post(task);
+        ctx_server.queue_results.add_waiting_task_id(id_task);
+
+        server_task_result result = ctx_server.queue_results.recv(id_task);
+        ctx_server.queue_results.remove_waiting_task_id(id_task);
+
+        res.set_content(result.data.dump(), "application/json");
+        res.status = 200; // HTTP OK
+    };
+
     auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
         return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
             res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
@@ -3363,7 +3426,6 @@ int main(int argc, char ** argv) {
 
     // register API routes
     svr->Get ("/health",              handle_health);
-    svr->Get ("/slots",               handle_slots);
     svr->Get ("/metrics",             handle_metrics);
     svr->Get ("/props",               handle_props);
     svr->Get ("/v1/models",           handle_models);
@@ -3378,6 +3440,11 @@ int main(int argc, char ** argv) {
     svr->Post("/v1/embeddings",       handle_embeddings);
     svr->Post("/tokenize",            handle_tokenize);
     svr->Post("/detokenize",          handle_detokenize);
+    // LoRA adapters hotswap
+    svr->Get ("/lora-adapters",       handle_lora_adapters_list);
+    svr->Post("/lora-adapters",       handle_lora_adapters_apply);
+    // Save & load slots
+    svr->Get ("/slots",               handle_slots);
     if (!params.slot_save_path.empty()) {
         // only enable slot endpoints if slot_save_path is set
         svr->Post("/slots/:id_slot",  handle_slots_action);

examples/server/tests/features/lora.feature

@@ -0,0 +1,36 @@
+@llama.cpp
+@lora
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model url https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/stories15M_MOE-F16.gguf
+    And   a model file stories15M_MOE-F16.gguf
+    And   a model alias stories15M_MOE
+    And   a lora adapter file from https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf
+    And   42 as server seed
+    And   1024 as batch size
+    And   1024 as ubatch size
+    And   2048 KV cache size
+    And   64 max tokens to predict
+    And   0.0 temperature
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Completion LoRA disabled
+    Given switch off lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching little|girl|three|years|old
+
+  Scenario: Completion LoRA enabled
+    Given switch on lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching eye|love|glass|sun

examples/server/tests/features/steps/steps.py

@@ -7,6 +7,7 @@ import subprocess
 import sys
 import threading
 import time
+import requests
 from collections.abc import Sequence
 from contextlib import closing
 from re import RegexFlag
@@ -70,6 +71,7 @@ def step_server_config(context, server_fqdn: str, server_port: str):
     context.user_api_key = None
     context.response_format = None
     context.temperature = None
+    context.lora_file = None
 
     context.tasks_result = []
     context.concurrent_tasks = []
@@ -82,6 +84,12 @@ def step_download_hf_model(context, hf_file: str, hf_repo: str):
     context.model_hf_file = hf_file
     context.model_file = os.path.basename(hf_file)
 
+@step('a lora adapter file from {lora_file_url}')
+def step_download_lora_file(context, lora_file_url: str):
+    file_name = lora_file_url.split('/').pop()
+    context.lora_file = f'../../../{file_name}'
+    with open(context.lora_file, 'wb') as f:
+        f.write(requests.get(lora_file_url).content)
 
 @step('a model file {model_file}')
 def step_model_file(context, model_file: str):
@@ -849,6 +857,17 @@ async def step_erase_slot(context, slot_id):
             context.response = response
 
 
+@step('switch {on_or_off} lora adapter {lora_id:d}')
+@async_run_until_complete
+async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
+    async with aiohttp.ClientSession() as session:
+        async with session.post(f'{context.base_url}/lora-adapters',
+                                json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
+                                headers={"Content-Type": "application/json"}) as response:
+            context.response = response
+            print([{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}])
+
+
 @step('the server responds with status code {status_code:d}')
 def step_server_responds_with_status_code(context, status_code):
     assert context.response.status == status_code
@@ -1326,6 +1345,8 @@ def start_server_background(context):
         server_args.extend(['--grp-attn-w', context.n_ga_w])
     if context.debug:
         server_args.append('--verbose')
+    if context.lora_file:
+        server_args.extend(['--lora', context.lora_file])
     if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
         server_args.extend(['--log-format', "text"])
 

examples/server/tests/requirements.txt

@@ -4,3 +4,4 @@ huggingface_hub~=0.20.3
 numpy~=1.26.4
 openai~=1.30.3
 prometheus-client~=0.20.0
+requests~=2.32.3

examples/server/utils.hpp

@@ -355,24 +355,6 @@ static json oaicompat_completion_params_parse(
 
     llama_params["__oaicompat"] = true;
 
-    // Map OpenAI parameters to llama.cpp parameters
-    //
-    // For parameters that are defined by the OpenAI documentation (e.g.
-    // temperature), we explicitly specify OpenAI's intended default; we
-    // need to do that because sometimes OpenAI disagrees with llama.cpp
-    //
-    // https://platform.openai.com/docs/api-reference/chat/create
-    llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body,   "model",             std::string("unknown"));
-    llama_params["frequency_penalty"] = json_value(body,   "frequency_penalty", 0.0);
-    llama_params["logit_bias"]        = json_value(body,   "logit_bias",        json::object());
-    llama_params["n_predict"]         = json_value(body,   "max_tokens",        -1);
-    llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
-    llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
-    llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
-    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
-
     // Apply chat template to the list of messages
     llama_params["prompt"] = format_chat(model, chat_template, body.at("messages"));
 

examples/simple/README.md

@@ -3,7 +3,7 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
 
 ```bash
-./simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
+./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
 
 ...
 

examples/speculative/speculative.cpp

@@ -66,7 +66,9 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
+    model_tgt = llama_init_tgt.model;
+    ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
     params.model = params.model_draft;
@@ -75,7 +77,9 @@ int main(int argc, char ** argv) {
         params.n_threads = params.n_threads_draft;
     }
     params.n_threads_batch = params.n_threads_batch_draft;
-    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
+    model_dft = llama_init_dft.model;
+    ctx_dft = llama_init_dft.context;
 
     const bool vocab_type_tgt = llama_vocab_type(model_tgt);
     LOG("vocab_type tgt: %d\n", vocab_type_tgt);

examples/sycl/README.md

@@ -12,9 +12,9 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
 
 List all SYCL devices with ID, compute capability, max work group size, ect.
 
-1. Build the llama.cpp for SYCL for all targets.
+1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
 
-2. Enable oneAPI running environment
+2. Enable oneAPI running environment *(if GGML_SYCL_TARGET is set to INTEL -default-)*
 
 ```
 source /opt/intel/oneapi/setvars.sh
@@ -29,19 +29,13 @@ source /opt/intel/oneapi/setvars.sh
 Check the ID in startup log, like:
 
 ```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+found 2 SYCL devices:
+|  |                   |                                       |       |Max    |        |Max  |Global |                     |
+|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
+|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
+|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
+| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29138|
+| 1| [level_zero:gpu:1]|                 Intel UHD Graphics 750|    1.3|     32|     512|   32| 62631M|            1.3.29138|
 
 ```
 
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|

examples/sycl/win-run-llama2.bat

@@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
 
-.\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0

ggml/include/ggml-metal.h

@@ -50,6 +50,8 @@ GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void
 
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
 
+GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family

ggml/include/ggml.h

@@ -1140,16 +1140,17 @@ extern "C" {
 
     // group normalize along ne0*ne1*n_groups
     // used in stable-diffusion
-    // TODO: eps is hardcoded to 1e-6 for now
     GGML_API struct ggml_tensor * ggml_group_norm(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_group_norm_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     // a - x
     // b - dy
@@ -1456,7 +1457,6 @@ extern "C" {
     // if mode & 2 == 1, GPT-NeoX style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
-    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1473,6 +1473,7 @@ extern "C" {
             int                   mode);
 
     // custom RoPE
+    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

ggml/src/ggml-aarch64.c

@@ -16,6 +16,8 @@
 
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Woverlength-strings"
+#elif defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
 #define UNUSED GGML_UNUSED

ggml/src/ggml-backend.c

@@ -351,15 +351,10 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
     }
 
     // an async copy would normally happen after all the queued operations on both backends are completed
-    // sync src, set_async dst
-    if (ggml_backend_buffer_is_host(src->buffer)) {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
-    } else {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_copy(src, dst);
-        ggml_backend_synchronize(backend_dst);
-    }
+    // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
+    ggml_backend_synchronize(backend_src);
+    ggml_backend_synchronize(backend_dst);
+    ggml_backend_tensor_copy(src, dst);
 }
 
 // events
@@ -1782,7 +1777,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                    ggml_backend_synchronize(input_backend);
+                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                    } else {
+                        ggml_backend_synchronize(split_backend);
+                    }
+                    ggml_backend_tensor_copy(input, input_cpy);
+                }
             }
         }
 

ggml/src/ggml-cann.cpp

@@ -627,7 +627,6 @@ GGML_CALL static void* ggml_backend_cann_buffer_get_base(
 GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
                                                        const void* src,
                                                        void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -679,7 +678,6 @@ GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
  */
 GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
     const ggml_tensor* tensor, void* src, void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -898,11 +896,10 @@ GGML_CALL static void ggml_backend_cann_buffer_init_tensor(
  * @param size Size of the data to be copied, in bytes.
  */
 GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
-    ggml_backend_buffer_t buffer, ggml_tensor* tensor, const void* data,
+    ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
     size_t offset, size_t size) {
-    // GGML_ASSERT(size == ggml_nbytes(tensor));
-    ggml_backend_cann_buffer_context* ctx =
-        (ggml_backend_cann_buffer_context*)buffer->context;
+    ggml_backend_cann_buffer_context *ctx =
+        (ggml_backend_cann_buffer_context *)buffer->context;
 
     ggml_cann_set_device(ctx->device);
     // TODO: refer to cann(#6017), it use thread's default stream.
@@ -910,22 +907,21 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
     // Why aclrtSynchronizeDevice?
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy(tensor->data, size, (const char*)data + offset,
-                              size, ACL_MEMCPY_HOST_TO_DEVICE));
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
     } else {
-        void* transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, (const char*)data + offset,
-                                    transform_buffer);
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
 
