server : fix duplicated n_predict key in the generation_settings (#8994)

Signed-off-by: Jiri Podivin <jpodivin@redhat.com>

.github/workflows/bench.yml.disabled

@@ -1,3 +1,6 @@
+# TODO: there have been some issues with the workflow, so disabling for now
+#       https://github.com/ggerganov/llama.cpp/issues/7893
+#
 # Benchmark
 name: Benchmark
 
@@ -129,6 +132,8 @@ jobs:
 
       - name: Server bench
         id: server_bench
+        env:
+            HEAD_REF: ${{ github.head_ref || github.ref_name }}
         run: |
           set -eux
 
@@ -137,7 +142,7 @@ jobs:
           python bench.py \
               --runner-label ${{ env.RUNNER_LABEL }} \
               --name ${{ github.job }} \
-              --branch ${{ github.head_ref || github.ref_name }} \
+              --branch $HEAD_REF \
               --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
               --scenario script.js \
               --duration ${{ github.event.inputs.duration || env.DURATION }} \

.github/workflows/build.yml

@@ -47,7 +47,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -105,7 +105,7 @@ jobs:
           sysctl -a
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -222,7 +222,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -696,22 +696,20 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'rpc-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'llvm-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'msvc-arm64'

.github/workflows/python-check-requirements.yml

@@ -6,15 +6,13 @@ on:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
+      - '**/requirements*.txt'
   pull_request:
     paths:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
+      - '**/requirements*.txt'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

Makefile

@@ -763,6 +763,10 @@ ifdef GGML_VULKAN_MEMORY_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
 endif
 
+ifdef GGML_VULKAN_PERF
+	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
+endif
+
 ifdef GGML_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
@@ -1454,26 +1458,20 @@ libllava.a: examples/llava/llava.cpp \
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
 
 llama-llava-cli: examples/llava/llava-cli.cpp \
-	examples/llava/clip.h \
-	examples/llava/clip.cpp \
-	examples/llava/llava.h \
 	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
-	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
 llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
-	examples/llava/clip.h \
-	examples/llava/clip.cpp \
-	examples/llava/llava.h \
 	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
-	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift

README.md

@@ -186,10 +186,12 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
+- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 
 **Infrastructure:**
 
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
+- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 
 **Games:**
 - [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.

common/common.cpp

@@ -1777,6 +1777,17 @@ std::string string_get_sortable_timestamp() {
     return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
 }
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    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();
+    }
+}
+
 void string_process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -2145,7 +2156,9 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
             tmp.clear();
             tmp.push_back(decoder_start_token_id);
         }
-        llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        if (llama_model_has_decoder(model)) {
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
         llama_reset_timings(lctx);
@@ -2689,12 +2702,6 @@ std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token>
     return text;
 }
 
-bool llama_should_add_bos_token(const llama_model * model) {
-    const int add_bos = llama_add_bos_token(model);
-
-    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
-}
-
 //
 // Chat template utils
 //

common/common.h

@@ -286,6 +286,8 @@ std::vector<std::string> string_split(std::string input, char separator);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
+
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     std::vector<T> values;
@@ -378,10 +380,6 @@ std::string llama_detokenize(
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
-
 //
 // Chat template utils
 //

common/grammar-parser.cpp

@@ -369,6 +369,9 @@ namespace grammar_parser {
             }
             // Validate the state to ensure that all rules are defined
             for (const auto & rule : state.rules) {
+                if (rule.empty()) {
+                    throw std::runtime_error("Undefined rule");
+                }
                 for (const auto & elem : rule) {
                     if (elem.type == LLAMA_GRETYPE_RULE_REF) {
                         // Ensure that the rule at that location exists

convert_hf_to_gguf.py

@@ -590,6 +590,12 @@ class Model:
         if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249":
             # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M
             res = "smollm"
+        if chkhsh == "3c30d3ad1d6b64202cd222813e7736c2db6e1bd6d67197090fc1211fbc612ae7":
+            # ref: https://huggingface.co/bigscience/bloom
+            res = "bloom"
+        if chkhsh == "bc01ce58980e1db43859146dc51b1758b3b88729b217a74792e9f8d43e479d21":
+            # ref: https://huggingface.co/TurkuNLP/gpt3-finnish-small
+            res = "gpt3-finnish"
 
         if res is None:
             logger.warning("\n")
@@ -893,7 +899,7 @@ class GPTNeoXModel(Model):
         return tensors
 
 
-@Model.register("BloomForCausalLM")
+@Model.register("BloomForCausalLM", "BloomModel")
 class BloomModel(Model):
     model_arch = gguf.MODEL_ARCH.BLOOM
 
@@ -3324,6 +3330,145 @@ class T5Model(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("T5EncoderModel")
+class T5EncoderModel(Model):
+    model_arch = gguf.MODEL_ARCH.T5ENCODER
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.shared_token_embeddings_found = False
+
+    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 / 'tokenizer.model'
+
+        # many older models use spiece.model tokenizer model filename
+        if not tokenizer_path.is_file():
+            tokenizer_path = self.dir_model / 'spiece.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())
+
+        # some models like Pile-T5 family use BPE tokenizer instead of Unigram
+        if sentencepiece_model.trainer_spec.model_type == 2:  # BPE
+            # assure the tokenizer model file name is correct
+            assert tokenizer_path.name == 'tokenizer.model'
+            return self._set_vocab_sentencepiece()
+        else:
+            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
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+                for key in added_tokens_json:
+                    token_id = added_tokens_json[key]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+        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)
+
+        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_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(False)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def set_gguf_parameters(self):
+        if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
+            logger.warning("Couldn't find context length in config.json, assuming default value of 512")
+            n_ctx = 512
+        self.gguf_writer.add_context_length(n_ctx)
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(self.hparams["num_heads"])
+        self.gguf_writer.add_key_length(self.hparams["d_kv"])
+        self.gguf_writer.add_value_length(self.hparams["d_kv"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_relative_attn_buckets_count(self.hparams["relative_attention_num_buckets"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight",
+        # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored
+        # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder
+        # and decoder and ignore the remaining ones.
+        if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]:
+            if not self.shared_token_embeddings_found:
+                name = "shared.weight"
+                self.shared_token_embeddings_found = True
+            else:
+                logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.")
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @Model.register("JAISLMHeadModel")
 class JaisModel(Model):
     model_arch = gguf.MODEL_ARCH.JAIS

convert_hf_to_gguf_update.py

@@ -94,6 +94,8 @@ models = [
     {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
     {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
     {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+    {'name': "bloom",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
+    {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
 ]
 
 

docs/backend/SYCL.md

@@ -80,7 +80,14 @@ The following release is verified with good quality:
 
 ### Intel GPU
 
-**Verified devices**
+SYCL backend supports Intel GPU Family:
+
+- Intel Data Center Max Series
+- Intel Flex Series, Arc Series
+- Intel Built-in Arc GPU
+- Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)).
+
+#### Verified devices
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
@@ -88,7 +95,7 @@ The following release is verified with good quality:
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
-| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
+| Intel iGPU                    | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -237,6 +244,13 @@ Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA devic
 ### II. Build llama.cpp
 
 #### Intel GPU
+
+```
+./examples/sycl/build.sh
+```
+
+or
+
 ```sh
 # Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
@@ -276,23 +290,26 @@ cmake --build build --config Release -j -v
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
 You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 ```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/llama-ls-sycl-device
 ```
+
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
@@ -304,12 +321,37 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 ```
 
+#### Choose level-zero devices
 
-4. Launch inference
+|Chosen Device ID|Setting|
+|-|-|
+|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+- Use device 0:
+
+```sh
+./examples/sycl/run_llama2.sh 0
+```
+- Use multiple devices:
+
+```sh
+./examples/sycl/run_llama2.sh
+```
+
+2. Command line
+Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device target specified by the user.
+- Single device: Use one device assigned by user. Default device id is 0.
 - Multiple devices: Automatically choose the devices with the same backend.
 
