move ndk code to a new library (#6951)

* Add left recursion check: quit early instead of going into an infinite loop

* Remove custom enum, rename left recursion check and move to "grammar internal" section, add handling for edge case where a leftmost nonterminal may be empty

* Remove unnecessary declaration

.github/workflows/build.yml

@@ -898,9 +898,9 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
-
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:
       - name: Clone
         id: checkout
@@ -932,6 +932,17 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
+          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload artifacts

convert-hf-to-gguf-update.py

@@ -74,9 +74,9 @@ models = [
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
     {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
-    {"name": "jina-en",        "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
-    {"name": "jina-es",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
-    {"name": "jina-de",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
+    {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
+    {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
 ]
 
 # make directory "models/tokenizers" if it doesn't exist

convert-hf-to-gguf.py

@@ -240,23 +240,6 @@ class Model:
         return False
 
     def write_tensors(self):
-        # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
-        def np_fp32_to_bf16(n: np.ndarray):
-            # force nan to quiet
-            n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
-            # flush subnormals to zero
-            n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
-            # round to nearest even
-            n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
-            return n.astype(np.int16)
-
-        # Doing this row-wise is much, much faster than element-wise, hence the signature
-        v_fp32_to_bf16 = np.vectorize(np_fp32_to_bf16, otypes=[np.int16], signature="(n)->(n)")
-        if self.lazy:
-            # TODO: find a way to implicitly wrap np.vectorize functions
-            # NOTE: the type is changed to reflect otypes passed to np.vectorize above
-            v_fp32_to_bf16 = gguf.LazyNumpyTensor._wrap_fn(v_fp32_to_bf16, meta_noop=np.int16)
-
         max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
 
         for name, data_torch in self.get_tensors():
@@ -309,27 +292,31 @@ class Model:
                 ))
 
                 if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
-                    if self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                    if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        data = gguf.quantize_bf16(data)
+                        assert data.dtype == np.int16
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
+                        data = gguf.quantize_q8_0(data)
+                        assert data.dtype == np.uint8
+                        data_qtype = gguf.GGMLQuantizationType.Q8_0
+
+                    else:  # default to float16 for quantized tensors
                         if data_dtype != np.float16:
                             data = data.astype(np.float16)
                         data_qtype = gguf.GGMLQuantizationType.F16
 
-                    elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                        if data_dtype != np.float32:
-                            data = data.astype(np.float32)
-                        data = v_fp32_to_bf16(data.view(np.int32))
-                        assert data.dtype == np.int16
-                        data_qtype = gguf.GGMLQuantizationType.BF16
-
-                else:  # by default, convert to float32
+                if data_qtype is None:  # by default, convert to float32
                     if data_dtype != np.float32:
                         data = data.astype(np.float32)
                     data_qtype = gguf.GGMLQuantizationType.F32
 
-                assert data_qtype is not None
-
+                block_size, type_size = gguf.GGML_QUANT_SIZES[data_qtype]
                 # reverse shape to make it similar to the internal ggml dimension order
-                shape_str = f"{{{', '.join(str(n) for n in reversed(data.shape))}}}"
+                shape_str = f"""{{{', '.join(str(n) for n in reversed(
+                    (*data.shape[:-1], data.shape[-1] * data.dtype.itemsize // type_size * block_size))
+                )}}}"""
 
                 # n_dims is implicit in the shape
                 logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
@@ -475,13 +462,13 @@ class Model:
             res = "dbrx"
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-en
-            res = "jina-en"
+            res = "jina-v2-en"
         if chkhsh == "171aeeedd6fb548d418a7461d053f11b6f1f1fc9b387bd66640d28a4b9f5c643":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-es
-            res = "jina-es"
+            res = "jina-v2-es"
         if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de
-            res = "jina-de"
+            res = "jina-v2-de"
 
         if res is None:
             logger.warning("\n")
@@ -859,6 +846,7 @@ class BaichuanModel(Model):
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
 
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
             if self.hparams["rope_scaling"].get("type") == "linear":
@@ -981,6 +969,7 @@ class XverseModel(Model):
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
 
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
             if self.hparams["rope_scaling"].get("type") == "linear":
@@ -1215,6 +1204,7 @@ class StableLMModel(Model):
         self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])
         self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
         self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
+        self.gguf_writer.add_file_type(self.ftype)
 
     _q_norms: list[dict[str, Tensor]] | None = None
     _k_norms: list[dict[str, Tensor]] | None = None
@@ -1591,6 +1581,7 @@ class QwenModel(Model):
         self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
         self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
 
 
 @Model.register("Qwen2ForCausalLM")
@@ -1828,6 +1819,7 @@ class PlamoModel(Model):
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(5)  # hparams["num_key_value_heads"]) is wrong
         self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
 
     def shuffle_attn_q_weight(self, data_torch):
         assert data_torch.size() == (5120, 5120)
@@ -2007,6 +1999,7 @@ in chat mode so that the conversation can end normally.")
         self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
+        self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]
@@ -2415,25 +2408,15 @@ class LazyTorchTensor(gguf.LazyBase):
     def numpy(self) -> gguf.LazyNumpyTensor:
         dtype = self._dtype_map[self.dtype]
         return gguf.LazyNumpyTensor(
-            meta=np.lib.stride_tricks.as_strided(np.zeros(1, dtype), self.shape, (0 for _ in self.shape)),
+            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape),
             lazy=self._lazy,
             args=(self,),
             func=(lambda s: s[0].numpy())
         )
 
     @classmethod
-    def eager_to_meta(cls, t: Tensor) -> Tensor:
-        if t.is_meta:
-            return t
-        return t.detach().to("meta")
-
-    @classmethod
-    def meta_with_dtype(cls, m: Tensor, dtype: torch.dtype) -> Tensor:
-        m = m.detach()
-        if not m.is_meta:
-            m = m.to("meta")
-        m.dtype = dtype
-        return m
+    def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: torch.Size) -> Tensor:
+        return torch.empty(size=shape, dtype=dtype, device="meta")
 
     @classmethod
     def __torch_function__(cls, func, types, args=(), kwargs=None):
@@ -2464,8 +2447,8 @@ def parse_args() -> argparse.Namespace:
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
     )
     parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "auto"], default="f16",
-        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
     )
     parser.add_argument(
         "--bigendian", action="store_true",
@@ -2523,6 +2506,7 @@ def main() -> None:
         "f32": gguf.LlamaFileType.ALL_F32,
         "f16": gguf.LlamaFileType.MOSTLY_F16,
         "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
         "auto": gguf.LlamaFileType.GUESSED,
     }
 

convert.py

@@ -24,7 +24,7 @@ from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable
+from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable, Optional
 
 import numpy as np
 from sentencepiece import SentencePieceProcessor
@@ -344,10 +344,47 @@ class Params:
         return params
 
 
+@dataclass
+class Metadata:
+    name: Optional[str] = None
+    author: Optional[str] = None
+    version: Optional[str] = None
+    url: Optional[str] = None
+    description: Optional[str] = None
+    licence: Optional[str] = None
+    source_url: Optional[str] = None
+    source_hf_repo: Optional[str] = None
+
+    @staticmethod
+    def load(metadata_path: Path) -> Metadata:
+        if metadata_path is None or not metadata_path.exists():
+            return Metadata()
+
+        with open(metadata_path, 'r') as file:
+            data = json.load(file)
+
+        # Create a new Metadata instance
+        metadata = Metadata()
+
+        # Assigning values to Metadata attributes if they exist in the JSON file
+        # This is based on LLM_KV_NAMES mapping in llama.cpp
+        metadata.name = data.get("general.name")
+        metadata.author = data.get("general.author")
+        metadata.version = data.get("general.version")
+        metadata.url = data.get("general.url")
+        metadata.description = data.get("general.description")
+        metadata.license = data.get("general.license")
+        metadata.source_url = data.get("general.source.url")
+        metadata.source_hf_repo = data.get("general.source.huggingface.repository")
+
+        return metadata
+
+
 #
 # vocab
 #
 
+
 @runtime_checkable
 class BaseVocab(Protocol):
     tokenizer_model: ClassVar[str]
@@ -1066,21 +1103,42 @@ class OutputFile:
     def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
-    def add_meta_arch(self, params: Params) -> None:
+    def add_meta_model(self, params: Params, metadata: Metadata) -> None:
+        # Metadata About The Model And Its Provenence
         name = "LLaMA"
-
-        # TODO: better logic to determine model name
-        if params.n_ctx == 4096:
-            name = "LLaMA v2"
+        if metadata is not None and metadata.name is not None:
+            name = metadata.name
         elif params.path_model is not None:
-            name = str(params.path_model.parent).split('/')[-1]
+            name = str(params.path_model.parent).split("/")[-1]
+        elif params.n_ctx == 4096:
+            # Heuristic detection of LLaMA v2 model
+            name = "LLaMA v2"
 