 #ifndef NDEBUG
-        void* check_buffer = malloc(size);
+        void *check_buffer = malloc(size);
         ggml_backend_cann_transform_back(tensor, transform_buffer,
                                          check_buffer);
-        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size) ==
-                    0);
+        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpy(tensor->data, size, transform_buffer, size,
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
+                              transform_buffer, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
         free(transform_buffer);
     }
@@ -947,21 +943,20 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
 GGML_CALL static void ggml_backend_cann_buffer_get_tensor(
     ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
     size_t offset, size_t size) {
-    GGML_ASSERT(size == ggml_nbytes(tensor));
     ggml_backend_cann_buffer_context* ctx =
         (ggml_backend_cann_buffer_context*)buffer->context;
 
     ggml_cann_set_device(ctx->device);
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy((char*)data + offset, size, tensor->data, size,
+        ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
     } else {
         void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpy(transform_buffer, size, tensor->data, size,
+        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
+                              (char*)tensor->data + offset, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         (char*)data + offset);
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
         free(transform_buffer);
     }
 }
@@ -1450,42 +1445,41 @@ ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
  * @param size Size of the data to copy in bytes.
  */
 GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
-                                                         ggml_tensor* tensor,
-                                                         const void* data,
+                                                         ggml_tensor *tensor,
+                                                         const void *data,
                                                          size_t offset,
                                                          size_t size) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(
-            tensor->data, size, (const char*)data + offset, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
+                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
+                                   cann_ctx->stream()));
     } else {
-        void* transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, (const char*)data + offset,
-                                    transform_buffer);
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
 
 #ifndef NDEBUG
-        void* check_buffer = malloc(size);
+        void *check_buffer = malloc(size);
         ggml_backend_cann_transform_back(tensor, transform_buffer,
                                          check_buffer);
-        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size));
+        GGML_ASSERT(memcmp(data, check_buffer, size));
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpyAsync(tensor->data, size, transform_buffer, size,
-                                   ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync(
+            (char *)tensor->data + offset, size, transform_buffer, size,
+            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
         free(transform_buffer);
     }
 }
 
 GGML_CALL static void ggml_backend_cann_get_tensor_async(
-    ggml_backend_t backend, const ggml_tensor* tensor, void* data,
+    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
     ggml_backend_buffer_t buf =
         tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
@@ -1493,17 +1487,16 @@ GGML_CALL static void ggml_backend_cann_get_tensor_async(
                 "unsupported buffer type");
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char*)data + offset, size, tensor->data,
+        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
                                    size, ACL_MEMCPY_DEVICE_TO_HOST,
                                    cann_ctx->stream()));
     } else {
-        void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(transform_buffer, size, tensor->data, size,
-                                   ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
+        void *transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpyAsync(
+            transform_buffer, size, (char *)tensor->data + offset, size,
+            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         (char*)data + offset);
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
         free(transform_buffer);
     }
 }
@@ -1666,10 +1659,13 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
-                // case GGML_TYPE_Q4_0:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
                 case GGML_TYPE_Q8_0:
+                    // TODO: fix me
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                case GGML_TYPE_Q4_0:
                     return true;
                 default:
                     return false;
@@ -1694,6 +1690,7 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
                     return true;
                 default:
                     return false;

ggml/src/ggml-cann/acl_tensor.cpp

@@ -37,6 +37,10 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
             return ACL_INT16;
         case GGML_TYPE_I32:
             return ACL_INT32;
+        case GGML_TYPE_Q4_0:
+            return ACL_INT4;
+        case GGML_TYPE_Q8_0:
+            return ACL_INT8;
         default:
             return ACL_DT_UNDEFINED;
     }
@@ -89,33 +93,6 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
     return false;
 }
 
-aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                                   size_t type_size, int64_t* ne, size_t* nb,
-                                   int64_t dims, aclFormat format,
-                                   size_t offset) {
-    int64_t tmp_ne[GGML_MAX_DIMS * 2];
-    int64_t tmp_stride[GGML_MAX_DIMS * 2];
-
-    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
-    for (int i = 0; i < dims; i++) {
-        tmp_stride[i] = nb[i] / type_size;
-    }
-
-    std::reverse(tmp_ne, tmp_ne + dims);
-    std::reverse(tmp_stride, tmp_stride + dims);
-
-    int64_t acl_storage_len = 0;
-    for (int i = 0; i < dims; i++) {
-        acl_storage_len += (ne[i] - 1) * nb[i];
-    }
-
-    aclTensor* acl_tensor =
-        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
-                        format, &acl_storage_len, 1, data_ptr);
-
-    return acl_tensor;
-}
-
 int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
                                   const ggml_tensor* src1,
                                   int64_t* bcast_src0_ne,

ggml/src/ggml-cann/acl_tensor.h

@@ -23,6 +23,9 @@
 #ifndef CANN_ACL_TENSOR_H
 #define CANN_ACL_TENSOR_H
 
+#include <algorithm>
+#include <cstring>
+
 #include <aclnn/aclnn_base.h>
 #include "common.h"
 
@@ -65,7 +68,8 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
                              size_t offset = 0);
 
 /**
- * @brief   Creates an ACL tensor from provided parameters.
+ * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
+ *          should be size_t or float.
  *
  * @details This function creates an ACL tensor using the provided data pointer,
  *          data type, dimensions, strides, format, offset, and additional parameters.
@@ -83,10 +87,34 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
+template<typename TYPE>
 aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                             size_t type_size, int64_t* ne, size_t* nb,
-                             int64_t dims, aclFormat format = ACL_FORMAT_ND,
-                             size_t offset = 0);
+                                   TYPE type_size, int64_t* ne, TYPE* nb,
+                                   int64_t dims,
+                                   aclFormat format = ACL_FORMAT_ND,
+                                   size_t offset = 0) {
+    int64_t tmp_ne[GGML_MAX_DIMS * 2];
+    int64_t tmp_stride[GGML_MAX_DIMS * 2];
+
+    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
+    for (int i = 0; i < dims; i++) {
+        tmp_stride[i] = nb[i] / type_size;
+    }
+
+    std::reverse(tmp_ne, tmp_ne + dims);
+    std::reverse(tmp_stride, tmp_stride + dims);
+
+    int64_t acl_storage_len = 0;
+    for (int i = 0; i < dims; i++) {
+        acl_storage_len += (ne[i] - 1) * nb[i];
+    }
+
+    aclTensor* acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
+                        format, &acl_storage_len, 1, data_ptr);
+
+    return acl_tensor;
+}
 
 /**
  * @brief   Checks if tensors require broadcasting based on their shapes.

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -464,9 +464,11 @@ void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclTensor* acl_src = ggml_cann_create_tensor(src);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    const float eps = 1e-6f;  // TODO: make this a parameter
     int n_groups = dst->op_params[0];
 
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -910,6 +912,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->ne);
             return;
         }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f16_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
         if (dst->type == GGML_TYPE_F16) {
             if (ggml_are_same_shape(src, dst)) {
                 cann_copy(ctx, acl_src, acl_dst);
@@ -971,6 +980,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->ne);
             return;
         }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f32_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
         if (dst->type == GGML_TYPE_F32) {
             if (ggml_are_same_shape(src, dst)) {
                 cann_copy(ctx, acl_src, acl_dst);
@@ -2463,21 +2479,33 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
  * @param dst The destination tensor where the result of the matrix
  * multiplication will be stored.
  */
-static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst) {
+static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
+                                   ggml_tensor* dst,
+                                   const enum ggml_type type) {
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
     // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
     // is regarded as batch. weight need transpose.
     int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
-    size_t weight_elem_size = sizeof(uint8_t);
-    size_t weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    }
+    else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    }
+    else {
+        GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
+    }
+    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+
     // size of one matrix is element_size * height * width.
     size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
     size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
 
     // scale stored at the end of weight. Also need transpose.
+    GGML_ASSERT(QK4_0 == QK8_0);
     int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
     size_t scale_elem_size = sizeof(uint16_t);
     size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
@@ -2541,8 +2569,9 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
                 (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                 input_elem_size, input_ne, input_nb, 2);
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
-                (char*)src0->data + batch0 * weight_stride, ACL_INT8,
-                weight_elem_size, weight_ne, weight_nb, 2);
+                (char*)src0->data + batch0 * weight_stride,
+                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                weight_nb, 2);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
                 scale_offset + batch0 * scale_stride, ACL_FLOAT16,
                 scale_elem_size, scale_ne, scale_nb, 2);
@@ -2596,11 +2625,9 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         case GGML_TYPE_F16:
             ggml_cann_mat_mul_fp(ctx, dst);
             break;
-        // case GGML_TYPE_Q4_0:
-        //     ggml_cann_mul_mat_q4_0(ctx, dst);
-        //     break;
+        case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
-            ggml_cann_mul_mat_q8_0(ctx, dst);
+            ggml_cann_mul_mat_quant(ctx, dst, type);
             break;
         default:
             GGML_ABORT("fatal error");

ggml/src/ggml-cann/kernels/CMakeLists.txt

@@ -9,6 +9,7 @@ file(GLOB SRC_FILES
     get_row_q8_0.cpp
     quantize_f32_q8_0.cpp
     quantize_f16_q8_0.cpp
+    quantize_float_to_q4_0.cpp
     dup.cpp
 )
 
@@ -29,4 +30,4 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
-#ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
+# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

ggml/src/ggml-cann/kernels/ascendc_kernels.h

@@ -8,6 +8,8 @@
 
 #include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
 #include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
+#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
+#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"
 
 #include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
 #include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"

ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp

@@ -0,0 +1,278 @@
+#include "kernel_operator.h"
+
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+#define Group_Size 32
+
+template <typename SRC_T>
+class QUANTIZE_FLOAT_TO_Q4_0 {
+   public:
+    __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *output_ne_ub) {
+        // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
+        //                         permute=[0,0,0,0]):
+        // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        // input stride of data elements
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+            output_ne[i] = output_ne_ub[i];
+        }
+
+        // output stride of data elements
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
+        }
+
+        // scale saved one by one after data:. [group1_scale, group2_scale, ...]
+        scale_ne = input_ne;
+        scale_stride[0] = 1;
+        scale_stride[1] = input_ne[0] / Group_Size;
+        for (int i = 2; i < 4; i++) {
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        // split input tensor by rows.
+        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
+        dr = nr / op_block_num;
+
+        uint64_t tails = nr % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        group_size_in_row = scale_stride[1];
+        int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
+                              output_ne[3] * sizeof(uint8_t) / 2;
+
+        input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
+        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
+                                                 group_size_in_row *
+                                                 sizeof(half)));
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
+        pipe.InitBuffer(output_queue, BUFFER_NUM,
+                            Group_Size * sizeof(int8_t) / 2);
+        pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
+        pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
+        pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
+        DataCopy(input_local, input_gm[offset], Group_Size);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
+        // and using DataCopyPad to avoid 32 bits align.
+        LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
+        LocalTensor<int8_t> output_int8_local =
+                                    output_local.ReinterpretCast<int8_t>();
+
+        DataCopyExtParams dataCopyParams;
+        dataCopyParams.blockCount = 1;
+        dataCopyParams.blockLen = Group_Size / 2  * sizeof(int8_t);
+        DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
+
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<float> input_local) {
+        DataCopy(cast_local, input_local, Group_Size);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<half> input_local) {
+        Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
+    }
+
+    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
+        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
+        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
+        const int64_t i1 =
+            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
+
+        const int64_t input_offset = i1 * input_stride[1] +
+                                     i2 * input_stride[2] +
+                                     i3 * input_stride[3] + Group_Size * group;
+
+        // output_offset is stride for output_gm which datatype is int8_t and
+        // divided by 2 is needed for int4b_t.
+        const int64_t output_offset = (i1 * output_stride[1] +
+                                       i2 * output_stride[2] +
+                                       i3 * output_stride[3] +
+                                       Group_Size * group) / 2;
+        copy_in(input_offset);
+
+        LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
+        LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
+        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
+        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
+        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
+        LocalTensor<float> min_local = min_queue.AllocTensor<float>();
+        LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
+        LocalTensor<half> half_local = half_queue.AllocTensor<half>();
+
+        input_to_cast(cast_local, input_local);
+
+        ReduceMax(max_local, cast_local, work_local, Group_Size);
+        ReduceMin(min_local, cast_local, work_local, Group_Size);
+        const float max_value = max_local.GetValue(0);
+        const float min_value = min_local.GetValue(0);
+        float d = max_value;
+        if (min_value < 0 && (-1 * min_value) > max_value) {
+            d = min_value;
+        }
+
+        d = d / (-8);
+        if (d != 0) {
+            Muls(cast_local, cast_local, 1.0f / d, Group_Size);
+        }
+
+        // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
+        float scalar = 8.5f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+        Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
+        scalar = 15.0f;
+        Mins(cast_local, cast_local, scalar, Group_Size);
+        scalar = -8.0f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+
+        // float->half->int4b
+        Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
+        Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
+
+        output_queue.EnQue(output_local);
+        copy_out(output_offset);
+
+        input_queue.FreeTensor(input_local);
+        work_queue.FreeTensor(work_local);
+        max_queue.FreeTensor(max_local);
+        min_queue.FreeTensor(min_local);
+        int8_queue.FreeTensor(int8_local);
+        half_queue.FreeTensor(half_local);
+        cast_queue.FreeTensor(cast_local);
+        return (half)d;
+    }
+
+    __aicore__ inline void calculate() {
+        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
+        uint32_t scale_local_offset = 0;
+        uint32_t scale_global_offset = 0;
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                half scale = calculate_group(i, j);
+                scale_local.SetValue(scale_local_offset++, scale);
+                // Copy Group_Size/2 length data each time.
+                if (scale_local_offset == Group_Size / 2) {
+                    scale_local_offset = 0;
+                    // TODO: OPTIMIZE ME
+                    pipe_barrier(PIPE_ALL);
+                    DataCopy(scale_gm[scale_global_offset], scale_local,
+                                      Group_Size / 2);
+                    pipe_barrier(PIPE_ALL);
+                    scale_global_offset += Group_Size / 2;
+                }
+            }
+        }
+
+        if (scale_local_offset != 0) {
+            pipe_barrier(PIPE_ALL);
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
+            DataCopyPad(scale_gm[scale_global_offset], scale_local,
+                        dataCopyParams);
+            pipe_barrier(PIPE_ALL);
+        }
+        scale_queue.FreeTensor(scale_local);
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t *scale_ne;
+    size_t scale_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t group_size_in_row;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<SRC_T> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int8_t> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<half> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<float> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}

ggml/src/ggml-cuda.cu

@@ -1501,7 +1501,7 @@ static void ggml_cuda_op_mul_mat(
         }
 
         // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
-        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
+        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
             const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
             const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
@@ -2358,33 +2358,35 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend,
 }
 
 GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));
-
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
-    if (!ggml_backend_buffer_is_cuda(src->buffer)) {
+    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
         return false;
     }
 
-    if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
+    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
         return false;
     }
 
-    // device -> device
+    // device -> device copy
     ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
     ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
 
+    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+
+    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+#ifndef NDEBUG
+        GGML_CUDA_LOG_WARN("%s: backend and buffer devices do not match\n", __func__);
+#endif
+        return false;
+    }
+
     if (backend_src != backend_dst) {
-        ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
-        ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
-
-        GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
-        GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
-
         // copy on src stream
         if (cuda_ctx_src->device == cuda_ctx_dst->device) {
-            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
         } else {
 #ifdef GGML_CUDA_NO_PEER_COPY
             return false;
@@ -2393,7 +2395,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
 #endif
         }
 
-        // record event on src stream
+        // record event on src stream after the copy
         if (!cuda_ctx_src->copy_event) {
             ggml_cuda_set_device(cuda_ctx_src->device);
             CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
@@ -2405,7 +2407,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
         CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
     } else {
         // src and dst are on the same backend
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
     }
     return true;
 }
@@ -2742,11 +2744,12 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_MUL_MAT_ID:
             {
                 struct ggml_tensor * a = op->src[0];
-                if (op->op == GGML_OP_MUL_MAT) {
-                    struct ggml_tensor * b = op->src[1];
-                    if (a->ne[3] != b->ne[3]) {
-                        return false;
-                    }
+                struct ggml_tensor * b = op->src[1];
+                if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
+                    return false;
+                }
+                if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) {
+                    return false;
                 }
                 switch (a->type) {
                     case GGML_TYPE_F32:
@@ -2877,7 +2880,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+            return (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) || op->src[0]->ne[0] == 128;
 #else
             if (op->src[0]->ne[0] == 128) {
                 return true;

ggml/src/ggml-cuda/norm.cu

@@ -142,8 +142,7 @@ 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 int group_size, const int ne_elements, cudaStream_t stream) {
-    static const float eps = 1e-6f;
+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);
@@ -196,8 +195,12 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     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], group_size, ggml_nelements(src0), stream);
+    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
 }
 
 void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-metal.m

@@ -210,7 +210,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COUNT
 };
 
-struct ggml_metal_context {
+struct ggml_backend_metal_context {
     int n_cb;
 
     id<MTLDevice>       device;
@@ -224,6 +224,10 @@ struct ggml_metal_context {
     bool support_simdgroup_mm;
 
     bool should_capture_next_compute;
+
+    // abort ggml_metal_graph_compute if callback returns true
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
 };
 
 // MSL code
@@ -289,7 +293,7 @@ static void * ggml_metal_host_malloc(size_t n) {
     return data;
 }
 
-static struct ggml_metal_context * ggml_metal_init(int n_cb) {
+static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
 
 #if TARGET_OS_OSX && !GGML_METAL_NDEBUG
@@ -306,7 +310,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
     // Configure context
-    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+    struct ggml_backend_metal_context * ctx = malloc(sizeof(struct ggml_backend_metal_context));
     ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
@@ -668,7 +672,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     return ctx;
 }
 
-static void ggml_metal_free(struct ggml_metal_context * ctx) {
+static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
     GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 
     for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
@@ -734,7 +738,7 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs
     return nil;
 }
 
-static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
+static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) {
     for (size_t i = 0, n = 3; i < n; ++i) {
         if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
             return false;
@@ -845,7 +849,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
 }
 
 static enum ggml_status ggml_metal_graph_compute(
-        struct ggml_metal_context * ctx,
+        struct ggml_backend_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
     @autoreleasepool {
@@ -878,8 +882,11 @@ static enum ggml_status ggml_metal_graph_compute(
         id<MTLCommandBuffer> command_buffer  = [ctx->queue commandBufferWithUnretainedReferences];
         command_buffer_builder[cb_idx] = command_buffer;
 
-        // enqueue the command buffers in order to specify their execution order
-        [command_buffer enqueue];
+        // always enqueue the first two command buffers
+        // enqueue all of the command buffers if we don't need to abort
+        if (cb_idx < 2 || ctx->abort_callback == NULL) {
+            [command_buffer enqueue];
+        }
     }
 
     const id<MTLCommandBuffer> *command_buffers = command_buffer_builder;
@@ -2229,10 +2236,8 @@ static enum ggml_status ggml_metal_graph_compute(
                         GGML_ASSERT(ne00 % 4 == 0);
                         GGML_ASSERT(ggml_is_contiguous(src0));
 
-                        //float eps;
-                        //memcpy(&eps, dst->op_params, sizeof(float));
-
-                        const float eps = 1e-6f; // TODO: temporarily hardcoded
+                        float eps;
+                        memcpy(&eps, dst->op_params + 1, sizeof(float));
 
                         const int32_t n_groups = ((int32_t *) dst->op_params)[0];
 
@@ -2829,7 +2834,9 @@ static enum ggml_status ggml_metal_graph_compute(
 
         [encoder endEncoding];
 
-        [command_buffer commit];
+        if (cb_idx < 2 || ctx->abort_callback == NULL) {
+            [command_buffer commit];
+        }
     });
 
     // Wait for completion and check status of each command buffer
@@ -2849,6 +2856,23 @@ static enum ggml_status ggml_metal_graph_compute(
 
             return GGML_STATUS_FAILED;
         }
+
+        id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? command_buffers[i + 1] : nil);
+        if (!next_buffer) {
+            continue;
+        }
+
+        bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued);
+        if (next_queued) {
+            continue;
+        }
+
+        if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) {
+            GGML_METAL_LOG_INFO("%s: command buffer %d aborted", __func__, i);
+            return GGML_STATUS_ABORTED;
+        }
+
+        [next_buffer commit];
     }
 
     if (should_capture) {
@@ -3152,7 +3176,7 @@ GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
 }
 
 GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
     ggml_metal_free(ctx);
     free(backend);
 }
@@ -3164,13 +3188,13 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe
 }
 
 GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return ggml_metal_graph_compute(metal_ctx, cgraph);
 }
 
 GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return ggml_metal_supports_op(metal_ctx, op);
 }
@@ -3215,9 +3239,9 @@ static ggml_guid_t ggml_backend_metal_guid(void) {
 }
 
 ggml_backend_t ggml_backend_metal_init(void) {
-    struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
-
+    struct ggml_backend_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
     if (ctx == NULL) {
+        GGML_METAL_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
         return NULL;
     }
 
@@ -3239,15 +3263,24 @@ bool ggml_backend_is_metal(ggml_backend_t backend) {
 void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
 
     ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
 }
 
+void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = user_data;
+}
+
 bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
 }
@@ -3255,7 +3288,7 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
 void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
     ctx->should_capture_next_compute = true;
 }
 

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -874,7 +874,7 @@ namespace dpct
         inline std::string get_preferred_gpu_platform_name() {
             std::string result;
 
