Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.github/workflows/build.yml
#	CMakeLists.txt
#	Makefile
#	ggml-metal.m

.github/workflows/server.yml

@@ -25,17 +25,14 @@ jobs:
     strategy:
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
+        build_type: [Debug]
         include:
           - build_type: Release
             sanitizer: ""
-        exclude:
+          - build_type: Debug
-          - build_type: Release
-            sanitizer: ADDRESS
-          - build_type: Release
             sanitizer: THREAD
-          - build_type: Release
+            disabled_on_pr: true
-            sanitizer: UNDEFINED
+      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
 
     container:
       image: ubuntu:latest
@@ -81,13 +78,14 @@ jobs:
 
       - name: Tests
         id: server_integration_tests
+        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
           cd examples/server/tests
           PORT=8888 ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
-        if: ${{ github.event.schedule != '' && matrix.build_type == 'Release' || github.event.inputs.slow_tests == 'true' }}
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd examples/server/tests
           PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
@@ -124,13 +122,14 @@ jobs:
 
       - name: Tests
         id: server_integration_tests
+        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
           cd examples/server/tests
           behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
 
       - name: Slow tests
         id: server_integration_tests_slow
-        if: ${{ github.event.schedule != '' || github.event.inputs.slow_tests == 'true' }}
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd examples/server/tests
           behave.exe --stop --no-skipped --no-capture --tags slow

.gitignore

@@ -13,6 +13,8 @@
 .vs/
 .vscode/
 
+ggml-metal-embed.metal
+
 lcov-report/
 gcovr-report/
 

common/common.cpp

@@ -1878,3 +1878,16 @@ void llama_embd_normalize(const float * inp, float * out, int n) {
     }
 }
 
+float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n){
+    double sum  = 0.0;
+    double sum1 = 0.0;
+    double sum2 = 0.0;
+
+    for (int i = 0; i < n; i++) {
+        sum  += embd1[i] * embd2[i];
+        sum1 += embd1[i] * embd1[i];
+        sum2 += embd2[i] * embd2[i];
+    }
+
+    return sum / (sqrt(sum1) * sqrt(sum2));
+}

common/common.h

@@ -282,3 +282,4 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40
 
 void llama_embd_normalize(const float * inp, float * out, int n);
 
+float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);

convert.py

@@ -332,6 +332,9 @@ class Params:
 #
 
 class BpeVocab:
+    tokenizer_model = "gpt2"
+    name = "bpe"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
         if isinstance(self.bpe_tokenizer.get('model'), dict):
@@ -390,6 +393,9 @@ class BpeVocab:
 
 
 class SentencePieceVocab:
+    tokenizer_model = "llama"
+    name = "spm"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
         added_tokens: dict[str, int]
@@ -453,6 +459,9 @@ class SentencePieceVocab:
 
 
 class HfVocab:
+    tokenizer_model = "llama"
+    name = "hfft"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
         try:
             from transformers import AutoTokenizer
@@ -553,7 +562,15 @@ class HfVocab:
         return f"<HfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab"
+class NoVocab:
+    tokenizer_model = "no_vocab"
+    name = "no_vocab"
+
+    def __repr__(self) -> str:
+        return "<NoVocab for a model without integrated vocabulary>"
+
+
+Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab | NoVocab"
 
 
 #
@@ -935,8 +952,10 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N
     # Handle special case where the model's vocab size is not set
     if params.n_vocab == -1:
         raise ValueError(
-            f"The model's vocab size is set to -1 in params.json. Please update it manually. Maybe {vocab.vocab_size}?"
+            f"The model's vocab size is set to -1 in params.json. Please update it manually.{f' Maybe {vocab.vocab_size}?' if hasattr(vocab, 'vocab_size') else ''}"
         )
+    if isinstance(vocab, NoVocab):
+        return  # model has no vocab
 
     # Check for a vocab size mismatch
     if params.n_vocab == vocab.vocab_size:
@@ -977,6 +996,7 @@ class OutputFile:
             name = str(params.path_model.parent).split('/')[-1]
 
         self.gguf.add_name                (name)
+        self.gguf.add_vocab_size          (params.n_vocab)
         self.gguf.add_context_length      (params.n_ctx)
         self.gguf.add_embedding_length    (params.n_embd)
         self.gguf.add_block_count         (params.n_layer)
@@ -1013,21 +1033,9 @@ class OutputFile:
         if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
 
