From 3fd62a6b1c9ca7b7c0093e984cc9c133c6f2726d Mon Sep 17 00:00:00 2001 From: compilade Date: Sun, 7 Jul 2024 15:04:39 -0400 Subject: [PATCH 01/31] py : type-check all Python scripts with Pyright (#8341) * py : type-check all Python scripts with Pyright * server-tests : use trailing slash in openai base_url * server-tests : add more type annotations * server-tests : strip "chat" from base_url in oai_chat_completions * server-tests : model metadata is a dict * ci : disable pip cache in type-check workflow The cache is not shared between branches, and it's 250MB in size, so it would become quite a big part of the 10GB cache limit of the repo. * py : fix new type errors from master branch * tests : fix test-tokenizer-random.py Apparently, gcc applies optimisations even when pre-processing, which confuses pycparser. * ci : only show warnings and errors in python type-check The "information" level otherwise has entries from 'examples/pydantic_models_to_grammar.py', which could be confusing for someone trying to figure out what failed, considering that these messages can safely be ignored even though they look like errors. --- .devops/nix/package.nix | 16 +++ .github/workflows/python-type-check.yml | 38 +++++++ convert_hf_to_gguf.py | 20 ++-- convert_llama_ggml_to_gguf.py | 3 +- examples/convert_legacy_llama.py | 21 ++-- .../convert_finetune_checkpoint_to_gguf.py | 2 +- examples/json_schema_pydantic_example.py | 5 +- examples/json_schema_to_grammar.py | 12 ++- .../llava/convert_image_encoder_to_gguf.py | 10 +- examples/llava/llava_surgery_v2.py | 10 +- examples/pydantic_models_to_grammar.py | 35 +++--- .../pydantic_models_to_grammar_examples.py | 7 +- examples/server/bench/bench.py | 11 +- examples/server/tests/features/steps/steps.py | 100 +++++++++--------- examples/server/tests/requirements.txt | 4 +- examples/server_embd.py | 4 +- .../convert_train_checkpoint_to_gguf.py | 2 +- ggml/ggml_vk_generate_shaders.py | 4 +- gguf-py/gguf/gguf_reader.py | 6 +- gguf-py/gguf/lazy.py | 31 +++--- gguf-py/scripts/__init__.py | 2 + gguf-py/scripts/gguf_hash.py | 6 +- gguf-py/scripts/gguf_new_metadata.py | 2 + gguf-py/tests/test_gguf.py | 2 +- pyrightconfig.json | 20 +++- requirements/requirements-all.txt | 12 +++ .../requirements-compare-llama-bench.txt | 2 + requirements/requirements-pydantic.txt | 2 + .../requirements-test-tokenizer-random.txt | 1 + scripts/check-requirements.sh | 12 +-- scripts/compare-llama-bench.py | 8 +- scripts/gen-unicode-data.py | 16 +-- tests/test-tokenizer-random.py | 44 ++++---- 33 files changed, 297 insertions(+), 173 deletions(-) create mode 100644 .github/workflows/python-type-check.yml create mode 100644 requirements/requirements-all.txt create mode 100644 requirements/requirements-compare-llama-bench.txt create mode 100644 requirements/requirements-pydantic.txt create mode 100644 requirements/requirements-test-tokenizer-random.txt diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 49e9b7528..00be596ce 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -89,6 +89,22 @@ let ps.tiktoken ps.torchWithoutCuda ps.transformers + + # server bench + ps.matplotlib + + # server tests + ps.openai + ps.behave + ps.prometheus-client + + # for examples/pydantic-models-to-grammar-examples.py + ps.docstring-parser + ps.pydantic + + # for scripts/compare-llama-bench.py + ps.gitpython + ps.tabulate ] ); diff --git a/.github/workflows/python-type-check.yml b/.github/workflows/python-type-check.yml new file mode 100644 index 000000000..e5ff5e6d7 --- /dev/null +++ b/.github/workflows/python-type-check.yml @@ -0,0 +1,38 @@ +name: Python Type-Check + +on: + push: + paths: + - '.github/workflows/python-type-check.yml' + - '**.py' + - '**/requirements*.txt' + pull_request: + paths: + - '.github/workflows/python-type-check.yml' + - '**.py' + - '**/requirements*.txt' + +concurrency: + group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} + cancel-in-progress: true + +jobs: + python-type-check: + runs-on: ubuntu-latest + name: pyright type-check + steps: + - name: Check out source repository + uses: actions/checkout@v4 + - name: Set up Python environment + uses: actions/setup-python@v5 + with: + python-version: "3.11" + - name: Install Python dependencies + # TODO: use a venv + run: pip install -r requirements/requirements-all.txt + - name: Type-check with Pyright + uses: jakebailey/pyright-action@v2 + with: + version: 1.1.370 + level: warning + warnings: true diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 6ee41d3a1..6cea73f08 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -265,7 +265,7 @@ class Model: break for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)): - data: np.ndarray = data # type hint + data: np.ndarray # type hint n_dims = len(data.shape) data_dtype = data.dtype data_qtype: gguf.GGMLQuantizationType | None = None @@ -599,10 +599,6 @@ class Model: tokenizer_path = self.dir_model / 'tokenizer.model' - tokens: list[bytes] = [] - scores: list[float] = [] - toktypes: list[int] = [] - if not tokenizer_path.is_file(): raise FileNotFoundError(f"File not found: {tokenizer_path}") @@ -2120,7 +2116,7 @@ class InternLM2Model(Model): logger.error(f'Error: Missing {tokenizer_path}') sys.exit(1) - sentencepiece_model = model.ModelProto() + sentencepiece_model = model.ModelProto() # pyright: ignore[reportAttributeAccessIssue] sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read()) add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix @@ -2972,16 +2968,16 @@ class T5Model(Model): if not tokenizer_path.is_file(): raise FileNotFoundError(f"File not found: {tokenizer_path}") - sentencepiece_model = model.ModelProto() + sentencepiece_model = model.ModelProto() # pyright: ignore[reportAttributeAccessIssue] sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read()) # some models like Pile-T5 family use BPE tokenizer instead of Unigram - if sentencepiece_model.trainer_spec.model_type == 2: # BPE + if sentencepiece_model.trainer_spec.model_type == 2: # BPE # assure the tokenizer model file name is correct assert tokenizer_path.name == 'tokenizer.model' return self._set_vocab_sentencepiece() else: - assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM + assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces @@ -3152,7 +3148,7 @@ class JaisModel(Model): # but Jais's PyTorch model simply precalculates the slope values and places them # in relative_pes.slopes n_head_closest_log2 = 2 ** math.floor(math.log2(self.hparams["n_head"])) - first_val = float(data_torch._data[0]) + first_val = float(data_torch[0].item()) self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2) return tensors @@ -3186,7 +3182,7 @@ class ChatGLMModel(Model): def set_vocab_chatglm3(self): dir_model = self.dir_model hparams = self.hparams - tokens: list[bytearray] = [] + tokens: list[bytes] = [] toktypes: list[int] = [] scores: list[float] = [] @@ -3335,7 +3331,7 @@ class ChatGLMModel(Model): special_vocab.add_to_gguf(self.gguf_writer) def set_gguf_parameters(self): - self.gguf_writer.add_name(self.hparams.get("_name_or_path").split("/")[1]) # THUDM/glm4-9b-chat or THUDM/chatglm3-6b + self.gguf_writer.add_name(self.hparams["_name_or_path"].split("/")[1]) # THUDM/glm4-9b-chat or THUDM/chatglm3-6b n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed")) n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads")) n_head_kv = self.hparams.get("multi_query_group_num", n_head) diff --git a/convert_llama_ggml_to_gguf.py b/convert_llama_ggml_to_gguf.py index 9349de3b3..95ea831a5 100755 --- a/convert_llama_ggml_to_gguf.py +++ b/convert_llama_ggml_to_gguf.py @@ -354,7 +354,8 @@ class GGMLToGGUF: def handle_metadata(cfg, hp): - import convert + import examples.convert_legacy_llama as convert + assert cfg.model_metadata_dir.is_dir(), 'Metadata dir is not a directory' hf_config_path = cfg.model_metadata_dir / "config.json" orig_config_path = cfg.model_metadata_dir / "params.json" diff --git a/examples/convert_legacy_llama.py b/examples/convert_legacy_llama.py index 721a57c00..c2c73e8ad 100755 --- a/examples/convert_legacy_llama.py +++ b/examples/convert_legacy_llama.py @@ -353,7 +353,7 @@ class Metadata: version: Optional[str] = None url: Optional[str] = None description: Optional[str] = None - licence: Optional[str] = None + license: Optional[str] = None source_url: Optional[str] = None source_hf_repo: Optional[str] = None @@ -492,12 +492,13 @@ class LazyTensor: LazyModel: TypeAlias = 'dict[str, LazyTensor]' +ModelFormat: TypeAlias = Literal['ggml', 'torch', 'safetensors', 'none'] @dataclass class ModelPlus: model: LazyModel paths: list[Path] # Where this was read from. - format: Literal['ggml', 'torch', 'safetensors', 'none'] + format: ModelFormat vocab: BaseVocab | None # For GGML models (which have vocab built in), the vocab. @@ -536,7 +537,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel: def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus: - formats = set(mp.format for mp in models_plus) + formats: set[ModelFormat] = set(mp.format for mp in models_plus) assert len(formats) == 1, "different formats?" format = formats.pop() paths = [path for mp in models_plus for path in mp.paths] @@ -555,7 +556,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus: else: model = merge_sharded([mp.model for mp in models_plus]) - return ModelPlus(model, paths, format, vocab) # pytype: disable=wrong-arg-types + return ModelPlus(model, paths, format, vocab) def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor: @@ -805,7 +806,7 @@ 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_model(self, params: Params, metadata: Metadata) -> None: + def add_meta_model(self, params: Params, metadata: Metadata | None) -> None: # Metadata About The Model And Its Provenence name = "LLaMA" if metadata is not None and metadata.name is not None: @@ -827,8 +828,8 @@ class OutputFile: 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.license is not None: + self.gguf.add_licence(metadata.license) if metadata.source_url is not None: self.gguf.add_source_url(metadata.source_url) if metadata.source_hf_repo is not None: @@ -943,7 +944,7 @@ 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, metadata: Metadata = None, + endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata | None = None, ) -> None: check_vocab_size(params, vocab, pad_vocab=pad_vocab) @@ -977,7 +978,7 @@ 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, + metadata: Metadata | None = None, ) -> None: check_vocab_size(params, vocab, pad_vocab=pad_vocab) @@ -1396,6 +1397,8 @@ def main(args_in: list[str] | None = None) -> None: if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab: vocab = model_plus.vocab + assert params is not None + logger.info(f"Vocab info: {vocab}") logger.info(f"Special vocab info: {special_vocab}") model = model_plus.model diff --git a/examples/finetune/convert_finetune_checkpoint_to_gguf.py b/examples/finetune/convert_finetune_checkpoint_to_gguf.py index c89090918..1b79d6995 100644 --- a/examples/finetune/convert_finetune_checkpoint_to_gguf.py +++ b/examples/finetune/convert_finetune_checkpoint_to_gguf.py @@ -74,7 +74,7 @@ class Tensor: if len(self.ne) == 0: self.nbytes = 0 else: - self.nbytes = int(np.product(self.ne)) * 4 + self.nbytes = int(np.prod(self.ne)) * 4 else: raise ValueError(f"Unhandled data type '{self.dtype}'") diff --git a/examples/json_schema_pydantic_example.py b/examples/json_schema_pydantic_example.py index c7ca7b8d9..19c0bdb5b 100644 --- a/examples/json_schema_pydantic_example.py +++ b/examples/json_schema_pydantic_example.py @@ -3,7 +3,7 @@ #! pip install pydantic #! python json_schema_pydantic_example.py -from pydantic import BaseModel, Extra, TypeAdapter +from pydantic import BaseModel, Field, TypeAdapter from annotated_types import MinLen from typing import Annotated, List, Optional import json, requests @@ -17,6 +17,9 @@ if True: The response_model param takes a type (+ supports Pydantic) and behaves just as w/ Instructor (see below) ''' + response_format = None + type_adapter = None + if response_model: type_adapter = TypeAdapter(response_model) schema = type_adapter.json_schema() diff --git a/examples/json_schema_to_grammar.py b/examples/json_schema_to_grammar.py index 072a230f7..a8779bf3b 100755 --- a/examples/json_schema_to_grammar.py +++ b/examples/json_schema_to_grammar.py @@ -1,4 +1,6 @@ #!