-        self.gguf.add_name                (name)
-        self.gguf.add_vocab_size          (params.n_vocab)
-        self.gguf.add_context_length      (params.n_ctx)
-        self.gguf.add_embedding_length    (params.n_embd)
-        self.gguf.add_block_count         (params.n_layer)
-        self.gguf.add_feed_forward_length (params.n_ff)
+        self.gguf.add_name(name)
+
+        if metadata is not None:
+            if metadata.author is not None:
+                self.gguf.add_author(metadata.author)
+            if metadata.version is not None:
+                self.gguf.add_version(metadata.version)
+            if metadata.url is not None:
+                self.gguf.add_url(metadata.url)
+            if metadata.description is not None:
+                self.gguf.add_description(metadata.description)
+            if metadata.licence is not None:
+                self.gguf.add_licence(metadata.licence)
+            if metadata.source_url is not None:
+                self.gguf.add_source_url(metadata.source_url)
+            if metadata.source_hf_repo is not None:
+                self.gguf.add_source_hf_repo(metadata.source_hf_repo)
+
+    def add_meta_arch(self, params: Params) -> None:
+        # Metadata About The Neural Architecture Itself
+        self.gguf.add_vocab_size(params.n_vocab)
+        self.gguf.add_context_length(params.n_ctx)
+        self.gguf.add_embedding_length(params.n_embd)
+        self.gguf.add_block_count(params.n_layer)
+        self.gguf.add_feed_forward_length(params.n_ff)
         self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
@@ -1183,13 +1241,14 @@ class OutputFile:
     @staticmethod
     def write_vocab_only(
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
+        of.add_meta_model(params, metadata)
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
         of.add_meta_special_vocab(svocab)
@@ -1216,12 +1275,14 @@ class OutputFile:
         fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
         concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
+        metadata: Metadata = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
+        of.add_meta_model(params, metadata)
         of.add_meta_arch(params)
         if isinstance(vocab, Vocab):
             of.add_meta_vocab(vocab)
@@ -1257,6 +1318,37 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
     raise ValueError(f"Unexpected combination of types: {name_to_type}")
 
 
+def model_parameter_count(model: LazyModel) -> int:
+    total_model_parameters = 0
+    for i, (name, lazy_tensor) in enumerate(model.items()):
+        sum_weights_in_tensor = 1
+        for dim in lazy_tensor.shape:
+            sum_weights_in_tensor *= dim
+        total_model_parameters += sum_weights_in_tensor
+    return total_model_parameters
+
+
+def model_parameter_count_rounded_notation(model_params_count: int) -> str:
+    if model_params_count > 1e12 :
+        # Trillions Of Parameters
+        scaled_model_params = model_params_count * 1e-12
+        scale_suffix = "T"
+    elif model_params_count > 1e9 :
+        # Billions Of Parameters
+        scaled_model_params = model_params_count * 1e-9
+        scale_suffix = "B"
+    elif model_params_count > 1e6 :
+        # Millions Of Parameters
+        scaled_model_params = model_params_count * 1e-6
+        scale_suffix = "M"
+    else:
+        # Thousands Of Parameters
+        scaled_model_params = model_params_count * 1e-3
+        scale_suffix = "K"
+
+    return f"{round(scaled_model_params)}{scale_suffix}"
+
+
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
     return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
             for (name, tensor) in model.items()}
@@ -1436,13 +1528,35 @@ class VocabFactory:
         return vocab, special_vocab
 
 
-def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
-    namestr = {
-        GGMLFileType.AllF32:    "f32",
-        GGMLFileType.MostlyF16: "f16",
-        GGMLFileType.MostlyQ8_0:"q8_0",
+def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str:
+    quantization = {
+        GGMLFileType.AllF32:    "F32",
+        GGMLFileType.MostlyF16: "F16",
+        GGMLFileType.MostlyQ8_0: "Q8_0",
     }[file_type]
-    ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
+
+    parameters = model_parameter_count_rounded_notation(model_params_count)
+
+    expert_count = ""
+    if params.n_experts is not None:
+        expert_count = f"{params.n_experts}x"
+
+    version = ""
+    if metadata is not None and metadata.version is not None:
+        version = f"-{metadata.version}"
+
+    name = "ggml-model"
+    if metadata is not None and metadata.name is not None:
+        name = metadata.name
+    elif params.path_model is not None:
+        name = params.path_model.name
+
+    return f"{name}{version}-{expert_count}{parameters}-{quantization}"
+
+
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path:
+    default_filename = default_convention_outfile(file_type, params, model_params_count, metadata)
+    ret = model_paths[0].parent / f"{default_filename}.gguf"
     if ret in model_paths:
         logger.error(
             f"Error: Default output path ({ret}) would overwrite the input. "
@@ -1480,17 +1594,30 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
     parser.add_argument("--verbose",      action="store_true",    help="increase output verbosity")
+    parser.add_argument("--metadata",     type=Path,              help="Specify the path for a metadata file")
+    parser.add_argument("--get-outfile",  action="store_true",    help="get calculated default outfile name")
 
     args = parser.parse_args(args_in)
 
     if args.verbose:
         logging.basicConfig(level=logging.DEBUG)
-    elif args.dump_single or args.dump:
+    elif args.dump_single or args.dump or args.get_outfile:
         # Avoid printing anything besides the dump output
         logging.basicConfig(level=logging.WARNING)
     else:
         logging.basicConfig(level=logging.INFO)
 
+    metadata = Metadata.load(args.metadata)
+
+    if args.get_outfile:
+        model_plus = load_some_model(args.model)
+        params = Params.load(model_plus)
+        model   = convert_model_names(model_plus.model, params, args.skip_unknown)
+        model_params_count = model_parameter_count(model_plus.model)
+        ftype   = pick_output_type(model, args.outtype)
+        print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100
+        return
+
     if args.no_vocab and args.vocab_only:
         raise ValueError("--vocab-only does not make sense with --no-vocab")
 
@@ -1504,6 +1631,9 @@ def main(args_in: list[str] | None = None) -> None:
     else:
         model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
 
+    model_params_count = model_parameter_count(model_plus.model)
+    logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})")
+
     if args.dump:
         do_dump_model(model_plus)
         return
@@ -1557,7 +1687,7 @@ def main(args_in: list[str] | None = None) -> None:
                 f_norm_eps = 1e-5,
             )
         OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
-                                    endianess=endianess, pad_vocab=args.pad_vocab)
+                                    endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
         logger.info(f"Wrote {outfile}")
         return
 
@@ -1570,13 +1700,13 @@ def main(args_in: list[str] | None = None) -> None:
     model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
-    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata)
 
     params.ftype = ftype
     logger.info(f"Writing {outfile}, format {ftype}")
 
     OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
-                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab)
+                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
     logger.info(f"Wrote {outfile}")
 
 

examples/llama.android/app/build.gradle.kts

@@ -7,8 +7,6 @@ android {
     namespace = "com.example.llama"
     compileSdk = 34
 
-    ndkVersion = "26.1.10909125"
-
     defaultConfig {
         applicationId = "com.example.llama"
         minSdk = 33
@@ -20,17 +18,6 @@ android {
         vectorDrawables {
             useSupportLibrary = true
         }
-        ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
-        }
-        externalNativeBuild {
-            cmake {
-                arguments += "-DCMAKE_BUILD_TYPE=Release"
-                cppFlags += listOf()
-                arguments += listOf()
-            }
-        }
     }
 
     buildTypes {
@@ -55,17 +42,6 @@ android {
     composeOptions {
         kotlinCompilerExtensionVersion = "1.5.1"
     }
-    packaging {
-        resources {
-            excludes += "/META-INF/{AL2.0,LGPL2.1}"
-        }
-    }
-    externalNativeBuild {
-        cmake {
-            path = file("src/main/cpp/CMakeLists.txt")
-            version = "3.22.1"
-        }
-    }
 }
 
 dependencies {
@@ -78,6 +54,7 @@ dependencies {
     implementation("androidx.compose.ui:ui-graphics")
     implementation("androidx.compose.ui:ui-tooling-preview")
     implementation("androidx.compose.material3:material3")
+    implementation(project(":llama"))
     testImplementation("junit:junit:4.13.2")
     androidTestImplementation("androidx.test.ext:junit:1.1.5")
     androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")

examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt

@@ -1,5 +1,6 @@
 package com.example.llama
 
+import android.llama.cpp.LLamaAndroid
 import android.util.Log
 import androidx.compose.runtime.getValue
 import androidx.compose.runtime.mutableStateOf
@@ -9,7 +10,7 @@ import androidx.lifecycle.viewModelScope
 import kotlinx.coroutines.flow.catch
 import kotlinx.coroutines.launch
 