-            std::string filter = "level-zero";
+            std::string filter = "";
             char* env = getenv("ONEAPI_DEVICE_SELECTOR");
             if (env) {
                 if (std::strstr(env, "level_zero")) {
@@ -892,11 +892,24 @@ namespace dpct
                 else {
                     throw std::runtime_error("invalid device filter: " + std::string(env));
                 }
+            } else {
+                auto default_device = sycl::device(sycl::default_selector_v);
+                auto default_platform_name = default_device.get_platform().get_info<sycl::info::platform::name>();
+
+                if (std::strstr(default_platform_name.c_str(), "Level-Zero") || default_device.is_cpu()) {
+                    filter = "level-zero";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "CUDA")) {
+                    filter = "cuda";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "HIP")) {
+                    filter = "hip";
+                }
             }
 
-            auto plaform_list = sycl::platform::get_platforms();
+            auto platform_list = sycl::platform::get_platforms();
 
-            for (const auto& platform : plaform_list) {
+            for (const auto& platform : platform_list) {
                 auto devices = platform.get_devices();
                 auto gpu_dev = std::find_if(devices.begin(), devices.end(), [](const sycl::device& d) {
                     return d.is_gpu();

ggml/src/ggml-sycl/norm.cpp

@@ -225,9 +225,8 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols,
 }
 
 static void group_norm_f32_sycl(const float* x, float* dst,
-    const int num_groups, const int group_size,
+    const int num_groups, const float eps, const int group_size,
     const int ne_elements, queue_ptr stream, int device) {
-    static const float eps = 1e-6f;
     if (group_size < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
@@ -343,8 +342,12 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor*
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
-    group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device);
+    group_norm_f32_sycl(src0_dd, dst_dd, num_groups, eps, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device);
 
     (void)src1;
     (void)dst;

ggml/src/ggml.c

@@ -56,6 +56,9 @@ int ggml_sve_cnt_b = 0;
 // disable POSIX deprecation warnings
 // these functions are never going away, anyway
 #pragma warning(disable: 4996)
+
+// unreachable code because of multiple instances of code after GGML_ABORT
+#pragma warning(disable: 4702)
 #endif
 
 #if defined(_WIN32)
@@ -2312,7 +2315,7 @@ inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) {
 inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
 inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
 inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
-inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
+inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
 inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
@@ -5374,6 +5377,7 @@ static struct ggml_tensor * ggml_group_norm_impl(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
     int n_groups,
+    float eps,
     bool inplace) {
 
     bool is_node = false;
@@ -5384,7 +5388,8 @@ static struct ggml_tensor * ggml_group_norm_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op_params[0] = n_groups;
+    ggml_set_op_params_i32(result, 0, n_groups);
+    ggml_set_op_params_f32(result, 1, eps);
 
     result->op = GGML_OP_GROUP_NORM;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5396,15 +5401,17 @@ static struct ggml_tensor * ggml_group_norm_impl(
 struct ggml_tensor * ggml_group_norm(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
-    int n_groups) {
-    return ggml_group_norm_impl(ctx, a, n_groups, false);
+    int n_groups,
+    float eps) {
+    return ggml_group_norm_impl(ctx, a, n_groups, eps, false);
 }
 
 struct ggml_tensor * ggml_group_norm_inplace(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
-    int n_groups) {
-    return ggml_group_norm_impl(ctx, a, n_groups, true);
+    int n_groups,
+    float eps) {
+    return ggml_group_norm_impl(ctx, a, n_groups, eps, true);
 }
 
 // ggml_mul_mat
@@ -12095,10 +12102,11 @@ static void ggml_compute_forward_group_norm_f32(
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    const float eps = 1e-6f; // TODO: make this a parameter
-
     // TODO: optimize
 
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int n_channels = src0->ne[2];
     int n_groups = dst->op_params[0];
     int n_channels_per_group = (n_channels + n_groups - 1) / n_groups;

ggml/src/vulkan-shaders/CMakeLists.txt

@@ -1,5 +1,7 @@
+find_package (Threads REQUIRED)
 
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)

ggml/src/vulkan-shaders/add.comp

@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) + FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)]));
 }

ggml/src/vulkan-shaders/clamp.comp

@@ -4,10 +4,12 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
 }

ggml/src/vulkan-shaders/concat.comp

@@ -0,0 +1,35 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+    const int dim = p.param3;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne22*p.ne21*p.ne20);
+    const uint i3_offset = i3 * p.ne22*p.ne21*p.ne20;
+    const uint i2 = (idx - i3_offset) / (p.ne21*p.ne20);
+    const uint i2_offset = i2*p.ne21*p.ne20;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne20;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne20;
+
+    uint o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? p.ne00 : (dim == 1 ? p.ne01 : (dim == 2 ? p.ne02 : p.ne03));
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint src1_idx = (i3 - o[3])*p.nb13 + (i2 - o[2])*p.nb12 + (i1 - o[1])*p.nb11 + (i0 - o[0])*p.nb10;
+    const uint dst_idx = i3*p.nb23 + i2*p.nb22 + i1*p.nb21 + i0*p.nb20;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
+#else
+    data_d[p.d_offset + dst_idx] = is_src0 ? data_a[src0_idx] : data_b[src1_idx];
+#endif
+}

ggml/src/vulkan-shaders/copy.comp

@@ -4,13 +4,15 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]);
 #else
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = data_a[src0_idx(gl_GlobalInvocationID.x)];
+    data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)];
 #endif
 }

ggml/src/vulkan-shaders/div.comp

@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) / FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) / FLOAT_TYPE(data_b[src1_idx(idx)]));
 }

ggml/src/vulkan-shaders/gelu.comp

@@ -13,7 +13,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 void main() {
     const float GELU_COEF_A    = 0.044715f;
     const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
     if (i >= p.KX) {
         return;

ggml/src/vulkan-shaders/gelu_quick.comp

@@ -0,0 +1,23 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const float GELU_QUICK_COEF = -1.702f;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(x * (1.0f / (1.0f + exp(GELU_QUICK_COEF * x))));
+}

ggml/src/vulkan-shaders/generic_binary_head.comp

@@ -7,7 +7,7 @@ layout (push_constant) uniform parameter
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
     uint d_offset;
-    float param1; float param2;
+    float param1; float param2; int param3;
 } p;
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
@@ -16,6 +16,10 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
 uint src0_idx(uint idx) {
     const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
     const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;

ggml/src/vulkan-shaders/generic_unary_head.comp

@@ -14,6 +14,10 @@ layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
 uint src0_idx(uint idx) {
     const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
     const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;

ggml/src/vulkan-shaders/group_norm.comp

@@ -0,0 +1,66 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+shared float tmp[BLOCK_SIZE];
+
+void main() {
+    const uint group_size = p.KX;
+    const float eps = p.param1;
+
+    const uint tid = gl_LocalInvocationID.x;
+    const uint start = gl_WorkGroupID.x * group_size + tid;
+    const uint end = start + group_size;
+
+    tmp[tid] = 0.0f;
+
+    // Calculate mean
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        tmp[tid] += float(data_a[col]);
+    }
+
+    // tmp up partial tmps and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float mean = tmp[0] / group_size;
+    barrier();
+    tmp[tid] = 0.0f;
+
+    // Calculate variance
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        const float xi = float(data_a[col]) - mean;
+        data_d[col] = D_TYPE(xi);
+        tmp[tid] += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float variance = tmp[0] / group_size;
+    const float scale = inversesqrt(variance + eps);
+
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        data_d[col] *= D_TYPE(scale);
+    }
+}

ggml/src/vulkan-shaders/im2col.comp

@@ -0,0 +1,57 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint batch_offset; uint offset_delta;
+    uint IC;
+    uint IW; uint IH;
+    uint OW; uint OH;
+    uint KW; uint KH;
+    uint pelements;
+    uint CHW;
+    int s0; int s1;
+    int p0; int p1;
+    int d0; int d1;
+} p;
+
+#include "types.comp"
+
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x;
+    if (i >= p.pelements) {
+        return;
+    }
+
+    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+    const uint kx = i / ksize;
+    const uint kd = kx * ksize;
+    const uint ky = (i - kd) / p.OW;
+    const uint ix = i % p.OW;
+
+    const uint oh = gl_GlobalInvocationID.y;
+    const uint batch = gl_GlobalInvocationID.z / p.IC;
+    const uint ic = gl_GlobalInvocationID.z % p.IC;
+
+    const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
+    const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
+
+    const uint offset_dst =
+        ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
+        (ic * (p.KW * p.KH) + ky * p.KW + kx);
+
+    if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
+        data_d[offset_dst] = D_TYPE(0.0f);
+    } else {
+        const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
+        data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
+    }
+}

ggml/src/vulkan-shaders/leaky_relu.comp

@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float val = float(data_a[i]);
+    data_d[i] = D_TYPE(max(val, 0.0f) + min(val, 0.0f) * p.param1);
+}

ggml/src/vulkan-shaders/mul.comp

@@ -4,9 +4,11 @@
 #include "generic_binary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(data_b[src1_idx(gl_GlobalInvocationID.x)]));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(data_b[src1_idx(idx)]));
 }

ggml/src/vulkan-shaders/mul_mat_vec.comp

@@ -16,6 +16,13 @@ void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
     const uint tid = gl_LocalInvocationID.x;
 
+    // There are not enough cols to use all threads
+    if (tid >= p.ncols) {
+        return;
+    }
+
+    const uint block_size = min(p.ncols, BLOCK_SIZE);
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -23,8 +30,8 @@ void main() {
 
     tmp[tid] = FLOAT_TYPE(0.0f);
 
-    [[unroll]] for (uint i = 0; i < p.ncols/BLOCK_SIZE; i += 2) {
-        const uint col = i*BLOCK_SIZE + 2*tid;
+    [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
+        const uint col = i*block_size + 2*tid;
         const uint ib = (row*p.ncols + col)/QUANT_K; // block index
         const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
         const uint iybs = col - col%QUANT_K; // y block start index
@@ -38,7 +45,7 @@ void main() {
 
     // sum up partial sums and write back result
     barrier();
-    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }

ggml/src/vulkan-shaders/norm.comp

@@ -14,7 +14,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 shared vec2 sum[BLOCK_SIZE];
 
 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
     sum[tid] = vec2(0.0f, 0.0f);

ggml/src/vulkan-shaders/pad.comp

@@ -0,0 +1,26 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i3_offset = i3 * p.ne12*p.ne11*p.ne10;
+    const uint i2 = (idx - i3_offset) / (p.ne11*p.ne10);
+    const uint i2_offset = i2*p.ne11*p.ne10;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne10;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne10;
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint dst_idx = i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0*p.nb10;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : 0.0f);
+}

ggml/src/vulkan-shaders/relu.comp

@@ -11,7 +11,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 void main() {
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
     if (i >= p.KX) {
         return;

ggml/src/vulkan-shaders/rms_norm.comp

@@ -14,7 +14,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 shared FLOAT_TYPE sum[BLOCK_SIZE];
 
