llm : add Falcon support (#2717)

* llama : refactor GGUF constants into static maps

* llama : check if model architecture is known

* llama : refactor llama_model_load_internal()

* gguf : add KV constant maps

* llm : read arch-specific KVs

* convert : add dummy scores + types

* falcon : load tensor data (CPU only)

* llama : fix loading progress bar

* llama : add arch member to llama_model

* falcon : CPU inference working

* falcon : support non-40B models

* falcon : minor

* llama : minor updates

ggml-ci

* convert-falcon-hf-to-gguf.py : fix special token mapping

* llama.cpp : llama default UNK token = id 0

* llama.cpp : fix bpe tokenizer

* llama.cpp : fix the fix of bpe tokenizer

* ggml : pass eps to ggml_norm

* metal : implement RoPE (mode = 2) + avoid ggml_repeat

* ggml : ggml_repeat always creates new tensor

* falcon : copy-paste self-attention from LLaMA

* metal : print extra compute pipeline info

* falcon : minor changes (still chasing the Metal problem)

* llama.cpp : fix linefeed token

* metal : fix GELU kernel numerical stability by using precise::tanh

* metal : temporary workaround for the concurrency optimization bug

* falcon : add CUDA offloading (#2739)

* llama : better model naming and size reporting

* llama : prep new tokenizer support

* llama : advanced BPE tokenizer based on ggllm.cpp imlpementation

* llama : remove oboslete comment

ggml-ci

* common : remove obsolete BPE API + disable test-tokenizer-1

* llama : revert BPE special-case in llama_byte_to_token()

* cuda : add TODOs for RoPE NeoX implementation

* llama : default special tokens based on vocab type

* perplexity : add log for start of tokenization

---------

Co-authored-by: klosax <131523366+klosax@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>

.devops/lamma-cpp-clblast.srpm.spec

@@ -0,0 +1,58 @@
+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp-clblast
+Version:        master
+Release:        1%{?dist}
+Summary:        OpenCL Inference of LLaMA model in pure C/C++
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git mesa-libOpenCL-devel
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j LLAMA_CLBLAST=1
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p main %{buildroot}%{_bindir}/llamacppclblast
+cp -p server %{buildroot}%{_bindir}/llamacppclblastserver
+cp -p simple %{buildroot}%{_bindir}/llamacppclblastsimple
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamacppclblast
+%{_bindir}/llamacppclblastserver
+%{_bindir}/llamacppclblastsimple
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

.devops/lamma-cpp-cublas.srpm.spec

@@ -0,0 +1,59 @@
+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp-cublas
+Version:        master
+Release:        1%{?dist}
+Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git cuda-toolkit
+Requires:       cuda-toolkit
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j LLAMA_CUBLAS=1
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p main %{buildroot}%{_bindir}/llamacppcublas
+cp -p server %{buildroot}%{_bindir}/llamacppcublasserver
+cp -p simple %{buildroot}%{_bindir}/llamacppcublassimple
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamacppcublas
+%{_bindir}/llamacppcublasserver
+%{_bindir}/llamacppcublassimple
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

.devops/llama-cpp.srpm.spec

@@ -0,0 +1,58 @@
+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp
+Version:        master
+Release:        1%{?dist}
+Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+
+%prep
+%autosetup
+
+%build
+make -j
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p main %{buildroot}%{_bindir}/llamacpp
+cp -p server %{buildroot}%{_bindir}/llamacppserver
+cp -p simple %{buildroot}%{_bindir}/llamacppsimple
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llamacpp
+%{_bindir}/llamacppserver
+%{_bindir}/llamacppsimple
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

.gitignore

@@ -3,6 +3,8 @@
 *.so
 *.gguf
 *.bin
+*.exe
+*.dll
 .DS_Store
 .build/
 .cache/
@@ -81,4 +83,3 @@ tests/test-quantize-fns
 tests/test-quantize-perf
 tests/test-sampling
 tests/test-tokenizer-0
-