 In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
@@ -326,11 +368,6 @@ Examples:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
-or run by script:
-
-```sh
-./examples/sycl/run_llama2.sh 0
-```
 
 - Use multiple devices:
 
@@ -338,12 +375,6 @@ or run by script:
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
-Otherwise, you can run the script:
-
-```sh
-./examples/sycl/run_llama2.sh
-```
-
 *Notes:*
 
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
@@ -390,7 +421,7 @@ c. Verify installation
 In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
-sycl-ls
+sycl-ls.exe
 ```
 
 There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
@@ -411,6 +442,18 @@ b. The new Visual Studio will install Ninja as default. (If not, please install
 
 ### II. Build llama.cpp
 
+You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
+
+Choose one of following methods to build from source code.
+
+1. Script
+
+```sh
+.\examples\sycl\win-build-sycl.bat
+```
+
+2. CMake
+
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
@@ -425,12 +468,8 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI
 cmake --build build --config Release -j
 ```
 
-Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
-```sh
-.\examples\sycl\win-build-sycl.bat
-```
-
 Or, use CMake presets to build:
+
 ```sh
 cmake --preset x64-windows-sycl-release
 cmake --build build-x64-windows-sycl-release -j --target llama-cli
@@ -442,7 +481,9 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+3. Visual Studio
+
+You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
 
 *Notes:*
 
@@ -450,23 +491,25 @@ Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choos
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 On the oneAPI command line window, run the following and step into the llama.cpp directory:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
-build\bin\ls-sycl-device.exe
+build\bin\llama-ls-sycl-device.exe
 ```
 
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
@@ -479,9 +522,27 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 
 ```
+#### Choose level-zero devices
 
+|Chosen Device ID|Setting|
+|-|-|
+|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
 
-4. Launch inference
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+```
+examples\sycl\win-run-llama2.bat
+```
+
+2. Command line
+
+Launch inference
 
 There are two device selection modes:
 
@@ -508,11 +569,7 @@ build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website ca
 ```
 build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
-Otherwise, run the following wrapper script:
 
-```
-.\examples\sycl\win-run-llama2.bat
-```
 
 Note:
 
@@ -526,17 +583,18 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
+
 ## Environment Variable
 
 #### Build
 
 | Name               | Value                             | Function                                    |
 |--------------------|-----------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
+| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
-| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
-| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
+| CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
+| CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
 #### Runtime
 
@@ -572,9 +630,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
   Otherwise, please double-check the GPU driver installation steps.
 
+- Can I report Ollama issue on Intel GPU to llama.cpp SYCL backend?
+
+  No. We can't support Ollama issue directly, because we aren't familiar with Ollama.
+
+  Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it.
+
+  It's same for other projects including llama.cpp SYCL backend.
+
+
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
 ## TODO
 
-- Support row layer split for multiple card runs.
+- NA

examples/cvector-generator/cvector-generator.cpp

@@ -271,7 +271,7 @@ struct tokenized_prompt {
     size_t max_seq_len;
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
         tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
         tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());

examples/embedding/embedding.cpp

@@ -31,13 +31,24 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 }
 
 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
+    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    const struct llama_model * model = llama_get_model(ctx);
+
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_cache_clear(ctx);
 
     // run model
     fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_decode(ctx, batch) < 0) {
-        fprintf(stderr, "%s : failed to decode\n", __func__);
+    if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
+        // encoder-only model
+        if (llama_encode(ctx, batch) < 0) {
+            fprintf(stderr, "%s : failed to encode\n", __func__);
+        }
+    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        // decoder-only model
+        if (llama_decode(ctx, batch) < 0) {
+            fprintf(stderr, "%s : failed to decode\n", __func__);
+        }
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -45,11 +56,22 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
             continue;
         }
 
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        const float * embd = nullptr;
+        int embd_pos = 0;
 
-        float * out = output + batch.seq_id[i][0] * n_embd;
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            // try to get token embeddings
+            embd = llama_get_embeddings_ith(ctx, i);
+            embd_pos = i;
+            GGML_ASSERT(embd != NULL && "failed to get token embeddings");
+        } else {
+            // try to get sequence embeddings - supported only when pooling_type is not NONE
+            embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
+            embd_pos = batch.seq_id[i][0];
+            GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        }
+
+        float * out = output + embd_pos * n_embd;
         llama_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
@@ -93,8 +115,9 @@ int main(int argc, char ** argv) {
     const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
+
+    if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__);
         return 1;
     }
 
@@ -153,13 +176,23 @@ int main(int argc, char ** argv) {
     const int n_prompts = prompts.size();
     struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
+    // count number of embeddings
+    int n_embd_count = 0;
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        for (int k = 0; k < n_prompts; k++) {
+            n_embd_count += inputs[k].size();
+        }
+    } else {
+        n_embd_count = n_prompts;
+    }
+
     // allocate output
     const int n_embd = llama_n_embd(model);
-    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    std::vector<float> embeddings(n_embd_count * n_embd, 0);
     float * emb = embeddings.data();
 
     // break into batches
-    int p = 0; // number of prompts processed already
+    int e = 0; // number of embeddings already stored
     int s = 0; // number of prompts in current batch
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
@@ -169,11 +202,11 @@ int main(int argc, char ** argv) {
 
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
-            float * out = emb + p * n_embd;
+            float * out = emb + e * n_embd;
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
-            llama_batch_clear(batch);
-            p += s;
+            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
             s = 0;
+            llama_batch_clear(batch);
         }
 
         // add to batch
@@ -182,40 +215,63 @@ int main(int argc, char ** argv) {
     }
 
     // final batch
-    float * out = emb + p * n_embd;
+    float * out = emb + e * n_embd;
     batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
 
     if (params.embd_out.empty()) {
-        // print the first part of the embeddings or for a single prompt, the full embedding
         fprintf(stdout, "\n");
-        for (int j = 0; j < n_prompts; j++) {
-            fprintf(stdout, "embedding %d: ", j);
-            for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-                if (params.embd_normalize == 0) {
-                    fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                } else {
-                    fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-                }
-            }
-            fprintf(stdout, "\n");
-        }
 
-        // print cosine similarity matrix
-        if (n_prompts > 1) {
-            fprintf(stdout, "\n");
-            printf("cosine similarity matrix:\n\n");
-            for (int i = 0; i < n_prompts; i++) {
-                fprintf(stdout, "%6.6s ", prompts[i].c_str());
-            }
-            fprintf(stdout, "\n");
-            for (int i = 0; i < n_prompts; i++) {
-                for (int j = 0; j < n_prompts; j++) {
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                    fprintf(stdout, "%6.2f ", sim);
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            for (int j = 0; j < n_embd_count; j++) {
+                fprintf(stdout, "embedding %d: ", j);
+                for (int i = 0; i < std::min(3, n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                fprintf(stdout, " ... ");
+                for (int i = n_embd - 3; i < n_embd; i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
                 }
-                fprintf(stdout, "%1.10s", prompts[i].c_str());
                 fprintf(stdout, "\n");
             }
+        } else {
+            // print the first part of the embeddings or for a single prompt, the full embedding
+            for (int j = 0; j < n_prompts; j++) {
+                fprintf(stdout, "embedding %d: ", j);
+                for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                fprintf(stdout, "\n");
+            }
+
+            // print cosine similarity matrix
+            if (n_prompts > 1) {
+                fprintf(stdout, "\n");
+                printf("cosine similarity matrix:\n\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    fprintf(stdout, "%6.6s ", prompts[i].c_str());
+                }
+                fprintf(stdout, "\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    for (int j = 0; j < n_prompts; j++) {
+                        float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                        fprintf(stdout, "%6.2f ", sim);
+                    }
+                    fprintf(stdout, "%1.10s", prompts[i].c_str());
+                    fprintf(stdout, "\n");
+                }
+            }
         }
     }
 
@@ -233,23 +289,23 @@ int main(int argc, char ** argv) {
             }
             fprintf(stdout, notArray ? "]\n    }" : "]");
             j++;
-            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
+            if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
         }
         fprintf(stdout, notArray ? "\n  ]" : "]\n");
 
         if (params.embd_out == "json+" && n_prompts > 1) {
             fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
-            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
+            for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
                 fprintf(stdout, "    [");
-                for (int j = 0;;) { // at least two iteration (n_prompts > 1)
+                for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
                     float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
                     fprintf(stdout, "%6.2f", sim);
                     j++;
-                    if (j < n_prompts) fprintf(stdout, ", "); else break;
+                    if (j < n_embd_count) fprintf(stdout, ", "); else break;
                 }
                 fprintf(stdout, " ]");
                 i++;
-                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
+                if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
             }
             fprintf(stdout, "\n  ]");
         }

examples/eval-callback/eval-callback.cpp

@@ -127,7 +127,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool run(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 

examples/export-lora/README.md

@@ -17,9 +17,9 @@ For example:
 
 ```bash
 ./bin/llama-export-lora \
-    -m open-llama-3b-v2-q8_0.gguf \
-    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
-    --lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.gguf
+    -m open-llama-3b-v2.gguf \
+    -o open-llama-3b-v2-english2tokipona-chat.gguf \
+    --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
 ```
 
 Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:

examples/export-lora/export-lora.cpp

@@ -10,6 +10,12 @@
 
 static bool g_verbose = false;
 
+struct tensor_transformation {
+    struct ggml_tensor * in;
+    struct ggml_tensor * out;
+    bool is_copy;
+};
+
 static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
     int id = gguf_find_key(ctx_gguf, key.c_str());
     return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
@@ -50,20 +56,6 @@ static struct gguf_context * load_gguf(std::string & fname, struct ggml_context
     return ctx_gguf;
 }
 
-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;
-    }
-    s = std::move(result);
-}
-
 struct file_input {
     struct ggml_context * ctx_meta = nullptr;
     struct gguf_context * ctx_gguf = nullptr;
@@ -212,8 +204,7 @@ struct lora_merge_ctx {
         }
 