-    def handle_tokenizer_model(self, vocab: Vocab) -> str:
-        # Map the vocab types to the supported tokenizer models
-        tokenizer_model = {
-            SentencePieceVocab: "llama",
-            HfVocab: "llama",
-            BpeVocab: "gpt2",
-        }.get(type(vocab))
-
-        # Block if vocab type is not predefined
-        if tokenizer_model is None:
-            raise ValueError("Unknown vocab type: Not supported")
-
-        return tokenizer_model
-
     def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
+        assert not isinstance(vocab, NoVocab)
+
         tokens = []
         scores = []
         toktypes = []
@@ -1043,11 +1051,8 @@ class OutputFile:
         return tokens, scores, toktypes
 
     def add_meta_vocab(self, vocab: Vocab) -> None:
-        # Handle the tokenizer model
-        tokenizer_model = self.handle_tokenizer_model(vocab)
-
         # Ensure that tokenizer_model is added to the GGUF model
-        self.gguf.add_tokenizer_model(tokenizer_model)
+        self.gguf.add_tokenizer_model(vocab.tokenizer_model)
 
         # Extract model vocabulary for model conversion
         tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab)
@@ -1074,6 +1079,26 @@ class OutputFile:
     def write_tensor_info(self) -> None:
         self.gguf.write_ti_data_to_file()
 
+    def write_tensor_data(self, ftype: GGMLFileType, model: LazyModel, concurrency: int) -> None:
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency=concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(
+                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
+                use_processpool_executor=True,
+            )
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
+        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
+            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            padi = len(str(len(model)))
+            print(
+                f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
+            )
+            self.gguf.write_tensor_data(ndarray)
+
     def close(self) -> None:
         self.gguf.close()
 
@@ -1082,7 +1107,7 @@ class OutputFile:
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
         endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
     ) -> None:
-        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
+        check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
@@ -1120,8 +1145,11 @@ class OutputFile:
 
         # meta data
         of.add_meta_arch(params)
-        of.add_meta_vocab(vocab)
+        if isinstance(vocab, NoVocab):
-        of.add_meta_special_vocab(svocab)
+            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
+        else:
+            of.add_meta_vocab(vocab)
+            of.add_meta_special_vocab(svocab)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -1131,24 +1159,7 @@ class OutputFile:
         of.write_tensor_info()
 
         # tensor data
-        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
+        of.write_tensor_data(ftype, model, concurrency)
-        if ftype == GGMLFileType.MostlyQ8_0:
-            ndarrays = bounded_parallel_map(
-                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
-                use_processpool_executor=True,
-            )
-        else:
-            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
-
-        start = time.time()
-        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
-            elapsed = time.time() - start
-            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
-            padi = len(str(len(model)))
-            print(
-                f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
-            )
-            of.gguf.write_tensor_data(ndarray)
 
         of.close()
 
@@ -1309,8 +1320,8 @@ class VocabFactory:
                 return vtype, path
         raise FileNotFoundError(f"Could not find any of {[self._FILES[vt] for vt in vocab_types]}")
 
-    def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
+    def _create_special_vocab(self, vocab: Vocab, model_parent_path: Path) -> gguf.SpecialVocab:
-        load_merges = vocabtype == "bpe"
+        load_merges = vocab.name == "bpe"
         n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
         return gguf.SpecialVocab(
             model_parent_path,
@@ -1319,30 +1330,34 @@ class VocabFactory:
             n_vocab=n_vocab,
         )
 
-    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+    def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
         vocab_type, path = self._select_file(vocab_types)
         print(f"Loading vocab file {path!r}, type {vocab_type!r}")
 
         added_tokens_path = path.parent / "added_tokens.json"
-        vocab: Vocab
         if vocab_type == "bpe":
-            vocab = BpeVocab(
+            return BpeVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocab_type == "spm":
+        if vocab_type == "spm":
-            vocab = SentencePieceVocab(
+            return SentencePieceVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocab_type == "hfft":
+        if vocab_type == "hfft":
-            vocab = HfVocab(
+            return HfVocab(
                 path.parent, added_tokens_path if added_tokens_path.exists() else None
             )
+        raise ValueError(vocab_type)
+
+    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+        vocab: Vocab
+        if len(vocab_types) == 1 and "no_vocab" in vocab_types:
+            vocab = NoVocab()
         else:
-            raise ValueError(vocab_type)
+            vocab = self._create_vocab_by_path(vocab_types)
         # FIXME: Respect --vocab-dir?
         special_vocab = self._create_special_vocab(
             vocab,
-            vocab_type,
             model_parent_path,
         )
         return vocab, special_vocab
@@ -1380,6 +1395,7 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
+    parser.add_argument("--no-vocab",     action="store_true",    help="store model without the vocab")
     parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--vocab-type",                           help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft")
@@ -1392,6 +1408,10 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
+    if args.no_vocab:
+        if args.vocab_only:
+            raise ValueError("no need to specify --vocab-only if using --no-vocab")
+        args.vocab_type = "no_vocab"
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
@@ -1442,7 +1462,7 @@ def main(args_in: list[str] | None = None) -> None:
         print(f"Wrote {outfile}")
         return
 