README.md

@@ -39,6 +39,7 @@ Last revision compatible with the old format: [dadbed9](https://github.com/ggerg
         <li><a href="#memorydisk-requirements">Memory/Disk Requirements</a></li>
         <li><a href="#quantization">Quantization</a></li>
         <li><a href="#interactive-mode">Interactive mode</a></li>
+        <li><a href="#constrained-output-with-grammars">Constrained output with grammars</a></li>
         <li><a href="#instruction-mode-with-alpaca">Instruction mode with Alpaca</a></li>
         <li><a href="#using-openllama">Using OpenLLaMA</a></li>
         <li><a href="#using-gpt4all">Using GPT4All</a></li>
@@ -604,6 +605,16 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
     CHAT_SAVE_DIR=./chat/bob ./examples/chat-persistent.sh
 ```
 
+### Constrained output with grammars
+
+`llama.cpp` supports grammars to constrain model output. For example, you can force the model to output JSON only:
+
+```bash
+./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
+```
+
+The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
+
 ### Instruction mode with Alpaca
 
 1. First, download the `ggml` Alpaca model into the `./models` folder
@@ -885,3 +896,4 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
 - [BLIS](./docs/BLIS.md)
 - [Performance troubleshooting](./docs/token_generation_performance_tips.md)
 - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+- [GBNF grammars](./grammars/README.md)

common/common.cpp

@@ -387,11 +387,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #else
             fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
 #endif // GGML_USE_CUBLAS
-        } else if (arg == "--mul-mat-q" || arg == "-mmq") {
+        } else if (arg == "--no-mul-mat-q" || arg == "-nommq") {
 #ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = true;
+            params.mul_mat_q = false;
 #else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--low-vram" || arg == "-lv") {
 #ifdef GGML_USE_CUBLAS
@@ -417,6 +417,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.antiprompt.push_back(argv[i]);
         } else if (arg == "--perplexity") {
             params.perplexity = true;
+        } else if (arg == "--ppl-stride") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.ppl_stride = std::stoi(argv[i]);
+        } else if (arg == "--ppl-output-type") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.ppl_output_type = std::stoi(argv[i]);
         } else if (arg == "--hellaswag") {
             params.hellaswag = true;
         } else if (arg == "--hellaswag-tasks") {
@@ -599,11 +611,11 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "                        number of layers to store in VRAM\n");
     fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
-    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
-    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
-    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
-    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
+    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
+    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
+    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif
     fprintf(stdout, "  --mtest               compute maximum memory usage\n");
     fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
@@ -732,35 +744,3 @@ std::string llama_token_to_str(const struct llama_context * ctx, llama_token tok
 
     return std::string(result.data(), result.size());
 }
-
-std::vector<llama_token> llama_tokenize_bpe(
-        struct llama_context * ctx,
-           const std::string & text,
-                        bool   add_bos) {
-    int n_tokens = text.length() + add_bos;
-    std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize_bpe(ctx, text.c_str(), result.data(), result.size(), add_bos);
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        int check = llama_tokenize_bpe(ctx, text.c_str(), result.data(), result.size(), add_bos);
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
-    }
-    return result;
-}
-
-std::string llama_token_to_str_bpe(const struct llama_context * ctx, llama_token token) {
-    std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str_bpe(ctx, token, result.data(), result.size());
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        const int check = llama_token_to_str_bpe(ctx, token, result.data(), result.size());
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
-    }
-
-    return std::string(result.data(), result.size());
-}
-

common/common.h

@@ -64,11 +64,15 @@ struct gpt_params {
     std::string lora_adapter = "";  // lora adapter path
     std::string lora_base    = "";  // base model path for the lora adapter
 
+    int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
+    int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
+                                    //                                       (which is more convenient to use for plotting)
+                                    //
     bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
     bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
-    bool mul_mat_q         = false; // if true, use experimental mul_mat_q kernels
+    bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
@@ -116,15 +120,6 @@ std::vector<llama_token> llama_tokenize(
            const std::string & text,
                         bool   add_bos);
 
-std::vector<llama_token> llama_tokenize_bpe(
-        struct llama_context * ctx,
-           const std::string & text,
-                        bool   add_bos);
-
 std::string llama_token_to_str(
         const struct llama_context * ctx,
                        llama_token   token);
-
-std::string llama_token_to_str_bpe(
-    const struct llama_context * ctx,
-                   llama_token   token);

convert-falcon-hf-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 # HF falcon--> gguf conversion
 
 import gguf
@@ -94,14 +95,17 @@ print("gguf: get model metadata")
 
 block_count = hparams["n_layer"]
 
-gguf_writer.add_name(last_dir)
+gguf_writer.add_name("Falcon")
 gguf_writer.add_context_length(2048) # not in config.json
 gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
 gguf_writer.add_embedding_length(hparams["hidden_size"])
 gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
 gguf_writer.add_block_count(block_count)
 gguf_writer.add_head_count(hparams["n_head"])
-if "n_head_kv" in hparams: gguf_writer.add_head_count_kv(hparams["n_head_kv"])
+if "n_head_kv" in hparams:
+    gguf_writer.add_head_count_kv(hparams["n_head_kv"])
+else:
+    gguf_writer.add_head_count_kv(1)
 gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
 
 # TOKENIZATION
@@ -109,6 +113,8 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
 print("gguf: get tokenizer metadata")
 
 tokens: List[str] = []
+scores: List[float] = []
+toktypes: List[int] = []
 merges: List[str] = []
 
 
@@ -152,41 +158,30 @@ if Path(dir_model + "/tokenizer.json").is_file():
             text = bytearray(pad_token)
 
         tokens.append(text)
+        scores.append(0.0)                      # dymmy
+        toktypes.append(gguf.TokenType.NORMAL)  # dummy
 
     gguf_writer.add_token_list(tokens)
+    gguf_writer.add_token_scores(scores)
+    gguf_writer.add_token_types(toktypes)
 
-    if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file():
-        print("gguf: get special token ids")
+print("gguf: get special token ids")
+# Look for special tokens in config.json
 
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
+if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
+    gguf_writer.add_bos_token_id(hparams["bos_token_id"])
 
-        # find special token ids
+if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
+    gguf_writer.add_eos_token_id(hparams["eos_token_id"])
 
-        if "bos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]:
-                    gguf_writer.add_bos_token_id(key["id"])
+if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
+    gguf_writer.add_unk_token_id(hparams["unk_token_id"])
 
-        if "eos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]:
-                    gguf_writer.add_eos_token_id(key["id"])
+if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
+    gguf_writer.add_sep_token_id(hparams["sep_token_id"])
 
-        if "unk_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]:
-                    gguf_writer.add_pad_token_id(key["id"])
+if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
+    gguf_writer.add_pad_token_id(hparams["pad_token_id"])
 
 
 # TENSORS
@@ -194,8 +189,9 @@ if Path(dir_model + "/tokenizer.json").is_file():
 tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 
 # params for qkv transform
-n_head = hparams["n_head"]
+n_head    = hparams["n_head"]
 n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+
 head_dim = hparams["hidden_size"] // n_head
 
 # tensor info

convert-gptneox-hf-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 # HF gptneox--> gguf conversion
 
 import gguf

convert-llama-7b-pth-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 # 7b pth llama --> gguf conversion
 # Only models with a single datafile are supported, like 7B
 # HF files required in the model dir: config.json tokenizer_config.json tokenizer.json tokenizer.model

convert-llama-ggmlv3-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys, struct, math, argparse
 from pathlib import Path
 
@@ -93,7 +94,7 @@ class Tensor:
         pad = ((offset + 31) & ~31) - offset
         offset += pad
         n_elems = np.prod(self.dims)
-        n_bytes = (n_elems * tysize) // blksize
+        n_bytes = np.int64(np.int64(n_elems) * np.int64(tysize)) // np.int64(blksize)
         self.start_offset = offset
         self.len_bytes = n_bytes
         offset += n_bytes
@@ -215,15 +216,10 @@ class GGMLToGGUF:
         if self.vocab_override is not None:
             vo = self.vocab_override
             print('* Adding vocab item(s)')
-            for (idx, vitem) in enumerate(vo.all_tokens()):
-                if len(vitem) == 3:
-                    tokens.append(vitem[0])
-                    scores.append(vitem[1])
-                    toktypes.append(vitem[2])
-                else:
-                    # Maybe try to guess the token type here?
-                    tokens.append(vitem[0])
-                    scores.append(vitem[1])
+            for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
+                tokens.append(vbytes)
+                scores.append(score)
+                toktypes.append(ttype)
             assert len(tokens) == hp.n_vocab, f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
             gguf_writer.add_token_list(tokens)
             gguf_writer.add_token_scores(scores)
@@ -231,9 +227,21 @@ class GGMLToGGUF:
                 gguf_writer.add_token_types(toktypes)
             return
         print(f'* Adding {hp.n_vocab} vocab item(s)')
+        assert len(self.model.vocab.items) >= 3, 'Cannot handle unexpectedly short model vocab'
         for (tokid, (vbytes, vscore)) in enumerate(self.model.vocab.items):
             tt = 1 # Normal
-            if len(vbytes) == 0:
+            # Special handling for UNK, BOS, EOS tokens.
+            if tokid <= 2:
+                if tokid == 0:
+                    vbytes = b'<unk>'
+                    tt = 2
+                elif tokid == 1:
+                    vbytes = b'<s>'
+                    tt = 3
+                else:
+                    vbytes = b'</s>'
+                    tt = 3
+            elif len(vbytes) == 0:
                 tt = 3 # Control
             elif tokid >= 3 and tokid <= 258 and len(vbytes) == 1:
                 vbytes = bytes(f'<0x{vbytes[0]:02X}>', encoding = 'UTF-8')
@@ -246,6 +254,9 @@ class GGMLToGGUF:
         gguf_writer.add_token_list(tokens)
         gguf_writer.add_token_scores(scores)
         gguf_writer.add_token_types(toktypes)
+        gguf_writer.add_unk_token_id(0)
+        gguf_writer.add_bos_token_id(1)
+        gguf_writer.add_eos_token_id(2)
 
     def add_tensors(self, gguf_writer):
         nm = self.name_map
@@ -330,4 +341,5 @@ def main():
     converter.save()
     print(f'* Successful completion. Output saved to: {cfg.output}')
 
-main()
+if __name__ == '__main__':
+    main()

convert-llama-hf-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 # HF llama --> gguf conversion
 
 import gguf

convert-lora-to-ggml.py

@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 import json
 import os
 import re
@@ -6,23 +6,22 @@ import struct
 import sys
 from typing import Any, Dict, Sequence, TextIO
 
+import numpy as np
 import torch
 
-from convert import DATA_TYPE_TO_FTYPE, NUMPY_TYPE_TO_DATA_TYPE, DataType
+NUMPY_TYPE_TO_FTYPE: Dict[str, int] = {"float32": 0, "float16": 1}
+
 
 HF_SUBLAYER_TO_GGML = {
-    "self_attn.q_proj": "attention.wq",
-    "self_attn.k_proj": "attention.wk",
-    "self_attn.v_proj": "attention.wv",
-    "self_attn.o_proj": "attention.wo",
-    "mlp.gate_proj": "feed_forward.w1",
-    "mlp.down_proj": "feed_forward.w2",
-    "mlp.up_proj": "feed_forward.w3",
-    "input_layernorm": "attention_norm",
+    "self_attn.q_proj": "attn_q",
+    "self_attn.k_proj": "attn_k",
+    "self_attn.v_proj": "attn_v",
+    "self_attn.o_proj": "attn_output",
+    "mlp.gate_proj": "ffn_gate",
+    "mlp.down_proj": "ffn_down",
+    "mlp.up_proj": "ffn_up",
+    "input_layernorm": "attn_norm",
     "post_attention_layernorm": "ffn_norm",
-    # "norm": "norm",
-    # "embed_tokens": "tok_embeddings",
-    # "lm_head": "output",
 }
 
 
@@ -39,7 +38,7 @@ def translate_tensor_name(t: str) -> str:
             sys.exit(1)
 
         output_string = (
-            f"layers.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
+            f"blk.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
         )
         return output_string
     else:
@@ -54,12 +53,14 @@ def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
     # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int
     # but some models ship a float value instead
     # let's convert to int, but fail if lossless conversion is not possible
-    assert int(params["lora_alpha"]) == params["lora_alpha"], "cannot convert float to int losslessly"
+    assert (
+        int(params["lora_alpha"]) == params["lora_alpha"]
+    ), "cannot convert float to int losslessly"
     fout.write(struct.pack("i", int(params["lora_alpha"])))
 
 
 def write_tensor_header(
-    self, name: str, shape: Sequence[int], data_type: DataType
+    self, name: str, shape: Sequence[int], data_type: np.dtype
 ) -> None:
     sname = name.encode("utf-8")
     fout.write(
@@ -67,7 +68,7 @@ def write_tensor_header(
             "iii",
             len(shape),
             len(sname),
-            DATA_TYPE_TO_FTYPE[NUMPY_TYPE_TO_DATA_TYPE[data_type]],
+            NUMPY_TYPE_TO_FTYPE[data_type.name],
         )
     )
     fout.write(struct.pack("i" * len(shape), *shape[::-1]))

convert.py

@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 
 import gguf
 import argparse
@@ -69,7 +69,10 @@ SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
     'I32': DT_I32,
 }
 
-class GGMLFileType(enum.Enum):
+# TODO: match this with `llama_ftype`
+# TODO: rename to LLAMAFileType
+# TODO: move to `gguf.py`
+class GGMLFileType(enum.IntEnum):
     AllF32    = 0
     MostlyF16 = 1  # except 1d tensors
 
@@ -101,6 +104,8 @@ class Params:
     n_head_kv:  int
     f_norm_eps: float
 
+    ftype: Optional[GGMLFileType] = None
+
     @staticmethod
     def find_n_mult(n_ff: int, n_embd: int) -> int:
         # hardcoded magic range
@@ -728,7 +733,11 @@ class OutputFile:
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
     def add_meta_arch(self, params: Params) -> None:
-        self.gguf.add_name                ("LLaMA")
+        ver = None
+        if (params.n_ctx == 4096):
+            ver = "v2"
+
+        self.gguf.add_name                ("LLaMA" if ver == None else "LLaMA " + ver)
         self.gguf.add_context_length      (params.n_ctx)
         self.gguf.add_embedding_length    (params.n_embd)
         self.gguf.add_block_count         (params.n_layer)
@@ -738,6 +747,9 @@ class OutputFile:
         self.gguf.add_head_count_kv       (params.n_head_kv)
         self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
 
+        if params.ftype:
+            self.gguf.add_file_type(params.ftype)
+
     def add_meta_vocab(self, vocab: Vocab) -> None:
         tokens = []
         scores = []
@@ -956,7 +968,7 @@ def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, Sentence
             path = path3
         else:
             raise FileNotFoundError(
-                f"Could not find tokenizer.model in {path} or its parent; "
+                f"Could not find {vocab_file} in {path} or its parent; "
                 "if it's in another directory, pass the directory as --vocab-dir")
 
     print(f"Loading vocab file '{path}', type '{vocabtype}'")
@@ -1020,6 +1032,12 @@ def main(args_in: Optional[List[str]] = None) -> None:
                             " - LLaMA v2: --ctx 4096\n")
         params.n_ctx = args.ctx
 
+    if args.outtype:
+        params.ftype = {
+            "f32": GGMLFileType.AllF32,
+            "f16": GGMLFileType.MostlyF16,
+        }[args.outtype]
+
     print(f"params = {params}")
 
     vocab: Vocab
@@ -1040,11 +1058,14 @@ def main(args_in: Optional[List[str]] = None) -> None:
             vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
             vocab = load_vocab(vocab_dir, args.vocabtype)
 
-        model       = model_plus.model
-        model       = convert_model_names(model, params)
-        output_type = pick_output_type(model, args.outtype)
-        model       = convert_to_output_type(model, output_type)
-        outfile     = args.outfile or default_outfile(model_plus.paths, output_type)
+        model   = model_plus.model
+        model   = convert_model_names(model, params)
+        ftype   = pick_output_type(model, args.outtype)
+        model   = convert_to_output_type(model, ftype)
+        outfile = args.outfile or default_outfile(model_plus.paths, ftype)
+
+        params.ftype = ftype
+        print(f"Writing {outfile}, format {ftype}")
 
         OutputFile.write_all(outfile, params, model, vocab)
         print(f"Wrote {outfile}")

examples/convert-llama2c-to-ggml/README.md

@@ -12,15 +12,19 @@ usage: ./convert-llama2c-to-ggml [options]
 
 options:
   -h, --help                       show this help message and exit
-  --copy-vocab-from-model FNAME    model path from which to copy vocab (default 'models/ggml-vocab.bin')
+  --copy-vocab-from-model FNAME    model path from which to copy vocab (default 'tokenizer.bin')
   --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model
   --llama2c-output-model FNAME     model path to save the converted llama2.c model (default ak_llama_model.bin')
 ```
 
