Merge branch 'master' into fix-tokenizer

convert.py

@@ -3,6 +3,7 @@
 import gguf
 import argparse
 import concurrent.futures
+from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor
 import copy
 import enum
 import faulthandler
@@ -17,13 +18,14 @@ import re
 import signal
 import struct
 import sys
+import time
 import zipfile
 import numpy as np
 
 from abc import ABCMeta, abstractmethod
 from dataclasses import dataclass
 from pathlib import Path
-from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Iterable, List, Literal, Optional, Sequence, Tuple, TypeVar, Union)
+from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, TypeVar, Union)
 from sentencepiece import SentencePieceProcessor  # type: ignore
 
 if TYPE_CHECKING:
@@ -37,30 +39,70 @@ NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
 ARCH=gguf.MODEL_ARCH.LLAMA
 NAMES=gguf.MODEL_TENSOR_NAMES[ARCH]
 
+DEFAULT_CONCURRENCY = 8
 #
 # data types
 #
 
 @dataclass(frozen=True)
-class UnquantizedDataType:
+class DataType:
     name: str
+    dtype: 'np.dtype[Any]'
+    valid_conversions: List[str]
 
-DT_F16  = UnquantizedDataType('F16')
+    def elements_to_bytes(self, n_elements: int) -> int:
-DT_F32  = UnquantizedDataType('F32')
+        return n_elements * self.dtype.itemsize
-DT_I32  = UnquantizedDataType('I32')
-DT_BF16 = UnquantizedDataType('BF16')
 
-DataType = Union[UnquantizedDataType]
+@dataclass(frozen=True)
+class UnquantizedDataType(DataType):
+    pass
 
-DATA_TYPE_TO_NUMPY: Dict[DataType, 'np.dtype[Any]'] = {
+DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
-    DT_BF16: np.dtype(np.uint16),
+DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
-    DT_F16:  np.dtype(np.float16),
+DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
-    DT_F32:  np.dtype(np.float32),
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
-    DT_I32:  np.dtype(np.int32),
-}
 
-NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = \
+@dataclass(frozen=True)
-    {dtype: data_type for (data_type, dtype) in DATA_TYPE_TO_NUMPY.items()}
+class QuantizedDataType(DataType):
+    block_size: int
+    quantized_dtype: 'np.dtype[Any]'
+    ggml_type: gguf.GGMLQuantizationType
+
+    def quantize(self, arr: NDArray) -> NDArray:
+        raise NotImplementedError(f'Quantization for {self.name} not implemented')
+
+    def elements_to_bytes(self, n_elements: int) -> int:
+        assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}'
+        return self.quantized_dtype.itemsize * (n_elements // self.block_size)
+
+@dataclass(frozen=True)
+class Q8_0QuantizedDataType(QuantizedDataType):
+    # Mini Q8_0 quantization in Python!
+    def quantize(self, arr: NDArray) -> NDArray:
+        assert arr.size % self.block_size == 0 and arr.size != 0, f'Bad array size {arr.size}'
+        assert arr.dtype == np.float32, f'Bad array type {arr.dtype}'
+        n_blocks = arr.size // self.block_size
+        blocks = arr.reshape((n_blocks, self.block_size))
+        # Much faster implementation of block quantization contributed by @Cebtenzzre
+        def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[Tuple[Any, Any]]:
+            d = abs(blocks).max(axis = 1) / np.float32(127)
+            with np.errstate(divide = 'ignore'):
+                qs = (blocks / d[:, None]).round()
+            qs[d == 0] = 0
+            yield from zip(d, qs)
+        return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype)
+
+DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
+    dtype = np.dtype(np.float32), valid_conversions = [],
+    ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
+    quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
+
+# Quantized types skipped here because they may also map to np.float32
+NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = {}
+for dt in (DT_BF16, DT_F16, DT_F32, DT_I32):
+    if dt.dtype in NUMPY_TYPE_TO_DATA_TYPE:
+        raise ValueError(f'Invalid duplicate data type {dt}')
+    NUMPY_TYPE_TO_DATA_TYPE[dt.dtype] = dt
 
 SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
     'BF16': DT_BF16,
@@ -73,20 +115,22 @@ SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
 # TODO: rename to LLAMAFileType
 # TODO: move to `gguf.py`
 class GGMLFileType(enum.IntEnum):
-    AllF32    = 0
+    AllF32     = 0
-    MostlyF16 = 1  # except 1d tensors
+    MostlyF16  = 1  # except 1d tensors
+    MostlyQ8_0 = 7  # except 1d tensors
 
     def type_for_tensor(self, name: str, tensor: 'LazyTensor') -> DataType:
-        if len(tensor.shape) == 1:
+        dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self)
-            # 1D tensors are always F32.
+        if dt is None:
-            return DT_F32
-        elif self == GGMLFileType.AllF32:
-            return DT_F32
-        elif self == GGMLFileType.MostlyF16:
-            return DT_F16
-        else:
             raise ValueError(self)
+        # 1D tensors are always F32.
+        return dt if len(tensor.shape) > 1 else DT_F32
 
+GGML_FILE_TYPE_TO_DATA_TYPE: Dict[GGMLFileType, DataType] = {
+    GGMLFileType.AllF32    : DT_F32,
+    GGMLFileType.MostlyF16 : DT_F16,
+    GGMLFileType.MostlyQ8_0: DT_Q8_0,
+}
 
 #
 # hparams loading
@@ -415,7 +459,7 @@ class UnquantizedTensor(Tensor):
         self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
 
     def astype(self, data_type: DataType) -> Tensor:
-        dtype = DATA_TYPE_TO_NUMPY[data_type]
+        dtype = data_type.dtype
         if self.data_type == DT_BF16:
             self.ndarray = bf16_to_fp32(self.ndarray)
         return UnquantizedTensor(self.ndarray.astype(dtype))
@@ -454,22 +498,6 @@ def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, conv
 GGMLCompatibleTensor = Union[UnquantizedTensor]
 
 
-class DeferredPermutedTensor(Tensor):
-    def __init__(self, base: Tensor, n_head: int, n_head_kv: int) -> None:
-        self.base = base
-        self.n_head = n_head
-        self.data_type = self.base.data_type
-
-    def astype(self, data_type: DataType) -> Tensor:
-        return self.base.astype(data_type).permute(self.n_head, self.n_head_kv)
-
-    def to_ggml(self) -> GGMLCompatibleTensor:
-        return self.base.to_ggml().permute(self.n_head, self.n_head_kv)
-
-    def permute(self, n_head: int, n_head_kv: int) -> Tensor:
-        raise Exception("shouldn't permute twice")
-
-
 @dataclass
 class LazyTensor:
     _load: Callable[[], Tensor]
@@ -479,7 +507,9 @@ class LazyTensor:
 
     def load(self) -> Tensor:
         ret = self._load()
-        assert ret.data_type == self.data_type, (self.data_type, ret.data_type, self.description)
+        # Should be okay if it maps to the same numpy type?
+        assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \
+                (self.data_type, ret.data_type, self.description)
         return ret
 
     def astype(self, data_type: DataType) -> 'LazyTensor':
@@ -490,8 +520,8 @@ class LazyTensor:
         return LazyTensor(load, self.shape, data_type, f'convert({data_type}) {self.description}')
 
     def validate_conversion_to(self, data_type: DataType) -> None:
-        if data_type == self.data_type:
+        if data_type != self.data_type and data_type.name not in self.data_type.valid_conversions:
-            return
+            raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.')
 