-    if model_plus.vocab is not None and args.vocab_dir is None:
+    if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
         vocab = model_plus.vocab
 
     print(f"Vocab info: {vocab}")

examples/embedding/embedding.cpp

@@ -113,13 +113,20 @@ int main(int argc, char ** argv) {
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        auto inp = ::llama_tokenize(ctx, prompt, true);
+        auto inp = ::llama_tokenize(ctx, prompt, true, false);
         if (inp.size() > n_batch) {
             inp.resize(n_batch);
         }
         inputs.push_back(inp);
     }
 
+    // add eos if not present
+    for (auto & inp : inputs) {
+        if (inp.empty() || inp.back() != llama_token_eos(model)) {
+            inp.push_back(llama_token_eos(model));
+        }
+    }
+
     // tokenization stats
     if (params.verbose_prompt) {
         for (int i = 0; i < (int) inputs.size(); i++) {
@@ -168,15 +175,26 @@ int main(int argc, char ** argv) {
     float * out = emb + p * n_embd;
     batch_decode(ctx, batch, out, s, n_embd);
 
-    // print first 3 embeddings
+    // print the first part of the embeddings
-    for (int j = 0; j < std::min(3, n_prompts); j++) {
+    fprintf(stdout, "\n");
-        fprintf(stderr, "embedding %d: ", j);
+    for (int j = 0; j < n_prompts; j++) {
-        for (int i = 0; i < n_embd; i++) {
+        fprintf(stdout, "embedding %d: ", j);
-            fprintf(stderr, "%f ", emb[j * n_embd + i]);
+        for (int i = 0; i < std::min(16, n_embd); i++) {
+            fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
         }
-        fprintf(stderr, "\n\n");
+        fprintf(stdout, "\n");
+    }
+
+    // print cosine similarity matrix
+    fprintf(stdout, "\n");
+    printf("cosine similarity matrix:\n\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, "\n");
     }
-    fprintf(stderr, "\n");
 
     // clean up
     llama_print_timings(ctx);

examples/gguf/gguf.cpp

@@ -211,6 +211,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
                 for (int j = 0; j < ggml_nelements(cur); ++j) {
                     if (data[j] != 100 + i) {
                         fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]);
+                        gguf_free(ctx);
                         return false;
                     }
                 }

examples/gritlm/gritlm.cpp

@@ -6,22 +6,6 @@
 
 // #define GRIT_DEBUG
 
-static float dot_product(const std::vector<float> & v1, const std::vector<float> & v2) {
-    float dot = 0.0f;
-    for (uint64_t i = 0; i < v1.size(); ++i) {
-        dot += v1[i] * v2[i];
-    }
-    return dot;
-}
-
-static float norm(const std::vector<float> & v) {
-    return std::sqrt(dot_product(v, v));
-}
-
-static float cosine_similarity(const std::vector<float> & v1, const std::vector<float> & v2) {
-    return dot_product(v1, v2) / (norm(v1) * norm(v2));
-}
-
 static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
     std::vector<std::vector<float>> result;
 
@@ -203,10 +187,12 @@ int main(int argc, char * argv[]) {
         const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
         const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
 
-        const float cosine_sim_q0_d0 = cosine_similarity(q_rep[0], d_rep[0]);
+        const int n_embd = llama_n_embd(mdl);
-        const float cosine_sim_q0_d1 = cosine_similarity(q_rep[0], d_rep[1]);
+
-        const float cosine_sim_q1_d0 = cosine_similarity(q_rep[1], d_rep[0]);
+        const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q1_d1 = cosine_similarity(q_rep[1], d_rep[1]);
+        const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
+        const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
 
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);

examples/llama-bench/llama-bench.cpp

@@ -104,6 +104,7 @@ static std::string get_cpu_info() {
                 }
             }
         }
+        fclose(f);
     }
 #endif
     // TODO: other platforms

examples/llava/README.md

@@ -63,12 +63,20 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
 ```console
 git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 ```
-2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+
+2) Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
 python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory
-3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
+
+4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
 ```console
 mkdir vit
 cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
@@ -76,18 +84,18 @@ cp ../llava-v1.6-vicuna-7b/llava.projector vit/
 curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
 ```
 
-4) Create the visual gguf model:
+5) Create the visual gguf model:
 ```console
 python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
 
-5) Then convert the model to gguf format:
+6) Then convert the model to gguf format:
 ```console
 python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
 ```
 
-6) And finally we can run the llava-cli using the 1.6 model version:
+7) And finally we can run the llava-cli using the 1.6 model version:
 ```console
 ./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
 ```

examples/llava/clip.cpp

@@ -995,6 +995,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         if (!new_clip->ctx_data) {
             fprintf(stderr, "%s: ggml_init() failed\n", __func__);
             clip_free(new_clip);
+            gguf_free(ctx);
             return nullptr;
         }
 