-An example command is as follows:
+An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
 
-`$ ./convert-llama2c-to-ggml --copy-vocab-from-model <ggml-vocab.bin> --llama2c-model <llama2.c model path> --llama2c-output-model <ggml output model path>`
+`$ ./convert-llama2c-to-ggml --copy-vocab-from-model ../llama2.c/tokenizer.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.ggmlv3.bin`
 
-Now you can use the model with command like:
+For now the generated model is in the legacy GGJTv3 format, so you need to convert it to gguf manually:
 
-`$ ./main -m <ggml output model path> -p "One day, Lily met a Shoggoth" -n 500 -c 256 -eps 1e-5`
+`$ python ./convert-llama-ggmlv3-to-gguf.py --eps 1e-5 --input stories42M.ggmlv3.bin --output stories42M.gguf.bin`
+
+Now you can use the model with a command like:
+
+`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -17,6 +17,9 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
+#define LLAMA_FILE_VERSION_GGJT_V3   3
+
 //////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc.
 typedef struct {
     int dim; // transformer dimension
@@ -49,10 +52,10 @@ typedef struct {
     // float* freq_cis_real; // (seq_len, dim/2)
     // float* freq_cis_imag; // (seq_len, dim/2)
     // (optional) classifier weights for the logits, on the last layer
-    //float* wcls;
+    float* wcls;
 } TransformerWeights;
 
-void malloc_weights(TransformerWeights* w, Config* p) {
+void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
     // we calloc instead of malloc to keep valgrind happy
     w->token_embedding_table = new float[p->vocab_size * p->dim]();
     printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
@@ -86,9 +89,16 @@ void malloc_weights(TransformerWeights* w, Config* p) {
 
     w->rms_final_weight = new float[p->dim]();
     printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+
+    if (shared_weights) {
+        w->wcls = NULL;
+    } else {
+        w->wcls = new float[p->vocab_size * p->dim]();
+        printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+    }
 }
 
-int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
+int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
     if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
     if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
     if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
@@ -100,6 +110,22 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
     if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
     if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
     if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+
+    // Skip freq_cis_real & freq_cis_imag
+    int head_size = p->dim / p->n_heads;
+    fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
+
+    if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+
+    // Check we didn't forget to read anything
+    auto curr = ftell(f);
+    fseek(f, 0, SEEK_END);
+    auto end = ftell(f);
+    if (curr != end) {
+        printf("Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", curr, end);
+        return 1;
+    }
+
     return 0;
 }
 
@@ -115,6 +141,7 @@ void free_weights(TransformerWeights* w) {
     delete w->w2;
     delete w->w3;
     delete w->rms_final_weight;
+    if (w->wcls) delete w->wcls;
 }
 
 void print_sample_weights(TransformerWeights *w){
@@ -131,6 +158,7 @@ void print_sample_weights(TransformerWeights *w){
     printf("%f\n", w->w2[0]);
     printf("%f\n", w->w3[0]);
     printf("%f\n", w->rms_att_weight[0]);
+    if (w->wcls) printf("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -509,26 +537,28 @@ bool is_ggml_file(const char *filename) {
 }
 
 void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
-    // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary
-    if (is_ggml_file(filename)) {
-
-        struct llama_context_params llama_params = llama_context_default_params();
-        llama_params.vocab_only = true;
-
-        struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
-        struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
-
-        const int n_vocab = llama_n_vocab(lctx);
-        vocab->id_to_token.resize(n_vocab);
-        for (int i=0; i<n_vocab; ++i) {
-            vocab->id_to_token[i].text  = llama_token_get_text(lctx, i);
-            vocab->id_to_token[i].score = llama_token_get_score(lctx, i);
-            vocab->id_to_token[i].type  = llama_token_get_type(lctx, i);
-            vocab->token_to_id.emplace(vocab->id_to_token[i].text, i);
-        }
-        llama_free(lctx);
-        llama_free_model(lmodel);
-    } else { // assume llama2.c vocabulary
+#pragma message("TODO: implement reading vocabulary using gguf")
+//    // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary
+//    if (is_ggml_file(filename)) {
+//
+//        struct llama_context_params llama_params = llama_context_default_params();
+//        llama_params.vocab_only = true;
+//
+//        struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
+//        struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
+//
+//        const int n_vocab = llama_n_vocab(lctx);
+//        vocab->id_to_token.resize(n_vocab);
+//        for (int i=0; i<n_vocab; ++i) {
+//            vocab->id_to_token[i].text  = llama_token_get_text(lctx, i);
+//            vocab->id_to_token[i].score = llama_token_get_score(lctx, i);
+//            vocab->id_to_token[i].type  = llama_token_get_type(lctx, i);
+//            vocab->token_to_id.emplace(vocab->id_to_token[i].text, i);
+//        }
+//        llama_free(lctx);
+//        llama_free_model(lmodel);
+//    } else
+    { // assume llama2.c vocabulary
         printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename);
         llama_file file(filename, "rb");
         const int  n_vocab = config->vocab_size;
@@ -538,6 +568,12 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
+            // Special-case handling of <0xXX> single byte tokens.
+            char byte_val;
+            if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
+                char cstr[2] = { byte_val, 0 };
+                text = cstr;
+            }
             vocab->id_to_token[i].text = text;
             vocab->id_to_token[i].score = score;
             vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED;
@@ -589,83 +625,80 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
     }
 
 #pragma message("TODO: implement file saving using gguf")
-    (void) vocab;
-    (void) model;
-    (void) w;
-//    // write_magic
-//    file.write_u32(LLAMA_FILE_MAGIC);   // magic
-//    file.write_u32(LLAMA_FILE_VERSION); // version
-//    // write_hparams
-//    file.write_u32(model->hparams.n_vocab);
-//    file.write_u32(model->hparams.n_embd);
-//    file.write_u32(model->hparams.n_mult);
-//    file.write_u32(model->hparams.n_head);
-//    file.write_u32(model->hparams.n_layer);
-//    file.write_u32(model->hparams.n_rot);
-//    file.write_u32(LLAMA_FTYPE_ALL_F32);
-//
-//    // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
-//    uint32_t n_vocab = model->hparams.n_vocab;
-//    for (uint32_t i = 0; i < n_vocab; i++) {
-//        const auto & token_data = vocab->id_to_token.at(i);
-//        file.write_u32((uint32_t) token_data.tok.size());
-//        file.write_raw(token_data.tok.data(), token_data.tok.size());
-//        file.write_raw(&token_data.score, sizeof(token_data.score));
-//    }
-//
-//    // stuff AK weights into GG weights one by one.
-//    // w->token_embedding_table -> model->tok_embeddings
-//    // float*                   -> struct ggml_tensor
-//    stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
-//    stuff_karpathy_weights_into_gg(model->output, w->token_embedding_table);
-//
-//    stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
-//    //print_row(model->norm, 0);
-//
-//    // for rms-att-weight
-//    int row_length = model->hparams.n_embd;
-//    const auto & hparams = model->hparams;
-//    //int n_ff = model->hparams.n_embd;
-//    int n_ff = get_n_ff(&hparams);
-//
-//    for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
-//        auto & layer = model->layers[i];
-//        // 1d
-//        stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.ffn_norm      , &w->rms_ffn_weight[i*row_length]);
-//
-//        // from 3d matrix layer x dim x dim to 2d matrix dim x dim
-//        stuff_karpathy_weights_into_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wv            , &w->wv[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wo            , &w->wo[i*row_length*row_length]);
-//
-//        stuff_karpathy_weights_into_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
-//        stuff_karpathy_weights_into_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
-//    }
-//    // write tensors
-//    write_tensor(&file, model->tok_embeddings);
-//    write_tensor(&file, model->norm);
-//    write_tensor(&file, model->output); // ?
-//    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
-//        auto & layer = model->layers[i];
-//
-//        write_tensor(&file, layer.attention_norm);
-//        write_tensor(&file, layer.wq);
-//        write_tensor(&file, layer.wk);
-//        write_tensor(&file, layer.wv);
-//        write_tensor(&file, layer.wo);
-//        write_tensor(&file, layer.ffn_norm);
-//        write_tensor(&file, layer.w1);
-//        write_tensor(&file, layer.w2);
-//        write_tensor(&file, layer.w3);
-//    }
+    // write_magic
+    file.write_u32(LLAMA_FILE_MAGIC_GGJT);   // magic
+    file.write_u32(LLAMA_FILE_VERSION_GGJT_V3); // version
+    // write_hparams
+    file.write_u32(model->hparams.n_vocab);
+    file.write_u32(model->hparams.n_embd);
+    file.write_u32(model->hparams.n_mult);
+    file.write_u32(model->hparams.n_head);
+    file.write_u32(model->hparams.n_layer);
+    file.write_u32(model->hparams.n_rot);
+    file.write_u32(LLAMA_FTYPE_ALL_F32);
+
+    // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
+    uint32_t n_vocab = model->hparams.n_vocab;
+    for (uint32_t i = 0; i < n_vocab; i++) {
+        const auto & token_data = vocab->id_to_token.at(i);
+        file.write_u32((uint32_t) token_data.text.size());
+        file.write_raw(token_data.text.data(), token_data.text.size());
+        file.write_raw(&token_data.score, sizeof(token_data.score));
+    }
+
+    // stuff AK weights into GG weights one by one.
+    // w->token_embedding_table -> model->tok_embeddings
+    // float*                   -> struct ggml_tensor
+    stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
+    stuff_karpathy_weights_into_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+
+    stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
+    //print_row(model->norm, 0);
+
+    // for rms-att-weight
+    int row_length = model->hparams.n_embd;
+    const auto & hparams = model->hparams;
+    //int n_ff = model->hparams.n_embd;
+    int n_ff = get_n_ff(&hparams);
+
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
+        auto & layer = model->layers[i];
+        // 1d
+        stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
+        stuff_karpathy_weights_into_gg(layer.ffn_norm      , &w->rms_ffn_weight[i*row_length]);
+
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim
+        stuff_karpathy_weights_into_gg(layer.wq            , &w->wq[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wk            , &w->wk[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wv            , &w->wv[i*row_length*row_length]);
+        stuff_karpathy_weights_into_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+
+        stuff_karpathy_weights_into_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
+        stuff_karpathy_weights_into_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
+        stuff_karpathy_weights_into_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
+    }
+    // write tensors
+    write_tensor(&file, model->tok_embeddings);
+    write_tensor(&file, model->norm);
+    write_tensor(&file, model->output); // ?
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        write_tensor(&file, layer.attention_norm);
+        write_tensor(&file, layer.wq);
+        write_tensor(&file, layer.wk);
+        write_tensor(&file, layer.wv);
+        write_tensor(&file, layer.wo);
+        write_tensor(&file, layer.ffn_norm);
+        write_tensor(&file, layer.w1);
+        write_tensor(&file, layer.w2);
+        write_tensor(&file, layer.w3);
+    }
 }
 
 struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "models/ggml-vocab.bin";
+    params.fn_vocab_model    = "tokenizer.bin";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
     params.fn_train_data     = "shakespeare.txt";
     params.fn_checkpoint_in  = "checkpoint.bin";
@@ -718,7 +751,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params)
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h, --help                       show this help message and exit\n");
-    fprintf(stderr, "  --copy-vocab-from-model FNAME    llama2.c vocabulary or ggml model path from which to copy vocab (default '%s')\n", params->fn_vocab_model);
+    fprintf(stderr, "  --copy-vocab-from-model FNAME    llama2.c vocabulary or ggmlv3 model path from which to copy vocab (default '%s')\n", params->fn_vocab_model);
     fprintf(stderr, "  --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model\n");
     fprintf(stderr, "  --llama2c-output-model FNAME     model path to save the converted llama2.c model (default %s')\n", params->fn_llama2c_output_model);
     fprintf(stderr, "\n");
@@ -791,9 +824,12 @@ int main(int argc, char ** argv) {
         if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
         // read in the config header
         if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        auto shared_weights = config.vocab_size > 0;
+        config.vocab_size = abs(config.vocab_size);
+
         // read in the Transformer weights
-        malloc_weights(&weights, &config);
-        if(checkpoint_init_weights(&weights, &config, file)) { return 1; }
+        malloc_weights(&weights, &config, shared_weights);
+        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
         fclose(file);
     }
 

examples/embd-input/embd_input.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import ctypes
 from ctypes import cdll, c_char_p, c_void_p, POINTER, c_float, c_int
 import numpy as np

examples/embd-input/llava.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys
 import os
 sys.path.insert(0, os.path.dirname(__file__))

examples/embd-input/minigpt4.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys
 import os
 sys.path.insert(0, os.path.dirname(__file__))

examples/embd-input/panda_gpt.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import sys
 import os
 sys.path.insert(0, os.path.dirname(__file__))

examples/jeopardy/graph.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import matplotlib.pyplot as plt
 import os
 import csv

examples/json-schema-to-grammar.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import argparse
 import json
 import re

examples/main/README.md

@@ -288,6 +288,10 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 -   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation.
 
+### Grammars
+
+-   `--grammar GRAMMAR`, `--grammar-file FILE`: Specify a grammar (defined inline or in a file) to constrain model output to a specific format. For example, you could force the model to output JSON or to speak only in emojis. See the [GBNF guide](../../grammars/README.md) for details on the syntax.
+
 ### Quantization
 
 For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-data--run).

examples/main/main.cpp

@@ -43,7 +43,7 @@ static bool is_interacting = false;
 void sigint_handler(int signo) {
     if (signo == SIGINT) {
         if (!is_interacting) {
-            is_interacting=true;
+            is_interacting = true;
         } else {
             console::cleanup();
             printf("\n");
@@ -189,23 +189,30 @@ int main(int argc, char ** argv) {
         }
     }
 
+    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+
     // tokenize the prompt
     std::vector<llama_token> embd_inp;
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
-        embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+        embd_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
     } else {
         embd_inp = session_tokens;
     }
 
+    // Should not run without any tokens
+    if (embd_inp.empty()) {
+        embd_inp.push_back(llama_token_bos(ctx));
+    }
+
     // Tokenize negative prompt
     std::vector<llama_token> guidance_inp;
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
         params.cfg_negative_prompt.insert(0, 1, ' ');
-        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, true);
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, is_spm);
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true);
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
     }
@@ -252,8 +259,8 @@ int main(int argc, char ** argv) {
     }
 
     // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true);
-    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false);
+    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", is_spm);
+    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
 
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {

examples/make-ggml.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 This script converts Hugging Face llama models to GGML and quantizes them.
 

examples/perplexity/perplexity.cpp

@@ -27,12 +27,136 @@ std::vector<float> softmax(const std::vector<float>& logits) {
     return probs;
 }
 
-void perplexity(llama_context * ctx, const gpt_params & params) {
+void perplexity_v2(llama_context * ctx, const gpt_params & params) {
     // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
-    auto tokens = ::llama_tokenize(ctx, params.prompt, true);
+
+    if (params.ppl_stride <= 0) {
+        fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride);
+        return;
+    }
+
+    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = is_spm;
+
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    const int calc_chunk = params.n_ctx;
+
+    fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
+
+    if (int(tokens.size()) <= calc_chunk) {
+        fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
+                tokens.size(), params.n_ctx, params.ppl_stride);
+        return;
+    }
+
+    const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1)  / params.ppl_stride;
+
+    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
+    const int n_vocab = llama_n_vocab(ctx);
+    const int n_batch = params.n_batch;
+
+    int count = 0;
+    double nll = 0.0;
+
+    fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
+
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * params.ppl_stride;
+        const int end   = start + calc_chunk;
+
+        const int num_batches = (calc_chunk + n_batch - 1) / n_batch;
+        //fprintf(stderr, "%s: evaluating %d...%d using %d batches\n", __func__, start, end, num_batches);
+
+        std::vector<float> logits;
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
+        for (int j = 0; j < num_batches; ++j) {
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            //fprintf(stderr, "    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
+            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+                //fprintf(stderr, "%s : failed to eval\n", __func__);
+                return;
+            }
+
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
+
+            // add BOS token for the first batch of each chunk
+            if (add_bos && j == 0) {
+                tokens[batch_start] = llama_token_bos(ctx);
+            }
+
+            const auto batch_logits = llama_get_logits(ctx);
+            logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+
+            if (j == 0) {
+                tokens[batch_start] = token_org;
+            }
+        }
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
+        if (i == 0) {
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
+            if (total_seconds >= 60*60) {
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
+        }
+
+        //fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
+        for (int j = params.n_ctx - params.ppl_stride - 1; j < params.n_ctx - 1; ++j) {
+
+            // Calculate probability of next token, given the previous ones.
+            const std::vector<float> tok_logits(
+                logits.begin() + (j + 0) * n_vocab,
+                logits.begin() + (j + 1) * n_vocab);
+
+            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
+
+            nll += -std::log(prob);
+            ++count;
+        }
+        // perplexity is e^(average negative log-likelihood)
+        if (params.ppl_output_type == 0) {
+            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        } else {
+            printf("%8d  %.4lf\n", i*params.ppl_stride, std::exp(nll / count));
+        }
+        fflush(stdout);
+    }
+    printf("\n");
+}
+
+void perplexity(llama_context * ctx, const gpt_params & params) {
+    if (params.ppl_stride > 0) {
+        perplexity_v2(ctx, params);
+        return;
+    }
+
+    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
+    // Output: `perplexity: 13.5106 [114/114]`
+    // BOS tokens will be added for each chunk before eval
+
+    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = is_spm;
+
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 
     const int n_chunk_max = tokens.size() / params.n_ctx;
 
@@ -63,7 +187,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
             const auto token_org = tokens[batch_start];
 
             // add BOS token for the first batch of each chunk
-            if (j == 0) {
+            if (add_bos && j == 0) {
                 tokens[batch_start] = llama_token_bos(ctx);
             }
 
@@ -116,7 +240,11 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
             ++count;
         }
         // perplexity is e^(average negative log-likelihood)
-        printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        if (params.ppl_output_type == 0) {
+            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        } else {
+            printf("%8d  %.4lf\n", i*params.n_ctx, std::exp(nll / count));
+        }
         fflush(stdout);
     }
     printf("\n");
@@ -177,8 +305,10 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     size_t hs_task_count = prompt_lines.size()/6;
     fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
 
+    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+
     // This is needed as usual for LLaMA models
-    bool prepend_bos = true;
+    const bool add_bos = is_spm;
 
     // Number of tasks to use when computing the score
     if ( params.hellaswag_tasks < hs_task_count  ) {
@@ -234,14 +364,13 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     std::vector<float> tok_logits(n_vocab);
 
     for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
-
         // Tokenize the context to count tokens
-        std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, prepend_bos);
+        std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos);
         size_t context_size = context_embd.size();
 
         // Do the 1st ending
         // In this case we include the context when evaluating
-        auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], prepend_bos);
+        auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
         auto query_size = query_embd.size();
         //printf("First query: %d\n",(int)query_size);
 
@@ -369,6 +498,12 @@ int main(int argc, char ** argv) {
     params.perplexity = true;
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
+    if (params.ppl_stride > 0) {
+        fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
+                params.n_ctx, params.n_ctx + params.ppl_stride/2);
+        params.n_ctx += params.ppl_stride/2;
+    }
+
     if (params.n_ctx > 2048) {
         fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
                 "expect poor results\n", __func__, params.n_ctx);

examples/quantize/quantize.cpp

@@ -14,25 +14,25 @@ struct quant_option {
 };
 
 static const std::vector<struct quant_option> QUANT_OPTIONS = {
-    { "Q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0,   " 3.50G, +0.2499 ppl @ 7B", },
-    { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1846 ppl @ 7B", },
-    { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.30G, +0.0796 ppl @ 7B", },
-    { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0415 ppl @ 7B", },
+    { "Q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0,   " 3.56G, +0.2166 ppl @ LLaMA-v1-7B", },
+    { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
+    { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
+    { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
 #ifdef GGML_USE_K_QUANTS
-    { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.67G, +0.8698 ppl @ 7B", },
+    { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
-    { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5505 ppl @ 7B", },
-    { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.06G, +0.2437 ppl @ 7B", },
-    { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1803 ppl @ 7B", },
+    { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
+    { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
+    { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
     { "Q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
-    { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.56G, +0.1149 ppl @ 7B", },
-    { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0535 ppl @ 7B", },
+    { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
+    { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
     { "Q5_K",   LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", },
-    { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0353 ppl @ 7B", },
-    { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0142 ppl @ 7B", },
-    { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, +0.0044 ppl @ 7B", },
+    { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
+    { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
+    { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", },
 #endif
-    { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ 7B", },
+    { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
 };

examples/reason-act.sh

@@ -1,4 +1,3 @@
-
 #!/bin/bash
 
 cd `dirname $0`

examples/server/README.md

@@ -126,7 +126,7 @@ node .
 