-class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
+class MainViewModel(private val llamaAndroid: LLamaAndroid = LLamaAndroid.instance()): ViewModel() {
     companion object {
         @JvmStatic
         private val NanosPerSecond = 1_000_000_000.0
@@ -28,7 +29,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
 
         viewModelScope.launch {
             try {
-                llm.unload()
+                llamaAndroid.unload()
             } catch (exc: IllegalStateException) {
                 messages += exc.message!!
             }
@@ -44,7 +45,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
         messages += ""
 
         viewModelScope.launch {
-            llm.send(text)
+            llamaAndroid.send(text)
                 .catch {
                     Log.e(tag, "send() failed", it)
                     messages += it.message!!
@@ -57,7 +58,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
         viewModelScope.launch {
             try {
                 val start = System.nanoTime()
-                val warmupResult = llm.bench(pp, tg, pl, nr)
+                val warmupResult = llamaAndroid.bench(pp, tg, pl, nr)
                 val end = System.nanoTime()
 
                 messages += warmupResult
@@ -70,7 +71,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
                     return@launch
                 }
 
-                messages += llm.bench(512, 128, 1, 3)
+                messages += llamaAndroid.bench(512, 128, 1, 3)
             } catch (exc: IllegalStateException) {
                 Log.e(tag, "bench() failed", exc)
                 messages += exc.message!!
@@ -81,7 +82,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
     fun load(pathToModel: String) {
         viewModelScope.launch {
             try {
-                llm.load(pathToModel)
+                llamaAndroid.load(pathToModel)
                 messages += "Loaded $pathToModel"
             } catch (exc: IllegalStateException) {
                 Log.e(tag, "load() failed", exc)

examples/llama.android/build.gradle.kts

@@ -2,4 +2,5 @@
 plugins {
     id("com.android.application") version "8.2.0" apply false
     id("org.jetbrains.kotlin.android") version "1.9.0" apply false
+    id("com.android.library") version "8.2.0" apply false
 }

examples/llama.android/llama/.gitignore

@@ -0,0 +1 @@
+/build

examples/llama.android/llama/CMakeLists.txt

@@ -37,7 +37,7 @@ FetchContent_MakeAvailable(llama)
 # used in the AndroidManifest.xml file.
 add_library(${CMAKE_PROJECT_NAME} SHARED
     # List C/C++ source files with relative paths to this CMakeLists.txt.
-    llama-android.cpp)
+        llama-android.cpp)
 
 # Specifies libraries CMake should link to your target library. You
 # can link libraries from various origins, such as libraries defined in this

examples/llama.android/llama/proguard-rules.pro

@@ -0,0 +1,21 @@
+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile

examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt

@@ -0,0 +1,24 @@
+package android.llama.cpp
+
+import androidx.test.platform.app.InstrumentationRegistry
+import androidx.test.ext.junit.runners.AndroidJUnit4
+
+import org.junit.Test
+import org.junit.runner.RunWith
+
+import org.junit.Assert.*
+
+/**
+ * Instrumented test, which will execute on an Android device.
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+@RunWith(AndroidJUnit4::class)
+class ExampleInstrumentedTest {
+    @Test
+    fun useAppContext() {
+        // Context of the app under test.
+        val appContext = InstrumentationRegistry.getInstrumentation().targetContext
+        assertEquals("android.llama.cpp.test", appContext.packageName)
+    }
+}

examples/llama.android/llama/src/main/AndroidManifest.xml

@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android">
+
+</manifest>

examples/llama.android/llama/src/main/cpp/CMakeLists.txt

@@ -0,0 +1,49 @@
+# For more information about using CMake with Android Studio, read the
+# documentation: https://d.android.com/studio/projects/add-native-code.html.
+# For more examples on how to use CMake, see https://github.com/android/ndk-samples.
+
+# Sets the minimum CMake version required for this project.
+cmake_minimum_required(VERSION 3.22.1)
+
+# Declares the project name. The project name can be accessed via ${ PROJECT_NAME},
+# Since this is the top level CMakeLists.txt, the project name is also accessible
+# with ${CMAKE_PROJECT_NAME} (both CMake variables are in-sync within the top level
+# build script scope).
+project("llama-android")
+
+include(FetchContent)
+FetchContent_Declare(
+        llama
+        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
+        GIT_TAG        master
+)
+
+# Also provides "common"
+FetchContent_MakeAvailable(llama)
+
+# Creates and names a library, sets it as either STATIC
+# or SHARED, and provides the relative paths to its source code.
+# You can define multiple libraries, and CMake builds them for you.
+# Gradle automatically packages shared libraries with your APK.
+#
+# In this top level CMakeLists.txt, ${CMAKE_PROJECT_NAME} is used to define
+# the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME}
+# is preferred for the same purpose.
+#
+# In order to load a library into your app from Java/Kotlin, you must call
+# System.loadLibrary() and pass the name of the library defined here;
+# for GameActivity/NativeActivity derived applications, the same library name must be
+# used in the AndroidManifest.xml file.
+add_library(${CMAKE_PROJECT_NAME} SHARED
+        # List C/C++ source files with relative paths to this CMakeLists.txt.
+        llama-android.cpp)
+
+# Specifies libraries CMake should link to your target library. You
+# can link libraries from various origins, such as libraries defined in this
+# build script, prebuilt third-party libraries, or Android system libraries.
+target_link_libraries(${CMAKE_PROJECT_NAME}
+        # List libraries link to the target library
+        llama
+        common
+        android
+        log)

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -81,7 +81,7 @@ static void log_callback(ggml_log_level level, const char * fmt, void * data) {
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
+Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring filename) {
     llama_model_params model_params = llama_model_default_params();
 
     auto path_to_model = env->GetStringUTFChars(filename, 0);
@@ -101,13 +101,13 @@ Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1model(JNIEnv *, jobject, jlong model) {
+Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
     llama_free_model(reinterpret_cast<llama_model *>(model));
 }
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
+Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmodel) {
     auto model = reinterpret_cast<llama_model *>(jmodel);
 
     if (!model) {
@@ -139,25 +139,25 @@ Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1context(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_free_1context(JNIEnv *, jobject, jlong context) {
     llama_free(reinterpret_cast<llama_context *>(context));
 }
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1free(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1free(JNIEnv *, jobject) {
     llama_backend_free();
 }
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_log_1to_1android(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_log_1to_1android(JNIEnv *, jobject) {
     llama_log_set(log_callback, NULL);
 }
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_bench_1model(
+Java_android_llama_cpp_LLamaAndroid_bench_1model(
         JNIEnv *env,
         jobject,
         jlong context_pointer,
@@ -271,13 +271,13 @@ Java_com_example_llama_Llm_bench_1model(
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
+Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
     llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
 }
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
+Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
 
     // Source: Copy of llama.cpp:llama_batch_init but heap-allocated.
 
@@ -313,19 +313,19 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) {
     llama_backend_init();
 }
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_system_1info(JNIEnv *env, jobject) {
+Java_android_llama_cpp_LLamaAndroid_system_1info(JNIEnv *env, jobject) {
     return env->NewStringUTF(llama_print_system_info());
 }
 
 extern "C"
 JNIEXPORT jint JNICALL
-Java_com_example_llama_Llm_completion_1init(
+Java_android_llama_cpp_LLamaAndroid_completion_1init(
         JNIEnv *env,
         jobject,
         jlong context_pointer,
@@ -376,7 +376,7 @@ Java_com_example_llama_Llm_completion_1init(
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_completion_1loop(
+Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         JNIEnv * env,
         jobject,
         jlong context_pointer,
@@ -438,6 +438,6 @@ Java_com_example_llama_Llm_completion_1loop(
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
     llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
 }

examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt

@@ -1,4 +1,4 @@
-package com.example.llama
+package android.llama.cpp
 
 import android.util.Log
 import kotlinx.coroutines.CoroutineDispatcher
@@ -10,7 +10,7 @@ import kotlinx.coroutines.withContext
 import java.util.concurrent.Executors
 import kotlin.concurrent.thread
 
-class Llm {
+class LLamaAndroid {
     private val tag: String? = this::class.simpleName
 
     private val threadLocalState: ThreadLocal<State> = ThreadLocal.withInitial { State.Idle }
@@ -165,8 +165,8 @@ class Llm {
         }
 
         // Enforce only one instance of Llm.
-        private val _instance: Llm = Llm()
+        private val _instance: LLamaAndroid = LLamaAndroid()
 
-        fun instance(): Llm = _instance
+        fun instance(): LLamaAndroid = _instance
     }
 }

examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt

@@ -0,0 +1,17 @@
+package android.llama.cpp
+
+import org.junit.Test
+
+import org.junit.Assert.*
+
+/**
+ * Example local unit test, which will execute on the development machine (host).
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+class ExampleUnitTest {
+    @Test
+    fun addition_isCorrect() {
+        assertEquals(4, 2 + 2)
+    }
+}

examples/llama.android/settings.gradle.kts

@@ -15,3 +15,4 @@ dependencyResolutionManagement {
 
 rootProject.name = "LlamaAndroid"
 include(":app")
+include(":llama")

examples/llava/llava-cli.cpp

@@ -300,14 +300,10 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    for (auto & image : params.image) {
+    if (prompt_contains_image(params.prompt)) {
         auto ctx_llava = llava_init_context(&params, model);
 