         // mapping base tensor to out tensor (same shape with base, but different type)
-        // if out_tensor == nullptr, we only copy it
-        std::vector<std::pair<struct ggml_tensor *, struct ggml_tensor *>> base_to_out_tensors;
+        std::vector<tensor_transformation> trans;
         for (auto & it : base_model.tensors) {
             bool t_a = true;
             bool t_b = true;
@@ -226,14 +217,22 @@ struct lora_merge_ctx {
                 // only copy
                 struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
                 ggml_set_name(cpy_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
+                trans.push_back({
+                    cpy_tensor,
+                    cpy_tensor,
+                    true,
+                });
                 gguf_add_tensor(ctx_out, cpy_tensor);
             } else if (t_a && t_b) {
                 // need merging
                 struct ggml_tensor * out_tensor = ggml_new_tensor(
                     ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
                 ggml_set_name(out_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
+                trans.push_back({
+                    base_tensor,
+                    out_tensor,
+                    false,
+                });
                 gguf_add_tensor(ctx_out, out_tensor);
             } else {
                 throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
@@ -248,12 +247,12 @@ struct lora_merge_ctx {
 
         // process base model tensors
         size_t n_merged = 0;
-        for (auto & it : base_to_out_tensors) {
-            if (it.second != nullptr) {
-                merge_tensor(it.first, it.second);
+        for (auto & it : trans) {
+            if (!it.is_copy) {
+                merge_tensor(it.in, it.out);
                 n_merged++;
             } else {
-                copy_tensor(it.first);
+                copy_tensor(it.in);
             }
         }
 
@@ -266,7 +265,7 @@ struct lora_merge_ctx {
         }
 
         printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
-        printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
+        printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
     }
 
     void copy_tensor(struct ggml_tensor * base) {
@@ -299,6 +298,10 @@ struct lora_merge_ctx {
         for (size_t i = 0; i < adapters.size(); ++i) {
             auto t_a = adapters[i]->get_tensor(name_lora_a);
             auto t_b = adapters[i]->get_tensor(name_lora_b);
+            // TODO: add support for quantized lora
+            if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) {
+                throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32");
+            }
             inp_a[i] = ggml_dup_tensor(ctx, t_a);
             inp_b[i] = ggml_dup_tensor(ctx, t_b);
         }

examples/imatrix/imatrix.cpp

@@ -433,8 +433,8 @@ static void process_logits(
 }
 
 static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
     const int n_ctx = llama_n_ctx(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();

examples/infill/infill.cpp

@@ -203,8 +203,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
         LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
     }
-    const bool add_bos = llama_should_add_bos_token(model);
-    GGML_ASSERT(llama_add_eos_token(model) != 1);
+    const bool add_bos = llama_add_bos_token(model);
+    GGML_ASSERT(!llama_add_eos_token(model));
     LOG("add_bos: %d\n", add_bos);
 
     std::vector<llama_token> embd_inp;

examples/llava/clip.cpp

@@ -210,17 +210,14 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
 }
 
 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 std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {

examples/llava/requirements.txt

@@ -2,4 +2,4 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 pillow~=10.2.0
 torch~=2.2.1
-torchvision==0.17.1
+torchvision~=0.17.1

examples/main/main.cpp

@@ -267,9 +267,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const bool add_bos = llama_should_add_bos_token(model);
+    const bool add_bos = llama_add_bos_token(model);
     if (!llama_model_has_encoder(model)) {
-        GGML_ASSERT(llama_add_eos_token(model) != 1);
+        GGML_ASSERT(!llama_add_eos_token(model));
     }
     LOG("add_bos: %d\n", add_bos);
 

examples/perplexity/perplexity.cpp

@@ -340,8 +340,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
 
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
@@ -480,8 +480,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
 
     std::ofstream logits_stream;
     if (!params.logits_file.empty()) {
@@ -1733,8 +1733,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     const int n_batch = params.n_batch;
     const int num_batches = (n_ctx + n_batch - 1)/n_batch;
     const int nv = 2*((n_vocab + 1)/2) + 4;
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
 
     std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
     std::vector<float>    kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);

examples/rpc/README.md

@@ -1,5 +1,9 @@
 ## Overview
 
+> [!IMPORTANT]
+> This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
+> insecure. **Never run the RPC server on an open network or in a sensitive environment!**
+
 The `rpc-server` allows  running `ggml` backend on a remote host.
 The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
 This can be used for distributed LLM inference with `llama.cpp` in the following way:

examples/rpc/rpc-server.cpp

@@ -16,7 +16,7 @@
 #include <stdio.h>
 
 struct rpc_server_params {
-    std::string host        = "0.0.0.0";
+    std::string host        = "127.0.0.1";
     int         port        = 50052;
     size_t      backend_mem = 0;
 };
@@ -114,6 +114,17 @@ int main(int argc, char * argv[]) {
         fprintf(stderr, "Invalid parameters\n");
         return 1;
     }
+
+    if (params.host != "127.0.0.1") {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "WARNING: Host ('%s') is != '127.0.0.1'\n", params.host.c_str());
+        fprintf(stderr, "         Never expose the RPC server to an open network!\n");
+        fprintf(stderr, "         This is an experimental feature and is not secure!\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "\n");
+    }
+
     ggml_backend_t backend = create_backend();
     if (!backend) {
         fprintf(stderr, "Failed to create backend\n");

examples/server/server.cpp

@@ -631,6 +631,7 @@ struct server_context {
 
     bool clean_kv_cache = true;
     bool add_bos_token  = true;
+    bool has_eos_token  = false;
 
     int32_t n_ctx; // total context for all clients / slots
 
@@ -692,9 +693,8 @@ struct server_context {
 
         n_ctx = llama_n_ctx(ctx);
 
-        add_bos_token = llama_should_add_bos_token(model);
-        GGML_ASSERT(llama_add_eos_token(model) != 1);
-
+        add_bos_token = llama_add_bos_token(model);
+        has_eos_token = !llama_add_eos_token(model);
         return true;
     }
 
@@ -753,13 +753,13 @@ struct server_context {
         default_generation_settings_for_props = get_formated_generation(slots.front());
         default_generation_settings_for_props["seed"] = -1;
 
-        // the update_slots() logic will always submit a maximum of n_batch tokens
+        // the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens
         // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
         {
             const int32_t n_batch = llama_n_batch(ctx);
 
             // only a single seq_id per token is needed
-            batch = llama_batch_init(n_batch, 0, 1);
+            batch = llama_batch_init(std::max(n_batch, params.n_parallel), 0, 1);
         }
 
         metrics.init();
@@ -1031,7 +1031,7 @@ struct server_context {
         {
             slot.sparams.logit_bias.clear();
 
-            if (json_value(data, "ignore_eos", false)) {
+            if (json_value(data, "ignore_eos", false) && has_eos_token) {
                 slot.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
             }
 
@@ -1136,28 +1136,19 @@ struct server_context {
         if (!system_prompt.empty()) {
             system_tokens = ::llama_tokenize(ctx, system_prompt, true);
 
-            llama_batch_clear(batch);
-
-            for (int i = 0; i < (int)system_tokens.size(); ++i) {
-                llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
-            }
-
             const int32_t n_batch = llama_n_batch(ctx);
+            const int32_t n_tokens_prompt = system_tokens.size();
 
-            for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
-                const int32_t n_tokens = std::min(params.n_batch, batch.n_tokens - i);
-                llama_batch batch_view = {
-                    n_tokens,
-                    batch.token    + i,
-                    nullptr,
-                    batch.pos      + i,
-                    batch.n_seq_id + i,
-                    batch.seq_id   + i,
-                    batch.logits   + i,
-                    0, 0, 0, // unused
-                };
+            for (int32_t i = 0; i < n_tokens_prompt; i += n_batch) {
+                const int32_t n_tokens = std::min(n_batch, n_tokens_prompt - i);
 
-                if (llama_decode(ctx, batch_view) != 0) {
+                llama_batch_clear(batch);
+
+                for (int32_t j = 0; j < n_tokens; ++j) {
+                    llama_batch_add(batch, system_tokens[i + j], i + j, { 0 }, false);
+                }
+
+                if (llama_decode(ctx, batch) != 0) {
                     LOG_ERROR("llama_decode() failed", {});
                     return;
                 }
@@ -1330,7 +1321,7 @@ struct server_context {
 
         return json {
             {"n_ctx",                     slot.n_ctx},
-            {"n_predict",                 slot.n_predict},
+            {"n_predict",                 slot.n_predict},     // Server configured n_predict
             {"model",                     params.model_alias},
             {"seed",                      slot.sparams.seed},
             {"temperature",               slot.sparams.temp},
@@ -1352,7 +1343,7 @@ struct server_context {
             {"mirostat_eta",              slot.sparams.mirostat_eta},
             {"penalize_nl",               slot.sparams.penalize_nl},
             {"stop",                      slot.params.antiprompt},
-            {"n_predict",                 slot.params.n_predict}, // TODO: fix duplicate key n_predict
+            {"max_tokens",                slot.params.n_predict}, // User configured n_predict
             {"n_keep",                    slot.params.n_keep},
             {"n_discard",                 slot.params.n_discard},
             {"ignore_eos",                ignore_eos},
@@ -1860,6 +1851,8 @@ struct server_context {
                     llama_lora_adapters_apply(ctx, lora_adapters);
                     server_task_result result;
                     result.id = task.id;
+                    result.stop = true;
+                    result.error = false;
                     result.data = json{{ "success", true }};
                     queue_results.send(result);
                 } break;
@@ -2044,7 +2037,7 @@ struct server_context {
                         slot.t_start_generation = 0;
 