     `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
 
-    `prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. A space is inserted in the front like main.cpp does.
+    `prompt`: Provide a prompt as a string, or as an array of strings and numbers representing tokens. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. If the prompt is a string, or an array with the first element given as a string, a space is inserted in the front like main.cpp does.
 
     `stop`: Specify a JSON array of stopping strings.
     These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).

examples/server/api_like_OAI.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import argparse
 from flask import Flask, jsonify, request, Response
 import urllib.parse

examples/server/server.cpp

@@ -190,6 +190,7 @@ struct llama_server_context
     size_t n_past = 0;
     size_t n_remain = 0;
 
+    json prompt;
     std::vector<llama_token> embd;
     std::vector<llama_token> last_n_tokens;
 
@@ -267,6 +268,53 @@ struct llama_server_context
         return true;
     }
 
+    std::vector<llama_token> tokenize(json json_prompt, bool add_bos)
+    {
+        // If `add_bos` is true, we only add BOS, when json_prompt is a string,
+        // or the first element of the json_prompt array is a string.
+        std::vector<llama_token> prompt_tokens;
+
+        if (json_prompt.is_array())
+        {
+            bool first = true;
+            for (const auto& p : json_prompt)
+            {
+                if (p.is_string())
+                {
+                    auto s = p.template get<std::string>();
+                    std::vector<llama_token> p;
+                    if (first)
+                    {
+                        s.insert(0, 1, ' '); // add a space if it's the first
+                        p = ::llama_tokenize(ctx, s, add_bos);
+                        first = false;
+                    }
+                    else
+                    {
+                        p = ::llama_tokenize(ctx, s, false);
+                    }
+                    prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
+                }
+                else
+                {
+                    if (first)
+                    {
+                        first = false;
+                    }
+                    prompt_tokens.push_back(p.template get<llama_token>());
+                }
+            }
+        }
+        else
+        {
+            auto s = json_prompt.template get<std::string>();
+            s.insert(0, 1, ' '); // always add a first space
+            prompt_tokens = ::llama_tokenize(ctx, s, add_bos);
+        }
+
+        return prompt_tokens;
+    }
+
     bool loadGrammar()
     {
         if (!params.grammar.empty()) {
@@ -294,8 +342,8 @@ struct llama_server_context
 
     void loadPrompt()
     {
-        params.prompt.insert(0, 1, ' '); // always add a first space
-        std::vector<llama_token> prompt_tokens = ::llama_tokenize(ctx, params.prompt, true);
+        auto prompt_tokens = tokenize(prompt, true);  // always add BOS
+
         num_prompt_tokens = prompt_tokens.size();
 
         if (params.n_keep < 0)
@@ -671,12 +719,11 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     fprintf(stdout, "                        number of layers to store in VRAM\n");
     fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
     fprintf(stdout, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
-    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
-    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
-    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
+    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
+    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif
     fprintf(stdout, "  -m FNAME, --model FNAME\n");
     fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
@@ -867,12 +914,12 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {});
 #endif // GGML_USE_CUBLAS
         }
-        else if (arg == "--mul-mat-q" || arg == "-mmq")
+        else if (arg == "--no-mul-mat-q" || arg == "-nommq")
         {
 #ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = true;
+            params.mul_mat_q = false;
 #else
-            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n", {});
+            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--main-gpu" || arg == "-mg")
@@ -1017,7 +1064,7 @@ static json format_final_response(llama_server_context &llama, const std::string
         {"tokens_predicted", llama.num_tokens_predicted},
         {"tokens_evaluated", llama.num_prompt_tokens},
         {"generation_settings", format_generation_settings(llama)},
-        {"prompt", llama.params.prompt},
+        {"prompt", llama.prompt},
         {"truncated", llama.truncated},
         {"stopped_eos", llama.stopped_eos},
         {"stopped_word", llama.stopped_word},
@@ -1086,10 +1133,18 @@ static void parse_options_completion(const json &body, llama_server_context &lla
     llama.params.penalize_nl = json_value(body, "penalize_nl", default_params.penalize_nl);
     llama.params.n_keep = json_value(body, "n_keep", default_params.n_keep);
     llama.params.seed = json_value(body, "seed", default_params.seed);
-    llama.params.prompt = json_value(body, "prompt", default_params.prompt);
     llama.params.grammar = json_value(body, "grammar", default_params.grammar);
     llama.params.n_probs = json_value(body, "n_probs", default_params.n_probs);
 
+    if (body.count("prompt") != 0)
+    {
+        llama.prompt = body["prompt"];
+    }
+    else
+    {
+        llama.prompt = "";
+    }
+
     llama.params.logit_bias.clear();
     if (json_value(body, "ignore_eos", false))
     {
@@ -1346,8 +1401,11 @@ int main(int argc, char **argv)
         auto lock = llama.lock();
 
         const json body = json::parse(req.body);
-        const std::string content = json_value<std::string>(body, "content", "");
-        const std::vector<llama_token> tokens = llama_tokenize(llama.ctx, content, false);
+        std::vector<llama_token> tokens;
+        if (body.count("content") != 0)
+        {
+            tokens = llama.tokenize(body["content"], false);
+        }
         const json data = format_tokenizer_response(tokens);
         return res.set_content(data.dump(), "application/json"); });
 
@@ -1359,7 +1417,14 @@ int main(int argc, char **argv)
 
         llama.rewind();
         llama_reset_timings(llama.ctx);
-        llama.params.prompt = json_value<std::string>(body, "content", "");
+        if (body.count("content") != 0)
+        {
+            llama.prompt = body["content"];
+        }
+        else
+        {
+            llama.prompt = "";
+        }
         llama.params.n_predict = 0;
         llama.loadPrompt();
         llama.beginCompletion();

ggml-alloc.c

@@ -238,7 +238,7 @@ static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_t
     alloc->n_free_blocks++;
 }
 
-void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n) {
+void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) {
     int pos = 0;
     for (int i = 0; i < n; i++) {
         if (list[i] != -1) {
@@ -547,7 +547,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                         struct ggml_tensor * view_src = get_view_source(parent);
                         struct hash_node * view_src_hn = hash_get(ht, view_src);
                         view_src_hn->n_views -= 1;
-                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
+                        AT_PRINTF("view_src %s\n", view_src->name);
                         if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
                             ggml_allocator_free_tensor(alloc, view_src);
                         }

ggml-alloc.h

@@ -12,7 +12,7 @@ GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
 
 // tell the allocator to parse nodes following the order described in the list
 // you should call this if your graph are optimized to execute out-of-order
-GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n);
+GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n);
 
 GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
 GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);

ggml-cuda.cu

@@ -287,7 +287,7 @@ static int g_device_count = -1;
 static int g_main_device = 0;
 static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES];
 static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
-static bool g_mul_mat_q = false;
+static bool g_mul_mat_q = true;
 
 static void * g_scratch_buffer = nullptr;
 static size_t g_scratch_size = 1024*1024*1024; // 1 GB by default
@@ -3907,6 +3907,29 @@ static __global__ void rope_f32(const float * x, float * dst, const int ncols, c
     dst[i + 1] = x0*sin_theta + x1*cos_theta;
 }
 
+// TODO: this implementation is wrong!
+//static __global__ void rope_neox_f32(const float * x, float * dst, const int ncols, const float p0,
+//                                const float p_delta, const int p_delta_rows, const float theta_scale) {
+//    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+//
+//    if (col >= ncols) {
+//        return;
+//    }
+//
+//    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+//    const int i = row*ncols + col/2;
+//
+//    const float theta = (p0 + p_delta * (row/p_delta_rows))*powf(theta_scale, col/2);
+//    const float sin_theta = sinf(theta);
+//    const float cos_theta = cosf(theta);
+//
+//    const float x0 = x[i + 0];
+//    const float x1 = x[i + ncols/2];
+//
+//    dst[i + 0]       = x0*cos_theta - x1*sin_theta;
+//    dst[i + ncols/2] = x0*sin_theta + x1*cos_theta;
+//}
+
 static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const float p, const float block_p, const float theta_scale) {
     const int col = blockDim.x*blockIdx.x + threadIdx.x;
     const int half_n_dims = ncols/4;
@@ -3979,24 +4002,29 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int
 
 // the CUDA soft max implementation differs from the CPU implementation
 // instead of doubles floats are used
-// values are also not normalized to the maximum value by subtracting it in the exponential function
-// theoretically these changes could cause problems with rounding error and arithmetic overflow but for LLaMa it seems to be fine
 static __global__ void soft_max_f32(const float * x, float * dst, const int ncols) {
     const int row = blockDim.x*blockIdx.x + threadIdx.x;
     const int block_size = blockDim.y;
     const int tid = threadIdx.y;
 
-    float tmp = 0.0;
-
-    for (int block_start = 0; block_start < ncols; block_start += block_size) {
-        const int col = block_start + tid;
-
-        if (col >= ncols) {
-            break;
-        }
+    float max_val = -INFINITY;
 