/usr/bin/env python3 +from __future__ import annotations + import argparse import itertools import json @@ -188,7 +190,7 @@ def _generate_min_max_int(min_value: Optional[int], max_value: Optional[int], ou raise RuntimeError("At least one of min_value or max_value must be set") class BuiltinRule: - def __init__(self, content: str, deps: list = None): + def __init__(self, content: str, deps: list | None = None): self.content = content self.deps = deps or [] @@ -248,7 +250,7 @@ class SchemaConverter: def _format_literal(self, literal): escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub( - lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal + lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)) or m.group(0), literal ) return f'"{escaped}"' @@ -403,11 +405,11 @@ class SchemaConverter: i = 0 length = len(pattern) - def to_rule(s: Tuple[str, bool]) -> str: + def to_rule(s: tuple[str, bool]) -> str: (txt, is_literal) = s return "\"" + txt + "\"" if is_literal else txt - def transform() -> Tuple[str, bool]: + def transform() -> tuple[str, bool]: ''' Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. ''' @@ -420,7 +422,7 @@ class SchemaConverter: # We only need a flat structure here to apply repetition operators to the last item, and # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially # (GBNF's syntax is luckily very close to regular expressions!) - seq: list[Tuple[str, bool]] = [] + seq: list[tuple[str, bool]] = [] def get_dot(): if self._dotall: diff --git a/examples/llava/convert_image_encoder_to_gguf.py b/examples/llava/convert_image_encoder_to_gguf.py index b00bf7c6d..36f6b92fb 100644 --- a/examples/llava/convert_image_encoder_to_gguf.py +++ b/examples/llava/convert_image_encoder_to_gguf.py @@ -185,6 +185,8 @@ else: fout.add_description("two-tower CLIP model") if has_text_encoder: + assert t_hparams is not None + assert tokens is not None # text_model hparams fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"]) fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"]) @@ -259,8 +261,8 @@ if has_vision_encoder: if processor is not None: - image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean - image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std + image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean # pyright: ignore[reportAttributeAccessIssue] + image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std # pyright: ignore[reportAttributeAccessIssue] else: image_mean = args.image_mean if args.image_mean is not None else default_image_mean image_std = args.image_std if args.image_std is not None else default_image_std @@ -272,7 +274,7 @@ fout.add_bool("clip.use_gelu", use_gelu) if has_llava_projector: - model.vision_model.encoder.layers.pop(-1) + model.vision_model.encoder.layers.pop(-1) # pyright: ignore[reportAttributeAccessIssue] projector = torch.load(args.llava_projector) for name, data in projector.items(): name = get_tensor_name(name) @@ -286,7 +288,7 @@ if has_llava_projector: print("Projector tensors added\n") -state_dict = model.state_dict() +state_dict = model.state_dict() # pyright: ignore[reportAttributeAccessIssue] for name, data in state_dict.items(): if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector): # we don't need this diff --git a/examples/llava/llava_surgery_v2.py b/examples/llava/llava_surgery_v2.py index eb56d6988..2d5b32fe6 100644 --- a/examples/llava/llava_surgery_v2.py +++ b/examples/llava/llava_surgery_v2.py @@ -2,7 +2,9 @@ import argparse import glob import os import torch -from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file +from safetensors import safe_open +from safetensors.torch import save_file +from typing import Any, ContextManager, cast # Function to determine if file is a SafeTensor file def is_safetensor_file(file_path): @@ -13,7 +15,7 @@ def is_safetensor_file(file_path): def load_model(file_path): if is_safetensor_file(file_path): tensors = {} - with safe_open(file_path, framework="pt", device="cpu") as f: + with cast(ContextManager[Any], safe_open(file_path, framework="pt", device="cpu")) as f: for key in f.keys(): tensors[key] = f.get_tensor(key).clone() # output shape @@ -134,7 +136,7 @@ if len(mm_tensors) == 0: if last_checkpoint is not None: for k, v in last_checkpoint.items(): print(k) - print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.") + print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint) if last_checkpoint is not None else 0} tensors.") print("No tensors found. Is this a LLaVA model?") exit() @@ -143,8 +145,10 @@ print(f"Found additional {len(first_mm_tensors)} tensors to extract.") # projector = {name: checkpoint.[name].float() for name in mm_tensors} projector = {} for name in mm_tensors: + assert last_checkpoint is not None projector[name] = last_checkpoint[name].float() for name in first_mm_tensors: + assert first_checkpoint is not None projector[name] = first_checkpoint[name].float() if len(projector) > 0: diff --git a/examples/pydantic_models_to_grammar.py b/examples/pydantic_models_to_grammar.py index f029c73a2..d8145710c 100644 --- a/examples/pydantic_models_to_grammar.py +++ b/examples/pydantic_models_to_grammar.py @@ -6,10 +6,10 @@ import re from copy import copy from enum import Enum from inspect import getdoc, isclass -from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin, get_type_hints +from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin from docstring_parser import parse -from pydantic import BaseModel, Field, create_model +from pydantic import BaseModel, create_model if TYPE_CHECKING: from types import GenericAlias @@ -17,6 +17,9 @@ else: # python 3.8 compat from typing import _GenericAlias as GenericAlias +# TODO: fix this +# pyright: reportAttributeAccessIssue=information + class PydanticDataType(Enum): """ @@ -234,8 +237,9 @@ def generate_gbnf_float_rules(max_digit=None, min_digit=None, max_precision=None # Define the integer part rule integer_part_rule = ( - "integer-part" + (f"-max{max_digit}" if max_digit is not None else "") + ( - f"-min{min_digit}" if min_digit is not None else "") + "integer-part" + + (f"-max{max_digit}" if max_digit is not None else "") + + (f"-min{min_digit}" if min_digit is not None else "") ) # Define the fractional part rule based on precision constraints @@ -458,7 +462,7 @@ def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[Bas if not issubclass(model, BaseModel): # For non-Pydantic classes, generate model_fields from __annotations__ or __init__ if hasattr(model, "__annotations__") and model.__annotations__: - model_fields = {name: (typ, ...) for name, typ in model.__annotations__.items()} + model_fields = {name: (typ, ...) for name, typ in model.__annotations__.items()} # pyright: ignore[reportGeneralTypeIssues] else: init_signature = inspect.signature(model.__init__) parameters = init_signature.parameters @@ -680,7 +684,7 @@ def generate_markdown_documentation( str: Generated text documentation. """ documentation = "" - pyd_models = [(model, True) for model in pydantic_models] + pyd_models: list[tuple[type[BaseModel], bool]] = [(model, True) for model in pydantic_models] for model, add_prefix in pyd_models: if add_prefix: documentation += f"{model_prefix}: {model.__name__}\n" @@ -700,7 +704,7 @@ def generate_markdown_documentation( # Indenting the fields section documentation += f" {fields_prefix}:\n" else: - documentation += f" Fields:\n" + documentation += f" Fields:\n" # noqa: F541 if isclass(model) and issubclass(model, BaseModel): for name, field_type in model.__annotations__.items(): # if name == "markdown_code_block": @@ -778,7 +782,7 @@ def generate_field_markdown( return field_text if field_description != "": - field_text += f" Description: " + field_description + "\n" + field_text += f" Description: {field_description}\n" # Check for and include field-specific examples if available if hasattr(model, "Config") and hasattr(model.Config, @@ -833,7 +837,7 @@ def generate_text_documentation( str: Generated text documentation. """ documentation = "" - pyd_models = [(model, True) for model in pydantic_models] + pyd_models: list[tuple[type[BaseModel], bool]] = [(model, True) for model in pydantic_models] for model, add_prefix in pyd_models: if add_prefix: documentation += f"{model_prefix}: {model.__name__}\n" @@ -1164,7 +1168,7 @@ def create_dynamic_model_from_function(func: Callable[..., Any]): dynamic_fields[param.name] = ( param.annotation if param.annotation != inspect.Parameter.empty else str, default_value) # Creating the dynamic model - dynamic_model = create_model(f"{func.__name__}", **dynamic_fields) # type: ignore[call-overload] + dynamic_model = create_model(f"{func.__name__}", **dynamic_fields) for name, param_doc in param_docs: dynamic_model.model_fields[name].description = param_doc.description @@ -1228,9 +1232,6 @@ def map_grammar_names_to_pydantic_model_class(pydantic_model_list): return output -from enum import Enum - - def json_schema_to_python_types(schema): type_map = { "any": Any, @@ -1275,7 +1276,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: if items != {}: array = {"properties": items} array_type = convert_dictionary_to_pydantic_model(array, f"{model_name}_{field_name}_items") - fields[field_name] = (List[array_type], ...) # type: ignore[valid-type] + fields[field_name] = (List[array_type], ...) else: fields[field_name] = (list, ...) elif field_type == "object": @@ -1285,7 +1286,8 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: required = field_data.get("enum", []) for key, field in fields.items(): if key not in required: - fields[key] = (Optional[fields[key][0]], ...) + optional_type = fields[key][0] + fields[key] = (Optional[optional_type], ...) else: field_type = json_schema_to_python_types(field_type) fields[field_name] = (field_type, ...) @@ -1305,6 +1307,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: required = dictionary.get("required", []) for key, field in fields.items(): if key not in required: - fields[key] = (Optional[fields[key][0]], ...) + optional_type = fields[key][0] + fields[key] = (Optional[optional_type], ...) custom_model = create_model(model_name, **fields) return custom_model diff --git a/examples/pydantic_models_to_grammar_examples.py b/examples/pydantic_models_to_grammar_examples.py index 160966649..8e7f46cf9 100644 --- a/examples/pydantic_models_to_grammar_examples.py +++ b/examples/pydantic_models_to_grammar_examples.py @@ -1,6 +1,7 @@ # Function calling example using pydantic models. +from __future__ import annotations + import datetime -import importlib import json from enum import Enum from typing import Optional, Union @@ -215,9 +216,9 @@ for call in json_data: if call["function"] == "Calculator": print(Calculator(**call["params"]).run()) elif call["function"] == "get_current_datetime": - print(current_datetime_model(**call["params"]).run()) + print(current_datetime_model(**call["params"]).run()) # pyright: ignore[reportAttributeAccessIssue] elif call["function"] == "get_current_weather": - print(current_weather_tool_model(**call["params"]).run()) + print(current_weather_tool_model(**call["params"]).run()) # pyright: ignore[reportAttributeAccessIssue] # Should output something like this: # 2024-01-14 13:36:06 # {"location": "London", "temperature": "42", "unit": "celsius"} diff --git a/examples/server/bench/bench.py b/examples/server/bench/bench.py index 4fbbb2032..2daac0884 100644 --- a/examples/server/bench/bench.py +++ b/examples/server/bench/bench.py @@ -1,3 +1,5 @@ +from __future__ import annotations + import argparse import json import os @@ -59,10 +61,11 @@ def main(args_in: list[str] | None = None) -> None: sys.exit(1) # start the benchmark + iterations = 0 + data = {} try: start_benchmark(args) - iterations = 0 with open("results.github.env", 'w') as github_env: # parse output with open('k6-results.json', 'r') as bench_results: @@ -129,7 +132,7 @@ def main(args_in: list[str] | None = None) -> None: timestamps, metric_values = zip(*values) metric_values = [float(value) for value in metric_values] prometheus_metrics[metric] = metric_values - timestamps_dt = [datetime.fromtimestamp(int(ts)) for ts in timestamps] + timestamps_dt = [str(datetime.fromtimestamp(int(ts))) for ts in timestamps] plt.figure(figsize=(16, 10), dpi=80) plt.plot(timestamps_dt, metric_values, label=metric) plt.xticks(rotation=0, fontsize=14, horizontalalignment='center', alpha=.7) @@ -156,7 +159,7 @@ def main(args_in: list[str] | None = None) -> None: plt.close() # Mermaid format in case images upload failed - with (open(f"{metric}.mermaid", 'w') as mermaid_f): + with open(f"{metric}.mermaid", 'w') as mermaid_f: mermaid = ( f"""--- config: @@ -278,7 +281,7 @@ def start_server_background(args): } server_process = subprocess.Popen( args, - **pkwargs) + **pkwargs) # pyright: ignore[reportArgumentType, reportCallIssue] def server_log(in_stream, out_stream): for line in iter(in_stream.readline, b''): diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py index 7b5dabb01..df0814cc9 100644 --- a/examples/server/tests/features/steps/steps.py +++ b/examples/server/tests/features/steps/steps.py @@ -1,5 +1,4 @@ import asyncio -import collections import json import os import re @@ -8,19 +7,23 @@ import subprocess import sys import threading import time +from collections.abc import Sequence from contextlib import closing from re import RegexFlag +from typing import Any, Literal, cast import aiohttp import numpy as np import openai -from behave import step +from openai.types.chat import ChatCompletionChunk +from behave import step # pyright: ignore[reportAttributeAccessIssue] from behave.api.async_step import async_run_until_complete from prometheus_client import parser +# pyright: reportRedeclaration=false @step("a server listening on {server_fqdn}:{server_port}") -def step_server_config(context, server_fqdn, server_port): +def step_server_config(context, server_fqdn: str, server_port: str): context.server_fqdn = server_fqdn context.server_port = int(server_port) context.n_threads = None @@ -74,34 +77,34 @@ def step_server_config(context, server_fqdn, server_port): @step('a model file {hf_file} from HF repo {hf_repo}') -def step_download_hf_model(context, hf_file, hf_repo): +def step_download_hf_model(context, hf_file: str, hf_repo: str): context.model_hf_repo = hf_repo context.model_hf_file = hf_file context.model_file = os.path.basename(hf_file) @step('a model file {model_file}') -def step_model_file(context, model_file): +def step_model_file(context, model_file: str): context.model_file = model_file @step('a model url {model_url}') -def step_model_url(context, model_url): +def step_model_url(context, model_url: str): context.model_url = model_url @step('a model alias {model_alias}') -def step_model_alias(context, model_alias): +def step_model_alias(context, model_alias: str): context.model_alias = model_alias @step('{seed:d} as server seed') -def step_seed(context, seed): +def step_seed(context, seed: int): context.server_seed = seed @step('{ngl:d} GPU offloaded layers') -def step_n_gpu_layer(context, ngl): +def step_n_gpu_layer(context, ngl: int): if 'N_GPU_LAYERS' in os.environ: new_ngl = int(os.environ['N_GPU_LAYERS']) if context.debug: @@ -111,37 +114,37 @@ def step_n_gpu_layer(context, ngl): @step('{n_threads:d} threads') -def step_n_threads(context, n_threads): +def step_n_threads(context, n_threads: int): context.n_thread = n_threads @step('{draft:d} as draft') -def step_draft(context, draft): +def step_draft(context, draft: int): context.draft = draft @step('{n_ctx:d} KV cache size') -def step_n_ctx(context, n_ctx): +def step_n_ctx(context, n_ctx: int): context.n_ctx = n_ctx @step('{n_slots:d} slots') -def step_n_slots(context, n_slots): +def step_n_slots(context, n_slots: int): context.n_slots = n_slots @step('{n_predict:d} server max tokens to predict') -def step_server_n_predict(context, n_predict): +def step_server_n_predict(context, n_predict: int): context.n_server_predict = n_predict @step('{slot_save_path} as slot save path') -def step_slot_save_path(context, slot_save_path): +def step_slot_save_path(context, slot_save_path: str): context.slot_save_path = slot_save_path @step('using slot id {id_slot:d}') -def step_id_slot(context, id_slot): +def step_id_slot(context, id_slot: int): context.id_slot = id_slot @@ -191,7 +194,7 @@ def step_start_server(context): @step("the server is {expecting_status}") @async_run_until_complete -async def step_wait_for_the_server_to_be_started(context, expecting_status): +async def step_wait_for_the_server_to_be_started(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str): match expecting_status: case 'healthy': await wait_for_health_status(context, context.base_url, 200, 'ok', @@ -221,7 +224,7 @@ async def step_wait_for_the_server_to_be_started(context, expecting_status): @step('all slots are {expected_slot_status_string}') @async_run_until_complete -async def step_all_slots_status(context, expected_slot_status_string): +async def step_all_slots_status(context, expected_slot_status_string: Literal['idle', 'busy'] | str): match expected_slot_status_string: case 'idle': expected_slot_status = 0 @@ -237,7 +240,7 @@ async def step_all_slots_status(context, expected_slot_status_string): @step('a completion request with {api_error} api error') @async_run_until_complete -async def step_request_completion(context, api_error): +async def step_request_completion(context, api_error: Literal['raised'] | str): expect_api_error = api_error == 'raised' seeds = await completions_seed(context, num_seeds=1) completion = await request_completion(context.prompts.pop(), @@ -777,8 +780,8 @@ def step_assert_metric_value(context, metric_name, metric_value): def step_available_models(context): # openai client always expects an api_key openai.api_key = context.user_api_key if context.user_api_key is not None else 'nope' - openai.api_base = f'{context.base_url}/v1' - context.models = openai.Model.list().data + openai.base_url = f'{context.base_url}/v1/' + context.models = openai.models.list().data @step('{n_model:d} models are supported') @@ -789,7 +792,7 @@ def step_supported_models(context, n_model): @step('model {i_model:d} is {param} {preposition} {param_value}') -def step_supported_models(context, i_model, param, preposition, param_value): +def step_supported_models(context, i_model: int, param: Literal['identified', 'trained'] | str, preposition: str, param_value: str): assert i_model < len(context.models) model = context.models[i_model] @@ -798,7 +801,7 @@ def step_supported_models(context, i_model, param, preposition, param_value): case 'identified': value = model.id case 'trained': - value = str(model.meta.n_ctx_train) + value = str(model.meta["n_ctx_train"]) case _: assert False, "param {param} not supported" assert param_value == value, f"model param {param} {value} != {param_value}" @@ -810,6 +813,7 @@ async def concurrent_requests(context, f_completion, *args, **kwargs): print(f"starting {context.n_prompts} concurrent completion requests...") assert context.n_prompts > 0 seeds = await completions_seed(context) + assert seeds is not None for prompt_no in range(context.n_prompts): shifted_args = [context.prompts.pop(), seeds[prompt_no], *args] context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs))) @@ -861,7 +865,7 @@ async def request_completion(prompt, id_slot=None, expect_api_error=None, user_api_key=None, - temperature=None): + temperature=None) -> int | dict[str, Any]: if debug: print(f"Sending completion request: {prompt}") origin = "my.super.domain" @@ -899,8 +903,8 @@ async def request_completion(prompt, async def oai_chat_completions(user_prompt, seed, system_prompt, - base_url, - base_path, + base_url: str, + base_path: str, async_client, debug=False, temperature=None, @@ -909,7 +913,7 @@ async def oai_chat_completions(user_prompt, enable_streaming=None, response_format=None, user_api_key=None, - expect_api_error=None): + expect_api_error=None) -> int | dict[str, Any]: if debug: print(f"Sending OAI Chat completions request: {user_prompt}") # openai client always expects an api key @@ -989,32 +993,35 @@ async def oai_chat_completions(user_prompt, else: try: openai.api_key = user_api_key - openai.api_base = f'{base_url}{base_path}' - chat_completion = openai.Completion.create( + openai.base_url = f'{base_url}{base_path.removesuffix("chat")}' + assert model is not None + chat_completion = openai.chat.completions.create( messages=payload['messages'], model=model, max_tokens=n_predict, stream=enable_streaming, - response_format=payload.get('response_format'), + response_format=payload.get('response_format') or openai.NOT_GIVEN, seed=seed, temperature=payload['temperature'] ) - except openai.error.AuthenticationError as e: + except openai.AuthenticationError as e: if expect_api_error is not None and expect_api_error: return 401 else: assert False, f'error raised: {e}' if enable_streaming: + chat_completion = cast(openai.Stream[ChatCompletionChunk], chat_completion) for chunk in chat_completion: assert len(chunk.choices) == 1 delta = chunk.choices[0].delta - if 'content' in delta: - completion_response['content'] += delta['content'] + if delta.content is not None: + completion_response['content'] += delta.content completion_response['timings']['predicted_n'] += 1 completion_response['truncated'] = chunk.choices[0].finish_reason != 'stop' else: assert len(chat_completion.choices) == 1 + assert chat_completion.usage is not None completion_response = { 'content': chat_completion.choices[0].message.content, 'timings': { @@ -1028,7 +1035,7 @@ async def oai_chat_completions(user_prompt, return completion_response -async def request_embedding(content, seed, base_url=None): +async def request_embedding(content, seed, base_url=None) -> list[list[float]]: async with aiohttp.ClientSession() as session: async with session.post(f'{base_url}/embedding', json={ @@ -1041,7 +1048,7 @@ async def request_embedding(content, seed, base_url=None): async def request_oai_embeddings(input, seed, base_url=None, user_api_key=None, - model=None, async_client=False): + model=None, async_client=False) -> list[list[float]]: # openai client always expects an api_key user_api_key = user_api_key if user_api_key is not None else 'nope' if async_client: @@ -1063,7 +1070,7 @@ async def request_oai_embeddings(input, seed, response_json = await response.json() assert response_json['model'] == model, f"invalid model received: {response_json['model']}" assert response_json['object'] == 'list' - if isinstance(input, collections.abc.Sequence): + if isinstance(input, Sequence): embeddings = [] for an_oai_embeddings in response_json['data']: embeddings.append(an_oai_embeddings['embedding']) @@ -1072,19 +1079,14 @@ async def request_oai_embeddings(input, seed, return embeddings else: openai.api_key = user_api_key - openai.api_base = f'{base_url}/v1' - oai_embeddings = openai.Embedding.create( + openai.base_url = f'{base_url}/v1/' + assert model is not None + oai_embeddings = openai.embeddings.create( model=model, input=input, ) - if isinstance(input, collections.abc.Sequence): - embeddings = [] - for an_oai_embeddings in oai_embeddings.data: - embeddings.append(an_oai_embeddings.embedding) - else: - embeddings = [oai_embeddings.data.embedding] - return embeddings + return [e.embedding for e in oai_embeddings.data] def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re_content=None): @@ -1122,7 +1124,7 @@ def assert_all_predictions_equal(completion_responses): if i == j: continue content_j = response_j['content'] - assert content_i == content_j, "contents not equal" + assert content_i == content_j, "contents not equal" def assert_all_predictions_different(completion_responses): @@ -1136,7 +1138,7 @@ def