@@ -1002,6 +1003,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         if (!fin) {
             printf("cannot open model file for loading tensors\n");
             clip_free(new_clip);
+            gguf_free(ctx);
             return nullptr;
         }
 
@@ -1023,6 +1025,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             if (!fin) {
                 printf("%s: failed to seek for tensor %s\n", __func__, name);
                 clip_free(new_clip);
+                gguf_free(ctx);
                 return nullptr;
             }
             int num_bytes = ggml_nbytes(cur);
@@ -1908,6 +1911,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 break;
             default:
                 printf("Please use an input file in f32 or f16\n");
+                gguf_free(ctx_out);
                 return false;
             }
 

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

@@ -119,6 +119,10 @@ def step_server_metrics(context):
 def step_start_server(context):
     start_server_background(context)
     attempts = 0
+    max_attempts = 20
+    if 'GITHUB_ACTIONS' in os.environ:
+        max_attempts *= 2
+
     while True:
         with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
             result = sock.connect_ex((context.server_fqdn, context.server_port))
@@ -126,7 +130,7 @@ def step_start_server(context):
                 print("\x1b[33;46mserver started!\x1b[0m")
                 return
             attempts += 1
-            if attempts > 20:
+            if attempts > max_attempts:
                 assert False, "server not started"
             print(f"waiting for server to start, connect error code = {result}...")
             time.sleep(0.1)
@@ -943,6 +947,9 @@ async def wait_for_health_status(context,
         print(f"Starting checking for health for expected_health_status={expected_health_status}\n")
     interval = 0.5
     counter = 0
+    if 'GITHUB_ACTIONS' in os.environ:
+        timeout *= 2
+
     async with aiohttp.ClientSession() as session:
         while True:
             async with await session.get(f'{base_url}/health', params=params) as health_response:

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -711,6 +711,7 @@ static bool load_checkpoint_file(const char * filename, struct my_llama_model *
 
     load_checkpoint_gguf(fctx, f_ggml_ctx, model, train);
 
+    gguf_free(fctx);
     return true;
 }
 

ggml-metal.m

@@ -280,6 +280,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     id<MTLLibrary> metal_library;
 
     // load library
+    //
+    // - first check if the library is embedded
+    // - then check if the library is in the bundle
+    // - if not found, load the source and compile it
+    // - if that fails, return NULL
     {
         NSBundle * bundle = nil;
 #ifdef SWIFT_PACKAGE
@@ -287,12 +292,21 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #else
         bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
 #endif
+
         NSError * error = nil;
-        NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
+
-        if (libPath != nil) {
+#if GGML_METAL_EMBED_LIBRARY
+        const bool try_metallib = false;
+#else
+        const bool try_metallib = true;
+#endif
+
+        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+        if (try_metallib && path_lib != nil) {
             // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:libPath];
+            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
             metal_library = [ctx->device newLibraryWithURL:libURL error:&error];
             if (error) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
@@ -305,31 +319,34 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
             extern const char ggml_metallib_start[];
             extern const char ggml_metallib_end[];
 
-            NSString * src  = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
+            NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
 #else
             GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
-            NSString * sourcePath;
+            NSString * path_source;
-            NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
 
-            GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, ggmlMetalPathResources ? [ggmlMetalPathResources UTF8String] : "nil");
+            GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
 
-            if (ggmlMetalPathResources) {
+            if (path_resource) {
-                sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal-merged.metal"];
+                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal-merged.metal"];
             } else {
-                sourcePath = [bundle pathForResource:@"ggml-metal-merged" ofType:@"metal"];
+                path_source = [bundle pathForResource:@"ggml-metal-merged" ofType:@"metal"];
             }
-            if (sourcePath == nil) {
+
+            if (path_source == nil) {
                 GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal-merged.metal, falling back to trying cwd\n", __func__);
-                sourcePath = @"ggml-metal-merged.metal";
+                path_source = @"ggml-metal.metal";
             }
-            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]);
+
-            NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+            NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
             if (error) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                 return NULL;
             }
-#endif
+#endif // GGML_METAL_EMBED_LIBRARY
 
             @autoreleasepool {
                 // dictionary of preprocessor macros

ggml.h

@@ -344,24 +344,24 @@ extern "C" {
     struct ggml_object;
     struct ggml_context;
 