+    for (int col = tid; col < ncols; col += block_size) {
         const int i = row*ncols + col;
-        const float val = expf(x[i]);
+        max_val = max(max_val, x[i]);
+    }
+
+    // find the max value in the block
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        max_val = max(max_val, __shfl_xor_sync(0xffffffff, max_val, mask, 32));
+    }
+
+    float tmp = 0.f;
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const int i = row*ncols + col;
+        const float val = expf(x[i] - max_val);
         tmp += val;
         dst[i] = val;
     }
@@ -4007,15 +4035,11 @@ static __global__ void soft_max_f32(const float * x, float * dst, const int ncol
         tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
     }
 
-    for (int block_start = 0; block_start < ncols; block_start += block_size) {
-        const int col = block_start + tid;
-
-        if (col >= ncols) {
-            break;
-        }
+    const float inv_tmp = 1.f / tmp;
 
+    for (int col = tid; col < ncols; col += block_size) {
         const int i = row*ncols + col;
-        dst[i] /= tmp;
+        dst[i] *= inv_tmp;
     }
 }
 
@@ -5514,7 +5538,8 @@ inline void ggml_cuda_op_rope(
 
     const float theta_scale = powf(freq_base, -2.0f/n_dims);
 
-    const bool is_glm = mode & 4;
+    const bool is_neox = mode & 2;
+    const bool is_glm  = mode & 4;
 
     // compute
     if (is_glm) {
@@ -5522,6 +5547,9 @@ inline void ggml_cuda_op_rope(
         const float id_p = min(p, n_ctx - 2.f);
         const float block_p = max(p - (n_ctx - 2.f), 0.f);
         rope_glm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, id_p, block_p, theta_scale, cudaStream_main);
+    } else if (is_neox) {
+        GGML_ASSERT(false && "RoPE NeoX not implemented yet");
+#pragma message("TODO: implement RoPE NeoX for CUDA")
     } else {
         const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale;
         rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main);

ggml-metal.m

@@ -167,7 +167,9 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #define GGML_METAL_ADD_KERNEL(name) \
         ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
         ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
-        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name); \
+        fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
+                (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
+                (int) ctx->pipeline_##name.threadExecutionWidth); \
         if (error) { \
             fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
             return NULL; \
@@ -218,12 +220,12 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #undef GGML_METAL_ADD_KERNEL
     }
 
-    fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-    fprintf(stderr, "%s: hasUnifiedMemory             = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    fprintf(stderr, "%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    fprintf(stderr, "%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
     if (ctx->device.maxTransferRate != 0) {
-        fprintf(stderr, "%s: maxTransferRate              = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+        fprintf(stderr, "%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
     } else {
-        fprintf(stderr, "%s: maxTransferRate              = built-in GPU\n", __func__);
+        fprintf(stderr, "%s: maxTransferRate               = built-in GPU\n", __func__);
     }
 
     return ctx;
@@ -537,8 +539,8 @@ void ggml_metal_graph_compute(
 
             id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
 
-            const int node_start =                                  (cb_idx + 0) * n_nodes_per_cb;
-            const int node_end   = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes);
 
             for (int ind = node_start; ind < node_end; ++ind) {
                 const int i = has_concur ? ctx->concur_list[ind] : ind;
@@ -744,32 +746,31 @@ void ggml_metal_graph_compute(
                                 [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
                                 ne00%32 == 0 &&
                                 ne11 > 1) {
-                                    switch (src0->type) {
-                                        case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
-                                        case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
-                                        case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
-                                        case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
-                                        case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
-                                        case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
-                                        case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
-                                        case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
-                                        default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
-                                    }
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
-                                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
-                                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
-                                    [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
-                                    [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
-                                    [encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
-                                    [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                                switch (src0->type) {
+                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
+                                    case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
+                                    case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
+                                    case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
+                                    case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
+                                    case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
+                                    case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
+                                    case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
+                                    default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                                 }
-                            else {
+                                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
+                                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
+                                [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
+                                [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
+                                [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
+                                [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
+                                [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
+                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:8];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:9];
+                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:10];
+                                [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                            } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
 
@@ -868,24 +869,24 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
-                                [encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
+                                [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
                                     src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #else
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #endif
                                 }
                                 else if (src0t == GGML_TYPE_Q5_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q6_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
                                     [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -938,16 +939,17 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_NORM:
                         {
-                            const float eps = 1e-5f;
+                            float eps;
+                            memcpy(&eps, dst->op_params, sizeof(float));
 
                             const int nth = 256;
 
                             [encoder setComputePipelineState:ctx->pipeline_norm];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
-                            [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
+                            [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
                             [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
@@ -990,7 +992,9 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
                             [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
                             [encoder setBytes:&m0  length:sizeof(    float) atIndex:18];
+
                             const int nth = 32;
+
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_ROPE:
@@ -1005,8 +1009,8 @@ void ggml_metal_graph_compute(
                             memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
                             [encoder setComputePipelineState:ctx->pipeline_rope];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
                             [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
                             [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
                             [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
@@ -1057,24 +1061,24 @@ void ggml_metal_graph_compute(
                                 default: GGML_ASSERT(false && "not implemented");
                             }
 
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;

ggml-metal.metal

@@ -87,7 +87,12 @@ kernel void kernel_gelu(
     device       float * dst,
     uint tpig[[thread_position_in_grid]]) {
     float x = src0[tpig];
-    dst[tpig] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+
+    // BEWARE !!!
+    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
+    // This was observed with Falcon 7B and 40B models
+    //
+    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
 }
 
 kernel void kernel_soft_max(
@@ -571,7 +576,25 @@ kernel void kernel_rope(
             dst_data[1] = x0*sin_theta + x1*cos_theta;
         }
     } else {
-        // TODO: implement
+        for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
+            for (int64_t ic = 0; ic < n_dims; ic += 2) {
+                const float cos_theta = cos(theta);
+                const float sin_theta = sin(theta);
+
+                theta *= theta_scale;
+
+                const int64_t i0 = ib*n_dims + ic/2;
+
+                device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+                const float x0 = src[0];
+                const float x1 = src[n_dims/2];
+
+                dst_data[0]        = x0*cos_theta - x1*sin_theta;
+                dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+            }
+        }
     }
 }
 

ggml.c

@@ -3554,9 +3554,9 @@ inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) {
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 
-static const float GELU_COEF_A    = 0.044715f;
-static const float GELU_QUICK_COEF    = -1.702f;
-static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+static const float GELU_COEF_A     = 0.044715f;
+static const float GELU_QUICK_COEF = -1.702f;
+static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
 
 inline static float ggml_gelu_f32(float x) {
     return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
@@ -5555,10 +5555,6 @@ struct ggml_tensor * ggml_repeat(
         is_node = true;
     }
 
-    if (ggml_are_same_shape(a, b) && !is_node) {
-        return a;
-    }
-
     struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne);
 
     result->op   = GGML_OP_REPEAT;
@@ -5789,6 +5785,7 @@ struct ggml_tensor * ggml_silu_back(
 static struct ggml_tensor * ggml_norm_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
+        float eps,
         bool inplace) {
     bool is_node = false;
 
@@ -5799,7 +5796,7 @@ static struct ggml_tensor * ggml_norm_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    // TODO: maybe store epsilon here?
+    ggml_set_op_params(result, &eps, sizeof(eps));
 
     result->op   = GGML_OP_NORM;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5810,14 +5807,16 @@ static struct ggml_tensor * ggml_norm_impl(
 
 struct ggml_tensor * ggml_norm(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a) {
-    return ggml_norm_impl(ctx, a, false);
+        struct ggml_tensor  * a,
+        float eps) {
+    return ggml_norm_impl(ctx, a, eps, false);
 }
 
 struct ggml_tensor * ggml_norm_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a) {
-    return ggml_norm_impl(ctx, a, true);
+        struct ggml_tensor  * a,
+        float eps) {
+    return ggml_norm_impl(ctx, a, eps, true);
 }
 
 // ggml_rms_norm
@@ -10619,7 +10618,8 @@ static void ggml_compute_forward_norm_f32(
 
     GGML_TENSOR_UNARY_OP_LOCALS;
 
-    const float eps = 1e-5f; // TODO: make this a parameter
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
 
     // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
@@ -12537,7 +12537,7 @@ static void ggml_compute_forward_rope_f32(
                         dst_data[1] = x0*sin_theta*zeta + x1*cos_theta*zeta;
                     }
                 } else {
-                    // TODO: this is probably wrong, but I can't figure it out ..
+                    // TODO: this might be wrong for ne0 != n_dims - need double check
                     // ref:  https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28
                     for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
                         for (int64_t ic = 0; ic < n_dims; ic += 2) {
@@ -12666,7 +12666,7 @@ static void ggml_compute_forward_rope_f16(
                         dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
                 } else {
-                    // TODO: this is probably wrong, but I can't figure it out ..
+                    // TODO: this might be wrong for ne0 != n_dims - need double check
                     // ref:  https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28
                     for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
                         for (int64_t ic = 0; ic < n_dims; ic += 2) {

ggml.h

@@ -909,14 +909,15 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // normalize along rows
-    // TODO: eps is hardcoded to 1e-5 for now
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_norm_inplace(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_rms_norm(
             struct ggml_context * ctx,

gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import shutil
 import sys
 import struct
@@ -26,14 +27,15 @@ KEY_GENERAL_DESCRIPTION          = "general.description"
 KEY_GENERAL_LICENSE              = "general.license"
 KEY_GENERAL_SOURCE_URL           = "general.source.url"
 KEY_GENERAL_SOURCE_HF_REPO       = "general.source.hugginface.repository"
+KEY_GENERAL_FILE_TYPE            = "general.file_type"
 
 # LLM
-KEY_LLM_CONTEXT_LENGTH        = "{arch}.context_length"
-KEY_LLM_EMBEDDING_LENGTH      = "{arch}.embedding_length"
-KEY_LLM_BLOCK_COUNT           = "{arch}.block_count"
-KEY_LLM_FEED_FORWARD_LENGTH   = "{arch}.feed_forward_length"
-KEY_LLM_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
-KEY_LLM_TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
+KEY_CONTEXT_LENGTH        = "{arch}.context_length"
+KEY_EMBEDDING_LENGTH      = "{arch}.embedding_length"
+KEY_BLOCK_COUNT           = "{arch}.block_count"
+KEY_FEED_FORWARD_LENGTH   = "{arch}.feed_forward_length"
+KEY_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
+KEY_TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
 