 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp

ggml/src/vulkan-shaders/scale.comp

@@ -4,9 +4,11 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]) * FLOAT_TYPE(p.param1));
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(p.param1));
 }

ggml/src/vulkan-shaders/silu.comp

@@ -11,7 +11,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 void main() {
-    const uint i = gl_GlobalInvocationID.x;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
     if (i >= p.KX) {
         return;

ggml/src/vulkan-shaders/soft_max.comp

@@ -28,7 +28,7 @@ shared FLOAT_TYPE vals[BLOCK_SIZE];
 
 void main() {
     const uint tid = gl_LocalInvocationID.x;
-    const uint rowx = gl_WorkGroupID.x;
+    const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint rowy = rowx % p.KY;
 
     float slope = 1.0f;

ggml/src/vulkan-shaders/square.comp

@@ -4,10 +4,12 @@
 #include "generic_unary_head.comp"
 
 void main() {
-    if (gl_GlobalInvocationID.x >= p.ne) {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(gl_GlobalInvocationID.x)]);
-    data_d[p.d_offset + dst_idx(gl_GlobalInvocationID.x)] = D_TYPE(val * val);
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val * val);
 }

ggml/src/vulkan-shaders/sum_rows.comp

@@ -14,7 +14,7 @@ layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {
-    const uint row = gl_WorkGroupID.x;
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
 
     tmp[col] = FLOAT_TYPE(0.0f);

ggml/src/vulkan-shaders/tanh.comp

@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    data_d[i] = D_TYPE(tanh(data_a[i]));
+}

ggml/src/vulkan-shaders/timestep_embedding.comp

@@ -0,0 +1,41 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint nb1;
+    uint dim;
+    uint max_period;
+} p;
+
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.y;
+    const uint j = gl_GlobalInvocationID.x;
+    const uint d_offset = i * p.nb1;
+
+    if (p.dim % 2 != 0 && j == ((p.dim + 1) / 2)) {
+        data_d[d_offset + p.dim] = 0.f;
+    }
+
+    const uint half_dim = p.dim / 2;
+    if (j >= half_dim) {
+        return;
+    }
+
+    const float timestep = float(data_a[i]);
+    const float freq = float(exp(-log(p.max_period) * j / half_dim));
+    const float arg = timestep * freq;
+    data_d[d_offset + j] = D_TYPE(cos(arg));
+    data_d[d_offset + j + half_dim] = D_TYPE(sin(arg));
+}

ggml/src/vulkan-shaders/types.comp

@@ -6,7 +6,7 @@
 #define QUANT_K 1
 #define QUANT_R 1
 
-#ifndef LOAD_VEC_A
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
 #define A_TYPE float
 #elif LOAD_VEC_A == 4
 #define A_TYPE vec4
@@ -19,7 +19,7 @@
 #define QUANT_K 1
 #define QUANT_R 1
 
-#ifndef LOAD_VEC_A
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
 #define A_TYPE float16_t
 #elif LOAD_VEC_A == 4
 #define A_TYPE f16vec4

ggml/src/vulkan-shaders/upscale.comp

@@ -0,0 +1,36 @@
+#version 450
+
+layout (push_constant) uniform parameter
+{
+    uint ne; uint d_offset;
+    uint nb00; uint nb01; uint nb02; uint nb03;
+    uint ne10; uint ne11; uint ne12; uint ne13;
+    float sf0; float sf1; float sf2; float sf3;
+} p;
+
+#include "types.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i10 = idx % p.ne10;
+    const uint i11 = (idx / p.ne10) % p.ne11;
+    const uint i12 = (idx / (p.ne10 * p.ne11)) % p.ne12;
+    const uint i13 = (idx / (p.ne10 * p.ne11 * p.ne12)) % p.ne13;
+
+    const uint i00 = uint(i10 / p.sf0);
+    const uint i01 = uint(i11 / p.sf1);
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+
+    data_d[p.d_offset + idx] = D_TYPE(data_a[i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
+}

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

@@ -22,6 +22,7 @@
 #ifdef _WIN32
     #include <windows.h>
     #include <direct.h> // For _mkdir on Windows
+    #include <algorithm> // For std::replace on w64devkit
 #else
     #include <unistd.h>
     #include <sys/wait.h>
@@ -30,20 +31,6 @@
 
 #define ASYNCIO_CONCURRENCY 64
 
-// define prototypes
-void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str);
-bool directory_exists(const std::string& path);
-bool create_directory(const std::string& path);
-std::string to_uppercase(const std::string& input);
-bool string_ends_with(const std::string& str, const std::string& suffix);
-std::string join_paths(const std::string& path1, const std::string& path2);
-std::string basename(const std::string &path);
-void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16);
-std::map<std::string, std::string> merge_maps(const std::map<std::string, std::string>& a, const std::map<std::string, std::string>& b);
-void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmul_id);
-void process_shaders(std::vector<std::future<void>>& tasks);
-void write_output_files();
-
 std::mutex lock;
 std::vector<std::pair<std::string, std::string>> shader_fnames;
 
@@ -52,7 +39,7 @@ std::string input_dir = "vulkan-shaders";
 std::string output_dir = "/tmp";
 std::string target_hpp = "ggml-vulkan-shaders.hpp";
 std::string target_cpp = "ggml-vulkan-shaders.cpp";
-bool clean = true;
+bool no_clean = false;
 
 const std::vector<std::string> type_names = {
     "f32",
@@ -193,11 +180,7 @@ bool string_ends_with(const std::string& str, const std::string& suffix) {
     return std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
 }
 
-#ifdef _WIN32
-    static const char path_separator = '\\';
-#else
-    static const char path_separator = '/';
-#endif
+static const char path_separator = '/';
 
 std::string join_paths(const std::string& path1, const std::string& path2) {
     return path1 + path_separator + path2;
@@ -212,7 +195,11 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
     std::string out_fname = join_paths(output_dir, name + ".spv");
     std::string in_path = join_paths(input_dir, in_fname);
 
-    std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname};
+    #ifdef _WIN32
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
+    #else
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o",  out_fname};
+    #endif
     for (const auto& define : defines) {
         cmd.push_back("-D" + define.first + "=" + define.second);
     }
@@ -283,9 +270,12 @@ void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmu
 
     for (const auto& tname : type_names) {
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+        // For unaligned, load one at a time for f32/f16, or two at a time for quants
+        std::string load_vec_a_unaligned = (tname == "f32" || tname == "f16") ? "1" : "2";
+        // For aligned matmul loads
         std::string load_vec_a = (tname == "f32" || tname == "f16") ? load_vec : "2";
         tasks.push_back(std::async(std::launch::async, [=] {
-            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
+            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
         }));
         tasks.push_back(std::async(std::launch::async, [=] {
             string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16);
@@ -354,6 +344,9 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
     tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
@@ -371,6 +364,9 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
+    }));
 
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
@@ -396,15 +392,42 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
         string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
 
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
 
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -438,6 +461,17 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
     tasks.push_back(std::async(std::launch::async, [=] {
         string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
 }
 
 void write_output_files() {
@@ -449,10 +483,16 @@ void write_output_files() {
 
     for (const auto& pair : shader_fnames) {
         const std::string& name = pair.first;
-        const std::string& path = pair.second;
+        #ifdef _WIN32
+            std::string path = pair.second;
+            std::replace(path.begin(), path.end(), '/', '\\' );
+        #else
+            const std::string& path = pair.second;
+        #endif
+
         FILE* spv = fopen(path.c_str(), "rb");
         if (!spv) {
-            std::cerr << "Error opening SPIR-V file: " << path << "\n";
+            std::cerr << "Error opening SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
             continue;
         }
 
@@ -464,7 +504,7 @@ void write_output_files() {
         size_t read_size = fread(data.data(), 1, size, spv);
         fclose(spv);
         if (read_size != size) {
-            std::cerr << "Error reading SPIR-V file: " << path << "\n";
+            std::cerr << "Error reading SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
             continue;
         }
 
@@ -478,9 +518,8 @@ void write_output_files() {
         }
         fprintf(src, "\n};\n\n");
 
-        if (clean) {
+        if (!no_clean) {
             std::remove(path.c_str());
-            // fprintf(stderr, "Removed: %s\n", path.c_str());
         }
     }
 
@@ -496,18 +535,6 @@ int main(int argc, char** argv) {
         }
     }
 
-    if (argc <= 1 || args.find("--help") != args.end()) {
-        std::cout << "Usage:\n"
-                     "\tvulkan-shaders-gen [options]\n\n"
-                     "Options:\n"
-                     "\t--glslc <path>        Path to glslc executable (default: /usr/bin/glslc)\n"
-                     "\t--input-dir           Directory containing shader sources (required)\n"
-                     "\t--output-dir          Output directory for generated SPIR-V files and optional C++ headers\n"
-                     "\t--target-hpp <path>   Path to generate a header file with shader declarations in C++ format\n"
-                     "\t--target-cpp <path>   Path to generate a source code file implementing the declared shaders (optional)\n"
-                     "\t--no-clean            Keep temporary SPIR-V files after build (default: remove them)\n";
-        return EXIT_SUCCESS;
-    }
     if (args.find("--glslc") != args.end()) {
         GLSLC = args["--glslc"]; // Path to glslc
     }
@@ -524,7 +551,7 @@ int main(int argc, char** argv) {
         target_cpp = args["--target-cpp"]; // Path to generated cpp file
     }
     if (args.find("--no-clean") != args.end()) {
-        clean = false; // Keep temporary SPIR-V files in output-dir after build
+        no_clean = true; // Keep temporary SPIR-V files in output-dir after build
     }
 
     if (!directory_exists(input_dir)) {

gguf-py/gguf/constants.py

@@ -161,6 +161,7 @@ class Keys:
         SUFFIX_ID            = "tokenizer.ggml.suffix_token_id"
         MIDDLE_ID            = "tokenizer.ggml.middle_token_id"
         EOT_ID               = "tokenizer.ggml.eot_token_id"
+        EOM_ID               = "tokenizer.ggml.eom_token_id"
 
     class Adapter:
         TYPE       = "adapter.type"
@@ -1145,6 +1146,9 @@ class GGMLQuantizationType(IntEnum):
     F64     = 28
     IQ1_M   = 29
     BF16    = 30
+    Q4_0_4_4 = 31
+    Q4_0_4_8 = 32
+    Q4_0_8_8 = 33
 