-        auto image_embed = load_image(ctx_llava, &params, image);
-        if (!image_embed) {
-            std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
-            return 1;
-        }
+        auto image_embed = load_image(ctx_llava, &params, "");
 
         // process the prompt
         process_prompt(ctx_llava, image_embed, &params, params.prompt);
@@ -316,7 +312,26 @@ int main(int argc, char ** argv) {
         llava_image_embed_free(image_embed);
         ctx_llava->model = NULL;
         llava_free(ctx_llava);
+    } else {
+        for (auto & image : params.image) {
+            auto ctx_llava = llava_init_context(&params, model);
+
+            auto image_embed = load_image(ctx_llava, &params, image);
+            if (!image_embed) {
+                std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
+                return 1;
+            }
+
+            // process the prompt
+            process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+            llama_print_timings(ctx_llava->ctx_llama);
+            llava_image_embed_free(image_embed);
+            ctx_llava->model = NULL;
+            llava_free(ctx_llava);
+        }
     }
+
     llama_free_model(model);
 
     return 0;

examples/perplexity/README.md

@@ -7,6 +7,8 @@ Also note that finetunes typically result in a higher perplexity value even thou
 
 Within llama.cpp the perplexity of base models is used primarily to judge the quality loss from e.g. quantized models vs. FP16.
 The convention among contributors is to use the Wikitext-2 test set for testing unless noted otherwise (can be obtained with `scripts/get-wikitext-2.sh`).
+When numbers are listed all command line arguments and compilation options are left at their defaults unless noted otherwise.
+llama.cpp numbers are **not** directly comparable to those of other projects because the exact values depend strongly on the implementation details.
 
 By default only the mean perplexity value and the corresponding uncertainty is calculated.
 The uncertainty is determined empirically by assuming a Gaussian distribution of the "correct" logits per and then applying error propagation.
@@ -32,7 +34,13 @@ In addition to the KL divergence the following statistics are calculated with `-
 
 ## LLaMA 3 8b Scoreboard
 
-Results are sorted by Kullback-Leibler divergence relative to FP16.
+| Revision | f364eb6f           |
+|:---------|:-------------------|
+| Backend  | CUDA               |
+| CPU      | AMD Epyc 7742      |
+| GPU      | 1x NVIDIA RTX 4090 |
+
+Results were generated using the CUDA backend and are sorted by Kullback-Leibler divergence relative to FP16.
 The "WT" importance matrices were created using varying numbers of Wikitext tokens and can be found [here](https://huggingface.co/JohannesGaessler/llama.cpp_importance_matrices/blob/main/imatrix-llama_3-8b-f16-2.7m_tokens.dat).
 
 | Quantization | imatrix | Model size [GiB] | PPL                    | ΔPPL                   | KLD                   | Mean Δp           | RMS Δp           |
@@ -89,6 +97,12 @@ K-quants score better on mean Δp than the legacy quants than e.g. KL divergence
 
 ## LLaMA 2 vs. LLaMA 3 Quantization comparison
 
+| Revision | f364eb6f           |
+|:---------|:-------------------|
+| Backend  | CUDA               |
+| CPU      | AMD Epyc 7742      |
+| GPU      | 1x NVIDIA RTX 4090 |
+
 | Metric          |          L2 7b q2_K |          L3 8b q2_K |        L2 7b q4_K_M |        L3 8b q4_K_M |          L2 7b q6_K |          L3 8b q6_K |          L2 7b q8_0 |          L3 8b q8_0 |
 |-----------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|
 | Mean PPL        | 5.794552 ± 0.032298 | 9.751568 ± 0.063312 | 5.877078 ± 0.032781 | 6.407115 ± 0.039119 | 5.808494 ± 0.032425 | 6.253382 ± 0.038078 | 5.798542 ± 0.032366 | 6.234284 ± 0.037878 |
@@ -107,6 +121,50 @@ K-quants score better on mean Δp than the legacy quants than e.g. KL divergence
 | RMS Δp          |     9.762 ± 0.053 % |    21.421 ± 0.079 % |     3.252 ± 0.024 % |     5.519 ± 0.050 % |     1.339 ± 0.010 % |     2.295 ± 0.019 % |     0.618 ± 0.011 % |     1.198 ± 0.007 % |
 | Same top p      |    85.584 ± 0.086 % |    71.138 ± 0.119 % |    94.665 ± 0.055 % |    91.901 ± 0.072 % |    97.520 ± 0.038 % |    96.031 ± 0.051 % |    98.846 ± 0.026 % |    97.674 ± 0.040 % |
 
+## LLaMA 3 BF16 vs. FP16 comparison
+
+| Revision | 83330d8c      |
+|:---------|:--------------|
+| Backend  | CPU           |
+| CPU      | AMD Epyc 7742 |
+| GPU      | N/A           |
+
+Results were calculated with LLaMA 3 8b BF16 as `--kl-divergence-base` and LLaMA 3 8b FP16 as the `--model` for comparison.
+
+| Metric                         |                    Value |
+|--------------------------------|--------------------------|
+| Mean PPL(Q)                    |      6.227711 ± 0.037833 |
+| Mean PPL(base)                 |      6.225194 ± 0.037771 |
+| Cor(ln(PPL(Q)), ln(PPL(base))) |                  99.990% |
+| Mean ln(PPL(Q)/PPL(base))      |      0.000404 ± 0.000086 |
+| Mean PPL(Q)/PPL(base)          |      1.000404 ± 0.000086 |
+| Mean PPL(Q)-PPL(base)          |      0.002517 ± 0.000536 |
+| Mean    KLD                    |  0.00002515 ± 0.00000020 |
+| Maximum KLD                    |                 0.012206 |
+| 99.9%   KLD                    |                 0.000799 |
+| 99.0%   KLD                    |                 0.000222 |
+| 99.0%   KLD                    |                 0.000222 |
+| Median  KLD                    |                 0.000013 |
+| 10.0%   KLD                    |                -0.000002 |
+| 5.0%   KLD                     |                -0.000008 |
+| 1.0%   KLD                     |                -0.000023 |
+| Minimum KLD                    |                -0.000059 |
+| Mean    Δp                     | -0.0000745 ± 0.0003952 % |
+| Maximum Δp                     |                   4.186% |
+| 99.9%   Δp                     |                   1.049% |
+| 99.0%   Δp                     |                   0.439% |
+| 95.0%   Δp                     |                   0.207% |
+| 90.0%   Δp                     |                   0.125% |
+| 75.0%   Δp                     |                   0.029% |
+| Median  Δp                     |                   0.000% |
+| 25.0%   Δp                     |                  -0.030% |
+| 10.0%   Δp                     |                  -0.126% |
+| 5.0%   Δp                      |                  -0.207% |
+| 1.0%   Δp                      |                  -0.434% |
+| 0.1%   Δp                      |                  -1.016% |
+| Minimum Δp                     |                  -4.672% |
+| RMS Δp                         |          0.150 ± 0.001 % |
+| Same top p                     |         99.739 ± 0.013 % |
 
 ## Old Numbers
 

examples/server/README.md

@@ -48,7 +48,7 @@ page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/
 - `--path`: Path from which to serve static files. Default: disabled
 - `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
 - `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s.
-- `--embedding`: Enable embedding extraction. Default: disabled
+- `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled
 - `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`
 - `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching).  Default: disabled
 - `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime)

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

@@ -887,6 +887,7 @@ async def oai_chat_completions(user_prompt,
                                base_path,
                                async_client,
                                debug=False,
+                               temperature=None,
                                model=None,
                                n_predict=None,
                                enable_streaming=None,
@@ -913,7 +914,8 @@ async def oai_chat_completions(user_prompt,
         "model": model,
         "max_tokens": n_predict,
         "stream": enable_streaming,
-        "seed": seed
+        "temperature": temperature if temperature is not None else 0.0,
+        "seed": seed,
     }
     if response_format is not None:
         payload['response_format'] = response_format
@@ -978,7 +980,8 @@ async def oai_chat_completions(user_prompt,
                 max_tokens=n_predict,
                 stream=enable_streaming,
                 response_format=payload.get('response_format'),
-                seed=seed
+                seed=seed,
+                temperature=payload['temperature']
             )
         except openai.error.AuthenticationError as e:
             if expect_api_error is not None and expect_api_error:

examples/server/utils.hpp

@@ -371,7 +371,7 @@ static json oaicompat_completion_params_parse(
     llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
     llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
     llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       0.0);
+    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
     llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
 