assert_all_predictions_different(completion_responses): if i == j: continue content_j = response_j['content'] - assert content_i != content_j, "contents not different" + assert content_i != content_j, "contents not different" def assert_all_token_probabilities_equal(completion_responses): @@ -1153,7 +1155,7 @@ def assert_all_token_probabilities_equal(completion_responses): if i == j: continue probs_j = response_j['completion_probabilities'][pos]['probs'] - assert probs_i == probs_j, "contents not equal" + assert probs_i == probs_j, "contents not equal" async def gather_tasks_results(context): @@ -1343,7 +1345,7 @@ def start_server_background(context): } context.server_process = subprocess.Popen( [str(arg) for arg in [context.server_path, *server_args]], - **pkwargs) + **pkwargs) # pyright: ignore[reportArgumentType, reportCallIssue] def server_log(in_stream, out_stream): for line in iter(in_stream.readline, b''): diff --git a/examples/server/tests/requirements.txt b/examples/server/tests/requirements.txt index 2e4f42ad2..2c741ea10 100644 --- a/examples/server/tests/requirements.txt +++ b/examples/server/tests/requirements.txt @@ -1,6 +1,6 @@ aiohttp~=3.9.3 behave~=1.2.6 huggingface_hub~=0.20.3 -numpy~=1.24.4 -openai~=0.25.0 +numpy~=1.26.4 +openai~=1.30.3 prometheus-client~=0.20.0 diff --git a/examples/server_embd.py b/examples/server_embd.py index a9a36a44c..0e34c6cea 100644 --- a/examples/server_embd.py +++ b/examples/server_embd.py @@ -1,13 +1,15 @@ import asyncio +import asyncio.threads import requests import numpy as np + n = 8 result = [] async def requests_post_async(*args, **kwargs): - return await asyncio.to_thread(requests.post, *args, **kwargs) + return await asyncio.threads.to_thread(requests.post, *args, **kwargs) async def main(): model_url = "http://127.0.0.1:6900" diff --git a/examples/train-text-from-scratch/convert_train_checkpoint_to_gguf.py b/examples/train-text-from-scratch/convert_train_checkpoint_to_gguf.py index ed93673bc..e045beb72 100644 --- a/examples/train-text-from-scratch/convert_train_checkpoint_to_gguf.py +++ b/examples/train-text-from-scratch/convert_train_checkpoint_to_gguf.py @@ -66,7 +66,7 @@ class Tensor: if len(self.ne) == 0: self.nbytes = 0 else: - self.nbytes = int(np.product(self.ne)) * 4 + self.nbytes = int(np.prod(self.ne)) * 4 else: raise ValueError(f"Unhandled data type '{self.dtype}'") diff --git a/ggml/ggml_vk_generate_shaders.py b/ggml/ggml_vk_generate_shaders.py index 38914eedb..41d5d9b8c 100644 --- a/ggml/ggml_vk_generate_shaders.py +++ b/ggml/ggml_vk_generate_shaders.py @@ -99,6 +99,8 @@ async def main(): tasks = [] + base_dict = {"FLOAT_TYPE": "float"} + for fp16 in (False, True): # MUL_MAT matmul_shaders(tasks, fp16, False) @@ -106,8 +108,6 @@ async def main(): matmul_shaders(tasks, fp16, True) for tname in type_names: - base_dict = {"FLOAT_TYPE": "float"} - # mul mat vec data_a_key = f"DATA_A_{tname.upper()}" shader = f"mul_mat_vec_{tname}.comp" if tname.endswith("_k") else "mul_mat_vec.comp" diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py index 20432bd25..e8e61abf8 100644 --- a/gguf-py/gguf/gguf_reader.py +++ b/gguf-py/gguf/gguf_reader.py @@ -67,7 +67,7 @@ class ReaderTensor(NamedTuple): class GGUFReader: # I - same as host, S - swapped - byte_order: Literal['I'] | Literal['S'] = 'I' + byte_order: Literal['I', 'S'] = 'I' alignment: int = GGUF_DEFAULT_ALIGNMENT data_offset: int @@ -86,7 +86,7 @@ class GGUFReader: GGUFValueType.BOOL: np.bool_, } - def __init__(self, path: os.PathLike[str] | str, mode: Literal['r'] | Literal['r+'] | Literal['c'] = 'r'): + def __init__(self, path: os.PathLike[str] | str, mode: Literal['r', 'r+', 'c'] = 'r'): self.data = np.memmap(path, mode = mode) offs = 0 @@ -140,7 +140,7 @@ class GGUFReader: return self.tensors[idx] def _get( - self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I'] | Literal['S'] | Literal['<'] = None, + self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I', 'S', '<'] = None, ) -> npt.NDArray[Any]: count = int(count) itemsize = int(np.empty([], dtype = dtype).itemsize) diff --git a/gguf-py/gguf/lazy.py b/gguf-py/gguf/lazy.py index 1167335b8..c50124cd9 100644 --- a/gguf-py/gguf/lazy.py +++ b/gguf-py/gguf/lazy.py @@ -16,16 +16,16 @@ logger = logging.getLogger(__name__) class LazyMeta(ABCMeta): def __new__(cls, name: str, bases: tuple[type, ...], namespace: dict[str, Any], **kwargs): - def __getattr__(self, __name: str) -> Any: - meta_attr = getattr(self._meta, __name) + def __getattr__(self, name: str) -> Any: + meta_attr = getattr(self._meta, name) if callable(meta_attr): return type(self)._wrap_fn( - (lambda s, *args, **kwargs: getattr(s, __name)(*args, **kwargs)), + (lambda s, *args, **kwargs: getattr(s, name)(*args, **kwargs)), use_self=self, ) elif isinstance(meta_attr, self._tensor_type): # e.g. self.T with torch.Tensor should still be wrapped - return type(self)._wrap_fn(lambda s: getattr(s, __name))(self) + return type(self)._wrap_fn(lambda s: getattr(s, name))(self) else: # no need to wrap non-tensor properties, # and they likely don't depend on the actual contents of the tensor @@ -141,19 +141,21 @@ class LazyBase(ABC, metaclass=LazyMeta): res = cls.meta_with_dtype_and_shape(meta_noop, res.shape) if isinstance(res, cls._tensor_type): - def collect_replace(t: LazyBase): - if collect_replace.shared_lazy is None: - collect_replace.shared_lazy = t._lazy - else: - collect_replace.shared_lazy.extend(t._lazy) - t._lazy = collect_replace.shared_lazy + class CollectSharedLazy: + # emulating a static variable + shared_lazy: None | deque[LazyBase] = None - # emulating a static variable - collect_replace.shared_lazy = None + @staticmethod + def collect_replace(t: LazyBase): + if CollectSharedLazy.shared_lazy is None: + CollectSharedLazy.shared_lazy = t._lazy + else: + CollectSharedLazy.shared_lazy.extend(t._lazy) + t._lazy = CollectSharedLazy.shared_lazy - LazyBase._recurse_apply(args, collect_replace) + LazyBase._recurse_apply(args, CollectSharedLazy.collect_replace) - shared_lazy = collect_replace.shared_lazy + shared_lazy = CollectSharedLazy.shared_lazy return cls(meta=cls.eager_to_meta(res), lazy=shared_lazy, args=args, func=lambda a: fn(*a, **kwargs)) else: @@ -184,6 +186,7 @@ class LazyBase(ABC, metaclass=LazyMeta): lt._args = cls._recurse_apply(lt._args, already_eager_to_eager) lt._data = lt._func(lt._args) # sanity check + assert lt._data is not None assert lt._data.dtype == lt._meta.dtype assert lt._data.shape == lt._meta.shape diff --git a/gguf-py/scripts/__init__.py b/gguf-py/scripts/__init__.py index f9d29cb69..e77f2e9c9 100644 --- a/gguf-py/scripts/__init__.py +++ b/gguf-py/scripts/__init__.py @@ -1,3 +1,5 @@ +# pyright: reportUnusedImport=false + from .gguf_convert_endian import main as gguf_convert_endian_entrypoint from .gguf_dump import main as gguf_dump_entrypoint from .gguf_set_metadata import main as gguf_set_metadata_entrypoint diff --git a/gguf-py/scripts/gguf_hash.py b/gguf-py/scripts/gguf_hash.py index 956775182..770b79a93 100755 --- a/gguf-py/scripts/gguf_hash.py +++ b/gguf-py/scripts/gguf_hash.py @@ -63,9 +63,9 @@ def gguf_hash(reader: GGUFReader, filename: str, disable_progress_bar) -> None: bar.update(sum_weights_in_tensor) sha1_layer = hashlib.sha1() - sha1_layer.update(tensor.data) - sha1.update(tensor.data) - uuidv5_sha1.update(tensor.data) + sha1_layer.update(tensor.data.data) + sha1.update(tensor.data.data) + uuidv5_sha1.update(tensor.data.data) print("sha1 {0} {1}:{2}".format(sha1_layer.hexdigest(), filename, tensor.name)) # noqa: NP100 # Flush Hash Progress Bar diff --git a/gguf-py/scripts/gguf_new_metadata.py b/gguf-py/scripts/gguf_new_metadata.py index c4b90d581..fce52a8c1 100755 --- a/gguf-py/scripts/gguf_new_metadata.py +++ b/gguf-py/scripts/gguf_new_metadata.py @@ -1,4 +1,6 @@ #!/usr/bin/env python3 +from __future__ import annotations + import logging import argparse import os diff --git a/gguf-py/tests/test_gguf.py b/gguf-py/tests/test_gguf.py index 0adeb7d55..76b52181e 100644 --- a/gguf-py/tests/test_gguf.py +++ b/gguf-py/tests/test_gguf.py @@ -1,4 +1,4 @@ -import gguf # noqa: F401 +import gguf # noqa: F401 # pyright: ignore[reportUnusedImport] # TODO: add tests diff --git a/pyrightconfig.json b/pyrightconfig.json index 020a71a4e..6016f4b6d 100644 --- a/pyrightconfig.json +++ b/pyrightconfig.json @@ -1,3 +1,21 @@ { "extraPaths": ["gguf-py"], -} + "pythonVersion": "3.9", + "pythonPlatform": "All", + "reportUnusedImport": "warning", + "reportDuplicateImport": "error", + "reportDeprecated": "warning", + "reportUnnecessaryTypeIgnoreComment": "warning", + "executionEnvironments": [ + { + // TODO: make this version override work correctly + "root": "gguf-py", + "pythonVersion": "3.8", + }, + { + // uses match expressions in steps.py + "root": "examples/server/tests", + "pythonVersion": "3.10", + }, + ], + } diff --git a/requirements/requirements-all.txt b/requirements/requirements-all.txt new file mode 100644 index 000000000..94de59d7e --- /dev/null +++ b/requirements/requirements-all.txt @@ -0,0 +1,12 @@ +-r ../examples/llava/requirements.txt +-r ../examples/server/bench/requirements.txt +-r ../examples/server/tests/requirements.txt + +-r ./requirements-compare-llama-bench.txt +-r ./requirements-pydantic.txt +-r ./requirements-test-tokenizer-random.txt + +-r ./requirements-convert_hf_to_gguf.txt +-r ./requirements-convert_hf_to_gguf_update.txt +-r ./requirements-convert_legacy_llama.txt +-r ./requirements-convert_llama_ggml_to_gguf.txt diff --git a/requirements/requirements-compare-llama-bench.txt b/requirements/requirements-compare-llama-bench.txt new file mode 100644 index 000000000..e0aaa3204 --- /dev/null +++ b/requirements/requirements-compare-llama-bench.txt @@ -0,0 +1,2 @@ +tabulate~=0.9.0 +GitPython~=3.1.43 diff --git a/requirements/requirements-pydantic.txt b/requirements/requirements-pydantic.txt new file mode 100644 index 000000000..2f9455b14 --- /dev/null +++ b/requirements/requirements-pydantic.txt @@ -0,0 +1,2 @@ +docstring_parser~=0.15 +pydantic~=2.6.3 diff --git a/requirements/requirements-test-tokenizer-random.txt b/requirements/requirements-test-tokenizer-random.txt new file mode 100644 index 000000000..2785e71a2 --- /dev/null +++ b/requirements/requirements-test-tokenizer-random.txt @@ -0,0 +1 @@ +cffi~=1.16.0 diff --git a/scripts/check-requirements.sh b/scripts/check-requirements.sh index 48f924c02..d3bbded13 100755 --- a/scripts/check-requirements.sh +++ b/scripts/check-requirements.sh @@ -108,6 +108,11 @@ check_convert_script() { fatal "$py missing requirements. Expected: $reqs" fi + # Check that all sub-requirements are added to top-level requirements.txt + if ! grep -qF "$reqs" requirements.txt; then + fatal "$reqs needs to be added to requirements.txt" + fi + local venv="$workdir/$pyname-venv" python3 -m venv "$venv" @@ -134,12 +139,7 @@ EOF readonly ignore_eq_eq='check_requirements: ignore "=="' -for req in "$reqs_dir"/*; do - # Check that all sub-requirements are added to top-level requirements.txt - if ! grep -qF "$req" requirements.txt; then - fatal "$req needs to be added to requirements.txt" - fi - +for req in */**/requirements*.txt; do # Make sure exact release versions aren't being pinned in the requirements # Filters out the ignore string if grep -vF "$ignore_eq_eq" "$req" | grep -q '=='; then diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py index 513dde5e1..92b9e682a 100755 --- a/scripts/compare-llama-bench.py +++ b/scripts/compare-llama-bench.py @@ -123,13 +123,13 @@ builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall() try: repo = git.Repo(".", search_parent_directories=True) -except git.exc.InvalidGitRepositoryError: +except git.InvalidGitRepositoryError: repo = None -def find_parent_in_data(commit): +def find_parent_in_data(commit: git.Commit): """Helper function to find the most recent parent measured in number of commits for which there is data.""" - heap = [(0, commit)] + heap: list[tuple[int, git.Commit]] = [(0, commit)] seen_hexsha8 = set() while heap: depth, current_commit = heapq.heappop(heap) @@ -144,7 +144,7 @@ def find_parent_in_data(commit): return None -def get_all_parent_hexsha8s(commit): +def get_all_parent_hexsha8s(commit: git.Commit): """Helper function to recursively get hexsha8 values for all parents of a commit.""" unvisited = [commit] visited = [] diff --git a/scripts/gen-unicode-data.py b/scripts/gen-unicode-data.py index 890e4d7c2..2d9bde01c 100644 --- a/scripts/gen-unicode-data.py +++ b/scripts/gen-unicode-data.py @@ -1,3 +1,5 @@ +from __future__ import annotations + import array import unicodedata import requests @@ -133,7 +135,7 @@ table_nfd.sort() # group ranges with same flags -ranges_flags = [(0, codepoint_flags[0])] # start, flags +ranges_flags: list[tuple[int, int]] = [(0, codepoint_flags[0])] # start, flags for codepoint, flags in enumerate(codepoint_flags): if flags != ranges_flags[-1][1]: ranges_flags.append((codepoint, flags)) @@ -141,11 +143,11 @@ ranges_flags.append((MAX_CODEPOINTS, 0x0000)) # group ranges with same nfd -ranges_nfd = [(0, 0, 0)] # start, last, nfd +ranges_nfd: list[tuple[int, int, int]] = [(0, 0, 0)] # start, last, nfd for codepoint, norm in table_nfd: start = ranges_nfd[-1][0] if ranges_nfd[-1] != (start, codepoint - 1, norm): - ranges_nfd.append(None) + ranges_nfd.append(None) # type: ignore[arg-type] # dummy, will be replaced below start = codepoint ranges_nfd[-1] = (start, codepoint, norm) @@ -179,13 +181,13 @@ for codepoint in table_whitespace: out("};\n") out("const std::unordered_map unicode_map_lowercase = {") -for tuple in table_lowercase: - out("{0x%06X, 0x%06X}," % tuple) +for tuple_lw in table_lowercase: + out("{0x%06X, 0x%06X}," % tuple_lw) out("};\n") out("const std::unordered_map unicode_map_uppercase = {") -for tuple in table_uppercase: - out("{0x%06X, 0x%06X}," % tuple) +for tuple_up in table_uppercase: + out("{0x%06X, 0x%06X}," % tuple_up) out("};\n") out("const std::vector unicode_ranges_nfd = { // start, last, nfd") diff --git a/tests/test-tokenizer-random.py b/tests/test-tokenizer-random.py index 48cab8a1e..c50a8ca32 100644 --- a/tests/test-tokenizer-random.py +++ b/tests/test-tokenizer-random.py @@ -6,6 +6,8 @@ # python3 tests/test-tokenizer-random.py ./models/ggml-vocab-llama-bpe.gguf ./models/tokenizers/llama-bpe # +from __future__ import annotations + import time import logging import argparse @@ -13,7 +15,9 @@ import subprocess import random import unicodedata -from typing import Iterator +from pathlib import Path +from typing import Any, Iterator, cast +from typing_extensions import Buffer import cffi from transformers import AutoTokenizer @@ -28,15 +32,15 @@ class LibLlama: DEFAULT_PATH_INCLUDES = ["./ggml/include/", "./include/"] DEFAULT_PATH_LIBLLAMA = "./build/src/libllama.so" # CMakeLists.txt: BUILD_SHARED_LIBS ON - def __init__(self, path_llama_h: str = None, path_includes: list[str] = [], path_libllama: str = None): + def __init__(self, path_llama_h: str | None = None, path_includes: list[str] = [], path_libllama: str | None = None): path_llama_h = path_llama_h or self.DEFAULT_PATH_LLAMA_H path_includes = path_includes or self.DEFAULT_PATH_INCLUDES path_libllama = path_libllama or self.DEFAULT_PATH_LIBLLAMA (self.ffi, self.lib) = self._load_libllama_cffi(path_llama_h, path_includes, path_libllama) self.lib.llama_backend_init() - def _load_libllama_cffi(self, path_llama_h: str, path_includes: list[str], path_libllama: str): - cmd = ["gcc", "-E", "-P", "-D__restrict=", "-D__attribute__(x)=", "-D__asm__(x)="] + def _load_libllama_cffi(self, path_llama_h: str, path_includes: list[str], path_libllama: str) -> tuple[cffi.FFI, Any]: + cmd = ["gcc", "-O0", "-E", "-P", "-D__restrict=", "-D__attribute__(x)=", "-D__asm__(x)="] cmd += ["-I" + path for path in path_includes] + [path_llama_h] res = subprocess.run(cmd, stdout=subprocess.PIPE) assert (res.returncode == 0) @@ -68,7 +72,7 @@ class LibLlama: class LibLlamaModel: def __init__(self, libllama: LibLlama, path_model: str, mparams={}, cparams={}): - self.lib = libllama.lib + self.lib: Any = libllama.lib self.ffi = libllama.ffi if isinstance(mparams, dict): mparams = libllama.model_default_params(**mparams) @@ -94,11 +98,11 @@ class LibLlamaModel: self.lib = None def tokenize(self, text: str, add_special: bool = False, parse_special: bool = False) -> list[int]: - text = text.encode("utf-8") - num = self.lib.llama_tokenize(self.model, text, len(text), self.token_ids, len(self.token_ids), add_special, parse_special) + encoded_text: bytes = text.encode("utf-8") + num = self.lib.llama_tokenize(self.model, encoded_text, len(encoded_text), self.token_ids, len(self.token_ids), add_special, parse_special) while num < 0 and len(self.token_ids) < (16 << 20): self.token_ids = self.ffi.new("llama_token[]", -2 * num) - num = self.lib.llama_tokenize(self.model, text, len(text), self.token_ids, len(self.token_ids), add_special, parse_special) + num = self.lib.llama_tokenize(self.model, encoded_text, len(encoded_text), self.token_ids, len(self.token_ids), add_special, parse_special) return list(self.token_ids[0:num]) def detokenize(self, ids: list[int], remove_special: bool = False, unparse_special: bool = False) -> str: @@ -110,7 +114,7 @@ class LibLlamaModel: while num < 0 and len(self.text_buff) < (16 << 20): self.text_buff = self.ffi.new("uint8_t[]", -2 * num) num = self.lib.llama_detokenize(self.model, self.token_ids, len(ids), self.text_buff, len(self.text_buff), remove_special, unparse_special) - return str(self.ffi.buffer(self.text_buff, num), encoding="utf-8", errors="replace") # replace errors with '\uFFFD' + return str(cast(Buffer, self.ffi.buffer(self.text_buff, num)), encoding="utf-8", errors="replace") # replace errors with '\uFFFD' class Tokenizer: @@ -152,7 +156,7 @@ class TokenizerGroundtruth (Tokenizer): class TokenizerLlamaCpp (Tokenizer): - libllama: LibLlama = None + libllama: LibLlama | None = None def __init__(self, vocab_file: str): if not self.libllama: @@ -404,7 +408,7 @@ def generator_random_vocab_words(tokenizer: TokenizerGroundtruth, iterations=100 def compare_tokenizers(tokenizer1: TokenizerGroundtruth, tokenizer2: TokenizerLlamaCpp, generator: Iterator[str]): - def find_first_mismatch(ids1: list[int], ids2: list[int]): + def find_first_mismatch(ids1: list[int] | str, ids2: list[int] | str): for i, (a, b) in enumerate(zip(ids1, ids2)): if a != b: return i @@ -433,7 +437,7 @@ def compare_tokenizers(tokenizer1: TokenizerGroundtruth, tokenizer2: TokenizerLl decode_errors = 0 MAX_ERRORS = 10 - logger.info("%s: %s" % (generator.__name__, "ini")) + logger.info("%s: %s" % (generator.__qualname__, "ini")) for text in generator: # print(repr(text), text.encode()) # print(repr(text), hex(ord(text[0])), text.encode()) @@ -472,13 +476,13 @@ def compare_tokenizers(tokenizer1: TokenizerGroundtruth, tokenizer2: TokenizerLl break t_total = time.perf_counter() - t_start - logger.info(f"{generator.__name__}: end, {t_encode1=:.3f} {t_encode2=:.3f} {t_decode1=:.3f} {t_decode2=:.3f} {t_total=:.3f}") + logger.info(f"{generator.__qualname__}: end, {t_encode1=:.3f} {t_encode2=:.3f} {t_decode1=:.3f} {t_decode2=:.3f} {t_total=:.3f}") -def main(argv: list[str] = None): +def main(argv: list[str] | None = None): parser = argparse.ArgumentParser() - parser.add_argument("vocab_file", help="path to vocab 'gguf' file") - parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file") + parser.add_argument("vocab_file", type=str, help="path to vocab 'gguf' file") + parser.add_argument("dir_tokenizer", type=str, help="directory containing 'tokenizer.model' file") parser.add_argument("--verbose", action="store_true", help="increase output verbosity") args = parser.parse_args(argv) @@ -520,7 +524,7 @@ if __name__ == "__main__": format = "%(levelname)s %(message)s", ) - path_tokenizers = "./models/tokenizers/" + path_tokenizers = Path("./models/tokenizers/") path_vocab_format = "./models/ggml-vocab-%s.gguf" tokenizers = [ @@ -556,6 +560,6 @@ if __name__ == "__main__": for tokenizer in tokenizers: logger.info("-" * 50) logger.info(f"TOKENIZER: '{tokenizer}'") - vocab_file = path_vocab_format % tokenizer - dir_tokenizer = path_tokenizers + "/" + tokenizer - main([vocab_file, dir_tokenizer, "--verbose"]) + vocab_file = Path(path_vocab_format % tokenizer) + dir_tokenizer = path_tokenizers / tokenizer + main([str(vocab_file), str(dir_tokenizer), "--verbose"]) From 04ce3a8b19256a155aea4d14eaa87edf274c93c3 Mon Sep 17 00:00:00 2001 From: toyer <2042519524@qq.com> Date: Mon, 8 Jul 2024 13:57:19 +0800 Subject: [PATCH 02/31] readme : add supported glm models (#8360) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index aba6a4c59..5498d5fc6 100644 --- a/README.md +++ b/README.md @@ -96,6 +96,7 @@ Typically finetunes of the base models below are supported as well. - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B) - [x] [OLMo](https://allenai.org/olmo) - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia) +- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) (instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md)) From ffd00797d81ef7db1528b9e10adbdc333ade6495 Mon Sep 17 00:00:00 2001 From: Kevin Wang Date: Mon, 8 Jul 2024 02:31:55 -0400 Subject: [PATCH 03/31] common : avoid unnecessary logits fetch (#8358) --- common/sampling.cpp | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/common/sampling.cpp b/common/sampling.cpp index 9f332fe57..e9db2715f 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -282,8 +282,6 @@ static llama_token llama_sampling_sample_impl( GGML_ASSERT(!original_logits.empty()); } llama_token id = 0; - // Get a pointer to the logits - float * logits = llama_get_logits_ith(ctx_main, idx); if (temp < 0.0) { // greedy sampling, with probs @@ -324,6 +322,9 @@ static llama_token llama_sampling_sample_impl( } if (ctx_sampling->grammar != NULL && !is_resampling) { + // Get a pointer to the logits + float * logits = llama_get_logits_ith(ctx_main, idx); + // Create an array with a single token data element for the sampled id llama_token_data single_token_data = {id, logits[id], 0.0f}; llama_token_data_array single_token_data_array = { &single_token_data, 1, false }; From 6f0dbf6ab087bcd286fb78560099ca0458316735 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 8 Jul 2024 09:34:35 +0300 Subject: [PATCH 04/31] infill : assert prefix/suffix tokens + remove old space logic (#8351) --- common/log.h | 2 +- examples/infill/infill.cpp | 25 ++++++++----------------- 2 files changed, 9 insertions(+), 18 deletions(-) diff --git a/common/log.h b/common/log.h index 09fa63c26..1bc5328ce 100644 --- a/common/log.h +++ b/common/log.h @@ -630,7 +630,7 @@ inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens) buf << "[ "; bool first = true; - for (const auto &token : tokens) + for (const auto & token : tokens) { if (!first) { buf << ", "; diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 0e682154d..dc93d2301 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -204,21 +204,17 @@ int main(int argc, char ** argv) { GGML_ASSERT(llama_add_eos_token(model) != 1); LOG("add_bos: %d\n", add_bos); - bool suff_rm_leading_spc = params.escape; - if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { - params.input_suffix.erase(0, 1); - suff_rm_leading_spc = false; - } std::vector embd_inp; std::vector embd_end; std::vector inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false); std::vector inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false); - const int space_token = 29871; - if (suff_rm_leading_spc && inp_sfx[0] == space_token) { - inp_sfx.erase(inp_sfx.begin()); - } + + GGML_ASSERT(llama_token_prefix(model) >= 0); + GGML_ASSERT(llama_token_suffix(model) >= 0); + inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); + embd_inp = params.spm_infill ? inp_sfx : inp_pfx; embd_end = params.spm_infill ? inp_pfx : inp_sfx; if (add_bos) { @@ -516,19 +512,14 @@ int main(int argc, char ** argv) { string_process_escapes(params.input_prefix); string_process_escapes(params.input_suffix); } - suff_rm_leading_spc = params.escape; - if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { - params.input_suffix.erase(0, 1); - suff_rm_leading_spc = false; - } + // tokenize new prefix and suffix std::vector inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false); std::vector inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false); - if (suff_rm_leading_spc && inp_sfx[0] == space_token) { - inp_sfx.erase(inp_sfx.begin()); - } + inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); + embd_inp = params.spm_infill ? inp_sfx : inp_pfx; embd_end = params.spm_infill ? inp_pfx : inp_sfx; if (add_bos) { From 470939d483d1c89b7292f78bac1fd27c42c171ce Mon Sep 17 00:00:00 2001 From: Kevin Wang Date: Mon, 8 Jul 2024 03:26:53 -0400 Subject: [PATCH 05/31] common : preallocate sampling token data vector (#8363) `emplace_back` repeatedly-called is slower than preallocating the vector to the vocab size and directly inserting the data. Some rudimentary profiling with `chrono` improves the performance of this block of code from ~500us/op to ~40us/op. Overall, this slightly improves the sampling performance which has a more substantial impact for the `examples/lookahead` implementation -- I am able to see a ~10% performance boost in lookahead inference. --- common/sampling.cpp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/common/sampling.cpp b/common/sampling.cpp index e9db2715f..6a483c815 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -378,7 +378,7 @@ static llama_token_data_array llama_sampling_prepare_impl( if (ctx_sampling->grammar != NULL && !apply_grammar) { GGML_ASSERT(original_logits != NULL); // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this. - *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))}; + *original_logits = {logits, logits + n_vocab}; } // apply params.logit_bias map @@ -391,10 +391,10 @@ static llama_token_data_array llama_sampling_prepare_impl( llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale); } - cur.clear(); + cur.resize(n_vocab); for (llama_token token_id = 0; token_id < n_vocab; token_id++) { - cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); + cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f}; } llama_token_data_array cur_p = { cur.data(), cur.size(), false }; From fde13b3bb9f569f07fd8af74696ee48a43d05131 Mon Sep 17 00:00:00 2001 From: John Balis Date: Tue, 2 Jul 2024 11:09:52 -0500 Subject: [PATCH 06/31] feat: cuda implementation for `ggml_conv_transpose_1d` (ggml/854) * conv transpose 1d passing test for 1d input and kernel * working for different input and output channel counts, added test for variable stride * initial draft appears to work with stride other than 1 * working with all old and new conv1d tests * added a test for large tensors * removed use cuda hardcoding * restored test-conv-transpose.c * removed unused arugments, and fixed bug where test failure would cause subsequent tests to fail * fixed accumulator bug * added test to test-backend-ops * fixed mistake * addressed review * fixed includes * removed blank lines * style and warning fixes * return failure when test fails * fix supports_op --------- Co-authored-by: slaren --- ggml/src/ggml-cuda.cu | 13 ++++ ggml/src/ggml-cuda/conv-transpose-1d.cu | 87 ++++++++++++++++++++++++ ggml/src/ggml-cuda/conv-transpose-1d.cuh | 5 ++ tests/test-backend-ops.cpp | 42 +++++++++++- 4 files changed, 146 insertions(+), 1 deletion(-) create mode 100644 ggml/src/ggml-cuda/conv-transpose-1d.cu create mode 100644 ggml/src/ggml-cuda/conv-transpose-1d.cuh diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 1c9ccc8a1..ed784ea1c 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -29,6 +29,7 @@ #include "ggml-cuda/tsembd.cuh" #include "ggml-cuda/unary.cuh" #include "ggml-cuda/upscale.cuh" +#include "ggml-cuda/conv-transpose-1d.cuh" #include #include @@ -2261,6 +2262,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_IM2COL: ggml_cuda_op_im2col(ctx, dst); break; + case GGML_OP_CONV_TRANSPOSE_1D: + ggml_cuda_op_conv_transpose_1d(ctx,dst); + break; case GGML_OP_POOL_2D: ggml_cuda_op_pool2d(ctx, dst); break; @@ -2804,6 +2808,15 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons ggml_type src0_type = op->src[0]->type; return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16; } break; + case GGML_OP_CONV_TRANSPOSE_1D: + { + ggml_type src0_type = op->src[0]->type; + ggml_type src1_type = op->src[1]->type; + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { + return true; + } + return false; + } break; case GGML_OP_NONE: case GGML_OP_RESHAPE: case GGML_OP_VIEW: diff --git a/ggml/src/ggml-cuda/conv-transpose-1d.cu b/ggml/src/ggml-cuda/conv-transpose-1d.cu new file mode 100644 index 000000000..b1e94d6f7 --- /dev/null +++ b/ggml/src/ggml-cuda/conv-transpose-1d.cu @@ -0,0 +1,87 @@ +#include "conv-transpose-1d.cuh" + +static __global__ void conv_transpose_1d_kernel( + const int s0, const int p0, const int d0, const int output_size, + const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3, + const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3, + const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3, + const float * src0, const float * src1, float * dst) { + int global_index = threadIdx.x + blockIdx.x * blockDim.x; + if (global_index >= output_size) { + return; + } + + int out_index = global_index / dst_ne0; + + float accumulator = 0; + + for (int c = 0; c < src0_ne2; c++) { + int idx = global_index % dst_ne0; + + int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0); + int input_offset = src1_ne0 * c; + + for (int i = 0; i < src1_ne0; i++) { + if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) { + continue; + } + int weight_idx = idx - i*s0; + + float kernel_weight = src0[kernel_offset + weight_idx]; + float input_value = src1[input_offset+i]; + + accumulator += kernel_weight * input_value; + } + } + dst[global_index] = accumulator; +} + +static void conv_transpose_1d_f32_f32_cuda( + const int s0, const int p0, const int d0, const int output_size, + const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3, + const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3, + const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3, + const float * src0, const float * src1, float * dst, + cudaStream_t stream) { + + const int num_blocks = (output_size + CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE; + conv_transpose_1d_kernel<<>>( + s0,p0,d0,output_size, + src0_ne0, src0_ne1, src0_ne2, src0_ne3, + src1_ne0, src1_ne1, src1_ne2, src1_ne3, + dst_ne0, dst_ne1, dst_ne2, dst_ne3, + src0,src1, dst); +} + +void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + + const ggml_tensor * src1 = dst->src[1]; + const float * src1_d = (const float *)src1->data; + + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + + const int32_t * opts = (const int32_t *)dst->op_params; + + const int s0 = opts[0]; + const int p0 = 0;//opts[3]; + const int d0 = 1;//opts[4]; + + const int64_t kernel_size = ggml_nelements(src0); + const int64_t input_size = ggml_nelements(src1); + const int64_t output_size = ggml_nelements(dst); + + conv_transpose_1d_f32_f32_cuda(s0, p0, d0, output_size, + src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], + src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], + dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], + src0_d, src1_d, dst_d, stream); +} diff --git a/ggml/src/ggml-cuda/conv-transpose-1d.cuh b/ggml/src/ggml-cuda/conv-transpose-1d.cuh new file mode 100644 index 000000000..6c2cf666b --- /dev/null +++ b/ggml/src/ggml-cuda/conv-transpose-1d.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE 256 + +void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 2bb71ac03..83d4a0eb1 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -1266,6 +1266,36 @@ struct test_pool2d : public test_case { } }; +// GGML_OP_CONV_TRANSPOSE_1D +struct test_conv_transpose_1d : public test_case { + + const std::array ne_input; + const std::array ne_kernel; + + // stride + const int s0; + // padding + const int p0; + // dilation + const int d0; + + std::string vars() override { + return VARS_TO_STR5(ne_input, ne_kernel, s0, p0, d0); + } + + test_conv_transpose_1d(std::array ne_input = {197, 32, 1, 1}, // [input_width, input_height, input_channels, 1] + std::array ne_kernel = {16, 32, 32, 1}, // [kernel_width, kernel_height, input_channels, 1] + int s0 = 1, int p0 = 0, int d0 = 1) + : ne_input(ne_input), ne_kernel(ne_kernel), s0(s0), p0(p0), d0(d0) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data()); + ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_kernel.data()); + ggml_tensor * out = ggml_conv_transpose_1d(ctx, kernel, input, s0, p0, d0); + return out; + } +}; + // GGML_OP_IM2COL struct test_im2col : public test_case { const ggml_type type_input; @@ -1279,7 +1309,7 @@ struct test_im2col : public test_case { // padding const int p0; const int p1; - // dilatation + // dilation const int d0; const int d1; // mode @@ -2098,6 +2128,16 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32)); test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16)); + test_cases.emplace_back(new test_conv_transpose_1d()); + test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1)); + test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1)); + test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 1, 0, 1)); + test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 2, 0, 1)); + test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 1, 0, 1)); + test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1)); + test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1)); + + test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 1})); test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {2, 1, 1, 1})); test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 2, 1, 1})); From 6847d54c4f72f8bf6767b6feae7a95394654335e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 8 Jul 2024 10:39:36 +0300 Subject: [PATCH 07/31] tests : fix whitespace (#0) --- tests/test-backend-ops.cpp | 10 +++------- 1 file changed, 3 insertions(+), 7 deletions(-) diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 83d4a0eb1..a1bfc8fdd 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -1268,16 +1268,12 @@ struct test_pool2d : public test_case { // GGML_OP_CONV_TRANSPOSE_1D struct test_conv_transpose_1d : public test_case { - const std::array ne_input; const std::array ne_kernel; - // stride - const int s0; - // padding - const int p0; - // dilation - const int d0; + const int s0; // stride + const int p0; // padding + const int d0; // dilation std::string vars() override { return VARS_TO_STR5(ne_input, ne_kernel, s0, p0, d0); From 2ee44c9a1865a928ccbbc16a2d7841d7513f31c1 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 8 Jul 2024 10:39:50 +0300 Subject: [PATCH 08/31] sync : ggml ggml-ci --- scripts/sync-ggml.last | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index 2da33e913..80159b70b 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -5378ea0d3c2f25bcd330ecb226ad2db454be86d0 +e3b3846976c94163f2b3dd128cc959782653edbb From 3f2d538b817112ad8429341c7e8657dcd660f4d3 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 8 Jul 2024 13:51:31 +0300 Subject: [PATCH 09/31] scripts : fix sync for sycl --- scripts/sync-ggml-am.sh | 3 +++ scripts/sync-ggml.sh | 1 + 2 files changed, 4 insertions(+) diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh index b05a33747..ba3bedf21 100755 --- a/scripts/sync-ggml-am.sh +++ b/scripts/sync-ggml-am.sh @@ -63,6 +63,7 @@ while read c; do src/ggml*.metal \ src/ggml*.cu \ src/ggml-cuda/* \ + src/ggml-sycl/* \ include/ggml*.h \ tests/test-opt.cpp \ tests/test-grad0.cpp \ @@ -113,6 +114,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then # src/ggml-quants.c -> ggml/src/ggml-quants.c # src/ggml-quants.h -> ggml/src/ggml-quants.h # src/ggml-rpc.cpp -> ggml/src/ggml-rpc.cpp + # src/ggml-sycl/* -> ggml/src/ggml-sycl/ # src/ggml-sycl.cpp -> ggml/src/ggml-sycl.cpp # src/ggml-vulkan.cpp -> ggml/src/ggml-vulkan.cpp # @@ -153,6 +155,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then -e 's/([[:space:]]|[ab]\/)src\/ggml-quants\.c/\1ggml\/src\/ggml-quants.c/g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-quants\.h/\1ggml\/src\/ggml-quants.h/g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-rpc\.cpp/\1ggml\/src\/ggml-rpc.cpp/g' \ + -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\.cpp/\1ggml\/src\/ggml-sycl.cpp/g' \ -e 's/([[:space:]]|[ab]\/)src\/ggml-vulkan\.cpp/\1ggml\/src\/ggml-vulkan.cpp/g' \ -e 's/([[:space:]]|[ab]\/)include\/ggml\.h/\1ggml\/include\/ggml.h/g' \ diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh index 2f32c1ce8..402446ef9 100755 --- a/scripts/sync-ggml.sh +++ b/scripts/sync-ggml.sh @@ -18,6 +18,7 @@ cp -rpv ../ggml/src/ggml-metal.metal ./ggml/src/ggml-metal.metal cp -rpv ../ggml/src/ggml-quants.c ./ggml/src/ggml-quants.c cp -rpv ../ggml/src/ggml-quants.h ./ggml/src/ggml-quants.h cp -rpv ../ggml/src/ggml-rpc.cpp ./ggml/src/ggml-rpc.cpp +cp -rpv ../ggml/src/ggml-sycl/* ./ggml/src/ggml-sycl/ cp -rpv ../ggml/src/ggml-sycl.cpp ./ggml/src/ggml-sycl.cpp cp -rpv ../ggml/src/ggml-vulkan.cpp ./ggml/src/ggml-vulkan.cpp From 2ec846d558f6385ea647f7b8e665eb249c1ebce7 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Alberto=20Cabrera=20P=C3=A9rez?