 
 # TODO: add GGMLFileType from ggml_ftype in ggml.h
@@ -1157,7 +1161,7 @@ class LlamaFileType(IntEnum):
     MOSTLY_F16           = 1   # except 1d tensors
     MOSTLY_Q4_0          = 2   # except 1d tensors
     MOSTLY_Q4_1          = 3   # except 1d tensors
-    MOSTLY_Q4_1_SOME_F16 = 4   # tok_embeddings.weight and output.weight are F16
+    # MOSTLY_Q4_1_SOME_F16 = 4   # tok_embeddings.weight and output.weight are F16
     # MOSTLY_Q4_2        = 5   # support has been removed
     # MOSTLY_Q4_3        = 6   # support has been removed
     MOSTLY_Q8_0          = 7   # except 1d tensors
@@ -1186,6 +1190,9 @@ class LlamaFileType(IntEnum):
     MOSTLY_IQ4_XS        = 30  # except 1d tensors
     MOSTLY_IQ1_M         = 31  # except 1d tensors
     MOSTLY_BF16          = 32  # except 1d tensors
+    MOSTLY_Q4_0_4_4      = 33  # except 1d tensors
+    MOSTLY_Q4_0_4_8      = 34  # except 1d tensors
+    MOSTLY_Q4_0_8_8      = 35  # except 1d tensors
 
     GUESSED              = 1024  # not specified in the model file
 
@@ -1259,6 +1266,9 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
     GGMLQuantizationType.F64:     (1, 8),
     GGMLQuantizationType.IQ1_M:   (256, QK_K // 8 + QK_K // 16  + QK_K // 32),
     GGMLQuantizationType.BF16:    (1, 2),
+    GGMLQuantizationType.Q4_0_4_4:(32, 2 + 16),
+    GGMLQuantizationType.Q4_0_4_8:(32, 2 + 16),
+    GGMLQuantizationType.Q4_0_8_8:(32, 2 + 16),
 }
 
 
@@ -1327,3 +1337,4 @@ KEY_TOKENIZER_PRIFIX_ID  = Keys.Tokenizer.PREFIX_ID
 KEY_TOKENIZER_SUFFIX_ID  = Keys.Tokenizer.SUFFIX_ID
 KEY_TOKENIZER_MIDDLE_ID  = Keys.Tokenizer.MIDDLE_ID
 KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID
+KEY_TOKENIZER_EOM_ID     = Keys.Tokenizer.EOM_ID

gguf-py/gguf/gguf_writer.py

@@ -828,6 +828,9 @@ class GGUFWriter:
     def add_eot_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOT_ID, id)
 
+    def add_eom_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.EOM_ID, id)
+
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:

gguf-py/gguf/lazy.py

@@ -191,6 +191,8 @@ class LazyBase(ABC, metaclass=LazyMeta):
 class LazyNumpyTensor(LazyBase):
     _tensor_type = np.ndarray
 
+    shape: tuple[int, ...]  # Makes the type checker happy in quants.py
+
     @classmethod
     def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: tuple[int, ...]) -> np.ndarray[Any, Any]:
         # The initial idea was to use np.nan as the fill value,

gguf-py/gguf/metadata.py

@@ -174,7 +174,7 @@ class Metadata:
             org_component, model_full_name_component = None, model_id
 
         # Check if we erroneously matched against './' or '../' etc...
-        if org_component is not None and org_component[0] == '.':
+        if org_component is not None and len(org_component) > 0 and org_component[0] == '.':
             org_component = None
 
         name_parts: list[str] = model_full_name_component.split('-')
@@ -284,20 +284,67 @@ class Metadata:
         ########################
         if model_card is not None:
 
-            if "model_name" in model_card and metadata.name is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.name = model_card.get("model_name")
+            def use_model_card_metadata(metadata_key: str, model_card_key: str):
+                if model_card_key in model_card and getattr(metadata, metadata_key, None) is None:
+                    setattr(metadata, metadata_key, model_card.get(model_card_key))
 
-            if "model_creator" in model_card and metadata.author is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.author = model_card.get("model_creator")
+            def use_array_model_card_metadata(metadata_key: str, model_card_key: str):
+                # Note: Will append rather than replace if already exist
+                tags_value = model_card.get(model_card_key, None)
+                if tags_value is None:
+                    return
 
-            if "model_type" in model_card and metadata.basename is None:
-                # Not part of huggingface model card standard but notice some model creator using it
-                # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
-                metadata.basename = model_card.get("model_type")
+                current_value = getattr(metadata, metadata_key, None)
+                if current_value is None:
+                    current_value = []
+
+                if isinstance(tags_value, str):
+                    current_value.append(tags_value)
+                elif isinstance(tags_value, list):
+                    current_value.extend(tags_value)
+
+                setattr(metadata, metadata_key, current_value)
+
+            # LLAMA.cpp's direct internal convention
+            # (Definitely not part of hugging face formal/informal standard)
+            #########################################
+            use_model_card_metadata("name", "name")
+            use_model_card_metadata("author", "author")
+            use_model_card_metadata("version", "version")
+            use_model_card_metadata("organization", "organization")
+            use_model_card_metadata("description", "description")
+            use_model_card_metadata("finetune", "finetune")
+            use_model_card_metadata("basename", "basename")
+            use_model_card_metadata("size_label", "size_label")
+            use_model_card_metadata("source_url", "url")
+            use_model_card_metadata("source_doi", "doi")
+            use_model_card_metadata("source_uuid", "uuid")
+            use_model_card_metadata("source_repo_url", "repo_url")
+
+            # LLAMA.cpp's huggingface style convention
+            # (Definitely not part of hugging face formal/informal standard... but with model_ appended to match their style)
+            ###########################################
+            use_model_card_metadata("name", "model_name")
+            use_model_card_metadata("author", "model_author")
+            use_model_card_metadata("version", "model_version")
+            use_model_card_metadata("organization", "model_organization")
+            use_model_card_metadata("description", "model_description")
+            use_model_card_metadata("finetune", "model_finetune")
+            use_model_card_metadata("basename", "model_basename")
+            use_model_card_metadata("size_label", "model_size_label")
+            use_model_card_metadata("source_url", "model_url")
+            use_model_card_metadata("source_doi", "model_doi")
+            use_model_card_metadata("source_uuid", "model_uuid")
+            use_model_card_metadata("source_repo_url", "model_repo_url")
+
+            # Hugging Face Direct Convention
+            #################################
+
+            # Not part of huggingface model card standard but notice some model creator using it
+            # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
+            use_model_card_metadata("name", "model_name")
+            use_model_card_metadata("author", "model_creator")
+            use_model_card_metadata("basename", "model_type")
 
             if "base_model" in model_card:
                 # This represents the parent models that this is based on
@@ -329,58 +376,18 @@ class Metadata:
                         base_model["repo_url"] = f"https://huggingface.co/{org_component}/{model_full_name_component}"
                     metadata.base_models.append(base_model)
 
-            if "license" in model_card and metadata.license is None:
-                metadata.license = model_card.get("license")
+            use_model_card_metadata("license", "license")
+            use_model_card_metadata("license_name", "license_name")
+            use_model_card_metadata("license_link", "license_link")
 
-            if "license_name" in model_card and metadata.license_name is None:
-                metadata.license_name = model_card.get("license_name")
+            use_array_model_card_metadata("tags", "tags")
+            use_array_model_card_metadata("tags", "pipeline_tag")
 
-            if "license_link" in model_card and metadata.license_link is None:
-                metadata.license_link = model_card.get("license_link")
+            use_array_model_card_metadata("languages", "languages")
+            use_array_model_card_metadata("languages", "language")
 
-            tags_value = model_card.get("tags", None)
-            if tags_value is not None:
-
-                if metadata.tags is None:
-                    metadata.tags = []
-
-                if isinstance(tags_value, str):
-                    metadata.tags.append(tags_value)
-                elif isinstance(tags_value, list):
-                    metadata.tags.extend(tags_value)
-
-            pipeline_tags_value = model_card.get("pipeline_tag", None)
-            if pipeline_tags_value is not None:
-
-                if metadata.tags is None:
-                    metadata.tags = []
-
-                if isinstance(pipeline_tags_value, str):
-                    metadata.tags.append(pipeline_tags_value)
-                elif isinstance(pipeline_tags_value, list):
-                    metadata.tags.extend(pipeline_tags_value)
-
-            language_value = model_card.get("languages", model_card.get("language", None))
-            if language_value is not None:
-
-                if metadata.languages is None:
-                    metadata.languages = []
-
-                if isinstance(language_value, str):
-                    metadata.languages.append(language_value)
-                elif isinstance(language_value, list):
-                    metadata.languages.extend(language_value)
-
-            dataset_value = model_card.get("datasets", model_card.get("dataset", None))
-            if dataset_value is not None:
-
-                if metadata.datasets is None:
-                    metadata.datasets = []
-
-                if isinstance(dataset_value, str):
-                    metadata.datasets.append(dataset_value)
-                elif isinstance(dataset_value, list):
-                    metadata.datasets.extend(dataset_value)
+            use_array_model_card_metadata("datasets", "datasets")
+            use_array_model_card_metadata("datasets", "dataset")
 