     // Apply chat template to the list of messages

ggml-cuda.cu

@@ -2713,6 +2713,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 }
 
 GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -2840,8 +2841,16 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
-        case GGML_OP_FLASH_ATTN_EXT:
             return true;
+        case GGML_OP_FLASH_ATTN_EXT:
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+#else
+            if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
+                return true;
+            }
+            return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         default:
             return false;
     }

ggml-cuda/common.cuh

@@ -321,6 +321,10 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 
 #define FP16_MMA_AVAILABLE !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 
+static bool fast_fp16_available(const int cc) {
+    return cc >= CC_PASCAL && cc != 610;
+}
+
 static bool fp16_mma_available(const int cc) {
     return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
 }

ggml-cuda/fattn-common.cuh

@@ -0,0 +1,47 @@
+#define FATTN_KQ_STRIDE       256
+#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
+#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
+
+template<int D, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_combine_results(
+        const float  * __restrict__ VKQ_parts,
+        const float2 * __restrict__ VKQ_meta,
+        float * __restrict__ dst) {
+    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
+    dst       +=                 D * gridDim.y*blockIdx.x;
+
+    const int tid = threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float2 meta[parallel_blocks];
+    if (tid < 2*parallel_blocks) {
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+    }
+
+    __syncthreads();
+
+    float kqmax = meta[0].x;
+#pragma unroll
+    for (int l = 1; l < parallel_blocks; ++l) {
+        kqmax = max(kqmax, meta[l].x);
+    }
+
+    float VKQ_numerator   = 0.0f;
+    float VKQ_denominator = 0.0f;
+#pragma unroll
+    for (int l = 0; l < parallel_blocks; ++l) {
+        const float diff = meta[l].x - kqmax;
+        const float KQ_max_scale = expf(diff);
+        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
+        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
+
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_denominator += KQ_max_scale * meta[l].y;
+    }
+
+    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+}

ggml-cuda/fattn-vec-f16.cu

@@ -0,0 +1,430 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f16.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    half  slopeh = __float2half(1.0f);
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = blockIdx.y;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slopeh = __float2half(powf(base, exph));
+    }
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ half KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -HALF_MAX_HALF;
+    }
+    half2 * KQ2 = (half2 *) KQ;
+
+    half kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -HALF_MAX_HALF;
+    }
+    half kqsum[ncols] = {0.0f};
+
+    __shared__ half kqmax_shared[ncols][WARP_SIZE];
+    __shared__ half kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    half2 VKQ[ncols] = {{0.0f, 0.0f}};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
+        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
+        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
+        half kqmax_new = kqmax[0];
+        half kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            half2 sum2[ncols] = {{0.0f, 0.0f}};
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum2[j] = warp_reduce_sum(sum2[j]);
+                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
+                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                if (ncols == 1) {
+                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
+                } else {
+                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
+                }
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= __half2half2(KQ_max_scale);
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < D; k0 += 2) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
+                break;
+            }
+
+            half2 V_k;
+            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
+            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid != 0) {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
+        }
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f16(
+        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
+        ggml_cuda_pool & pool, cudaStream_t main_stream
+) {
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
+    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const     int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>
+        <<<blocks_num, block_dim, shmem, main_stream>>> (
+                (const char *) Q->data,
+                (const char *) K->data,
+                (const char *) V->data,
+                mask ? ((const char *) mask->data) : nullptr,
+                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+                scale, max_bias, m0, m1, n_head_log2,
+                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+                Q->nb[1], Q->nb[2], Q->nb[3],
+                K->nb[1], K->nb[2], K->nb[3],
+                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+                );
+    CUDA_CHECK(cudaGetLastError());
+
+    if (parallel_blocks == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    constexpr int cols_per_block = 1;
+    constexpr int parallel_blocks = 4;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 256:
+            launch_fattn_vec_f16<256, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block = 1;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block = 2;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block = 4;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block = 8;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 8;
+    constexpr int parallel_blocks = 1;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}

ggml-cuda/fattn-vec-f16.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn-vec-f32.cu

@@ -0,0 +1,384 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f32.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = blockIdx.y;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = powf(base, exph);
+    }
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -FLT_MAX/2.0f;
+    }
+
+    float kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols] = {0.0f};
+
+    __shared__ float kqmax_shared[ncols][WARP_SIZE];
+    __shared__ float kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    float2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            Q_h2[j][i0/WARP_SIZE]    = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE].x *= scale;
+            Q_h2[j][i0/WARP_SIZE].y *= scale;
+        }
+    }
+
+    float VKQ[ncols] = {0.0f};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            float sum[ncols] = {0.0f};
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum[j] +=  __low2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].x;
+                    sum[j] += __high2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].y;
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum[j] = warp_reduce_sum(sum[j]);
+                sum[j] += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum[j]);
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum[j];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const float val = expf(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= KQ_max_scale;
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < D; ++k) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k >= ne11) {
+                break;
+            }
+
+            const float V_ki = __half2float(V_h[(k_VKQ_0 + k)*stride_KV + tid]);
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_ki*KQ[j*D + k];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        float dst_val = VKQ[j_VKQ];
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid != 0) {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
+        }
+    }
+}
+
+template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f32(
+        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
+        ggml_cuda_pool & pool, cudaStream_t main_stream
+) {
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
+    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const     int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>
+        <<<blocks_num, block_dim, shmem, main_stream>>> (
+                (const char *) Q->data,
+                (const char *) K->data,
+                (const char *) V->data,
+                mask ? ((const char *) mask->data) : nullptr,
+                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+                scale, max_bias, m0, m1, n_head_log2,
+                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+                Q->nb[1], Q->nb[2], Q->nb[3],
+                K->nb[1], K->nb[2], K->nb[3],
+                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+                );
+    CUDA_CHECK(cudaGetLastError());
+
+    if (parallel_blocks == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}
+
+void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block = 1;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block = 2;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block = 4;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block = 8;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 8;
+    constexpr int parallel_blocks = 1;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}

ggml-cuda/fattn-vec-f32.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn.cu

@@ -1,4 +1,7 @@
 #include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f16.cuh"
+#include "fattn-vec-f32.cuh"
 #include "fattn.cuh"
 
 #include <cstdint>
@@ -7,251 +10,6 @@
 #include <mma.h>
 #endif
 
-#define FATTN_KQ_STRIDE       256
-#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
-#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
-
-template<int D, int ncols, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-static __global__ void flash_attn_vec_ext_f16(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const float max_bias,
-        const float m0,
-        const float m1,
-        const uint32_t n_head_log2,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int nb31,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
-#if FP16_AVAILABLE
-    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
-
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic0;
-
-    const int stride_KV  = nb11 / sizeof(half);
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    half slopeh = __float2half(1.0f);
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int h = blockIdx.y;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slopeh = __float2half(powf(base, exph));
-    }
-
-    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
-    constexpr int nwarps = D / WARP_SIZE;
-    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ half KQ[ncols*D];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        KQ[j*D + tid] = -HALF_MAX_HALF;
-    }
-    half2 * KQ2 = (half2 *) KQ;
-
-    half kqmax[ncols];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqmax[j] = -HALF_MAX_HALF;
-    }
-    half kqsum[ncols] = {0.0f};
-
-    __shared__ half kqmax_shared[ncols][WARP_SIZE];
-    __shared__ half kqsum_shared[ncols][WARP_SIZE];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        if (threadIdx.y == 0) {
-            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
-            kqsum_shared[j][threadIdx.x] = 0.0f;
-        }
-    }
-    __syncthreads();
-
-    // Convert Q to half2 and store in registers:
-    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-#pragma unroll
-        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-            const int i = i0 + threadIdx.x;
-
-            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
-            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
-        }
-    }
-
-    half2 VKQ[ncols] = {{0.0f, 0.0f}};
-
-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
-        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
-        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
-        half kqmax_new = kqmax[0];
-        half kqmax_new_arr[ncols];
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            kqmax_new_arr[j] = kqmax[j];
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
-                break;
-            }
-
-            half2 sum2[ncols] = {{0.0f, 0.0f}};
-#pragma unroll
-            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
-                const int k_KQ = k_KQ_0 + threadIdx.x;
-
-                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
-#pragma unroll
-                for (int j = 0; j < ncols; ++j) {
-                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
-                }
-            }
-
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                sum2[j] = warp_reduce_sum(sum2[j]);
-                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
-                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
-
-                if (ncols == 1) {
-                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
-                } else {
-                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
-                }
-
-                if (threadIdx.x == 0) {
-                    KQ[j*D + i_KQ] = sum;
-                }
-            }
-        }
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
-
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-            if (threadIdx.x == 0) {
-                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
-            kqmax[j] = kqmax_new_j;
-
-            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
-            kqsum[j] = kqsum[j]*KQ_max_scale + val;
-            KQ[j*D + tid] = val;
-
-            VKQ[j] *= __half2half2(KQ_max_scale);
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k0 = 0; k0 < D; k0 += 2) {
-            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
-                break;
-            }
-
-            half2 V_k;
-            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
-            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
-            }
-        }
-
-        __syncthreads();
-    }
-
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqsum[j] = warp_reduce_sum(kqsum[j]);
-        if (threadIdx.x == 0) {
-            kqsum_shared[j][threadIdx.y] = kqsum[j];
-        }
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
-        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
-        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
-
-        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
-        if (parallel_blocks == 1) {
-            dst_val /= kqsum[j_VKQ];
-        }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
-    }
-
-    if (parallel_blocks != 1 && tid != 0) {
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
-        }
-    }
-#else
-   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
-}
-
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
 template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t>
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
@@ -655,54 +413,6 @@ static __global__ void flash_attn_ext_f16(
 #endif // FP16_MMA_AVAILABLE
 }
 