+    // NOTE: always add types at the end of the enum to keep backward compatibility
     enum ggml_type {
-        GGML_TYPE_F32  = 0,
+        GGML_TYPE_F32     = 0,
-        GGML_TYPE_F16  = 1,
+        GGML_TYPE_F16     = 1,
-        GGML_TYPE_Q4_0 = 2,
+        GGML_TYPE_Q4_0    = 2,
-        GGML_TYPE_Q4_1 = 3,
+        GGML_TYPE_Q4_1    = 3,
         // GGML_TYPE_Q4_2 = 4, support has been removed
-        // GGML_TYPE_Q4_3 (5) support has been removed
+        // GGML_TYPE_Q4_3 = 5, support has been removed
-        GGML_TYPE_Q5_0 = 6,
+        GGML_TYPE_Q5_0    = 6,
-        GGML_TYPE_Q5_1 = 7,
+        GGML_TYPE_Q5_1    = 7,
-        GGML_TYPE_Q8_0 = 8,
+        GGML_TYPE_Q8_0    = 8,
-        GGML_TYPE_Q8_1 = 9,
+        GGML_TYPE_Q8_1    = 9,
-        // k-quantizations
+        GGML_TYPE_Q2_K    = 10,
-        GGML_TYPE_Q2_K = 10,
+        GGML_TYPE_Q3_K    = 11,
-        GGML_TYPE_Q3_K = 11,
+        GGML_TYPE_Q4_K    = 12,
-        GGML_TYPE_Q4_K = 12,
+        GGML_TYPE_Q5_K    = 13,
-        GGML_TYPE_Q5_K = 13,
+        GGML_TYPE_Q6_K    = 14,
-        GGML_TYPE_Q6_K = 14,
+        GGML_TYPE_Q8_K    = 15,
-        GGML_TYPE_Q8_K = 15,
         GGML_TYPE_IQ2_XXS = 16,
         GGML_TYPE_IQ2_XS  = 17,
         GGML_TYPE_IQ3_XXS = 18,
@@ -370,9 +370,9 @@ extern "C" {
         GGML_TYPE_IQ3_S   = 21,
         GGML_TYPE_IQ2_S   = 22,
         GGML_TYPE_IQ4_XS  = 23,
-        GGML_TYPE_I8,
+        GGML_TYPE_I8      = 24,
-        GGML_TYPE_I16,
+        GGML_TYPE_I16     = 25,
-        GGML_TYPE_I32,
+        GGML_TYPE_I32     = 26,
         GGML_TYPE_COUNT,
     };
 
@@ -390,20 +390,20 @@ extern "C" {
 
     // model file types
     enum ggml_ftype {
-        GGML_FTYPE_UNKNOWN     = -1,
+        GGML_FTYPE_UNKNOWN        = -1,
-        GGML_FTYPE_ALL_F32     = 0,
+        GGML_FTYPE_ALL_F32        = 0,
-        GGML_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_F16     = 1,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0    = 2,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_1    = 3,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q8_0    = 7,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_0    = 8,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_1    = 9,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_K    = 10, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q3_K    = 11, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_K    = 12, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_K    = 13, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q6_K    = 14, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors

gguf-py/gguf/constants.py

@@ -32,6 +32,7 @@ class Keys:
         FILE_TYPE            = "general.file_type"
 
     class LLM:
+        VOCAB_SIZE            = "{arch}.vocab_size"
         CONTEXT_LENGTH        = "{arch}.context_length"
         EMBEDDING_LENGTH      = "{arch}.embedding_length"
         BLOCK_COUNT           = "{arch}.block_count"
@@ -661,6 +662,9 @@ class GGMLQuantizationType(IntEnum):
     IQ3_S   = 21
     IQ2_S   = 22
     IQ4_XS  = 23
+    I8      = 24
+    I16     = 25
+    I32     = 26
 
 
 class GGUFEndian(IntEnum):
@@ -727,6 +731,9 @@ GGML_QUANT_SIZES = {
     GGMLQuantizationType.IQ3_S:   (256, 2 + QK_K // 4 + QK_K // 8 + QK_K // 32 + 4),
     GGMLQuantizationType.IQ2_S:   (256, 2 + QK_K // 4 + QK_K // 16),
     GGMLQuantizationType.IQ4_XS:  (256, 2 + 2 + QK_K // 2 + QK_K // 64),
+    GGMLQuantizationType.I8:      (1, 1),
+    GGMLQuantizationType.I16:     (1, 2),
+    GGMLQuantizationType.I32:     (1, 4),
 }
 
 
@@ -746,6 +753,7 @@ KEY_GENERAL_SOURCE_HF_REPO       = Keys.General.SOURCE_HF_REPO
 KEY_GENERAL_FILE_TYPE            = Keys.General.FILE_TYPE
 