                         if (slot.infill) {
-                            const bool add_bos = llama_should_add_bos_token(model);
+                            const bool add_bos = llama_add_bos_token(model);
                             bool suff_rm_leading_spc = true;
                             if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
                                 params.input_suffix.erase(0, 1);

examples/tokenize/tokenize.cpp

@@ -362,7 +362,7 @@ int main(int raw_argc, char ** raw_argv) {
         prompt = stdin_buffer.str();
     }
 
-    const bool model_wants_add_bos = llama_should_add_bos_token(model);
+    const bool model_wants_add_bos = llama_add_bos_token(model);
     const bool add_bos = model_wants_add_bos && !no_bos;
     const bool parse_special = !no_parse_special;
 

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1722421184,
-        "narHash": "sha256-/DJBI6trCeVnasdjUo9pbnodCLZcFqnVZiLUfqLH4jA=",
+        "lastModified": 1723175592,
+        "narHash": "sha256-M0xJ3FbDUc4fRZ84dPGx5VvgFsOzds77KiBMW/mMTnI=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "9f918d616c5321ad374ae6cb5ea89c9e04bf3e58",
+        "rev": "5e0ca22929f3342b19569b21b2f3462f053e497b",
         "type": "github"
       },
       "original": {

ggml/CMakeLists.txt

@@ -129,13 +129,13 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_USE_GRAPHS                 "ggml: use CUDA graphs (llama.cpp only)"          OFF)
 
-option(GGML_CURL                            "ggml: use libcurl to download model from an URL" OFF)
 option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
 option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
+option(GGML_VULKAN_PERF                     "ggml: enable Vulkan perf output"                 OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)

ggml/include/ggml.h

@@ -244,6 +244,8 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
+#define GGML_ROPE_TYPE_NEOX 2
+
 #define GGUF_MAGIC "GGUF"
 
 #define GGUF_VERSION 3
@@ -1453,8 +1455,8 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
-    // if mode & 2 == 1, GPT-NeoX style
+    // if (mode & 1) - skip n_past elements (NOT SUPPORTED)
+    // if (mode & GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
     GGML_API struct ggml_tensor * ggml_rope(

ggml/src/CMakeLists.txt

@@ -602,6 +602,10 @@ if (GGML_VULKAN)
             add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
         endif()
 
+        if (GGML_VULKAN_PERF)
+            add_compile_definitions(GGML_VULKAN_PERF)
+        endif()
+
         if (GGML_VULKAN_VALIDATE)
             add_compile_definitions(GGML_VULKAN_VALIDATE)
         endif()

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2881,7 +2881,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast,
                              beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     // init cos/sin cache
     ggml_cann_pool_alloc sin_allocator(

ggml/src/ggml-cuda/rope.cu

@@ -226,7 +226,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     const int32_t * pos = (const int32_t *) src1_d;
 

ggml/src/ggml-metal.m

@@ -310,7 +310,7 @@ static struct ggml_backend_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_backend_metal_context * ctx = malloc(sizeof(struct ggml_backend_metal_context));
+    struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context));
     ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
@@ -2313,7 +2313,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
                         memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-                        const bool is_neox = mode & 2;
+                        const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
                         id<MTLComputePipelineState> pipeline = nil;
 

ggml/src/ggml-rpc.cpp

@@ -197,6 +197,10 @@ static std::shared_ptr<socket_t> create_server_socket(const char * host, int por
         fprintf(stderr, "Failed to set SO_REUSEADDR\n");
         return nullptr;
     }
+    if (inet_addr(host) == INADDR_NONE) {
+        fprintf(stderr, "Invalid host address: %s\n", host);
+        return nullptr;
+    }
     struct sockaddr_in serv_addr;
     serv_addr.sin_family = AF_INET;
     serv_addr.sin_addr.s_addr = inet_addr(host);
@@ -879,6 +883,14 @@ ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rp
     if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
         return nullptr;
     }
+
+    // require that the tensor data does not go beyond the buffer end
+    uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
+    uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
+    uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
+    GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
+    GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
+
     result->op = (ggml_op) tensor->op;
     for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
         result->op_params[i] = tensor->op_params[i];
@@ -898,7 +910,7 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
     const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
     uint64_t offset;
     memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
-    size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
+    const size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
 
     struct ggml_init_params params {
         /*.mem_size   =*/ ggml_tensor_overhead(),
@@ -913,6 +925,17 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+
     const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
     ggml_backend_tensor_set(tensor, data, offset, size);
     ggml_free(ctx);
@@ -943,6 +966,17 @@ bool rpc_server::get_tensor(const std::vector<uint8_t> & input, std::vector<uint
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+
     // output serialization format: | data (size bytes) |
     output.resize(size, 0);
     ggml_backend_tensor_get(tensor, output.data(), offset, size);

ggml/src/ggml-sycl/rope.cpp

@@ -226,7 +226,7 @@ void ggml_sycl_op_rope(
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     const int32_t * pos = (const int32_t *) src1_dd;
 

ggml/src/ggml.c

@@ -3724,7 +3724,8 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         struct ggml_tensor  * view_src,
         size_t                view_offs) {
 
-    assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
+    GGML_ASSERT(type >= 0 && type < GGML_TYPE_COUNT);
+    GGML_ASSERT(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
 
     // find the base tensor and absolute offset
     if (view_src != NULL && view_src->view_src != NULL) {
@@ -14093,7 +14094,7 @@ static void ggml_compute_forward_rope_f32(
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -14218,7 +14219,7 @@ static void ggml_compute_forward_rope_f16(
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -21128,7 +21129,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 (int64_t) info->ne[2] *
                 (int64_t) info->ne[3];
 
-            if (ne % ggml_blck_size(info->type) != 0) {
+            if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
                 fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
                         __func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
                 fclose(file);

ggml/src/kompute-shaders/op_rope_f16.comp

@@ -11,7 +11,7 @@ void main() {
     const uint i2 = gl_WorkGroupID.y;
     const uint i1 = gl_WorkGroupID.x;
 
-    const bool is_neox = (pcs.mode & 2) != 0;
+    const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;
 
     float corr_dims[2];
     rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

ggml/src/kompute-shaders/op_rope_f32.comp

@@ -11,7 +11,7 @@ void main() {
     const uint i2 = gl_WorkGroupID.y;
     const uint i1 = gl_WorkGroupID.x;
 
-    const bool is_neox = (pcs.mode & 2) != 0;
+    const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;
 
     float corr_dims[2];
     rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

ggml/src/kompute-shaders/rope_common.comp

@@ -1,5 +1,7 @@
 #include "common.comp"
 
+#define GGML_ROPE_TYPE_NEOX 2
+
 // TODO: use a local size of 32 or more (Metal uses 1024)
 layout(local_size_x = 1) in;
 

ggml/src/vulkan-shaders/concat.comp

@@ -30,6 +30,10 @@ void main() {
 #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];
+    if (is_src0) {
+        data_d[p.d_offset + dst_idx] = data_a[src0_idx];
+    } else {
+        data_d[p.d_offset + dst_idx] = data_b[src1_idx];
+    }
 #endif
 }

ggml/src/vulkan-shaders/mul_mat_vec.comp

@@ -39,8 +39,7 @@ void main() {
         vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);
 
         // matrix multiplication
-        tmp[tid] += FLOAT_TYPE(v.x) * FLOAT_TYPE(data_b[b_offset + iybs + iqs]) +
-                    FLOAT_TYPE(v.y) * FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
+        tmp[tid] = fma(FLOAT_TYPE(v.x), FLOAT_TYPE(data_b[b_offset + iybs + iqs]), fma(FLOAT_TYPE(v.y), FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]), tmp[tid]));
     }
 
     // sum up partial sums and write back result

ggml/src/vulkan-shaders/mul_mat_vec_nc.comp

@@ -53,7 +53,7 @@ void main() {
 
         const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
 
-        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
+        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
     }
 
     // sum up partial sums and write back result

ggml/src/vulkan-shaders/mul_mat_vec_p021.comp

@@ -52,7 +52,7 @@ void main() {
         // y is not transposed but permuted
         const uint iy = channel*nrows_y + row_y;
 
-        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
+        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
     }
 