-template<int D, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-static __global__ void flash_attn_combine_results(
-        const float  * __restrict__ VKQ_parts,
-        const float2 * __restrict__ VKQ_meta,
-        float * __restrict__ dst) {
-#if FP16_AVAILABLE
-    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
-    dst       +=                 D * gridDim.y*blockIdx.x;
-
-    const int tid = threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ float2 meta[parallel_blocks];
-    if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
-    }
-
-    __syncthreads();
-
-    float kqmax = meta[0].x;
-#pragma unroll
-    for (int l = 1; l < parallel_blocks; ++l) {
-        kqmax = max(kqmax, meta[l].x);
-    }
-
-    float VKQ_numerator   = 0.0f;
-    float VKQ_denominator = 0.0f;
-#pragma unroll
-    for (int l = 0; l < parallel_blocks; ++l) {
-        const float diff = meta[l].x - kqmax;
-        const float KQ_max_scale = expf(diff);
-        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
-        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
-
-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
-        VKQ_denominator += KQ_max_scale * meta[l].y;
-    }
-
-    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
-#else
-   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
-}
-
 constexpr int get_max_power_of_2(int x) {
     return x % 2 == 0 ? 2*get_max_power_of_2(x/2) : 1;
 }
@@ -727,66 +437,6 @@ static_assert(get_VKQ_stride( 80, 1, 16) ==  16, "Test failed.");
 static_assert(get_VKQ_stride( 80, 2, 16) ==  16, "Test failed.");
 static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
 
-template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f16(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
-    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
-    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
-
-    if (parallel_blocks > 1) {
-        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
-        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
-    }
-
-    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
-    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
-    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
-    const     int  shmem = 0;
-
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
-
-    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
-
-    const uint32_t n_head      = Q->ne[2];
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
-
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-    flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>
-        <<<blocks_num, block_dim, shmem, main_stream>>> (
-                (const char *) Q->data,
-                (const char *) K->data,
-                (const char *) V->data,
-                mask ? ((const char *) mask->data) : nullptr,
-                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale, max_bias, m0, m1, n_head_log2,
-                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
-                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
-                Q->nb[1], Q->nb[2], Q->nb[3],
-                K->nb[1], K->nb[2], K->nb[3],
-                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
-                );
-    CUDA_CHECK(cudaGetLastError());
-
-    if (parallel_blocks == 1) {
-        return;
-    }
-
-    const dim3 block_dim_combine(D, 1, 1);
-    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
-    const int  shmem_combine = 0;
-
-    flash_attn_combine_results<D, parallel_blocks>
-        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
-        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
-    CUDA_CHECK(cudaGetLastError());
-}
-
 template <int D, int cols_per_block, int nwarps, int parallel_blocks, typename KQ_acc_t> void launch_fattn_f16_impl(
         const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
         ggml_cuda_pool & pool, cudaStream_t main_stream
@@ -891,95 +541,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     const int32_t precision = KQV->op_params[2];
 
+    if (!fast_fp16_available(cc)) {
+        ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        return;
+    }
+
     if (!fp16_mma_available(cc)) {
-        GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-        GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
-
-        if (Q->ne[1] == 1) {
-            constexpr int cols_per_block = 1;
-            constexpr int parallel_blocks = 4;
-            switch (Q->ne[0]) {
-                case 64:
-                    launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                case 128:
-                    launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                default:
-                    GGML_ASSERT(false);
-                    break;
-            }
-            return;
-        }
-
-        if (Q->ne[1] == 2) {
-            constexpr int cols_per_block = 2;
-            constexpr int parallel_blocks = 4;
-            switch (Q->ne[0]) {
-                case 64:
-                    launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                case 128:
-                    launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                default:
-                    GGML_ASSERT(false);
-                    break;
-            }
-            return;
-        }
-
-        if (Q->ne[1] <= 4) {
-            constexpr int cols_per_block = 4;
-            constexpr int parallel_blocks = 4;
-            switch (Q->ne[0]) {
-                case 64:
-                    launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                case 128:
-                    launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                default:
-                    GGML_ASSERT(false);
-                    break;
-            }
-            return;
-        }
-
-        if (Q->ne[1] <= 8) {
-            constexpr int cols_per_block = 8;
-            constexpr int parallel_blocks = 4;
-            switch (Q->ne[0]) {
-                case 64:
-                    launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                case 128:
-                    launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                default:
-                    GGML_ASSERT(false);
-                    break;
-            }
-            return;
-        }
-
-        constexpr int cols_per_block = 8;
-        constexpr int parallel_blocks = 1;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
         return;
     }
 
     if (precision != GGML_PREC_DEFAULT) {
+        if (Q->ne[1] == 1 && (Q->ne[0] == 64 || Q->ne[0] == 128)) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            return;
+        }
+
         if (Q->ne[1] <= 32 || Q->ne[0] > 128) {
             constexpr int cols_per_block = 16;
             constexpr int nwarps         =  4;
@@ -1037,22 +614,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     }
 
     if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) {
-        constexpr int cols_per_block = 1;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 256:
-                launch_fattn_vec_f16<256, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
         return;
     }
 

ggml-sycl.cpp

@@ -15564,26 +15564,6 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
     const int64_t r2 = ne12/ne02;
     const int64_t r3 = ne13/ne03;
 
-#if 0
-    // use syclGemmEx
-    {
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                int i03 = i13 / r3;
-                int i02 = i12 / r2;
-
-                SYCL_CHECK(
-                        syclGemmEx(g_sycl_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , SYCL_R_16F,   nb01/sizeof(half),
-                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, SYCL_R_16F,   nb11/sizeof(float),
-                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
-                            cu_compute_type,
-                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-            }
-        }
-    }
-#else
     if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
@@ -15595,7 +15575,6 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
             nb11 / nb10, nb12 / nb10, beta,
             (char *)dst_t, cu_data_type, ne01, nb2 / nb0,
             ne12 * ne13, cu_compute_type)));
-        g_sycl_handles[g_main_device]->wait();
     } else {
         const int ne23 = ne12*ne13;
 
@@ -15626,7 +15605,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
                                          nb02, nb03, nb12_scaled, nb13_scaled,
                                          nbd2, nbd3, r2, r3, item_ct1);
                                  });
-            }).wait();
+            });
         }
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
             *g_sycl_handles[g_main_device], oneapi::mkl::transpose::trans,
@@ -15637,9 +15616,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
             dpct::library_data_t::real_half, nb11 / nb10, beta,
             (void **)(ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23,
             cu_compute_type)));
-        g_sycl_handles[g_main_device]->wait();
     }
-#endif
 
     if (no_mixed_dtypes) {
         const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);

gguf-py/gguf/__init__.py

@@ -2,5 +2,6 @@ from .constants import *
 from .lazy import *
 from .gguf_reader import *
 from .gguf_writer import *
+from .quants import *
 from .tensor_mapping import *
 from .vocab import *

gguf-py/gguf/gguf_writer.py

@@ -13,6 +13,7 @@ from string import ascii_letters, digits
 import numpy as np
 
 from .constants import (
+    GGML_QUANT_SIZES,
     GGUF_DEFAULT_ALIGNMENT,
     GGUF_MAGIC,
     GGUF_VERSION,
@@ -195,7 +196,7 @@ class GGUFWriter:
         return ((x + n - 1) // n) * n
 
     def add_tensor_info(
-        self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype[np.float16] | np.dtype[np.float32],
+        self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype,
         tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None,
     ) -> None:
         if self.state is not WriterState.EMPTY:
@@ -208,10 +209,6 @@ class GGUFWriter:
         encoded_name = name.encode("utf-8")
         self.ti_data += self._pack("Q", len(encoded_name))
         self.ti_data += encoded_name
-        n_dims = len(tensor_shape)
-        self.ti_data += self._pack("I", n_dims)
-        for i in range(n_dims):
-            self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
         if raw_dtype is None:
             if tensor_dtype == np.float16:
                 dtype = GGMLQuantizationType.F16
@@ -231,6 +228,15 @@ class GGUFWriter:
                 raise ValueError("Only F16, F32, F64, I8, I16, I32, I64 tensors are supported for now")
         else:
             dtype = raw_dtype
+            if tensor_dtype == np.uint8:
+                block_size, type_size = GGML_QUANT_SIZES[raw_dtype]
+                if tensor_shape[-1] % type_size != 0:
+                    raise ValueError(f"Quantized tensor row size ({tensor_shape[-1]}) is not a multiple of {dtype.name} type size ({type_size})")
+                tensor_shape = tuple(tensor_shape[:-1]) + (tensor_shape[-1] // type_size * block_size,)
+        n_dims = len(tensor_shape)
+        self.ti_data += self._pack("I", n_dims)
+        for i in range(n_dims):
+            self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
         self.ti_data += self._pack("I", dtype)
         self.ti_data += self._pack("Q", self.offset_tensor)
         self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)

gguf-py/gguf/lazy.py

@@ -6,6 +6,7 @@ from typing import Any, Callable
 from collections import deque
 
 import numpy as np
+from numpy._typing import _Shape
 from numpy.typing import DTypeLike
 