 
 LazyModel = Dict[str, LazyTensor]
@@ -617,9 +647,7 @@ class LazyUnpickler(pickle.Unpickler):
         info = self.zip_file.getinfo(filename)
 
         def load(offset: int, elm_count: int) -> NDArray:
-            dtype = DATA_TYPE_TO_NUMPY.get(data_type)
+            dtype = data_type.dtype
-            if dtype is None:
-                raise Exception("tensor stored in unsupported format")
             fp = self.zip_file.open(info)
             fp.seek(offset * dtype.itemsize)
             size = elm_count * dtype.itemsize
@@ -683,7 +711,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
 
     def convert(info: Dict[str, Any]) -> LazyTensor:
         data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
-        numpy_dtype = DATA_TYPE_TO_NUMPY[data_type]
+        numpy_dtype = data_type.dtype
         shape: List[int] = info['shape']
         begin, end = info['data_offsets']
         assert 0 <= begin <= end <= len(byte_buf)
@@ -723,23 +751,35 @@ def lazy_load_file(path: Path) -> ModelPlus:
 In = TypeVar('In')
 Out = TypeVar('Out')
 
-def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int) -> Iterable[Out]:
+def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, factory: Callable = ThreadPoolExecutor) -> Iterable[Out]:
     '''Parallel map, but with backpressure.  If the caller doesn't call `next`
     fast enough, this will stop calling `func` at some point rather than
     letting results pile up in memory.  Specifically, there is a max of one
     output value buffered per thread.'''
-    with concurrent.futures.ThreadPoolExecutor() as executor:
+    if concurrency < 2:
+        yield from map(func, iterable)
+        # Not reached.
+    iterable = iter(iterable)
+    with factory(max_workers = max_workers) as executor:
         futures: List[concurrent.futures.Future[Out]] = []
-        items_rev = list(iterable)[::-1]
+        done = False
-        for i in range(min(concurrency, len(items_rev))):
+        for _ in range(concurrency):
-            futures.append(executor.submit(func, items_rev.pop()))
+            try:
+                futures.append(executor.submit(func, next(iterable)))
+            except StopIteration:
+                done = True
+                break
+
         while futures:
             result = futures.pop(0).result()
-            if items_rev:
+            while not done and len(futures) < concurrency:
-                futures.append(executor.submit(func, items_rev.pop()))
+                try:
+                    futures.append(executor.submit(func, next(iterable)))
+                except StopIteration:
+                    done = True
+                    break
             yield result
 
-
 def check_vocab_size(params: Params, vocab: Vocab) -> None:
     if params.n_vocab != vocab.vocab_size:
         assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
@@ -804,12 +844,11 @@ class OutputFile:
         self.gguf.add_token_types(toktypes)
 
     def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
-        n_elements = 1
+        n_elements = int(np.prod(tensor.shape))
-        for dim in tensor.shape:
+        raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
-            n_elements *= dim
+        data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype
-        data_type = DATA_TYPE_TO_NUMPY[tensor.data_type]
+        data_nbytes = tensor.data_type.elements_to_bytes(n_elements)
-        data_nbytes = n_elements * data_type.itemsize
+        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype = raw_dtype)
-        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes)
 
     def write_meta(self) -> None:
         self.gguf.write_header_to_file()
@@ -835,7 +874,20 @@ class OutputFile:
         of.close()
 
     @staticmethod
-    def write_all(fname_out: Path, params: Params, model: LazyModel, vocab: Vocab) -> None:
+    def do_item(item: Tuple[str, LazyTensor]) -> Tuple[DataType, NDArray]:
+        name, lazy_tensor = item
+        tensor = lazy_tensor.load().to_ggml()
+        return (lazy_tensor.data_type, tensor.ndarray)
+
+    @staticmethod
+    def maybe_do_quantize(item: Tuple[DataType, NDArray]) -> NDArray:
+        dt, arr = item
+        if not isinstance(dt, QuantizedDataType):
+            return arr
+        return dt.quantize(arr)
+
+    @staticmethod
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
         check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out)
@@ -851,16 +903,19 @@ class OutputFile:
         of.write_meta()
         of.write_tensor_info()
 