     // dst is not transposed and not permuted

ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -39,24 +39,25 @@ void main() {
         FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
         FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
         for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3);
-            sum2 += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF)
-                  + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF);
+            sum1 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3), sum1))))))));
+            sum2 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF), sum2))))))));
         }
-        tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, sum1, fma(-dmin, sum2, tmp[tmp_idx]));
     }
 
     // sum up partial sums and write back result

ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -40,16 +40,17 @@ void main() {
 
         FLOAT_TYPE sum = FLOAT_TYPE(0.0);
         for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4))
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4));
+            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
         }
-        tmp[16 * ix + tid] += d * sum;
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(d, sum, tmp[tmp_idx]);
     }
 
     // sum up partial sums and write back result

ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -67,17 +67,17 @@ void main() {
         const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66]  >> 4);
         const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67]  >> 4);
 
-        const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1 + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * q4_3);
-        const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_5 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7);
-        const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx]) * q4_8 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_9 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * q4_10 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * q4_11);
-        const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_12 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_13 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * q4_14 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15);
-        const FLOAT_TYPE smin = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y1_idx    ]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx    ]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * sc7
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7
-        );
-        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
+        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx]),      q4_0,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]),  q4_1,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]),  q4_2,  FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) *  q4_3)));
+        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_4,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), q4_5,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), q4_6,  FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7)));
+        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx]),      q4_8,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]),  q4_9,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]),  q4_10, FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) *  q4_11)));
+        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_12, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), q4_13, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), q4_14, FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15)));
+        const FLOAT_TYPE smin =
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 3]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 35]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 3]), sc6,     FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7)))))))))))))));
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
 #else
         const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
         const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
@@ -88,16 +88,19 @@ void main() {
         const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
         const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
 
-        const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) * q4_0  + FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * q4_1);
-        const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_2  + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
-        const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) * q4_4  + FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * q4_5);
-        const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
-        const FLOAT_TYPE smin = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y1_idx]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
-        );
+        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), q4_0, FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * q4_1);
+        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_2, FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
+        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), q4_4, FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * q4_5);
+        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
+        const FLOAT_TYPE smin =
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
+          + fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7)))))));
 
-        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) + sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
+        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) +
+                        sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, (fma(sx, FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f), fma(sy, FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f),
+                       fma(sz, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)), fma(sw, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))))))), fma(-dmin, smin, tmp[tmp_idx]));
 #endif
     }
 

ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -66,35 +66,33 @@ void main() {
         const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80]  >> 4);
         const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81]  >> 4);
 
-        const FLOAT_TYPE sx = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y1_idx     ]) * (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) * (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0))
-        );
-        const FLOAT_TYPE sy = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) * (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0))
-        );
-        const FLOAT_TYPE sz = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y2_idx     ]) * (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) * (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0))
-        );
-        const FLOAT_TYPE sw = FLOAT_TYPE(
-            FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) * (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0))
-          + FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0))
-        );
-        const FLOAT_TYPE smin = FLOAT_TYPE(
-            (FLOAT_TYPE(data_b[b_offset + y1_idx]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17])) * sc2 + (FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49])) * sc3
-          + (FLOAT_TYPE(data_b[b_offset + y2_idx]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17])) * sc6 + (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7
-        );
-        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
+        const FLOAT_TYPE sx =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx +  1]), (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sy =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sz =
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx +  1]), (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sw =
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE smin =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2,
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3,
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6,
+              (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7)));
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
     }
 
     // sum up partial sums and write back result

ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -44,22 +44,22 @@ void main() {
         const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
 #if K_QUANTS_PER_ITERATION == 1
-        FLOAT_TYPE sum = FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32)
-                       + FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32);
-        tmp[16 * ix + tid] += sum;
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx]))))))));
 #else
         FLOAT_TYPE sum = FLOAT_TYPE(0.0);
         [[unroll]] for (int l = 0; l < 4; ++l) {
-            sum += FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32)
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32)
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32)
-                 + FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d * FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32);
+            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum))));
         }
         tmp[16 * ix + tid] += sum;
 #endif

ggml/src/vulkan-shaders/mul_mm.comp

@@ -326,10 +326,10 @@ void main() {
                 mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
             }
             const float d = loadd.x * sc;
-            const float m = loadd.y * mbyte;
+            const float m = -loadd.y * mbyte;
 
-            buf_a[buf_idx    ] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) - m);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) - m);
+            buf_a[buf_idx    ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF), m));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
 #elif defined(DATA_A_Q5_K)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
             const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
@@ -357,10 +357,10 @@ void main() {
                 mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
             }
             const float d = loadd.x * sc;
-            const float m = loadd.y * mbyte;
+            const float m = -loadd.y * mbyte;
 
-            buf_a[buf_idx    ] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0)) - m);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0)) - m);
+            buf_a[buf_idx    ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0), m));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
 #elif defined(DATA_A_Q6_K)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
             const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
@@ -463,7 +463,8 @@ void main() {
                 [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                     [[unroll]] for (uint cc = 0; cc < TN; cc++) {
                         [[unroll]] for (uint cr = 0; cr < TM; cr++) {
-                            sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr] += float(cache_a[wsir * TM + cr]) * float(cache_b[wsic * TN + cc]);
+                            const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
+                            sums[sums_idx] = fma(float(cache_a[wsir * TM + cr]), float(cache_b[wsic * TN + cc]), sums[sums_idx]);
                         }
                     }
                 }

ggml/src/vulkan-shaders/repeat.comp

@@ -0,0 +1,24 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+uint src0_idx_mod(uint idx) {
+    const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+    return (i13 % p.ne03)*p.nb03 + (i12 % p.ne02)*p.nb02 + (i11 % p.ne01)*p.nb01 + (i10 % p.ne00)*p.nb00;
+}
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx_mod(idx)]);
+}

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

@@ -380,6 +380,10 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
         string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
 
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));

gguf-py/examples/writer.py

@@ -15,7 +15,6 @@ def writer_example() -> None:
     # Example usage with a file
     gguf_writer = GGUFWriter("example.gguf", "llama")
 
-    gguf_writer.add_architecture()
     gguf_writer.add_block_count(12)
     gguf_writer.add_uint32("answer", 42)  # Write a 32-bit integer
     gguf_writer.add_float32("answer_in_float", 42.0)  # Write a 32-bit float

gguf-py/gguf/constants.py

@@ -217,6 +217,7 @@ class MODEL_ARCH(IntEnum):
     CHATGLM      = auto()
     BITNET       = auto()
     T5           = auto()
+    T5ENCODER    = auto()
     JAIS         = auto()
 
 
@@ -344,6 +345,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.CHATGLM:        "chatglm",
     MODEL_ARCH.BITNET:         "bitnet",
     MODEL_ARCH.T5:             "t5",
+    MODEL_ARCH.T5ENCODER:      "t5encoder",
     MODEL_ARCH.JAIS:           "jais",
 }
 
@@ -1036,6 +1038,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ENC_FFN_UP,
         MODEL_TENSOR.ENC_OUTPUT_NORM,
     ],
+    MODEL_ARCH.T5ENCODER: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ENC_ATTN_NORM,
+        MODEL_TENSOR.ENC_ATTN_Q,
+        MODEL_TENSOR.ENC_ATTN_K,
+        MODEL_TENSOR.ENC_ATTN_V,
+        MODEL_TENSOR.ENC_ATTN_OUT,
+        MODEL_TENSOR.ENC_ATTN_REL_B,
+        MODEL_TENSOR.ENC_FFN_NORM,
+        MODEL_TENSOR.ENC_FFN_GATE,
+        MODEL_TENSOR.ENC_FFN_DOWN,
+        MODEL_TENSOR.ENC_FFN_UP,
+        MODEL_TENSOR.ENC_OUTPUT_NORM,
+    ],
     MODEL_ARCH.JAIS: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

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

@@ -93,15 +93,14 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
         LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
         LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,
+        LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
+        LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
     };
 
-    // note: these values should be synchronized with ggml_rope
-    // TODO: maybe move this enum to ggml.h (ggml_rope_type)
     enum llama_rope_type {
         LLAMA_ROPE_TYPE_NONE = -1,
-        LLAMA_ROPE_TYPE_NORM =  0,
-        LLAMA_ROPE_TYPE_NEOX =  2,
-        LLAMA_ROPE_TYPE_GLM  =  4,
+        LLAMA_ROPE_TYPE_NORM = 0,
+        LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX,
     };
 
     enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
@@ -504,6 +503,9 @@ extern "C" {
     // Returns true if the model contains an encoder that requires llama_encode() call
     LLAMA_API bool llama_model_has_encoder(const struct llama_model * model);
 
+    // Returns true if the model contains a decoder that requires llama_decode() call
+    LLAMA_API bool llama_model_has_decoder(const struct llama_model * model);
+
     // For encoder-decoder models, this function returns id of the token that must be provided
     // to the decoder to start generating output sequence. For other models, it returns -1.
     LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model);
@@ -912,11 +914,8 @@ extern "C" {
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
     LLAMA_API llama_token llama_token_pad(const struct llama_model * model); // padding
 