         # Hugging Face Parameter Heuristics
         ####################################

gguf-py/tests/test_quants.py

@@ -0,0 +1,237 @@
+#!/usr/bin/env python3
+
+# Test gguf.quants so that it exactly matches the C implementation of the (de)quantization
+
+# NOTE: this is kind of a mess, but at least it worked for initially testing the Python implementations.
+
+from __future__ import annotations
+
+import argparse
+from math import prod
+import os
+import sys
+from pathlib import Path
+import ctypes
+import logging
+import numpy as np
+
+# Necessary to load the local gguf package
+if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
+    sys.path.insert(0, str(Path(__file__).parent.parent))
+
+import gguf
+from gguf.constants import GGMLQuantizationType
+
+
+logger = logging.getLogger("test-quants")
+
+
+c_float_p = ctypes.POINTER(ctypes.c_float)
+
+
+class ggml_init_params(ctypes.Structure):
+    _fields_ = [
+        ("mem_size", ctypes.c_size_t),
+        ("mem_buffer", ctypes.c_void_p),
+        ("no_alloc", ctypes.c_bool),
+    ]
+
+
+class GGMLQuants:
+    libggml: ctypes.CDLL
+
+    def __init__(self, libggml: Path):
+        self.libggml = ctypes.CDLL(str(libggml))
+        self.libggml.ggml_quantize_chunk.restype = ctypes.c_size_t
+        # enum ggml_type   type,
+        #    const float * src,
+        #           void * dst,
+        #        int64_t   start,
+        #        int64_t   nrows,
+        #        int64_t   n_per_row,
+        #    const float * imatrix) {
+        self.libggml.ggml_quantize_chunk.argtypes = (
+            ctypes.c_int,
+            ctypes.POINTER(ctypes.c_float),
+            ctypes.c_void_p,
+            ctypes.c_int64,
+            ctypes.c_int64,
+            ctypes.c_int64,
+            ctypes.POINTER(ctypes.c_float),
+        )
+
+        self.libggml.ggml_quantize_requires_imatrix.restype = ctypes.c_bool
+        self.libggml.ggml_quantize_requires_imatrix.argtypes = (ctypes.c_int,)
+
+        for t in (
+            "q4_0", "q4_1", "q5_0", "q5_1", "q8_0",
+            "q2_K", "q3_K", "q4_K", "q5_K", "q6_K",
+            "iq2_xxs", "iq2_xs", "iq2_s", "iq3_xxs", "iq3_s", "iq1_s", "iq1_m",
+            "iq4_nl", "iq4_xs",
+        ):
+            dequant_func: ctypes._NamedFuncPointer = getattr(self.libggml, "dequantize_row_" + t)
+            dequant_func.restype = None
+            dequant_func.argtypes = (ctypes.c_void_p, ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
+
+        self.libggml.ggml_fp16_to_fp32_row.restype = None
+        self.libggml.ggml_fp16_to_fp32_row.argtypes = (ctypes.POINTER(ctypes.c_uint16), ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
+        self.libggml.ggml_bf16_to_fp32_row.restype = None
+        self.libggml.ggml_bf16_to_fp32_row.argtypes = (ctypes.POINTER(ctypes.c_uint16), ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
+
+        self.libggml.ggml_init.argtypes = (ggml_init_params,)
+
+        self.libggml.ggml_init(ggml_init_params(1 * 1024 * 1024, 0, False))
+
+    def dequantize(self, tensor: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
+        result = np.zeros(gguf.quant_shape_from_byte_shape(tensor.shape, qtype), dtype=np.float32, order="C")
+        if qtype == GGMLQuantizationType.F32:
+            # no-op
+            result = tensor.view(np.float32)
+        elif qtype == GGMLQuantizationType.F16:
+            self.libggml.ggml_fp16_to_fp32_row(tensor.ctypes.data_as(ctypes.POINTER(ctypes.c_uint16)), result.ctypes.data_as(c_float_p), result.size)
+        elif qtype == GGMLQuantizationType.BF16:
+            self.libggml.ggml_bf16_to_fp32_row(tensor.ctypes.data_as(ctypes.POINTER(ctypes.c_uint16)), result.ctypes.data_as(c_float_p), result.size)
+        else:
+            lw_qname = qtype.name.lower()
+            if lw_qname[-1] == "k":
+                lw_qname = lw_qname[:-1] + "K"
+            dequant_func: ctypes._NamedFuncPointer = getattr(self.libggml, "dequantize_row_" + lw_qname)
+            dequant_func(tensor.ctypes.data_as(ctypes.c_void_p), result.ctypes.data_as(c_float_p), result.size)
+        return result
+
+    def quantize(self, data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
+        result = np.zeros(gguf.quant_shape_to_byte_shape(data.shape, qtype), dtype=np.uint8, order="C")
+        if self.libggml.ggml_quantize_requires_imatrix(qtype.value):
+            # TODO: is a column-wise sum of squares appropriate?
+            qw = np.sum((data * data).reshape((-1, data.shape[-1])), axis=0).ctypes.data_as(c_float_p)
+        else:
+            qw = ctypes.cast(0, c_float_p)
+        result_size = self.libggml.ggml_quantize_chunk(qtype.value, data.ctypes.data_as(c_float_p), result.ctypes.data_as(ctypes.c_void_p), 0, prod(data.shape[:-1]), data.shape[-1], qw)
+        assert result.size == result_size
+        return result
+
+
+def compare_tensors(t1: np.ndarray, t2: np.ndarray, qtype: GGMLQuantizationType) -> bool:
+    same = np.array_equal(t1, t2)
+    if same:
+        return True
+    else:
+        block_size, type_size = gguf.GGML_QUANT_SIZES[qtype]
+        if t1.dtype == np.float32:
+            t1 = t1.reshape((-1, block_size))
+            t2 = t2.reshape((-1, block_size))
+        else:
+            t1 = t1.reshape((-1, type_size))
+            t2 = t2.reshape((-1, type_size))
+        x = t1.view(np.uint8) ^ t2.view(np.uint8)
+        diff_bits = np.count_nonzero(np.unpackbits(x, axis=-1), axis=-1)
+        num_bad_blocks = np.count_nonzero(diff_bits, axis=0)
+        if num_bad_blocks == 0 and t1.shape == t2.shape:
+            logger.debug("Bits are equal, but arrays don't match, likely contains NANs")
+            return True
+        logger.debug(f"{num_bad_blocks} bad blocks ({100 * num_bad_blocks / x.shape[0]:.6f}%)")
+        bad_block_id = np.argmax(diff_bits, axis=0)
+        logger.debug(f"Worst block id: {bad_block_id}")
+        logger.debug(f"Sample bad block ({diff_bits[bad_block_id]} differing bits):\n{t1[bad_block_id]}\nReference:\n{t2[bad_block_id]}")
+
+        sum_diff_bits = np.sum(diff_bits)
+        logger.debug(f"{sum_diff_bits} bits differ ({100 * sum_diff_bits/(x.size * 8):.6f}%)")
+        return False
+
+
+def do_test(libggml_path: Path, quick: bool = False):
+    ggml_quants = GGMLQuants(libggml_path)
+
+    np.set_printoptions(precision=None, threshold=(4 * 256) + 1, formatter={"int": lambda n: "0x%02X" % n})
+
+    r = np.random.randn(8, 1024, 1024).astype(np.float32, copy=False)
+
+    for qtype in (GGMLQuantizationType.F16, *gguf.quants._type_traits.keys()):
+        has_dequantize = False
+        has_quantize = False
+
+        try:
+            gguf.dequantize(np.zeros((gguf.GGML_QUANT_SIZES[qtype][1]), dtype=np.uint8), qtype)
+            has_dequantize = True
+        except (NotImplementedError, AssertionError) as e:
+            if isinstance(e, AssertionError):
+                logger.error(f"Error with {qtype.name}: {e}")
+                raise e
+        try:
+            gguf.quantize(np.zeros((gguf.GGML_QUANT_SIZES[qtype][0]), dtype=np.float32), qtype)
+            has_quantize = True
+        except (NotImplementedError, AssertionError) as e:
+            if isinstance(e, AssertionError):
+                logger.error(f"Error with {qtype.name}: {e}")
+                raise e
+
+        if not has_dequantize and not has_quantize:
+            continue
+
+        logger.info(f"Testing {qtype.name}")
+
+        rc = r.copy(order="C")
+
+        pyq = None
+        ggq = None
+
+        if has_quantize:
+            logger.debug(f"Quantizing to {qtype.name} with Python")
+            pyq = gguf.quants.quantize(rc, qtype)
+
+            logger.debug(f"Quantizing to {qtype.name} with C")
+            ggq = ggml_quants.quantize(rc, qtype)
+
+            if qtype == GGMLQuantizationType.F16:
+                pyq = pyq.view(np.uint8)
+            quant_equal = compare_tensors(pyq, ggq, qtype)
+
+            if not quant_equal:
+                logger.error(f"Quantization to {qtype.name} does not match ❌")
+            else:
+                logger.info(f"Quantization to {qtype.name} matches exactly ✅")
+
+        if has_dequantize:
+            if ggq is None and not quick:
+                logger.debug(f"Quantizing to {qtype.name} with C")
+                ggq = ggml_quants.quantize(rc, qtype)
+
+            if ggq is not None:
+                logger.debug(f"Dequantizing from {qtype.name} with Python")
+                pydq = gguf.quants.dequantize(ggq, qtype)
+                logger.debug(f"Dequantizing from {qtype.name} with C")
+                ggdq = ggml_quants.dequantize(ggq, qtype)
+
+                dequant_equal = compare_tensors(pydq, ggdq, qtype)
+
+                if not dequant_equal:
+                    logger.error(f"Dequantization from {qtype.name} does not match ❌")
+                else:
+                    logger.info(f"Dequantization from {qtype.name} matches exactly ✅")
+
+            rq_shape = gguf.quants.quant_shape_to_byte_shape((8, 1024, 1024 // 2), qtype)
+            rq = np.random.random(rq_shape).astype(np.float16).view(np.uint8)
+
+            logger.debug(f"Dequantizing random f16 data as {qtype.name} with Python")
+            pydq = gguf.quants.dequantize(rq, qtype)
+            logger.debug(f"Dequantizing random f16 data as {qtype.name} with C")
+            ggdq = ggml_quants.dequantize(rq, qtype)
+
+            dequant_equal = compare_tensors(pydq, ggdq, qtype)
+
+            if not dequant_equal:
+                logger.error(f"Dequantization from random f16 data as {qtype.name} does not match ❌")
+            else:
+                logger.info(f"Dequantization from random f16 data as {qtype.name} matches exactly ✅")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Test Python (de)quantization against the reference C implementation")
+    parser.add_argument("--libggml", type=Path, default=Path(__file__).parent.parent.parent / "build" / "ggml" / "src" / "libggml.so", help="The path to libggml.so")
+    parser.add_argument("--quick", action="store_true", help="Don't quantize with C when it's not strictly necessary")
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.DEBUG)
+
+    do_test(args.libggml, args.quick)

include/llama.h

@@ -345,7 +345,7 @@ extern "C" {
         int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
         enum llama_ftype ftype;              // quantize to this llama_ftype
         enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // itoken embeddings tensor type
+        enum ggml_type token_embedding_type; // token embeddings tensor type
         bool allow_requantize;               // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor;         // quantize output.weight
         bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored

scripts/sync-ggml-am.sh

@@ -62,8 +62,10 @@ while read c; do
         src/ggml*.m \
         src/ggml*.metal \
         src/ggml*.cu \
+        src/ggml-cann/* \
         src/ggml-cuda/* \
         src/ggml-sycl/* \
+        src/vulkan-shaders/* \
         include/ggml*.h \
         tests/test-opt.cpp \
         tests/test-grad0.cpp \
@@ -107,6 +109,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-alloc.c        -> ggml/src/ggml-alloc.c
     # src/ggml-backend-impl.h -> ggml/src/ggml-backend-impl.h
     # src/ggml-backend.c      -> ggml/src/ggml-backend.c
+    # src/ggml-cann/*         -> ggml/src/ggml-cann/
+    # src/ggml-cann.cpp       -> ggml/src/ggml-cann.cpp
     # src/ggml-common.h       -> ggml/src/ggml-common.h
     # src/ggml-cuda/*         -> ggml/src/ggml-cuda/
     # src/ggml-cuda.cu        -> ggml/src/ggml-cuda.cu
@@ -125,6 +129,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # include/ggml-alloc.h   -> ggml/include/ggml-alloc.h
     # include/ggml-backend.h -> ggml/include/ggml-backend.h
     # include/ggml-blas.h    -> ggml/include/ggml-blas.h
+    # include/ggml-cann.h    -> ggml/include/ggml-cann.h
     # include/ggml-cuda.h    -> ggml/include/ggml-cuda.h
     # include/ggml-kompute.h -> ggml/include/ggml-kompute.h
     # include/ggml-metal.h   -> ggml/include/ggml-metal.h
@@ -151,6 +156,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)src\/ggml-alloc\.c/\1ggml\/src\/ggml-alloc.c/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-backend-impl\.h/\1ggml\/src\/ggml-backend-impl.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-backend\.c/\1ggml\/src\/ggml-backend.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\.cpp/\1ggml\/src\/ggml-cann.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-common\.h/\1ggml\/src\/ggml-common.h/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\.cu/\1ggml\/src\/ggml-cuda.cu/g' \
@@ -168,18 +175,19 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)include\/ggml-alloc\.h/\1ggml\/include\/ggml-alloc.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-backend\.h/\1ggml\/include\/ggml-backend.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-blas\.h/\1ggml\/include\/ggml-blas.h/g' \
+        -e 's/([[:space:]]|[ab]\/)include\/ggml-cann\.h/\1ggml\/include\/ggml-cann.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-cuda\.h/\1ggml\/include\/ggml-cuda.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-kompute\.h/\1ggml\/include\/ggml-kompute.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-metal\.h/\1ggml\/include\/ggml-metal.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-rpc\.h/\1ggml\/include\/ggml-rpc.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-sycl\.h/\1ggml\/include\/ggml-sycl.h/g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml-vulkan\.h/\1ggml\/include\/ggml-vulkan.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common\.h/examples\/common.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common\.cpp/examples\/common.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.h/examples\/common-ggml.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.cpp/examples\/common-ggml.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)LICENSE/LICENSE/g' \
-        -e 's/([[:space:]]|[ab]\/)scripts\/gen-authors\.sh/scripts\/gen-authors.sh/g' \
+        -e 's/([[:space:]]|[ab]\/)examples\/common\.h/\1examples\/common.h/g' \
+        -e 's/([[:space:]]|[ab]\/)examples\/common\.cpp/\1examples\/common.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.h/\1examples\/common-ggml.h/g' \
+        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.cpp/\1examples\/common-ggml.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)LICENSE/\1LICENSE/g' \
+        -e 's/([[:space:]]|[ab]\/)scripts\/gen-authors\.sh/\1scripts\/gen-authors.sh/g' \
         > ggml-src.patch.tmp
     mv ggml-src.patch.tmp ggml-src.patch
 

scripts/sync-ggml.last

@@ -1 +1 @@
-31d544f87835a55602883fe09156bb85a4c163d8
+6c71d5a071d842118fb04c03c4b15116dff09621

scripts/sync-ggml.sh

@@ -10,6 +10,8 @@ cp -rpv ../ggml/src/ggml-aarch64.h      ./ggml/src/ggml-aarch64.h
 cp -rpv ../ggml/src/ggml-alloc.c        ./ggml/src/ggml-alloc.c
 cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml/src/ggml-backend-impl.h
 cp -rpv ../ggml/src/ggml-backend.c      ./ggml/src/ggml-backend.c
+cp -rpv ../ggml/src/ggml-cann/*         ./ggml/src/ggml-cann/
+cp -rpv ../ggml/src/ggml-cann.cpp       ./ggml/src/ggml-cann.cpp
 cp -rpv ../ggml/src/ggml-common.h       ./ggml/src/ggml-common.h
 cp -rpv ../ggml/src/ggml-cuda/*         ./ggml/src/ggml-cuda/
 cp -rpv ../ggml/src/ggml-cuda.cu        ./ggml/src/ggml-cuda.cu
@@ -29,6 +31,7 @@ cp -rpv ../ggml/include/ggml.h         ./ggml/include/ggml.h
 cp -rpv ../ggml/include/ggml-alloc.h   ./ggml/include/ggml-alloc.h
 cp -rpv ../ggml/include/ggml-backend.h ./ggml/include/ggml-backend.h
 cp -rpv ../ggml/include/ggml-blas.h    ./ggml/include/ggml-blas.h
+cp -rpv ../ggml/include/ggml-cann.h    ./ggml/include/ggml-cann.h
 cp -rpv ../ggml/include/ggml-cuda.h    ./ggml/include/ggml-cuda.h
 cp -rpv ../ggml/include/ggml-kompute.h ./ggml/include/ggml-kompute.h
 cp -rpv ../ggml/include/ggml-metal.h   ./ggml/include/ggml-metal.h

src/llama-vocab.cpp

@@ -816,6 +816,9 @@ struct llm_tokenizer_ugm {
      * the best tokenization.
     */
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+        // get current size of output (for reversal later)
+        size_t output_size = output.size();
+
         // normalize the input first
         std::string normalized;
         normalize(text, &normalized);
@@ -895,7 +898,7 @@ struct llm_tokenizer_ugm {
         }
 
         // reverse the output since we added tokens starting from the end of the input
-        std::reverse(output.begin(), output.end());
+        std::reverse(output.begin() + output_size, output.end());
     }
 
 private:
@@ -1444,7 +1447,8 @@ llama_token_attr llama_token_get_attr_impl(const struct llama_vocab & vocab, lla
 bool llama_token_is_eog_impl(const struct llama_vocab & vocab, llama_token token) {
     return token != -1 && (
         token == llama_token_eos_impl(vocab) ||
-        token == llama_token_eot_impl(vocab)
+        token == llama_token_eot_impl(vocab) ||
+        token == llama_token_eom_impl(vocab)
     );
 }
 
@@ -1500,6 +1504,10 @@ llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
     return vocab.special_eot_id;
 }
 
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eom_id;
+}
+
 int32_t llama_tokenize_impl(
     const struct llama_vocab & vocab,
                   const char * text,

src/llama-vocab.h

@@ -45,6 +45,7 @@ struct llama_vocab {
     id special_suffix_id = -1;
     id special_middle_id = -1;
     id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
+    id special_eom_id    = -1;
 
     // tokenizer flags
     bool tokenizer_add_space_prefix = false;
@@ -101,6 +102,7 @@ llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_eot_impl   (const struct llama_vocab & vocab);
+llama_token llama_token_eom_impl   (const struct llama_vocab & vocab);
 
 int32_t llama_tokenize_impl(
         const struct llama_vocab & vocab,

src/llama.cpp

@@ -122,17 +122,14 @@ static std::string trim(const std::string & str) {
 }
 
 static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
+    if (search.empty()) {
+        return; // Avoid infinite loop if 'search' is an empty string
+    }
+    size_t pos = 0;
+    while ((pos = s.find(search, pos)) != std::string::npos) {
+        s.replace(pos, search.length(), replace);
+        pos += replace.length();
     }
-    s = std::move(result);
 }
 
 static bool is_float_close(float a, float b, float abs_tol) {
@@ -362,6 +359,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_SUFFIX_ID,
     LLM_KV_TOKENIZER_MIDDLE_ID,
     LLM_KV_TOKENIZER_EOT_ID,
+    LLM_KV_TOKENIZER_EOM_ID,
 
     LLM_KV_ADAPTER_TYPE,
     LLM_KV_ADAPTER_LORA_ALPHA,
@@ -459,6 +457,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_SUFFIX_ID,            "tokenizer.ggml.suffix_token_id"          },
     { LLM_KV_TOKENIZER_MIDDLE_ID,            "tokenizer.ggml.middle_token_id"          },
     { LLM_KV_TOKENIZER_EOT_ID,               "tokenizer.ggml.eot_token_id"             },
+    { LLM_KV_TOKENIZER_EOM_ID,               "tokenizer.ggml.eom_token_id"             },
 
     { LLM_KV_ADAPTER_TYPE,                  "adapter.type"       },
     { LLM_KV_ADAPTER_LORA_ALPHA,            "adapter.lora.alpha" },
@@ -5586,6 +5585,7 @@ static void llm_load_vocab(
             { LLM_KV_TOKENIZER_SUFFIX_ID, vocab.special_suffix_id },
             { LLM_KV_TOKENIZER_MIDDLE_ID, vocab.special_middle_id },
             { LLM_KV_TOKENIZER_EOT_ID,    vocab.special_eot_id    },
+            { LLM_KV_TOKENIZER_EOM_ID,    vocab.special_eom_id    },
         };
 
         for (const auto & it : special_token_types) {
@@ -5638,6 +5638,17 @@ static void llm_load_vocab(
                 }
             }
         }
+
+        // find EOM token: "<|eom_id|>"
+        //
+        // TODO: convert scripts should provide this token through the KV metadata LLAMA_KV_TOKENIZER_EOM_ID
+        //       for now, we apply this workaround to find the EOM token based on its text
+        if (vocab.special_eom_id == -1) {
+            const auto & t = vocab.token_to_id.find("<|eom_id|>");
+            if (t != vocab.token_to_id.end()) {
+                vocab.special_eom_id = t->second;
+            }
+        }
     }
 
     // build special tokens cache

tests/test-backend-ops.cpp

@@ -1511,6 +1511,7 @@ struct test_group_norm : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const int32_t num_groups;
+    const float eps;
 
     std::string vars() override {
         return VARS_TO_STR3(type, ne, num_groups);
@@ -1518,12 +1519,13 @@ struct test_group_norm : public test_case {
 
     test_group_norm(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 64, 320, 1},
-            int32_t num_groups = 32)
-        : type(type), ne(ne), num_groups(num_groups) {}
+            int32_t num_groups = 32,
+            float eps = 1e-6f)
+        : type(type), ne(ne), num_groups(num_groups), eps(eps) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * out = ggml_group_norm(ctx, a, num_groups);
+        ggml_tensor * out = ggml_group_norm(ctx, a, num_groups, eps);
         return out;
     }
 };
@@ -2271,9 +2273,10 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     for (ggml_type type_a : other_types) {
         for (ggml_type type_b : {GGML_TYPE_F32}) {
-
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, ggml_blck_size(type_a), { 1,  1}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1,  1}, {1, 1}));
+            if (ggml_blck_size(type_a) != 256) {
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, ggml_blck_size(type_a), {1,  1}, {1, 1}));
+            }
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1,  1}, {1, 1}));
         }
     }