-        def do_item(item: Tuple[str, LazyTensor]) -> NDArray:
-            name, lazy_tensor = item
-            return lazy_tensor.load().to_ggml().ndarray
-
         # tensor data
-        ndarrays = bounded_parallel_map(do_item, model.items(), concurrency=8)
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, factory = ProcessPoolExecutor)
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
         for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
             size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
             padi = len(str(len(model)))
-            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type}")
+            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}")
             of.gguf.write_tensor_data(ndarray)
 
         of.close()
@@ -872,6 +927,8 @@ def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFi
         return GGMLFileType.AllF32
     if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)):
         return GGMLFileType.MostlyF16
+    if output_type_str == "q8_0":
+        return GGMLFileType.MostlyQ8_0
 
     name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
 
@@ -918,7 +975,7 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
             print(f"skipping tensor {name_new}")
             continue
         else:
-            print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type} | {lazy_tensor.shape}")
+            print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
             out[name_new] = lazy_tensor
 
     return out
@@ -1023,6 +1080,7 @@ def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path:
     namestr = {
         GGMLFileType.AllF32:    "f32",
         GGMLFileType.MostlyF16: "f16",
+        GGMLFileType.MostlyQ8_0:"q8_0",
     }[file_type]
     ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
     if ret in model_paths:
@@ -1046,12 +1104,13 @@ def main(args_in: Optional[List[str]] = None) -> None:
     parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outtype",     choices=["f32", "f16"], help="output format (default: based on input)")
+    parser.add_argument("--outtype",     choices=["f32", "f16", "q8_0"], help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
     parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
     parser.add_argument("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
     parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
     args = parser.parse_args(args_in)
 
     if args.dump_single:
@@ -1073,6 +1132,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
         params.ftype = {
             "f32": GGMLFileType.AllF32,
             "f16": GGMLFileType.MostlyF16,
+            "q8_0": GGMLFileType.MostlyQ8_0,
         }[args.outtype]
 
     print(f"params = {params}")
@@ -1104,7 +1164,7 @@ def main(args_in: Optional[List[str]] = None) -> None:
         params.ftype = ftype
         print(f"Writing {outfile}, format {ftype}")
 
-        OutputFile.write_all(outfile, params, model, vocab)
+        OutputFile.write_all(outfile, ftype, params, model, vocab, concurrency = args.concurrency)
         print(f"Wrote {outfile}")
 
 

examples/main/main.cpp

@@ -598,7 +598,12 @@ int main(int argc, char ** argv) {
                     last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
                     last_n_repeat, alpha_frequency, alpha_presence);
                 if (!penalize_nl) {
-                    logits[llama_token_nl(ctx)] = nl_logit;
+                    for (size_t idx = 0; idx < candidates_p.size; idx++) {
+                        if (candidates_p.data[idx].id == llama_token_nl(ctx)) {
+                            candidates_p.data[idx].logit = nl_logit;
+                            break;
+                        }
+                    }
                 }
 
                 if (grammar != NULL) {

examples/perplexity/perplexity.cpp

@@ -351,6 +351,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     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;
+    fprintf(stderr, "================================= is_spm = %d\n", is_spm);
 
     // This is needed as usual for LLaMA models
     const bool add_bos = is_spm;
@@ -406,6 +407,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     double acc = 0.0f;
     const int n_vocab = llama_n_vocab(ctx);
 
+    std::vector<std::vector<int>> ending_tokens(4);
+
     std::vector<float> tok_logits(n_vocab);
 
     for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
@@ -413,11 +416,21 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos);
         size_t context_size = context_embd.size();
 
+        for (int i = 0; i < 4; ++i) {
+            ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[i], add_bos);
+            for (int k = 0; k < int(context_size); ++k) {
+                if (ending_tokens[i][k] != context_embd[k]) {
+                    fprintf(stderr, "Oops: ending %d of task %d differs from context at position %d\n",i,int(task_idx),k);
+                    break;
+                }
+            }
+        }
+
         // 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], add_bos);
+        //auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
+        auto query_embd = ending_tokens[0];
         auto query_size = query_embd.size();
-        //printf("First query: %d\n",(int)query_size);
 
         // Stop if query wont fit the ctx window
         if (query_size > (size_t)params.n_ctx) {
@@ -462,7 +475,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         for (size_t ending_idx = 1; ending_idx < 4; ending_idx++) {
 