-    // Returns -1 if unknown, 1 for true or 0 for false.
-    LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model);
-
-    // Returns -1 if unknown, 1 for true or 0 for false.
-    LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model);
+    LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
+    LLAMA_API bool llama_add_eos_token(const struct llama_model * model);
 
     // Codellama infill tokens
     LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix

src/llama-impl.h

@@ -24,3 +24,18 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 #define LLAMA_LOG_INFO(...)  llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
 #define LLAMA_LOG_WARN(...)  llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
 #define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+//
+// helpers
+//
+
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    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();
+    }
+}

src/llama-sampling.cpp

@@ -85,14 +85,14 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra
             constexpr float bucket_low   = -10.0f;
             constexpr float bucket_high  =  10.0f;
             constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
-            constexpr float bucker_inter = -bucket_low * bucket_scale;
+            constexpr float bucket_inter = -bucket_low * bucket_scale;
 
             std::vector<int> bucket_idx(candidates->size);
             std::vector<int> histo(nbuckets, 0);
 
             for (int i = 0; i < (int)candidates->size; ++i) {
                 const float val = candidates->data[i].logit;
-                int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
+                int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
                 ib = std::max(0, std::min(nbuckets-1, ib));
                 bucket_idx[i] = ib;
                 ++histo[ib];

src/llama-vocab.cpp

@@ -16,20 +16,6 @@
 // helpers
 //
 
-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;
-    }
-    s = std::move(result);
-}
-
 LLAMA_ATTRIBUTE_FORMAT(1, 2)
 static std::string format(const char * fmt, ...) {
     va_list ap;
@@ -424,6 +410,8 @@ struct llm_tokenizer_bpe {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_PORO:
+            case LLAMA_VOCAB_PRE_TYPE_BLOOM:
+            case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH:
                 regex_exprs = {
                     " ?[^(\\s|.,!?…。，、।۔،)]+",
                 };
@@ -1480,11 +1468,11 @@ llama_token llama_token_pad_impl(const struct llama_vocab & vocab) {
     return vocab.special_pad_id;
 }
 
-int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab) {
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab) {
     return vocab.tokenizer_add_bos;
 }
 
-int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab) {
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab) {
     return vocab.tokenizer_add_eos;
 }
 

src/llama-vocab.h

@@ -95,8 +95,8 @@ llama_token llama_token_sep_impl(const struct llama_vocab & vocab);
 llama_token llama_token_nl_impl (const struct llama_vocab & vocab);
 llama_token llama_token_pad_impl(const struct llama_vocab & vocab);
 
-int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab);
-int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
 
 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);

src/llama.cpp

@@ -121,17 +121,6 @@ static std::string trim(const std::string & str) {
     return str.substr(start, end - start);
 }
 
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    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();
-    }
-}
-
 static bool is_float_close(float a, float b, float abs_tol) {
     // Check for non-negative tolerance
     if (abs_tol < 0.0) {
@@ -219,6 +208,7 @@ enum llm_arch {
     LLM_ARCH_CHATGLM,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
+    LLM_ARCH_T5ENCODER,
     LLM_ARCH_JAIS,
     LLM_ARCH_UNKNOWN,
 };
@@ -263,6 +253,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_CHATGLM,         "chatglm"      },
     { LLM_ARCH_BITNET,          "bitnet"       },
     { LLM_ARCH_T5,              "t5"           },
+    { LLM_ARCH_T5ENCODER,       "t5encoder"    },
     { LLM_ARCH_JAIS,            "jais"         },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
@@ -1272,6 +1263,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_ENC_FFN_UP,           "enc.blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_T5ENCODER,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,           "token_embd" },
+            { LLM_TENSOR_OUTPUT,               "output" },
+            { LLM_TENSOR_ENC_OUTPUT_NORM,      "enc.output_norm" },
+            { LLM_TENSOR_ENC_ATTN_NORM,        "enc.blk.%d.attn_norm" },
+            { LLM_TENSOR_ENC_ATTN_Q,           "enc.blk.%d.attn_q" },
+            { LLM_TENSOR_ENC_ATTN_K,           "enc.blk.%d.attn_k" },
+            { LLM_TENSOR_ENC_ATTN_V,           "enc.blk.%d.attn_v" },
+            { LLM_TENSOR_ENC_ATTN_OUT,         "enc.blk.%d.attn_o" },
+            { LLM_TENSOR_ENC_ATTN_REL_B,       "enc.blk.%d.attn_rel_b" },
+            { LLM_TENSOR_ENC_FFN_NORM,         "enc.blk.%d.ffn_norm" },
+            { LLM_TENSOR_ENC_FFN_GATE,         "enc.blk.%d.ffn_gate" },
+            { LLM_TENSOR_ENC_FFN_DOWN,         "enc.blk.%d.ffn_down" },
+            { LLM_TENSOR_ENC_FFN_UP,           "enc.blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_JAIS,
         {
@@ -3566,13 +3575,8 @@ namespace GGUFMeta {
 
 using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
 
-// TODO: update when needed or think of some clever automatic way to do this
-static size_t llama_model_max_nodes(const llama_model & /*model*/) {
-    //if (model.arch == LLM_ARCH_LLAMA && model.hparams.n_layer > ??) { // llama-3 405B
-    //    return 32768;
-    //}
-
-    return 8192;
+static size_t llama_model_max_nodes(const llama_model & model) {
+    return std::max<size_t>(8192, model.tensors_by_name.size()*5);
 }
 
 struct llama_model_loader {
@@ -4892,7 +4896,6 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_PHI3:
             {
-                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_layer) {
@@ -4901,6 +4904,22 @@ static void llm_load_hparams(
                     case 40: model.type = e_model::MODEL_14B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
+
+                // for backward compatibility ; see: https://github.com/ggerganov/llama.cpp/pull/8931
+                if ((hparams.n_layer == 32 || hparams.n_layer == 40) && hparams.n_ctx_train == 4096) {
+                    // default value for Phi-3-mini-4k-instruct and Phi-3-medium-4k-instruct
+                    hparams.n_swa = 2047;
+                } else if (hparams.n_layer == 32 && hparams.n_head_kv(0) == 32 && hparams.n_ctx_train == 131072) {
+                    // default value for Phi-3-mini-128k-instruct
+                    hparams.n_swa = 262144;
+                } else if (hparams.n_layer == 40 && hparams.n_ctx_train == 131072) {
+                    // default value for Phi-3-medium-128k-instruct
+                    hparams.n_swa = 131072;
+                }
+                bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+                if (!found_swa && hparams.n_swa == 0) {
+                    throw std::runtime_error("invalid value for sliding_window");
+                }
             } break;
         case LLM_ARCH_PLAMO:
             {
@@ -5198,6 +5217,12 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                }
             } break;
+        case LLM_ARCH_T5ENCODER:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts);
+                model.type = e_model::MODEL_UNKNOWN;
+            } break;
         case LLM_ARCH_JAIS:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -5442,6 +5467,12 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "codeshell") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL;
+            } else if (
+                tokenizer_pre == "bloom") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_BLOOM;
+            } else if (
+                tokenizer_pre == "gpt3-finnish") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -7432,6 +7463,42 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_UP,   "weight", i), {n_embd,   n_ff});
                     }
                 } break;
+            case LLM_ARCH_T5ENCODER:
+                {
+                    const auto n_rel_attn_bkts = hparams.n_rel_attn_bkts;
+
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm_enc = ml.create_tensor(ctx_output, tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm_enc  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_ATTN_NORM,  "weight", i), {n_embd});
+                        layer.attn_rel_b_enc = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wq_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wk_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd});
+
+                        layer.ffn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_GATE, "weight", i), {n_embd,   n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_down_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up_enc   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
             case LLM_ARCH_JAIS:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@@ -13146,7 +13213,7 @@ struct llm_build_context {
         return gf;
     }
 
-    struct ggml_cgraph * build_t5() {
+    struct ggml_cgraph * build_t5_encoder() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
@@ -13161,303 +13228,323 @@ struct llm_build_context {
 
         inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
 
-        if (lctx.is_encoding) {
-            struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
+        GGML_ASSERT(lctx.is_encoding);
+        struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
 
-            // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-            struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false);
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false);
 
-            for (int il = 0; il < n_layer; ++il) {
-                struct ggml_tensor * inpSA = inpL;
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
 
-                // norm
-                cur = llm_build_norm(ctx0, inpL, hparams,
-                        model.layers[il].attn_norm_enc, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_norm", il);
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm_enc, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
 
-                // self-attention
-                {
-                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_enc, cur);
-                    cb(Qcur, "Qcur", il);
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_enc, cur);
+                cb(Qcur, "Qcur", il);
 
-                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_enc, cur);
-                    cb(Kcur, "Kcur", il);
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_enc, cur);
+                cb(Kcur, "Kcur", il);
 