 
@@ -110,7 +111,7 @@ class LazyBase(ABC, metaclass=LazyMeta):
             return o
 
     @classmethod
-    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike = False) -> Callable[[Any], Any]:
+    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike | tuple[DTypeLike, Callable[[tuple[int, ...]], tuple[int, ...]]] = False) -> Callable[[Any], Any]:
         def wrapped_fn(*args, **kwargs):
             if kwargs is None:
                 kwargs = {}
@@ -130,9 +131,14 @@ class LazyBase(ABC, metaclass=LazyMeta):
                 res = args[0]
                 assert isinstance(res, cls)
                 res = res._meta
-                # allow operations to override the dtype
+                # allow operations to override the dtype and shape
                 if meta_noop is not True:
-                    res = cls.meta_with_dtype(res, meta_noop)
+                    if isinstance(meta_noop, tuple):
+                        dtype, shape = meta_noop
+                        assert callable(shape)
+                        res = cls.meta_with_dtype_and_shape(dtype, shape(res.shape))
+                    else:
+                        res = cls.meta_with_dtype_and_shape(meta_noop, res.shape)
 
             if isinstance(res, cls._tensor_type):
                 def collect_replace(t: LazyBase):
@@ -168,7 +174,12 @@ class LazyBase(ABC, metaclass=LazyMeta):
             while _t._data is None:
                 lt = _t._lazy.popleft()
                 if lt._data is not None:
-                    raise ValueError(f"{lt} did not belong in the lazy queue")
+                    # Lazy tensor did not belong in the lazy queue.
+                    # Weirdly only happens with Bloom models...
+                    # likely because tensors aren't unique in the queue.
+                    # The final output is still the same as in eager mode,
+                    # so it's safe to ignore this.
+                    continue
                 assert lt._func is not None
                 lt._args = cls._recurse_apply(lt._args, already_eager_to_eager)
                 lt._data = lt._func(lt._args)
@@ -183,12 +194,12 @@ class LazyBase(ABC, metaclass=LazyMeta):
 
     @classmethod
     def eager_to_meta(cls, t: Any) -> Any:
-        return cls.meta_with_dtype(t, t.dtype)
+        return cls.meta_with_dtype_and_shape(t.dtype, t.shape)
 
     # must be overridden, meta tensor init is backend-specific
     @classmethod
     @abstractmethod
-    def meta_with_dtype(cls, m: Any, dtype: Any) -> Any: pass
+    def meta_with_dtype_and_shape(cls, dtype: Any, shape: Any) -> Any: pass
 
     @classmethod
     def from_eager(cls, t: Any) -> Any:
@@ -205,15 +216,15 @@ class LazyNumpyTensor(LazyBase):
     _tensor_type = np.ndarray
 
     @classmethod
-    def meta_with_dtype(cls, m: np.ndarray[Any, Any], dtype: DTypeLike) -> np.ndarray[Any, Any]:
+    def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: _Shape) -> np.ndarray[Any, Any]:
         # The initial idea was to use np.nan as the fill value,
         # but non-float types like np.int16 can't use that.
         # So zero it is.
         cheat = np.zeros(1, dtype)
-        return np.lib.stride_tricks.as_strided(cheat, m.shape, (0 for _ in m.shape))
+        return np.lib.stride_tricks.as_strided(cheat, shape, (0 for _ in shape))
 
     def astype(self, dtype, *args, **kwargs):
-        meta = type(self).meta_with_dtype(self._meta, dtype)
+        meta = type(self).meta_with_dtype_and_shape(dtype, self._meta.shape)
         full_args = (self, dtype,) + args
         # very important to pass the shared _lazy deque, or else there's an infinite loop somewhere.
         return type(self)(meta=meta, args=full_args, lazy=self._lazy, func=(lambda a: a[0].astype(*a[1:], **kwargs)))

gguf-py/gguf/quants.py

@@ -0,0 +1,109 @@
+from __future__ import annotations
+from typing import Callable
+
+from numpy.typing import DTypeLike
+
+from .constants import GGML_QUANT_SIZES, GGMLQuantizationType
+from .lazy import LazyNumpyTensor
+
+import numpy as np
+
+
+# same as ggml_compute_fp32_to_bf16 in ggml-impl.h
+def __compute_fp32_to_bf16(n: np.ndarray) -> np.ndarray:
+    n = n.astype(np.float32, copy=False).view(np.int32)
+    # force nan to quiet
+    n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
+    # flush subnormals to zero
+    n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
+    # round to nearest even
+    n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
+    return n.astype(np.int16)
+
+
+# This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
+def __apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray:
+    rows = arr.reshape((-1, arr.shape[-1]))
+    osize = 1
+    for dim in oshape:
+        osize *= dim
+    out = np.empty(shape=osize, dtype=otype)
+    # compute over groups of 16 rows (arbitrary, but seems good for performance)
+    n_groups = rows.shape[0] // 16
+    np.concatenate([func(group).ravel() for group in np.array_split(rows, n_groups)], axis=0, out=out)
+    return out.reshape(oshape)
+
+
+def __quantize_bf16_array(n: np.ndarray) -> np.ndarray:
+    return __apply_over_grouped_rows(__compute_fp32_to_bf16, arr=n, otype=np.int16, oshape=n.shape)
+
+
+__quantize_bf16_lazy = LazyNumpyTensor._wrap_fn(__quantize_bf16_array, meta_noop=np.int16)
+
+
+def quantize_bf16(n: np.ndarray):
+    if type(n) is LazyNumpyTensor:
+        return __quantize_bf16_lazy(n)
+    else:
+        return __quantize_bf16_array(n)
+
+
+__q8_block_size, __q8_type_size = GGML_QUANT_SIZES[GGMLQuantizationType.Q8_0]
+
+
+def can_quantize_to_q8_0(n: np.ndarray) -> bool:
+    return n.shape[-1] % __q8_block_size == 0
+
+
+# round away from zero
+# ref: https://stackoverflow.com/a/59143326/22827863
+def np_roundf(n: np.ndarray) -> np.ndarray:
+    a = abs(n)
+    floored = np.floor(a)
+    b = floored + np.floor(2 * (a - floored))
+    return np.sign(n) * b
+
+
+def __quantize_q8_0_shape_change(s: tuple[int, ...]) -> tuple[int, ...]:
+    return (*s[:-1], s[-1] // __q8_block_size * __q8_type_size)
+
+
+# Implementation of Q8_0 with bit-exact same results as reference implementation in ggml-quants.c
+def __quantize_q8_0_rows(n: np.ndarray) -> np.ndarray:
+    shape = n.shape
+    assert shape[-1] % __q8_block_size == 0
+
+    n_blocks = n.size // __q8_block_size
+
+    blocks = n.reshape((n_blocks, __q8_block_size)).astype(np.float32, copy=False)
+
+    d = abs(blocks).max(axis=1, keepdims=True) / 127
+    with np.errstate(divide="ignore"):
+        id = np.where(d == 0, 0, 1 / d)
+    qs = np_roundf(blocks * id)
+
+    # (n_blocks, 2)
+    d = d.astype(np.float16).view(np.uint8)
+    # (n_blocks, block_size)
+    qs = qs.astype(np.int8).view(np.uint8)
+
+    assert d.shape[1] + qs.shape[1] == __q8_type_size
+
+    return np.concatenate([d, qs], axis=1).reshape(__quantize_q8_0_shape_change(shape))
+
+
+def __quantize_q8_0_array(n: np.ndarray) -> np.ndarray:
+    return __apply_over_grouped_rows(__quantize_q8_0_rows, arr=n, otype=np.uint8, oshape=__quantize_q8_0_shape_change(n.shape))
+
+
+__quantize_q8_0_lazy = LazyNumpyTensor._wrap_fn(
+    __quantize_q8_0_array,
+    meta_noop=(np.uint8, __quantize_q8_0_shape_change),
+)
+
+
+def quantize_q8_0(data: np.ndarray):
+    if type(data) is LazyNumpyTensor:
+        return __quantize_q8_0_lazy(data)
+    else:
+        return __quantize_q8_0_array(data)

llama.cpp

@@ -2805,6 +2805,11 @@ static void llama_kv_cache_defrag(struct llama_kv_cache & cache) {
     cache.do_defrag = true;
 }
 