 # attention
 KEY_ATTENTION_HEAD_COUNT        = "{arch}.attention.head_count"
@@ -581,7 +583,7 @@ class GGUFWriter:
         self.add_string(KEY_GENERAL_AUTHOR, author)
 
     def add_tensor_data_layout(self, layout: str):
-        self.add_string(KEY_LLM_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
+        self.add_string(KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
 
     def add_url(self, url: str):
         self.add_string(KEY_GENERAL_URL, url)
@@ -595,6 +597,9 @@ class GGUFWriter:
     def add_source_hf_repo(self, repo: str):
         self.add_string(KEY_GENERAL_SOURCE_HF_REPO, repo)
 
+    def add_file_type(self, ftype: int):
+        self.add_uint32(KEY_GENERAL_FILE_TYPE, ftype)
+
     def add_name(self, name: str):
         self.add_string(KEY_GENERAL_NAME, name)
 
@@ -608,27 +613,27 @@ class GGUFWriter:
 
     def add_context_length(self, length: int):
         self.add_uint32(
-            KEY_LLM_CONTEXT_LENGTH.format(arch=self.arch), length)
+            KEY_CONTEXT_LENGTH.format(arch=self.arch), length)
 
     def add_embedding_length(self, length: int):
         self.add_uint32(
-            KEY_LLM_EMBEDDING_LENGTH.format(arch=self.arch), length)
+            KEY_EMBEDDING_LENGTH.format(arch=self.arch), length)
 
     def add_block_count(self, length: int):
         self.add_uint32(
-            KEY_LLM_BLOCK_COUNT.format(arch=self.arch), length)
+            KEY_BLOCK_COUNT.format(arch=self.arch), length)
 
     def add_feed_forward_length(self, length: int):
         self.add_uint32(
-            KEY_LLM_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+            KEY_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
 
     def add_parallel_residual(self, use: bool):
         self.add_bool(
-            KEY_LLM_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
+            KEY_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
 
     def add_tensor_data_layout(self, layout: str):
         self.add_string(
-            KEY_LLM_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
+            KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
 
     def add_head_count(self, count: int):
         self.add_uint32(

grammars/README.md

@@ -0,0 +1,91 @@
+# GBNF Guide
+
+GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `examples/main` and `examples/server`.
+
+## Background
+
+[Bakus-Naur Form (BNF)](https://en.wikipedia.org/wiki/Backus%E2%80%93Naur_form) is a notation for describing the syntax of formal languages like programming languages, file formats, and protocols. GBNF is an extension of BNF that primarily adds a few modern regex-like features.
+
+## Basics
+
+In GBNF, we define *production rules* that specify how a *non-terminal* (rule name) can be replaced with sequences of *terminals* (characters, specifically Unicode [code points](https://en.wikipedia.org/wiki/Code_point)) and other non-terminals. The basic format of a production rule is `nonterminal ::= sequence...`.
+
+## Example
+
+Before going deeper, let's look at some of the features demonstrated in `grammars/chess.gbnf`, a small chess notation grammar:
+```
+# `root` specifies the pattern for the overall output
+root ::= (
+    # it must start with the characters "1. " followed by a sequence
+    # of characters that match the `move` rule, followed by a space, followed
+    # by another move, and then a newline
+    "1. " move " " move "\n"
+
+    # it's followed by one or more subsequent moves, numbered with one or two digits
+    ([1-9] [0-9]? ". " move " " move "\n")+
+)
+
+# `move` is an abstract representation, which can be a pawn, nonpawn, or castle.
+# The `[+#]?` denotes the possibility of checking or mate signs after moves
+move ::= (pawn | nonpawn | castle) [+#]?
+
+pawn ::= ...
+nonpawn ::= ...
+castle ::= ...
+```
+
+## Non-Terminals and Terminals
+
+Non-terminal symbols (rule names) stand for a pattern of terminals and other non-terminals. They are required to be a dashed lowercase word, like `move`, `castle`, or `check-mate`.
+
+Terminals are actual characters ([code points](https://en.wikipedia.org/wiki/Code_point)). They can be specified as a sequence like `"1"` or `"O-O"` or as ranges like `[1-9]` or `[NBKQR]`.
+
+## Characters and character ranges
+
+Terminals support the full range of Unicode. Unicode characters can be specified directly in the grammar, for example `hiragana ::= [ぁ-ゟ]`, or with escapes: 8-bit (`\xXX`), 16-bit (`\uXXXX`) or 32-bit (`\UXXXXXXXX`).
+
+Character ranges can be negated with `^`:
+```
+single-line ::= [^\n]+ "\n"`
+```
+
+## Sequences and Alternatives
+
+The order of symbols in a sequence matter. For example, in `"1. " move " " move "\n"`, the `"1. "` must come before the first `move`, etc.
+
+Alternatives, denoted by `|`, give different sequences that are acceptable. For example, in `move ::= pawn | nonpawn | castle`, `move` can be a `pawn` move, a `nonpawn` move, or a `castle`.
+
+Parentheses `()` can be used to group sequences, which allows for embedding alternatives in a larger rule or applying repetition and optptional symbols (below) to a sequence.
+
+## Repetition and Optional Symbols
+
+- `*` after a symbol or sequence means that it can be repeated zero or more times.
+- `+` denotes that the symbol or sequence should appear one or more times.
+- `?` makes the preceding symbol or sequence optional.
+
+## Comments and newlines
+
+Comments can be specified with `#`:
+```
+# defines optional whitspace
+ws ::= [ \t\n]+
+```
+
+Newlines are allowed between rules and between symbols or sequences nested inside parentheses. Additionally, a newline after an alternate marker `|` will continue the current rule, even outside of parentheses.
+
+## The root rule
+
+In a full grammar, the `root` rule always defines the starting point of the grammar. In other words, it specifies what the entire output must match.
+
+```
+# a grammar for lists
+root ::= ("- " item)+
+item ::= [^\n]+ "\n"
+```
+
+## Next steps
+
+This guide provides a brief overview. Check out the GBNF files in this directory (`grammars/`) for examples of full grammars. You can try them out with:
+```
+./main -m <model> --grammar-file grammars/some-grammar.gbnf -p 'Some prompt'
+```

k_quants.c

@@ -77,6 +77,11 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         }
         return 1/iscale;
     }
+    bool return_early = false;
+    if (rmse_type < 0) {
+        rmse_type = -rmse_type;
+        return_early = true;
+    }
     int weight_type = rmse_type%2;
     float sumlx = 0;
     float suml2 = 0;
@@ -89,56 +94,9 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         suml2 += w*l*l;
     }
     float scale = sumlx/suml2;
+    if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
     float best = scale * sumlx;
-    for (int itry = 0; itry < 3; ++itry) {
-        iscale = 1/scale;
-        float slx = 0;
-        float sl2 = 0;
-        bool changed = false;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            l = MAX(-nmax, MIN(nmax-1, l));
-            if (l + nmax != L[i]) { changed = true; }
-            float w = weight_type == 1 ? x[i] * x[i] : 1.f;
-            slx += w*x[i]*l;
-            sl2 += w*l*l;
-        }
-        if (!changed || sl2 == 0 || slx*slx <= best*sl2) { break; }
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
-        }
-        sumlx = slx; suml2 = sl2;
-        scale = sumlx/suml2;
-        best = scale * sumlx;
-    }
-    for (int itry = 0; itry < 5; ++itry) {
-        int n_changed = 0;
-        for (int i = 0; i < n; ++i) {
-            float w = weight_type == 1 ? x[i]*x[i] : 1;
-            int l = L[i] - nmax;
-            float slx = sumlx - w*x[i]*l;
-            if (slx > 0) {
-                float sl2 = suml2 - w*l*l;
-                int new_l = nearest_int(x[i] * sl2 / slx);
-                new_l = MAX(-nmax, MIN(nmax-1, new_l));
-                if (new_l != l) {
-                    slx += w*x[i]*new_l;
-                    sl2 += w*new_l*new_l;
-                    if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
-                        L[i] = nmax + new_l; sumlx = slx; suml2 = sl2;
-                        scale = sumlx / suml2; best = scale * sumlx;
-                        ++n_changed;
-                    }
-                }
-            }
-        }
-        if (!n_changed) { break; }
-    }
-    if (rmse_type < 3) {
-        return scale;
-    }
-    for (int is = -4; is <= 4; ++is) {
+    for (int is = -9; is <= 9; ++is) {
         if (is == 0) {
             continue;
         }
@@ -221,12 +179,17 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t *
     return 1/iscale;
 }
 
-static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min, int ntry) {
+static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
+        int ntry, float alpha) {
     float min = x[0];
     float max = x[0];
+    float sum_x = 0;
+    float sum_x2 = 0;
     for (int i = 1; i < n; ++i) {
         if (x[i] < min) min = x[i];
         if (x[i] > max) max = x[i];
+        sum_x += x[i];
+        sum_x2 += x[i]*x[i];
     }
     if (max == min) {
         for (int i = 0; i < n; ++i) L[i] = 0;
@@ -254,7 +217,7 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
         for (int i = 0; i < n; ++i) {
             sum += x[i] - scale*L[i];
         }
-        min = sum/n;
+        min = alpha*min + (1 - alpha)*sum/n;
         if (min > 0) min = 0;
         iscale = 1/scale;
         if (!did_change) break;
@@ -263,6 +226,82 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
     return scale;
 }
 