 # LLM
+KEY_VOCAB_SIZE            = Keys.LLM.VOCAB_SIZE
 KEY_CONTEXT_LENGTH        = Keys.LLM.CONTEXT_LENGTH
 KEY_EMBEDDING_LENGTH      = Keys.LLM.EMBEDDING_LENGTH
 KEY_BLOCK_COUNT           = Keys.LLM.BLOCK_COUNT

gguf-py/gguf/gguf_reader.py

@@ -248,6 +248,15 @@ class GGUFReader:
             elif ggml_type == GGMLQuantizationType.F16:
                 item_count = n_elems
                 item_type = np.float16
+            elif ggml_type == GGMLQuantizationType.I8:
+                item_count = n_elems
+                item_type = np.int8
+            elif ggml_type == GGMLQuantizationType.I16:
+                item_count = n_elems
+                item_type = np.int16
+            elif ggml_type == GGMLQuantizationType.I32:
+                item_count = n_elems
+                item_type = np.int32
             else:
                 item_count = n_bytes
                 item_type = np.uint8

gguf-py/gguf/gguf_writer.py

@@ -196,9 +196,6 @@ class GGUFWriter:
         if self.state is not WriterState.EMPTY:
             raise ValueError(f'Expected output file to be empty, got {self.state}')
 
-        if raw_dtype is None and tensor_dtype not in (np.float32, np.float16):
-            raise ValueError("Only F32 and F16 tensors are supported for now")
-
         encoded_name = name.encode("utf8")
         self.ti_data += self._pack("Q", len(encoded_name))
         self.ti_data += encoded_name
@@ -207,7 +204,18 @@ class GGUFWriter:
         for i in range(n_dims):
             self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
         if raw_dtype is None:
-            dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
+            if tensor_dtype == np.float32:
+                dtype = GGMLQuantizationType.F32
+            elif tensor_dtype == np.float16:
+                dtype = GGMLQuantizationType.F16
+            elif tensor_dtype == np.int8:
+                dtype = GGMLQuantizationType.I8
+            elif tensor_dtype == np.int16:
+                dtype = GGMLQuantizationType.I16
+            elif tensor_dtype == np.int32:
+                dtype = GGMLQuantizationType.I32
+            else:
+                raise ValueError("Only F32, F16, I8, I16, I32 tensors are supported for now")
         else:
             dtype = raw_dtype
         self.ti_data += self._pack("I", dtype)
@@ -313,6 +321,9 @@ class GGUFWriter:
         self.data_alignment = alignment
         self.add_uint32(Keys.General.ALIGNMENT, alignment)
 
+    def add_vocab_size(self, size: int) -> None:
+        self.add_uint32(Keys.LLM.VOCAB_SIZE.format(arch=self.arch), size)
+
     def add_context_length(self, length: int) -> None:
         self.add_uint32(Keys.LLM.CONTEXT_LENGTH.format(arch=self.arch), length)
 

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.7.0"
+version = "0.8.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

koboldcpp.py

@@ -616,7 +616,7 @@ maxhordelen = 256
 modelbusy = threading.Lock()
 requestsinqueue = 0
 defaultport = 5001
-KcppVersion = "1.61.2"
+KcppVersion = "1.62"
 showdebug = True
 showsamplerwarning = True
 showmaxctxwarning = True

llama.cpp

@@ -282,6 +282,7 @@ enum llm_kv {
     LLM_KV_GENERAL_SOURCE_URL,
     LLM_KV_GENERAL_SOURCE_HF_REPO,
 
+    LLM_KV_VOCAB_SIZE,
     LLM_KV_CONTEXT_LENGTH,
     LLM_KV_EMBEDDING_LENGTH,
     LLM_KV_BLOCK_COUNT,
@@ -345,6 +346,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_GENERAL_SOURCE_URL,            "general.source.url"                    },
     { LLM_KV_GENERAL_SOURCE_HF_REPO,        "general.source.huggingface.repository" },
 
+    { LLM_KV_VOCAB_SIZE,                    "%s.vocab_size"            },
     { LLM_KV_CONTEXT_LENGTH,                "%s.context_length"        },
     { LLM_KV_EMBEDDING_LENGTH,              "%s.embedding_length"      },
     { LLM_KV_BLOCK_COUNT,                   "%s.block_count"           },
@@ -3288,10 +3290,11 @@ static const char * llama_model_type_name(e_model type) {
 
 static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
     switch (type) {
-        case LLAMA_VOCAB_TYPE_SPM: return "SPM";
+        case LLAMA_VOCAB_TYPE_NONE: return "no vocab";
-        case LLAMA_VOCAB_TYPE_BPE: return "BPE";
+        case LLAMA_VOCAB_TYPE_SPM:  return "SPM";
-        case LLAMA_VOCAB_TYPE_WPM: return "WPM";
+        case LLAMA_VOCAB_TYPE_BPE:  return "BPE";
-        default:                   return "unknown";
+        case LLAMA_VOCAB_TYPE_WPM:  return "WPM";
+        default:                    return "unknown";
     }
 }
 
@@ -3323,14 +3326,14 @@ static void llm_load_hparams(
     ml.get_key(LLM_KV_GENERAL_NAME, model.name, false);
 