-                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_enc, cur);
-                    cb(Vcur, "Vcur", il);
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_enc, cur);
+                cb(Vcur, "Vcur", il);
 
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 
-                    struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-                    struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+                struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+                struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
 
-                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                    cb(kq, "kq", il);
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                cb(kq, "kq", il);
 
-                    struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
-                    struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b);
-                    struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
-                    cb(kq_b, "kq_b", il);
+                struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
+                struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b);
+                struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
+                cb(kq_b, "kq_b", il);
 
-                    kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias);
-                    cb(kq, "kq_soft_max_ext", il);
+                kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias);
+                cb(kq, "kq_soft_max_ext", il);
 
-                    struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
-                    cb(v, "v", il);
+                struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
+                cb(v, "v", il);
 
-                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
-                    cb(kqv, "kqv", il);
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
+                cb(kqv, "kqv", il);
 
-                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                    cb(kqv_merged, "kqv_merged", il);
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
 
-                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                    cb(cur, "kqv_merged_cont", il);
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
 
-                    ggml_build_forward_expand(gf, cur);
+                ggml_build_forward_expand(gf, cur);
 
-                    cur = ggml_mul_mat(ctx0, model.layers[il].wo_enc, cur);
-                    cb(cur, "kqv_out", il);
-                }
-
-                if (il == n_layer - 1) {
-                    // skip computing output for unused tokens
-                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                    n_tokens = n_outputs;
-                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-                }
-
-                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-                cb(ffn_inp, "ffn_inp", il);
-
-                // feed-forward network
-                {
-                    cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                            model.layers[il].ffn_norm_enc, NULL,
-                            LLM_NORM_RMS, cb, il);
-                    cb(cur, "ffn_norm", il);
-
-                    // T5 uses relu, flan-T5 uses gelu-gated
-                    cur = llm_build_ffn(ctx0, lctx, cur,
-                            model.layers[il].ffn_up_enc,   NULL, NULL,
-                            model.layers[il].ffn_gate_enc, NULL, NULL,
-                            model.layers[il].ffn_down_enc, NULL, NULL,
-                            NULL,
-                            model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
-                            model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
-                            cb, il);
-                    cb(cur, "ffn_out", il);
-                }
-
-                cur = ggml_add(ctx0, cur, ffn_inp);
-                cb(cur, "ffn_out", il);
-
-                ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
-                if (layer_dir != nullptr) {
-                    cur = ggml_add(ctx0, cur, layer_dir);
-                }
-                cb(cur, "l_out", il);
-
-                // input for next layer
-                inpL = cur;
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_enc, cur);
+                cb(cur, "kqv_out", il);
             }
 
-            cur = inpL;
-            cb(cur, "result_embd", -1);
-
-            cur = llm_build_norm(ctx0, cur, hparams,
-                    model.output_norm_enc, NULL,
-                    LLM_NORM_RMS, cb, -1);
-            cb(cur, "result_norm", -1);
-        } else {
-            GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
-
-            struct ggml_tensor * embd_enc       = llm_build_inp_embd_enc();
-            struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true);
-
-            struct ggml_tensor * KQ_mask_dec   = build_inp_KQ_mask();
-            struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross();
-
-            for (int il = 0; il < n_layer; ++il) {
-                struct ggml_tensor * inpSA = inpL;
-
-                // norm
-                cur = llm_build_norm(ctx0, inpL, hparams,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_norm", il);
-
-                // self-attention
-                {
-                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-                    cb(Qcur, "Qcur", il);
-
-                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-                    cb(Kcur, "Kcur", il);
-
-                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
-                    cb(Vcur, "Vcur", il);
-
-                    llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il);
-
-                    struct ggml_tensor * k =
-                        ggml_view_3d(ctx0, kv_self.k_l[il],
-                                n_embd_head_k, n_kv, n_head_kv,
-                                ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
-                                ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
-                                0);
-                    cb(k, "k", il);
-
-                    struct ggml_tensor * v =
-                        ggml_view_3d(ctx0, kv_self.v_l[il],
-                                n_kv, n_embd_head_v, n_head_kv,
-                                ggml_element_size(kv_self.v_l[il])*n_ctx,
-                                ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v,
-                                0);
-                    cb(v, "v", il);
-
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                    struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-
-                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                    cb(kq, "kq", il);
-
-                    struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
-                    struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b);
-                    struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
-                    cb(kq_b, "kq_b", il);
-
-                    kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias);
-                    cb(kq, "kq_soft_max_ext", il);
-
-                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
-                    cb(kqv, "kqv", il);
-
-                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                    cb(kqv_merged, "kqv_merged", il);
-
-                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                    cb(cur, "kqv_merged_cont", il);
-
-                    ggml_build_forward_expand(gf, cur);
-
-                    cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur);
-                    cb(cur, "kqv_out", il);
-                }
-
-                cur = ggml_add(ctx0, cur, inpSA);
-                cb(cur, "cross_inp", il);
-
-                struct ggml_tensor * inpCA = cur;
-
-                // norm
-                cur = llm_build_norm(ctx0, cur, hparams,
-                        model.layers[il].attn_norm_cross, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_norm_cross", il);
-
-                // cross-attention
-                {
-                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_cross, cur);
-                    cb(Qcur, "Qcur", il);
-
-                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_cross, embd_enc);
-                    cb(Kcur, "Kcur", il);
-
-                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_cross, embd_enc);
-                    cb(Vcur, "Vcur", il);
-
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
-
-                    struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-                    struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
-
-                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                    cb(kq, "kq", il);
-
-                    kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
-                    cb(kq, "kq_soft_max_ext", il);
-
-                    struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
-                    cb(v, "v", il);
-
-                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
-                    cb(kqv, "kqv", il);
-
-                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                    cb(kqv_merged, "kqv_merged", il);
-
-                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                    cb(cur, "kqv_merged_cont", il);
-
-                    ggml_build_forward_expand(gf, cur);
-
-                    cur = ggml_mul_mat(ctx0, model.layers[il].wo_cross, cur);
-                    cb(cur, "kqv_out", il);
-                }
-
-                if (il == n_layer - 1) {
-                    // skip computing output for unused tokens
-                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                    n_tokens = n_outputs;
-                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-                    inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
-                }
-
-                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
-                cb(ffn_inp, "ffn_inp", il);
-
-                // feed-forward network
-                {
-                    cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                            model.layers[il].ffn_norm, NULL,
-                            LLM_NORM_RMS, cb, il);
-                    cb(cur, "ffn_norm", il);
-
-                    // T5 uses relu, flan-T5 uses gelu-gated
-                    cur = llm_build_ffn(ctx0, lctx, cur,
-                            model.layers[il].ffn_up,   NULL, NULL,
-                            model.layers[il].ffn_gate, NULL, NULL,
-                            model.layers[il].ffn_down, NULL, NULL,
-                            NULL,
-                            model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
-                            model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ,
-                            cb, il);
-                    cb(cur, "ffn_out", il);
-                }
-
-                cur = ggml_add(ctx0, cur, ffn_inp);
-                cb(cur, "ffn_out", il);
-
-                ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
-                if (layer_dir != nullptr) {
-                    cur = ggml_add(ctx0, cur, layer_dir);
-                }
-                cb(cur, "l_out", il);
-
-                // input for next layer
-                inpL = cur;
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
-            cur = inpL;
-            cb(cur, "result_embd", -1);
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
 
-            cur = llm_build_norm(ctx0, cur, hparams,
-                    model.output_norm, NULL,
-                    LLM_NORM_RMS, cb, -1);
-            cb(cur, "result_norm", -1);
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm_enc, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
 
-            // lm_head
-            cur = ggml_mul_mat(ctx0, model.output, cur);
-            cb(cur, "result_output", -1);
+                // T5 uses relu, flan-T5 uses gelu-gated
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up_enc,   NULL, NULL,
+                        model.layers[il].ffn_gate_enc, NULL, NULL,
+                        model.layers[il].ffn_down_enc, NULL, NULL,
+                        NULL,
+                        model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
+                        model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
+                        cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+            if (layer_dir != nullptr) {
+                cur = ggml_add(ctx0, cur, layer_dir);
+            }
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
         }
 