+static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
+        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+        float rmin, float rdelta, int nstep, bool use_mad) {
+    float min = x[0];
+    float max = x[0];
+    float sum_w = weights[0];
+    float sum_x = sum_w * x[0];
+    for (int i = 1; i < n; ++i) {
+        if (x[i] < min) min = x[i];
+        if (x[i] > max) max = x[i];
+        float w = weights[i];
+        sum_w += w;
+        sum_x += w * x[i];
+    }
+    if (min > 0) min = 0;
+    if (max == min) {
+        for (int i = 0; i < n; ++i) L[i] = 0;
+        *the_min = -min;
+        return 0.f;
+    }
+    float iscale = nmax/(max - min);
+    float scale = 1/iscale;
+    float best_mad = 0;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale*(x[i] - min));
+        L[i] = MAX(0, MIN(nmax, l));
+        float diff = scale * L[i] + min - x[i];
+        diff = use_mad ? fabsf(diff) : diff * diff;
+        float w = weights[i];
+        best_mad += w * diff;
+    }
+    if (nstep < 1) {
+        *the_min = -min;
+        return scale;
+    }
+    for (int is = 0; is <= nstep; ++is) {
+        iscale = (rmin + rdelta*is + nmax)/(max - min);
+        float sum_l = 0, sum_l2 = 0, sum_xl = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale*(x[i] - min));
+            l = MAX(0, MIN(nmax, l));
+            Laux[i] = l;
+            float w = weights[i];
+            sum_l += w*l;
+            sum_l2 += w*l*l;
+            sum_xl += w*l*x[i];
+        }
+        float D = sum_w * sum_l2 - sum_l * sum_l;
+        if (D > 0) {
+            float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
+            float this_min   = (sum_l2 * sum_x - sum_l * sum_xl)/D;
+            if (this_min > 0) {
+                this_min = 0;
+                this_scale = sum_xl / sum_l2;
+            }
+            float mad = 0;
+            for (int i = 0; i < n; ++i) {
+                float diff = this_scale * Laux[i] + this_min - x[i];
+                diff = use_mad ? fabsf(diff) : diff * diff;
+                float w = weights[i];
+                mad += w * diff;
+            }
+            if (mad < best_mad) {
+                for (int i = 0; i < n; ++i) {
+                    L[i] = Laux[i];
+                }
+                best_mad = mad;
+                scale = this_scale;
+                min = this_min;
+            }
+        }
+    }
+    *the_min = -min;
+    return scale;
+}
+
 #if QK_K == 256
 static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
     if (j < 4) {
@@ -281,6 +320,8 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict
     const int nb = k / QK_K;
 
     uint8_t L[QK_K];
+    uint8_t Laux[16];
+    float   weights[16];
     float mins[QK_K/16];
     float scales[QK_K/16];
 
@@ -291,7 +332,8 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict
         float max_scale = 0; // as we are deducting the min, scales are always positive
         float max_min = 0;
         for (int j = 0; j < QK_K/16; ++j) {
-            scales[j] = make_qkx1_quants(16, 3, x + 16*j, L + 16*j, &mins[j], 5);
+            for (int l = 0; l < 16; ++l) weights[l] = fabsf(x[16*j + l]);
+            scales[j] = make_qkx2_quants(16, 3, x + 16*j, weights, L + 16*j, &mins[j], Laux, -0.5f, 0.1f, 15, true);
             float scale = scales[j];
             if (scale > max_scale) {
                 max_scale = scale;
@@ -637,6 +679,8 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
     const int nb = k / QK_K;
 
     uint8_t L[QK_K];
+    uint8_t Laux[32];
+    float   weights[32];
     float mins[QK_K/32];
     float scales[QK_K/32];
 
@@ -645,7 +689,12 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
         float max_scale = 0; // as we are deducting the min, scales are always positive
         float max_min = 0;
         for (int j = 0; j < QK_K/32; ++j) {
-            scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 5);
+            //scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
+            float sum_x2 = 0;
+            for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
+            float av_x = sqrtf(sum_x2/32);
+            for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+            scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
             float scale = scales[j];
             if (scale > max_scale) {
                 max_scale = scale;
@@ -798,6 +847,8 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
     uint8_t L[QK_K];
     float mins[QK_K/32];
     float scales[QK_K/32];
+    float weights[32];
+    uint8_t Laux[32];
 #else
     int8_t L[QK_K];
     float scales[QK_K/16];
@@ -810,7 +861,12 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
         float max_scale = 0; // as we are deducting the min, scales are always positive
         float max_min = 0;
         for (int j = 0; j < QK_K/32; ++j) {
-            scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 5);
+            //scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
+            float sum_x2 = 0;
+            for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
+            float av_x = sqrtf(sum_x2/32);
+            for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+            scales[j] = make_qkx2_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.5f, 0.1f, 15, false);
             float scale = scales[j];
             if (scale > max_scale) {
                 max_scale = scale;

llama.h

@@ -103,6 +103,8 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
+
+        LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
 
     typedef struct llama_token_data {
@@ -245,6 +247,8 @@ extern "C" {
     LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
     LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_context * ctx);
+
     LLAMA_API int llama_model_n_vocab(const struct llama_model * model);
     LLAMA_API int llama_model_n_ctx  (const struct llama_model * model);
     LLAMA_API int llama_model_n_embd (const struct llama_model * model);
@@ -366,13 +370,6 @@ extern "C" {
                              int   n_max_tokens,
                             bool   add_bos);
 
-    LLAMA_API int llama_tokenize_bpe(
-            struct llama_context * ctx,
-                      const char * text,
-                     llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
-
     LLAMA_API int llama_tokenize_with_model(
         const struct llama_model * model,
                       const char * text,
@@ -388,12 +385,6 @@ extern "C" {
                                   char * buf,
                                   int    length);
 
-    LLAMA_API int llama_token_to_str_bpe(
-            const struct llama_context * ctx,
-                           llama_token   token,
-                                  char * buf,
-                                  int    length);
-
     LLAMA_API int llama_token_to_str_with_model(
               const struct llama_model * model,
                            llama_token   token,

tests/CMakeLists.txt

@@ -28,7 +28,8 @@ llama_build_and_test_executable(test-sampling.cpp)
 llama_build_executable(test-tokenizer-0.cpp)
 llama_test_executable (test-tokenizer-0.llama test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
 llama_build_executable(test-tokenizer-1.cpp)
-llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+# test-tokenizer-1 requires a BPE vocab. re-enable when we have one.
+#llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
 #llama_test_executable(test-tokenizer-1.aquila test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)

tests/test-tokenizer-0.cpp

@@ -17,6 +17,8 @@ static std::string unescape_whitespace(llama_context* ctx, const std::vector<lla
 static const std::map<std::string, std::vector<llama_token>> & k_tests() {
     static std::map<std::string, std::vector<llama_token>> _k_tests = {
         { " ",                      {1,    259, }, },
+        { "  ",                     { 1,    1678, }, },
+        { "   ",                    { 1,     268, }, },
         { "\t",                     { 1,    29871,   12, }, },
         { "\n",                     { 1,    29871,   13, }, },
         { "\t\n",                   { 1,    29871,   12,     13, }, },
@@ -38,6 +40,12 @@ static const std::map<std::string, std::vector<llama_token>> & k_tests() {
                 243,    162,    155,    185,  30722,    243,    162,    143,    174,  30598,
                 313,  20787,    953,   3848,    275,  16125,    630,  29897,  29871,  31681,
                 313,   6194,    953,  29877,   2397,    393,    756,    967,   1914,   5993,  29897, }, },
+        { "Hello",                  { 1,    15043 }, },
+        { " Hello",                 { 1,    29871,  15043 }, },
+        { "  Hello",                { 1,    259,    15043 }, },
+        { "   Hello",               { 1,    1678,   15043 }, },
+        { "    Hello",              { 1,    268,    15043 }, },
+        { "    Hello\n    Hello",   { 1,    268,    15043,  13,     1678,   15043 }, },
     };
 
     return _k_tests;
@@ -106,7 +114,8 @@ int main(int argc, char **argv) {
 
         if (!correct) {
             fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
-            fprintf(stderr, "%s : detokenized to: '%s'\n", __func__, unescape_whitespace(ctx, test_kv.second).c_str());
+            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                unescape_whitespace(ctx, res).c_str(), unescape_whitespace(ctx, test_kv.second).c_str());
             fprintf(stderr, "%s : expected tokens: ", __func__);
             for (const auto & t : test_kv.second) {
                 fprintf(stderr, "%6d, ", t);

tests/test-tokenizer-1.cpp

@@ -11,18 +11,11 @@
 #include <locale>
 
 static std::string escape_whitespace(const std::string& text) {
-    std::string result;
-    bool escaping = false;
-    result += "\xe2\x96\x81";
+    std::string result = "\xe2\x96\x81";
     for (size_t offs = 0; offs < text.length(); ++offs) {
         if (text[offs] == ' ') {
-            if (!escaping) {
-                result += "\xe2\x96\x81";
-                escaping = true;
-            }
-        }
-        else {
-            escaping = false;
+            result += "\xe2\x96\x81";
+        } else {
             result += text[offs];
         }
     }
@@ -74,11 +67,13 @@ int main(int argc, char **argv) {
         }
     }
 
+    GGML_ASSERT(llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_BPE);
+
     const int n_vocab = llama_n_vocab(ctx);
 
     for (int i = 0; i < n_vocab; ++i) {
-        std::string forward = llama_token_to_str_bpe(ctx, i);
-        std::vector<llama_token> tokens = llama_tokenize_bpe(ctx, forward, false);
+        std::string forward = llama_token_to_str(ctx, i);
+        std::vector<llama_token> tokens = llama_tokenize(ctx, forward, false);
         if (tokens.size() == 1) {
             if (i != tokens[0]) {
                 std::string backward = llama_token_to_str(ctx, tokens[0]);
@@ -86,16 +81,6 @@ int main(int argc, char **argv) {
                     __func__, i, llama_token_to_str(ctx, i).c_str(), tokens[0], backward.c_str());
                 return 2;
             }
-        } else {
-            llama_token_type type = llama_token_get_type(ctx, i);
-            if (type == LLAMA_TOKEN_TYPE_UNKNOWN || type == LLAMA_TOKEN_TYPE_CONTROL || type == LLAMA_TOKEN_TYPE_BYTE) {
-                fprintf(stderr, "%s : info: token %d is string %s and bpe returns tokens %s\n",
-                    __func__, i, llama_token_to_str(ctx, i).c_str(), unescape_whitespace(ctx, tokens).c_str());
-            } else {
-                fprintf(stderr, "%s : error: token %d is string %s but bpe returns tokens %s\n",
-                    __func__, i, llama_token_to_str(ctx, i).c_str(), unescape_whitespace(ctx, tokens).c_str());
-                return 2;
-            }
         }
     }