     // get hparams kv
-    ml.get_arr_n(LLM_KV_TOKENIZER_LIST,       hparams.n_vocab);
+    ml.get_key(LLM_KV_VOCAB_SIZE,           hparams.n_vocab,       false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab);
-    ml.get_key  (LLM_KV_CONTEXT_LENGTH,       hparams.n_ctx_train);
+    ml.get_key(LLM_KV_CONTEXT_LENGTH,       hparams.n_ctx_train);
-    ml.get_key  (LLM_KV_EMBEDDING_LENGTH,     hparams.n_embd);
+    ml.get_key(LLM_KV_EMBEDDING_LENGTH,     hparams.n_embd);
-    ml.get_key  (LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff);
+    ml.get_key(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff);
-    ml.get_key  (LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head);
+    ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head);
-    ml.get_key  (LLM_KV_BLOCK_COUNT,          hparams.n_layer);
+    ml.get_key(LLM_KV_BLOCK_COUNT,          hparams.n_layer);
-    ml.get_key  (LLM_KV_EXPERT_COUNT,         hparams.n_expert,      false);
+    ml.get_key(LLM_KV_EXPERT_COUNT,         hparams.n_expert,      false);
-    ml.get_key  (LLM_KV_EXPERT_USED_COUNT,    hparams.n_expert_used, false);
+    ml.get_key(LLM_KV_EXPERT_USED_COUNT,    hparams.n_expert_used, false);
 
     GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS);
     GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert);
@@ -3692,30 +3695,25 @@ static void llm_load_vocab(
 
     const auto kv = LLM_KV(model.arch);
 
-    const int token_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_LIST).c_str());
-    if (token_idx == -1) {
-        throw std::runtime_error("cannot find tokenizer vocab in model file\n");
-    }
-
-    const float * scores = nullptr;
-    const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str());
-    if (score_idx != -1) {
-        scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
-    }
-
-    const int * toktypes = nullptr;
-    const int toktype_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE).c_str());
-    if (toktype_idx != -1) {
-        toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
-    }
-
     // determine vocab type
     {
         std::string tokenizer_name;
 
         ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_name);
 
-        if (tokenizer_name == "llama") {
+        if (tokenizer_name == "no_vocab") {
+            vocab.type = LLAMA_VOCAB_TYPE_NONE;
+
+            // default special tokens
+            vocab.special_bos_id = -1;
+            vocab.special_eos_id = -1;
+            vocab.special_unk_id = -1;
+            vocab.special_sep_id = -1;
+            vocab.special_pad_id = -1;
+            vocab.linefeed_id    = -1;
+
+            return;
+        } else if (tokenizer_name == "llama") {
             vocab.type = LLAMA_VOCAB_TYPE_SPM;
 
             // default special tokens
@@ -3790,6 +3788,23 @@ static void llm_load_vocab(
         }
     }
 
+    const int token_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_LIST).c_str());
+    if (token_idx == -1) {
+        throw std::runtime_error("cannot find tokenizer vocab in model file\n");
+    }
+
+    const float * scores = nullptr;
+    const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str());
+    if (score_idx != -1) {
+        scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
+    }
+
+    const int * toktypes = nullptr;
+    const int toktype_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE).c_str());
+    if (toktype_idx != -1) {
+        toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
+    }
+
     const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
 
     vocab.id_to_token.resize(n_vocab);
@@ -3997,7 +4012,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     LLAMA_LOG_INFO("%s: n_ff             = %u\n",     __func__, hparams.n_ff);
     LLAMA_LOG_INFO("%s: n_expert         = %u\n",     __func__, hparams.n_expert);
     LLAMA_LOG_INFO("%s: n_expert_used    = %u\n",     __func__, hparams.n_expert_used);
-    LLAMA_LOG_INFO("%s: causal attm      = %d\n",     __func__, hparams.causal_attn);
+    LLAMA_LOG_INFO("%s: causal attn      = %d\n",     __func__, hparams.causal_attn);
     LLAMA_LOG_INFO("%s: pooling type     = %d\n",     __func__, hparams.pooling_type);
     LLAMA_LOG_INFO("%s: rope type        = %d\n",     __func__, hparams.rope_type);
     LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type);
@@ -5095,7 +5110,8 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
 
         llm_load_print_meta(ml, model);
 
-        if (model.hparams.n_vocab != model.vocab.id_to_token.size()) {
+        if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
+            model.hparams.n_vocab != model.vocab.id_to_token.size()) {
             throw std::runtime_error("vocab size mismatch");
         }
 
@@ -9108,8 +9124,8 @@ static int llama_decode_internal(
     //llama_synchronize(&lctx);
 