             // Tokenize the query
-            query_embd = ::llama_tokenize(ctx, hs_data[task_idx].ending[ending_idx], false);
+            query_embd.resize(ending_tokens[ending_idx].size() - context_size);
+            std::memcpy(query_embd.data(), ending_tokens[ending_idx].data() + context_size, query_embd.size()*sizeof(int));
             query_size = query_embd.size();
 
             // Stop if query wont fit the ctx window

flake.lock

@@ -5,11 +5,11 @@
         "systems": "systems"
       },
       "locked": {
-        "lastModified": 1685518550,
+        "lastModified": 1692799911,
-        "narHash": "sha256-o2d0KcvaXzTrPRIo0kOLV0/QXHhDQ5DTi+OxcjO8xqY=",
+        "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "a1720a10a6cfe8234c0e93907ffe81be440f4cef",
+        "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1685931219,
+        "lastModified": 1692913444,
-        "narHash": "sha256-8EWeOZ6LKQfgAjB/USffUSELPRjw88A+xTcXnOUvO5M=",
+        "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "7409480d5c8584a1a83c422530419efe4afb0d19",
+        "rev": "18324978d632ffc55ef1d928e81630c620f4f447",
         "type": "github"
       },
       "original": {

flake.nix

@@ -6,6 +6,9 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
+        name = "llama.cpp";
+        src = ./.;
+        meta.mainProgram = "llama";
         inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
         buildInputs = with pkgs; [ openmpi ];
         osSpecific = with pkgs; buildInputs ++
@@ -21,11 +24,17 @@
               CoreGraphics
               CoreVideo
             ]
+          else if isDarwin then
+            with pkgs.darwin.apple_sdk.frameworks; [
+              Accelerate
+              CoreGraphics
+              CoreVideo
+            ]
           else
             with pkgs; [ openblas ]
         );
         pkgs = import nixpkgs { inherit system; };
-        nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
+        nativeBuildInputs = with pkgs; [ cmake ninja pkgconfig ];
         llama-python =
           pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
         postPatch = ''
@@ -38,35 +47,35 @@
           mv $out/bin/server $out/bin/llama-server
         '';
         cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
-      in {
+      in
+      {
         packages.default = pkgs.stdenv.mkDerivation {
-          name = "llama.cpp";
+          inherit name src meta postPatch nativeBuildInputs buildInputs postInstall;
-          src = ./.;
-          postPatch = postPatch;
-          nativeBuildInputs = nativeBuildInputs;
-          buildInputs = osSpecific;
           cmakeFlags = cmakeFlags
             ++ (if isAarch64 && isDarwin then [
-              "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
+            "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-              "-DLLAMA_METAL=ON"
+            "-DLLAMA_METAL=ON"
-            ] else [
+          ] else [
-              "-DLLAMA_BLAS=ON"
+            "-DLLAMA_BLAS=ON"
-              "-DLLAMA_BLAS_VENDOR=OpenBLAS"
+            "-DLLAMA_BLAS_VENDOR=OpenBLAS"
           ]);
-          postInstall = postInstall;
-          meta.mainProgram = "llama";
         };
         packages.opencl = pkgs.stdenv.mkDerivation {
-          name = "llama.cpp";
+          inherit name src meta postPatch nativeBuildInputs postInstall;
-          src = ./.;
-          postPatch = postPatch;
-          nativeBuildInputs = nativeBuildInputs;
           buildInputs = with pkgs; buildInputs ++ [ clblast ];
           cmakeFlags = cmakeFlags ++ [
             "-DLLAMA_CLBLAST=ON"
           ];
-          postInstall = postInstall;
+        };
-          meta.mainProgram = "llama";
+        packages.rocm = pkgs.stdenv.mkDerivation {
+          inherit name src meta postPatch nativeBuildInputs postInstall;
+          buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ];
+          cmakeFlags = cmakeFlags ++ [
+            "-DLLAMA_HIPBLAS=1"
+            "-DCMAKE_C_COMPILER=hipcc"
+            "-DCMAKE_CXX_COMPILER=hipcc"
+            "-DCMAKE_POSITION_INDEPENDENT_CODE=ON"
+          ];
         };
         apps.llama-server = {
           type = "app";
@@ -80,8 +89,13 @@
           type = "app";
           program = "${self.packages.${system}.default}/bin/llama";
         };
+        apps.quantize = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/quantize";
+        };
         apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
+          buildInputs = [ llama-python ];
           packages = nativeBuildInputs ++ osSpecific;
         };
       });