+        cur = inpL;
+        cb(cur, "result_embd", -1);
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm_enc, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_t5_decoder() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        GGML_ASSERT(!lctx.is_encoding);
+        GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
+
+        struct ggml_tensor * embd_enc       = llm_build_inp_embd_enc();
+        struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true);
+
+        struct ggml_tensor * KQ_mask_dec   = build_inp_KQ_mask();
+        struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il);
+
+                struct ggml_tensor * k =
+                    ggml_view_3d(ctx0, kv_self.k_l[il],
+                            n_embd_head_k, n_kv, n_head_kv,
+                            ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
+                            ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
+                            0);
+                cb(k, "k", il);
+
+                struct ggml_tensor * v =
+                    ggml_view_3d(ctx0, kv_self.v_l[il],
+                            n_kv, n_embd_head_v, n_head_kv,
+                            ggml_element_size(kv_self.v_l[il])*n_ctx,
+                            ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v,
+                            0);
+                cb(v, "v", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+                struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                cb(kq, "kq", il);
+
+                struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
+                struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b);
+                struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
+                cb(kq_b, "kq_b", il);
+
+                kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias);
+                cb(kq, "kq_soft_max_ext", il);
+
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
+                cb(kqv, "kqv", il);
+
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
+
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
+
+                ggml_build_forward_expand(gf, cur);
+
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur);
+                cb(cur, "kqv_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, inpSA);
+            cb(cur, "cross_inp", il);
+
+            struct ggml_tensor * inpCA = cur;
+
+            // norm
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_norm_cross, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm_cross", il);
+
+            // cross-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_cross, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_cross, embd_enc);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_cross, embd_enc);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
+
+                struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+                struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                cb(kq, "kq", il);
+
+                kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
+                cb(kq, "kq_soft_max_ext", il);
+
+                struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
+                cb(v, "v", il);
+
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
+                cb(kqv, "kqv", il);
+
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
+
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
+
+                ggml_build_forward_expand(gf, cur);
+
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_cross, cur);
+                cb(cur, "kqv_out", il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                // T5 uses relu, flan-T5 uses gelu-gated
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
+                        model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ,
+                        cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+            if (layer_dir != nullptr) {
+                cur = ggml_add(ctx0, cur, layer_dir);
+            }
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+        cb(cur, "result_embd", -1);
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
         ggml_build_forward_expand(gf, cur);
 
         return gf;
@@ -13909,7 +13996,15 @@ static struct ggml_cgraph * llama_build_graph(
             } break;
         case LLM_ARCH_T5:
             {
-                result = llm.build_t5();
+                if (lctx.is_encoding) {
+                    result = llm.build_t5_encoder();
+                } else {
+                    result = llm.build_t5_decoder();
+                }
+            } break;
+        case LLM_ARCH_T5ENCODER:
+            {
+                result = llm.build_t5_encoder();
             } break;
         case LLM_ARCH_JAIS:
             {
@@ -14357,7 +14452,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
 
     // TODO: use a per-batch flag for logits presence instead
     const bool has_logits = !cparams.embeddings;
-    const bool has_embd   =  lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE));
+    const bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
 
     const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
     const size_t embd_size   = has_embd   ?  n_embd*n_outputs_max : 0;
@@ -14628,12 +14723,15 @@ static int llama_decode_internal(
             res  = nullptr;
             embd = nullptr;
         } else if (cparams.embeddings) {
-            res = nullptr; // do not extract logits for embedding case
-            embd = gf->nodes[gf->n_nodes - 1];
-            if (strcmp(embd->name, "result_embd_pooled") != 0) {
-                embd = gf->nodes[gf->n_nodes - 2];
+            res  = nullptr; // do not extract logits for embedding case
+            embd = nullptr;
+            for (int i = gf->n_nodes - 1; i >= 0; --i) {
+                if (strcmp(gf->nodes[i]->name, "result_embd_pooled") == 0) {
+                    embd = gf->nodes[i];
+                    break;
+                }
             }
-            GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+            GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
         } else {
             embd = nullptr; // do not extract embeddings when not needed
             GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
@@ -14840,9 +14938,24 @@ static int llama_encode_internal(
     ggml_cgraph * gf = llama_build_graph(lctx, batch, false);
 
     // the output embeddings after the final encoder normalization
-    struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embd = nullptr;
 
-    GGML_ASSERT(strcmp(embd->name, "result_norm") == 0);
+    // there are two cases here
+    if (llama_model_has_decoder(&lctx.model)) {
+        // first case is an encoder-decoder T5 model where embeddings are passed to decoder
+        embd = gf->nodes[gf->n_nodes - 1];
+        GGML_ASSERT(strcmp(embd->name, "result_norm") == 0 && "missing result_output tensor");
+    } else {
+        // second case is an encoder-only T5 model
+        if (cparams.embeddings) {
+            // only output embeddings if required
+            embd = gf->nodes[gf->n_nodes - 1];
+            if (strcmp(embd->name, "result_embd_pooled") != 0) {
+                embd = gf->nodes[gf->n_nodes - 2];
+            }
+            GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+        }
+    }
 
     ggml_backend_sched_alloc_graph(lctx.sched, gf);
 
@@ -14855,20 +14968,54 @@ static int llama_encode_internal(
         ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd);
         GGML_ASSERT(backend_embd != nullptr);
 
-        // extract token embeddings
-        GGML_ASSERT(lctx.embd != nullptr);
+        if (llama_model_has_decoder(&lctx.model)) {
+            lctx.embd_enc.resize(n_tokens*n_embd);
+            float * embd_out = lctx.embd_enc.data();
 
-        lctx.embd_enc.resize(n_tokens*n_embd);
-        float * embd_out = lctx.embd_enc.data();
+            ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
 
-        ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
+            // remember the sequence ids used during the encoding - needed for cross attention later
+            lctx.seq_ids_enc.resize(n_tokens);
+            for (uint32_t i = 0; i < n_tokens; i++) {
+                for (int s = 0; s < batch.n_seq_id[i]; s++) {
+                    llama_seq_id seq_id = batch.seq_id[i][s];
+                    lctx.seq_ids_enc[i].insert(seq_id);
+                }
+            }
+        } else {
+            GGML_ASSERT(lctx.embd != nullptr);
 
-        // remember the sequence ids used during the encoding - needed for cross attention later
-        lctx.seq_ids_enc.resize(n_tokens);
-        for (uint32_t i = 0; i < n_tokens; i++) {
-            for (int s = 0; s < batch.n_seq_id[i]; s++) {
-                llama_seq_id seq_id = batch.seq_id[i][s];
-                lctx.seq_ids_enc[i].insert(seq_id);
+            switch (cparams.pooling_type) {
+                case LLAMA_POOLING_TYPE_NONE:
+                    {
+                        // extract token embeddings
+                        GGML_ASSERT(lctx.embd != nullptr);
+                        float * embd_out = lctx.embd;
+
+                        GGML_ASSERT(n_tokens*n_embd <= (int64_t) lctx.embd_size);
+                        ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
+                    } break;
+                case LLAMA_POOLING_TYPE_MEAN:
+                case LLAMA_POOLING_TYPE_CLS:
+                case LLAMA_POOLING_TYPE_LAST:
+                    {
+                        // extract sequence embeddings
+                        auto & embd_seq_out = lctx.embd_seq;
+                        embd_seq_out.clear();
+
+                        for (uint32_t i = 0; i < n_tokens; i++) {
+                            const llama_seq_id seq_id = batch.seq_id[i][0];
+                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                                continue;
+                            }
+                            embd_seq_out[seq_id].resize(n_embd);
+                            ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
+                        }
+                    } break;
+                case LLAMA_POOLING_TYPE_UNSPECIFIED:
+                    {
+                        GGML_ABORT("unknown pooling type");
+                    }
             }
         }
     }
@@ -16578,6 +16725,8 @@ struct llama_context * llama_new_context_with_model(
 
     ctx->sampling.rng = std::mt19937(params.seed);
     ctx->logits_all   = params.logits_all;
+    // build worst-case graph for encoder if a model contains encoder
+    ctx->is_encoding  = llama_model_has_encoder(model);
 
     uint32_t kv_size = cparams.n_ctx;
     ggml_type type_k = params.type_k;
@@ -16892,6 +17041,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_MAMBA:
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_T5:
+        case LLM_ARCH_T5ENCODER:
         case LLM_ARCH_JAIS:
             return LLAMA_ROPE_TYPE_NONE;
 
@@ -17039,8 +17189,16 @@ struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const ch
 
 bool llama_model_has_encoder(const struct llama_model * model) {
     switch (model->arch) {
-        case LLM_ARCH_T5: return true;
-        default:          return false;
+        case LLM_ARCH_T5:        return true;
+        case LLM_ARCH_T5ENCODER: return true;
+        default:                 return false;
+    }
+}
+
+bool llama_model_has_decoder(const struct llama_model * model) {
+    switch (model->arch) {
+        case LLM_ARCH_T5ENCODER: return false;
+        default:                 return true;
     }
 }
 
@@ -18547,11 +18705,11 @@ llama_token llama_token_pad(const struct llama_model * model) {
     return llama_token_pad_impl(model->vocab);
 }
 
-int32_t llama_add_bos_token(const struct llama_model * model) {
+bool llama_add_bos_token(const struct llama_model * model) {
     return llama_add_bos_token_impl(model->vocab);
 }
 
-int32_t llama_add_eos_token(const struct llama_model * model) {
+bool llama_add_eos_token(const struct llama_model * model) {
     return llama_add_eos_token_impl(model->vocab);
 }