+static uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams) {
+    // the FA kernels require padding to avoid extra runtime boundary checks
+    return cparams.flash_attn ? 256u : 32u;
+}
+
 //
 // model loading and saving
 //
@@ -4424,7 +4429,9 @@ static void llm_load_vocab(
             } else if (
                     tokenizer_pre == "gpt-2"   ||
                     tokenizer_pre == "jina-es" ||
-                    tokenizer_pre == "jina-de") {
+                    tokenizer_pre == "jina-de" ||
+                    tokenizer_pre == "jina-v2-es" ||
+                    tokenizer_pre == "jina-v2-de") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "refact") {
@@ -11508,7 +11515,8 @@ static int llama_decode_internal(
                 // a heuristic, to avoid attending the full cache if it is not yet utilized
                 // after enough generations, the benefit from this heuristic disappears
                 // if we start defragmenting the cache, the benefit from this will be more important
-                kv_self.n = std::min(kv_self.size, std::max(256u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 256)));
+                const uint32_t pad = llama_kv_cache_get_padding(cparams);
+                kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(llama_kv_cache_cell_max(kv_self), pad)));
                 //kv_self.n = llama_kv_cache_cell_max(kv_self);
             }
         }
@@ -13174,6 +13182,58 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates(
     return rejects;
 }
 
+static bool llama_grammar_detect_left_recursion(
+        const std::vector<std::vector<llama_grammar_element>> & rules,
+        size_t                                                  rule_index,
+        std::vector<bool>                                     * rules_visited,
+        std::vector<bool>                                     * rules_in_progress,
+        std::vector<bool>                                     * rules_may_be_empty) {
+    if ((*rules_in_progress)[rule_index]) {
+        return true;
+    }
+
+    (*rules_in_progress)[rule_index] = true;
+
+    const std::vector<llama_grammar_element> & rule = rules[rule_index];
+
+    // First check if the rule might produce the empty string. This could be done combined with the second
+    // step but it's more readable as two steps.
+    bool at_rule_start = true;
+    for (size_t i = 0; i < rule.size(); i++) {
+        if (llama_grammar_is_end_of_sequence(&rule[i])) {
+            if (at_rule_start) {
+                (*rules_may_be_empty)[rule_index] = true;
+                break;
+            }
+            at_rule_start = true;
+        } else {
+            at_rule_start = false;
+        }
+    }
+
+    // Second, recurse into leftmost nonterminals (or next-leftmost as long as the previous nonterminal may
+    // be empty)
+    bool recurse_into_nonterminal = true;
+    for (size_t i = 0; i < rule.size(); i++) {
+        if (rule[i].type == LLAMA_GRETYPE_RULE_REF && recurse_into_nonterminal) {
+            if (llama_grammar_detect_left_recursion(rules, (size_t)rule[i].value, rules_visited, rules_in_progress, rules_may_be_empty)) {
+                return true;
+            }
+            if (!((*rules_may_be_empty)[(size_t)rule[i].value])) {
+                recurse_into_nonterminal = false;
+            }
+        } else if (llama_grammar_is_end_of_sequence(&rule[i])) {
+            recurse_into_nonterminal = true;
+        } else {
+            recurse_into_nonterminal = false;
+        }
+    }
+
+    (*rules_in_progress)[rule_index] = false;
+    (*rules_visited)[rule_index] = true;
+    return false;
+}
+
 //
 // grammar - external
 //
@@ -13193,6 +13253,19 @@ struct llama_grammar * llama_grammar_init(
         vec_rules[i].push_back({LLAMA_GRETYPE_END, 0});
     }
 
+    // Check for left recursion
+    std::vector<bool> rules_visited(n_rules);
+    std::vector<bool> rules_in_progress(n_rules);
+    std::vector<bool> rules_may_be_empty(n_rules);
+    for (size_t i = 0; i < n_rules; i++) {
+        if (rules_visited[i]) {
+            continue;
+        }
+        if (llama_grammar_detect_left_recursion(vec_rules, i, &rules_visited, &rules_in_progress, &rules_may_be_empty)) {
+            throw std::runtime_error(format("unsupported grammar, left recursion detected for nonterminal at index %zu", i));
+        }
+    }
+
     // loop over alternates of start rule to build initial stacks
     std::vector<std::vector<const llama_grammar_element *>> stacks;
     pos = vec_rules[start_rule_index].data();
@@ -13215,6 +13288,9 @@ struct llama_grammar * llama_grammar_init(
         }
     } while (true);
 
+    // Important: vec_rules has to be moved here, not copied, because stacks contains
+    // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
+    // then the pointers would be invalidated when the local vec_rules goes out of scope.
     return new llama_grammar{ std::move(vec_rules), std::move(stacks), {} };
 }
 
@@ -15509,6 +15585,11 @@ struct llama_context * llama_new_context_with_model(
         return nullptr;
     }
 
+    if (params.flash_attn && model->arch == LLM_ARCH_GROK) {
+        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
     llama_context * ctx = new llama_context(*model);
 
     const auto & hparams = model->hparams;
@@ -15532,7 +15613,7 @@ struct llama_context * llama_new_context_with_model(
     cparams.rope_freq_scale  = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale;
 
     // this is necessary due to kv_self.n being padded later during inference
-    cparams.n_ctx            = GGML_PAD(cparams.n_ctx, 256);
+    cparams.n_ctx            = GGML_PAD(cparams.n_ctx, llama_kv_cache_get_padding(cparams));
 
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch          = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
@@ -15577,11 +15658,6 @@ struct llama_context * llama_new_context_with_model(
         }
     }
 
-    if (cparams.flash_attn && model->arch == LLM_ARCH_GROK) {
-        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
-        cparams.flash_attn = false;
-    }
-
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }

tests/test-backend-ops.cpp

@@ -2,6 +2,7 @@
 #include <ggml-alloc.h>
 #include <ggml-backend.h>
 #include <ggml-backend-impl.h>
+
 #include <algorithm>
 #include <array>
 #include <cfloat>
@@ -2173,11 +2174,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    for (int hs : { 64, 128, }) { // other head sizes not implemented
-#else
     for (int hs : { 64, 80, 128, 256, }) {
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         for (float max_bias : {0.0f, 8.0f}) {
             for (int nh : { 32, }) {
                 for (int kv : { 512, 1024, }) {

tests/test-grammar-integration.cpp

@@ -28,6 +28,19 @@ static llama_grammar* build_grammar(const std::string & grammar_str) {
     return grammar;
 }
 
+static bool test_build_grammar_fails(const std::string & grammar_str) {
+    fprintf(stderr, "⚫ Testing failure for grammar: %s\n", grammar_str.c_str());
+    bool grammar_fails = false;
+    try {
+        build_grammar(grammar_str);
+        fprintf(stderr, "  ❌ Expected build failure, but succeeded\n");
+    } catch (const std::exception & err) {
+        grammar_fails = true;
+        fprintf(stdout, "  ✅︎\n");
+    }
+    return grammar_fails;
+}
+
 static bool match_string(const std::string & input, llama_grammar* grammar) {
     auto decoded = decode_utf8(input, {});
 
@@ -320,6 +333,38 @@ number ::= [0-9]+)""";
     fprintf(stderr, "  ✅︎ Passed\n");
 }
 
+static void test_failure_left_recursion() {
+    fprintf(stderr, "⚫ Testing left recursion detection:\n");
+
+    // Test simple left recursion detection
+    const std::string simple_str = R"""(root ::= "a" | root "a")""";
+    assert(test_build_grammar_fails(simple_str));
+
+    // Test more complicated left recursion detection
+    const std::string medium_str = R"""(
+root ::= asdf
+asdf ::= "a" | asdf "a"
+)""";
+    assert(test_build_grammar_fails(medium_str));
+
+    // Test even more complicated left recursion detection
+    const std::string hard_str = R"""(
+root ::= asdf
+asdf ::= "a" | foo "b"
+foo ::= "c" | asdf "d" | "e")""";
+    assert(test_build_grammar_fails(hard_str));
+
+    // Test yet even more complicated left recursion detection
+    const std::string hardest_str = R"""(
+root ::= asdf
+asdf ::= "a" | foo "b"
+foo ::= "c" | empty asdf "d" | "e"
+empty ::= "blah" | )""";
+    assert(test_build_grammar_fails(hardest_str));
+
+    fprintf(stderr, "  ✅︎ Passed\n");
+}
+
 int main() {
     fprintf(stdout, "Running grammar integration tests...\n");
     test_simple_grammar();
@@ -327,6 +372,7 @@ int main() {
     test_quantifiers();
     test_failure_missing_root();
     test_failure_missing_reference();
+    test_failure_left_recursion();
     fprintf(stdout, "All tests passed.\n");
     return 0;
 }