llama.cpp

@@ -1,9 +1,6 @@
 // Defines fileno on msys:
 #ifndef _GNU_SOURCE
 #define _GNU_SOURCE
-#include <cstddef>
-#include <cstdint>
-#include <cstdio>
 #endif
 
 #include "llama.h"
@@ -62,6 +59,9 @@
 #include <cinttypes>
 #include <climits>
 #include <cstdarg>
+#include <cstddef>
+#include <cstdint>
+#include <cstdio>
 #include <cstring>
 #include <ctime>
 #include <fstream>
@@ -955,10 +955,10 @@ struct llama_vocab {
     id linefeed_id = 13;
 
     int find_bpe_rank(std::string token_left, std::string token_right) const {
-        replace_all(token_left,  " ",  "Ġ");
+        replace_all(token_left,  " ",  "\u0120");
-        replace_all(token_left,  "\n", "Ċ");
+        replace_all(token_left,  "\n", "\u010A");
-        replace_all(token_right, " ",  "Ġ");
+        replace_all(token_right, " ",  "\u0120");
-        replace_all(token_right, "\n", "Ċ");
+        replace_all(token_right, "\n", "\u010A");
 
         auto it = bpe_ranks.find(std::make_pair(token_left, token_right));
         if (it == bpe_ranks.end()) {
@@ -3894,7 +3894,7 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array *
 
     // Calculate absolute value of second derivatives
     for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        second_derivatives[i] = abs(second_derivatives[i]);
+        second_derivatives[i] = std::abs(second_derivatives[i]);
     }
 
     // Normalize the second derivatives
@@ -4660,6 +4660,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
     std::unique_ptr<llama_model_loader> ml(new llama_model_loader(fname_inp, /*use_mmap*/ false));
 
+    llama_model model;
+    llm_load_arch(*ml, model);
+    llm_load_hparams(*ml, model, 0, 0, 0);
+
     const size_t align = GGUF_DEFAULT_ALIGNMENT;
     struct gguf_context * ctx_out = gguf_init_empty();
 
@@ -4685,6 +4689,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             ++n_feed_forward_w2;
         }
     }
+    if (n_attention_wv != n_feed_forward_w2 || (uint32_t)n_attention_wv != model.hparams.n_layer) {
+        LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n",
+                __func__, n_attention_wv, n_feed_forward_w2, model.hparams.n_layer);
+    }
 
     int i_attention_wv = 0;
     int i_feed_forward_w2 = 0;
@@ -4761,8 +4769,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
             if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
                 int nx = tensor->ne[0];
-                int ny = tensor->ne[1];
+                if (nx % QK_K == 0) {
-                if (nx % QK_K == 0 && ny % QK_K == 0) {
                     new_type = GGML_TYPE_Q6_K;
                 }
             } else if (name.find("attn_v.weight") != std::string::npos) {
@@ -4776,6 +4783,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
                 else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
                         (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
+                if (model.type == MODEL_70B) {
+                    // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
+                    // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
+                    // nearly negligible increase in model size by quantizing this tensor with more bits:
+                    if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
+                }
                 ++i_attention_wv;
             } else if (name.find("ffn_down.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
@@ -4805,8 +4818,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
                 int nx = tensor->ne[0];
                 int ny = tensor->ne[1];
-                if (nx % QK_K != 0 || ny % QK_K != 0) {
+                if (nx % QK_K != 0) {
-                    LLAMA_LOG_INFO("\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
+                    LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K);
                     convert_incompatible_tensor = true;
                 }
             }