     // decide if we need to defrag the kv cache
-    if (cparams.defrag_thold >= 0.0f) {
+    if (cparams.causal_attn && cparams.defrag_thold >= 0.0f) {
-        const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used + n_tokens_all)/float(kv_self.n) : 0.0f;
+        const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used)/float(kv_self.n) : 0.0f;
 
         // queue defragmentation for next llama_kv_cache_update
         if (fragmentation > cparams.defrag_thold) {
@@ -9141,6 +9157,11 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
     // number of cells moved
     uint32_t n_moves = 0;
 
+    // each move requires 6*n_layer tensors (see build_defrag)
+    //   - source view, destination view, copy operation
+    //   - x2 for keys and values
+    const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer);
+
     // determine which KV cells to move where
     //
     //  cell i moves to ids[i]
@@ -9167,15 +9188,6 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
             nh++;
         }
 
-        // each move requires 6*n_layer tensors (see build_defrag)
-        //   - source view, destination view, copy operation
-        //   - x2 for keys and values
-        //
-        if (6*(n_moves + nh)*n_layer >= LLAMA_MAX_NODES) {
-            // the graph is too big, we cannot move more cells
-            break;
-        }
-
         uint32_t nf = 0;
         uint32_t is = n_kv - 1;
 
@@ -9205,11 +9217,19 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
         // are we moving a continuous block of memory?
         bool cont = false;
 
+        // should we stop searching for the next move?
+        bool stop = false;
+
         // go back and move the nf cells to the hole
         for (; i1 < n_kv; ++i1) {
             auto & cell1 = kv_self.cells[i1];
 
             if (cell1.is_empty() || ids[i1] != n_kv) {
+                if (n_moves == max_moves) {
+                    stop = true;
+                    break;
+                }
+
                 cont = false;
                 continue;
             }
@@ -9236,6 +9256,10 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
             }
         }
 
+        if (stop || n_moves == max_moves) {
+            break;
+        }
+
         //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh);
 
         i0 += nh - 1;
@@ -9425,26 +9449,32 @@ static enum llama_vocab_type llama_vocab_get_type(const llama_vocab & vocab) {
 }
 
 static bool llama_is_normal_token(const llama_vocab & vocab, llama_token id) {
+    GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_NORMAL;
 }
 
 static bool llama_is_unknown_token(const llama_vocab & vocab, llama_token id) {
+    GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_UNKNOWN;
 }
 
 static bool llama_is_control_token(const llama_vocab & vocab, llama_token id) {
+    GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_CONTROL;
 }
 
 static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) {
+    GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_BYTE;
 }
 
 static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id) {
+    GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_USER_DEFINED;
 }
 
 static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
+    GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE);
     GGML_ASSERT(llama_is_byte_token(vocab, id));
     const auto& token_data = vocab.id_to_token.at(id);
     switch (llama_vocab_get_type(vocab)) {
@@ -9466,6 +9496,7 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
 }
 
 static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
+    GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE);
     static const char * hex = "0123456789ABCDEF";
     switch (llama_vocab_get_type(vocab)) {
         case LLAMA_VOCAB_TYPE_SPM: {
@@ -10527,6 +10558,8 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                     }
                 }
             } break;
+        case LLAMA_VOCAB_TYPE_NONE:
+            GGML_ASSERT(false);
     }
 
     return output;
@@ -12261,7 +12294,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
     return new_type;
 }
 
-static int32_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
+static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
     std::mutex mutex;
     int counter = 0;
     size_t new_size = 0;
@@ -13437,7 +13470,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
 }
 
 int32_t llama_n_vocab(const struct llama_model * model) {
-    return model->vocab.id_to_token.size();
+    return model->hparams.n_vocab;
 }
 
 int32_t llama_n_ctx_train(const struct llama_model * model) {
@@ -14271,14 +14304,17 @@ float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id
 }
 
 const char * llama_token_get_text(const struct llama_model * model, llama_token token) {
+    GGML_ASSERT(model->vocab.type != LLAMA_VOCAB_TYPE_NONE);
     return model->vocab.id_to_token[token].text.c_str();
 }
 
 float llama_token_get_score(const struct llama_model * model, llama_token token) {
+    GGML_ASSERT(model->vocab.type != LLAMA_VOCAB_TYPE_NONE);
     return model->vocab.id_to_token[token].score;
 }
 
 llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token) {
+    GGML_ASSERT(model->vocab.type != LLAMA_VOCAB_TYPE_NONE);
     return model->vocab.id_to_token[token].type;
 }
 

llama.h

@@ -59,9 +59,10 @@ extern "C" {
     typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
+        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
-        LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_SPM  = 1, // SentencePiece
-        LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
+        LLAMA_VOCAB_TYPE_BPE  = 2, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM  = 3, // WordPiece
     };
 
     // note: these values should be synchronized with ggml_rope