= Date: Mon, 8 Jul 2024 14:22:41 +0100 Subject: [PATCH 10/31] sycl : fix powf call in device code (#8368) --- ggml/src/ggml-sycl/rope.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp index eabf1693e..6f507941a 100644 --- a/ggml/src/ggml-sycl/rope.cpp +++ b/ggml/src/ggml-sycl/rope.cpp @@ -55,7 +55,7 @@ static void rope_norm( const int i = row*ne0 + i0; const int i2 = row/p_delta_rows; - const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f); + const float theta_base = pos[i2] * sycl::pow(theta_scale, i0 / 2.0f); const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f; @@ -98,7 +98,7 @@ static void rope_neox( const int i = row*ne0 + i0/2; const int i2 = row/p_delta_rows; - const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f); + const float theta_base = pos[i2] * sycl::pow(theta_scale, i0 / 2.0f); const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f; From c4dd11d1d3903e1922c06242e189f6310fc4d8c3 Mon Sep 17 00:00:00 2001 From: b4b4o Date: Mon, 8 Jul 2024 22:19:24 +0800 Subject: [PATCH 11/31] readme : fix web link error [no ci] (#8347) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 5498d5fc6..aa7070811 100644 --- a/README.md +++ b/README.md @@ -98,7 +98,7 @@ Typically finetunes of the base models below are supported as well. - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia) - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) -(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md)) +(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md)) **Multimodal models:** From a130eccef42b75a84da270411cefeed45c153e30 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Alberto=20Cabrera=20P=C3=A9rez?= Date: Mon, 8 Jul 2024 21:35:17 +0100 Subject: [PATCH 12/31] labeler : updated sycl to match docs and code refactor (#8373) --- .github/labeler.yml | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/.github/labeler.yml b/.github/labeler.yml index 9c0397d16..89436740d 100644 --- a/.github/labeler.yml +++ b/.github/labeler.yml @@ -16,7 +16,9 @@ SYCL: - any-glob-to-any-file: - ggml/include/ggml-sycl.h - ggml/src/ggml-sycl.cpp - - README-sycl.md + - ggml/src/ggml-sycl/** + - docs/backend/SYCL.md + - examples/sycl/** Nvidia GPU: - changed-files: - any-glob-to-any-file: From 7fdb6f73e35605c8dbc39e9f19cd9ed84dbc87f2 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 9 Jul 2024 01:36:38 +0300 Subject: [PATCH 13/31] flake.lock: Update (#8342) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Flake lock file updates: • Updated input 'flake-parts': 'github:hercules-ci/flake-parts/2a55567fcf15b1b1c7ed712a2c6fadaec7412ea8?narHash=sha256-iKzJcpdXih14qYVcZ9QC9XuZYnPc6T8YImb6dX166kw%3D' (2024-06-01) → 'github:hercules-ci/flake-parts/9227223f6d922fee3c7b190b2cc238a99527bbb7?narHash=sha256-pQMhCCHyQGRzdfAkdJ4cIWiw%2BJNuWsTX7f0ZYSyz0VY%3D' (2024-07-03) • Updated input 'flake-parts/nixpkgs-lib': 'https://github.com/NixOS/nixpkgs/archive/eb9ceca17df2ea50a250b6b27f7bf6ab0186f198.tar.gz?narHash=sha256-lIbdfCsf8LMFloheeE6N31%2BBMIeixqyQWbSr2vk79EQ%3D' (2024-06-01) → 'https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz?narHash=sha256-Fm2rDDs86sHy0/1jxTOKB1118Q0O3Uc7EC0iXvXKpbI%3D' (2024-07-01) • Updated input 'nixpkgs': 'github:NixOS/nixpkgs/b2852eb9365c6de48ffb0dc2c9562591f652242a?narHash=sha256-C8e9S7RzshSdHB7L%2Bv9I51af1gDM5unhJ2xO1ywxNH8%3D' (2024-06-27) → 'github:NixOS/nixpkgs/9f4128e00b0ae8ec65918efeba59db998750ead6?narHash=sha256-rwz8NJZV%2B387rnWpTYcXaRNvzUSnnF9aHONoJIYmiUQ%3D' (2024-07-03) Co-authored-by: github-actions[bot] --- flake.lock | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/flake.lock b/flake.lock index 973ff4e56..5001fbf91 100644 --- a/flake.lock +++ b/flake.lock @@ -5,11 +5,11 @@ "nixpkgs-lib": "nixpkgs-lib" }, "locked": { - "lastModified": 1717285511, - "narHash": "sha256-iKzJcpdXih14qYVcZ9QC9XuZYnPc6T8YImb6dX166kw=", + "lastModified": 1719994518, + "narHash": "sha256-pQMhCCHyQGRzdfAkdJ4cIWiw+JNuWsTX7f0ZYSyz0VY=", "owner": "hercules-ci", "repo": "flake-parts", - "rev": "2a55567fcf15b1b1c7ed712a2c6fadaec7412ea8", + "rev": "9227223f6d922fee3c7b190b2cc238a99527bbb7", "type": "github" }, "original": { @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1719506693, - "narHash": "sha256-C8e9S7RzshSdHB7L+v9I51af1gDM5unhJ2xO1ywxNH8=", + "lastModified": 1720031269, + "narHash": "sha256-rwz8NJZV+387rnWpTYcXaRNvzUSnnF9aHONoJIYmiUQ=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "b2852eb9365c6de48ffb0dc2c9562591f652242a", + "rev": "9f4128e00b0ae8ec65918efeba59db998750ead6", "type": "github" }, "original": { @@ -36,14 +36,14 @@ }, "nixpkgs-lib": { "locked": { - "lastModified": 1717284937, - "narHash": "sha256-lIbdfCsf8LMFloheeE6N31+BMIeixqyQWbSr2vk79EQ=", + "lastModified": 1719876945, + "narHash": "sha256-Fm2rDDs86sHy0/1jxTOKB1118Q0O3Uc7EC0iXvXKpbI=", "type": "tarball", - "url": "https://github.com/NixOS/nixpkgs/archive/eb9ceca17df2ea50a250b6b27f7bf6ab0186f198.tar.gz" + "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz" }, "original": { "type": "tarball", - "url": "https://github.com/NixOS/nixpkgs/archive/eb9ceca17df2ea50a250b6b27f7bf6ab0186f198.tar.gz" + "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz" } }, "root": { From 7d0e23d72ef4540d0d4409cb63ae682c17d53926 Mon Sep 17 00:00:00 2001 From: compilade Date: Tue, 9 Jul 2024 01:04:49 -0400 Subject: [PATCH 14/31] gguf-py : do not use internal numpy types (#7472) --- gguf-py/gguf/lazy.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/gguf-py/gguf/lazy.py b/gguf-py/gguf/lazy.py index c50124cd9..6e266f34f 100644 --- a/gguf-py/gguf/lazy.py +++ b/gguf-py/gguf/lazy.py @@ -6,7 +6,6 @@ from typing import Any, Callable from collections import deque import numpy as np -from numpy._typing import _Shape from numpy.typing import DTypeLike @@ -219,7 +218,7 @@ class LazyNumpyTensor(LazyBase): _tensor_type = np.ndarray @classmethod - def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: _Shape) -> np.ndarray[Any, Any]: + def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: tuple[int, ...]) -> 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. From 9beb2dda038e2107a39a95def94a4cf971e048ea Mon Sep 17 00:00:00 2001 From: daghanerdonmez <44506702+daghanerdonmez@users.noreply.github.com> Date: Tue, 9 Jul 2024 09:16:00 +0300 Subject: [PATCH 15/31] readme : fix typo [no ci] (#8389) Bakus-Naur --> Backus-Naur --- grammars/README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/grammars/README.md b/grammars/README.md index 886023f77..01b02abb4 100644 --- a/grammars/README.md +++ b/grammars/README.md @@ -4,7 +4,7 @@ GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia. ## Background -[Bakus-Naur Form (BNF)](https://en.wikipedia.org/wiki/Backus%E2%80%93Naur_form) is a notation for describing the syntax of formal languages like programming languages, file formats, and protocols. GBNF is an extension of BNF that primarily adds a few modern regex-like features. +[Backus-Naur Form (BNF)](https://en.wikipedia.org/wiki/Backus%E2%80%93Naur_form) is a notation for describing the syntax of formal languages like programming languages, file formats, and protocols. GBNF is an extension of BNF that primarily adds a few modern regex-like features. ## Basics From 9925ca4087a34ab973b07bf06c0b770cb586830b Mon Sep 17 00:00:00 2001 From: Borislav Stanimirov Date: Tue, 9 Jul 2024 11:38:00 +0300 Subject: [PATCH 16/31] cmake : allow external ggml (#8370) --- CMakeLists.txt | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 4f6cd6872..cb97d8791 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -115,7 +115,10 @@ llama_option_depr(WARNING LLAMA_SYCL_F16 GGML_SYCL_F16) # build the library # -add_subdirectory(ggml) +if (NOT TARGET ggml) + add_subdirectory(ggml) + # ... otherwise assume ggml is added by a parent CMakeLists.txt +endif() add_subdirectory(src) # From 5b0b8d8cfb5ddf2118f686ba6c30fab3f71b384b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Alberto=20Cabrera=20P=C3=A9rez?= Date: Tue, 9 Jul 2024 15:03:15 +0100 Subject: [PATCH 17/31] sycl : Reenabled mmvq path for the SYCL Nvidia Backend (#8372) * SYCL : Reenabled mmvq path for the SYCL Nvidia Backend * Reduced verbosity of comment --- ggml/src/ggml-sycl.cpp | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index 21006cd7b..9c419ba89 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -3658,6 +3658,10 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE); #endif // SYCL_USE_XMX + // mmvq path is faster in the CUDA backend. + if (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda) + use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q; + if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { // KQ single-batch ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst); From a03e8dd99d3954e7547137936c5a2a3b348bdb7f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Tue, 9 Jul 2024 17:11:07 +0200 Subject: [PATCH 18/31] make/cmake: LLAMA_NO_CCACHE -> GGML_NO_CCACHE (#8392) --- CMakeLists.txt | 4 ---- Makefile | 12 +++++++++--- 2 files changed, 9 insertions(+), 7 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index cb97d8791..befc4433c 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -50,9 +50,6 @@ endif() # option list # -# general -option(LLAMA_CCACHE "llama: use ccache if available" ON) - # debug option(LLAMA_ALL_WARNINGS "llama: enable all compiler warnings" ON) option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF) @@ -77,7 +74,6 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF) include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake) # override ggml options -set(GGML_CCACHE ${LLAMA_CCACHE}) set(GGML_SANITIZE_THREAD ${LLAMA_SANITIZE_THREAD}) set(GGML_SANITIZE_ADDRESS ${LLAMA_SANITIZE_ADDRESS}) set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED}) diff --git a/Makefile b/Makefile index bb6e2f968..be7fb9401 100644 --- a/Makefile +++ b/Makefile @@ -245,17 +245,22 @@ MK_CFLAGS = -std=c11 -fPIC MK_CXXFLAGS = -std=c++11 -fPIC MK_NVCCFLAGS = -std=c++11 -ifndef LLAMA_NO_CCACHE +ifdef LLAMA_NO_CCACHE +GGML_NO_CCACHE := 1 +DEPRECATE_WARNING := 1 +endif + +ifndef GGML_NO_CCACHE CCACHE := $(shell which ccache) ifdef CCACHE export CCACHE_SLOPPINESS = time_macros -$(info I ccache found, compilation results will be cached. Disable with LLAMA_NO_CCACHE.) +$(info I ccache found, compilation results will be cached. Disable with GGML_NO_CCACHE.) CC := $(CCACHE) $(CC) CXX := $(CCACHE) $(CXX) else $(info I ccache not found. Consider installing it for faster compilation.) endif # CCACHE -endif # LLAMA_NO_CCACHE +endif # GGML_NO_CCACHE # clock_gettime came in POSIX.1b (1993) # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional @@ -926,6 +931,7 @@ $(info - LLAMA_NO_LLAMAFILE) $(info - LLAMA_NO_ACCELERATE) $(info - LLAMA_NO_OPENMP) $(info - LLAMA_NO_METAL) +$(info - LLAMA_NO_CCACHE) $(info ) endif From e500d6135ac42c4a23baf6a9a1a934dc023e5c7d Mon Sep 17 00:00:00 2001 From: Clint Herron Date: Tue, 9 Jul 2024 11:54:43 -0400 Subject: [PATCH 19/31] Deprecation warning to assist with migration to new binary names (#8283) * Adding a simple program to provide a deprecation warning that can exist to help people notice the binary name change from #7809 and migrate to the new filenames. * Build legacy replacement binaries only if they already exist. Check for their existence every time so that they are not ignored. --- Makefile | 78 ++++++++++++++++++- examples/deprecation-warning/README.md | 51 ++++++++++++ .../deprecation-warning.cpp | 35 +++++++++ 3 files changed, 160 insertions(+), 4 deletions(-) create mode 100644 examples/deprecation-warning/README.md create mode 100644 examples/deprecation-warning/deprecation-warning.cpp diff --git a/Makefile b/Makefile index be7fb9401..104120090 100644 --- a/Makefile +++ b/Makefile @@ -64,10 +64,14 @@ TEST_TARGETS = \ tests/test-tokenizer-1-spm # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned -LEGACY_TARGETS = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ +LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \ retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm +# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them. +# We don't want to clutter things too much, so we only build replacements for the most commonly used binaries. +LEGACY_TARGETS_BUILD = main quantize perplexity embedding server finetune + # Deprecation aliases ifdef LLAMA_CUBLAS $(error LLAMA_CUBLAS is removed. Use GGML_CUDA instead.) @@ -193,7 +197,7 @@ ifdef GGML_RPC BUILD_TARGETS += rpc-server endif -default: $(BUILD_TARGETS) +default: $(BUILD_TARGETS) $(LEGACY_TARGETS_BUILD) test: $(TEST_TARGETS) @failures=0; \ @@ -228,7 +232,7 @@ test: $(TEST_TARGETS) fi @echo 'All tests passed.' -all: $(BUILD_TARGETS) $(TEST_TARGETS) +all: $(BUILD_TARGETS) $(TEST_TARGETS) $(LEGACY_TARGETS_BUILD) ifdef RISCV_CROSS_COMPILE CC := riscv64-unknown-linux-gnu-gcc @@ -1098,7 +1102,7 @@ clean: rm -vrf ggml/src/ggml-cuda/template-instances/*.o rm -rvf $(BUILD_TARGETS) rm -rvf $(TEST_TARGETS) - rm -rvf $(LEGACY_TARGETS) + rm -rvf $(LEGACY_TARGETS_CLEAN) find examples pocs -type f -name "*.o" -delete # @@ -1494,3 +1498,69 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \ $(OBJ_GGML) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + +# +# Deprecated binaries that we want to keep around long enough for people to migrate to the new filenames, then these can be removed. +# +# Mark legacy binary targets as .PHONY so that they are always checked. +.PHONY: main quantize perplexity embedding server finetune + +main: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard main)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'main' binary is deprecated. Please use 'llama-cli' instead." + @echo " Remove the 'main' binary to remove this warning." + @echo "#########" +endif + +quantize: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard quantize)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead." + @echo " Remove the 'quantize' binary to remove this warning." + @echo "#########" +endif + +perplexity: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard perplexity)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead." + @echo " Remove the 'perplexity' binary to remove this warning." + @echo "#########" +endif + +embedding: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard embedding)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead." + @echo " Remove the 'embedding' binary to remove this warning." + @echo "#########" +endif + +server: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard server)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'server' binary is deprecated. Please use 'llama-server' instead." + @echo " Remove the 'server' binary to remove this warning." + @echo "#########" +endif + +finetune: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard finetune)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'finetune' binary is deprecated. Please use 'llama-finetune' instead." + @echo " Remove the 'finetune' binary to remove this warning." + @echo "#########" +endif diff --git a/examples/deprecation-warning/README.md b/examples/deprecation-warning/README.md new file mode 100644 index 000000000..1e20feb4a --- /dev/null +++ b/examples/deprecation-warning/README.md @@ -0,0 +1,51 @@ +# Migration notice for binary filenames + +> [!IMPORTANT] +[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809) + +This migration was important, but it is a breaking change that may not always be immediately obvious to users. + +Please update all scripts and workflows to use the new binary names. + +| Old Filename | New Filename | +| ---- | ---- | +| main | llama-cli | +| server | llama-server | +| llama-bench | llama-bench | +| embedding | llama-embedding | +| finetune | llama-finetune | +| quantize | llama-quantize | +| tokenize | llama-tokenize | +| export-lora | llama-export-lora | +| libllava.a | libllava.a | +| baby-llama | llama-baby-llama | +| batched | llama-batched | +| batched-bench | llama-batched-bench | +| benchmark-matmult | llama-benchmark-matmult | +| convert-llama2c-to-ggml | llama-convert-llama2c-to-ggml | +| eval-callback | llama-eval-callback | +| gbnf-validator | llama-gbnf-validator | +| gguf | llama-gguf | +| gguf-split | llama-gguf-split | +| gritlm | llama-gritlm | +| imatrix | llama-imatrix | +| infill | llama-infill | +| llava-cli | llama-llava-cli | +| lookahead | llama-lookahead | +| lookup | llama-lookup | +| lookup-create | llama-lookup-create | +| lookup-merge | llama-lookup-merge | +| lookup-stats | llama-lookup-stats | +| parallel | llama-parallel | +| passkey | llama-passkey | +| perplexity | llama-perplexity | +| q8dot | llama-q8dot | +| quantize-stats | llama-quantize-stats | +| retrieval | llama-retrieval | +| save-load-state | llama-save-load-state | +| simple | llama-simple | +| speculative | llama-speculative | +| train-text-from-scratch | llama-train-text-from-scratch | +| vdot | llama-vdot | +| tests/test-c.o | tests/test-c.o | + diff --git a/examples/deprecation-warning/deprecation-warning.cpp b/examples/deprecation-warning/deprecation-warning.cpp new file mode 100644 index 000000000..11b35d2c2 --- /dev/null +++ b/examples/deprecation-warning/deprecation-warning.cpp @@ -0,0 +1,35 @@ +// Warns users that this filename was deprecated, and provides a link for more information. + +#include +#include +#include + +// Main +int main(int argc, char** argv) { + std::string filename = "main"; + if (argc >= 1) { + filename = argv[0]; + } + + // Get only the program name from the full path + auto pos = filename.find_last_of('/'); + if (pos != std::string::npos) { + filename = filename.substr(pos+1); + } + + // Append "llama-" to the beginning of filename to get the replacemnt filename + auto replacement_filename = "llama-" + filename; + + // The exception is if the filename is "main", then our replacement filename is "llama-cli" + if (filename == "main") { + replacement_filename = "llama-cli"; + } + + fprintf(stdout, "\n"); + fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str()); + fprintf(stdout, " Please use '%s' instead.\n", replacement_filename.c_str()); + fprintf(stdout, " See https://github.com/ggerganov/llama.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n"); + fprintf(stdout, "\n"); + + return EXIT_FAILURE; +} From fd560fe680c72fd0a0af2bc8881add20ad919071 Mon Sep 17 00:00:00 2001 From: Andy Salerno Date: Tue, 9 Jul 2024 11:58:44 -0700 Subject: [PATCH 20/31] Update README.md to fix broken link to docs (#8399) Update the "Performance troubleshooting" doc link to be correct - the file was moved into a dir called 'development' --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index aa7070811..058919068 100644 --- a/README.md +++ b/README.md @@ -453,7 +453,7 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio - [How to build](./docs/build.md) - [Running on Docker](./docs/docker.md) - [Build on Android](./docs/android.md) -- [Performance troubleshooting](./docs/token_generation_performance_tips.md) +- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md) - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks) **Seminal papers and background on the models** From a59f8fdc85e1119d470d8766e29617962549d993 Mon Sep 17 00:00:00 2001 From: Clint Herron Date: Tue, 9 Jul 2024 18:26:40 -0400 Subject: [PATCH 21/31] Server: Enable setting default sampling parameters via command-line (#8402) * Load server sampling parameters from the server context by default. * Wordsmithing comment --- examples/server/server.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 47bea1591..8feff6702 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -884,7 +884,8 @@ struct server_context { bool launch_slot_with_task(server_slot & slot, const server_task & task) { slot_params default_params; - llama_sampling_params default_sparams; + // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them) + llama_sampling_params default_sparams = params.sparams; auto & data = task.data; if (data.count("__oaicompat") != 0) { From 8f0fad42b9589f9b11362c74ea5338c51e4c7e61 Mon Sep 17 00:00:00 2001 From: laik Date: Wed, 10 Jul 2024 19:19:10 +0800 Subject: [PATCH 22/31] py : fix extra space in convert_hf_to_gguf.py (#8407) --- convert_hf_to_gguf.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 6cea73f08..ecb807262 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -1356,7 +1356,7 @@ class LlamaModel(Model): def set_vocab(self): try: - self. _set_vocab_sentencepiece() + self._set_vocab_sentencepiece() except FileNotFoundError: try: self._set_vocab_llama_hf() From e4dd31ff892627665d3c53c5d1a03c0d282d9d45 Mon Sep 17 00:00:00 2001 From: RunningLeon Date: Wed, 10 Jul 2024 19:26:40 +0800 Subject: [PATCH 23/31] py : fix converter for internlm2 (#8321) * update internlm2 * remove unused file * fix lint --- convert_hf_to_gguf.py | 68 +++++++++++++++++++++++++++++++++---------- 1 file changed, 52 insertions(+), 16 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index ecb807262..ebb5ca376 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -2144,6 +2144,9 @@ class InternLM2Model(Model): toktype = SentencePieceTokenTypes.UNUSED elif tokenizer.IsByte(token_id): toktype = SentencePieceTokenTypes.BYTE + # take care of ununsed raw token + if piece.startswith('[UNUSED'): + toktype = SentencePieceTokenTypes.UNKNOWN tokens.append(text) scores.append(score) @@ -2159,6 +2162,47 @@ class InternLM2Model(Model): scores.append(-1000.0) toktypes.append(SentencePieceTokenTypes.USER_DEFINED) + chat_eos_token = '<|im_end|>' + chat_eos_token_id = None + + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {}) + for token_id, foken_data in added_tokens_decoder.items(): + token_id = int(token_id) + token = foken_data["content"] + if token == chat_eos_token: + chat_eos_token_id = token_id + token = token.encode("utf-8") + if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN: + assert(tokens[token_id] == token) + tokens[token_id] = token + scores[token_id] = -1000.0 + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + if foken_data.get("special"): + toktypes[token_id] = SentencePieceTokenTypes.CONTROL + + tokenizer_file = self.dir_model / 'tokenizer.json' + if tokenizer_file.is_file(): + with open(tokenizer_file, "r", encoding="utf-8") as f: + tokenizer_json = json.load(f) + added_tokens = tokenizer_json.get("added_tokens", []) + for foken_data in added_tokens: + token_id = int(foken_data["id"]) + token = foken_data["content"] + if token == chat_eos_token: + chat_eos_token_id = token_id + token = token.encode("utf-8") + if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN: + assert(tokens[token_id] == token) + tokens[token_id] = token + scores[token_id] = -1000.0 + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + if foken_data.get("special"): + toktypes[token_id] = SentencePieceTokenTypes.CONTROL + self.gguf_writer.add_tokenizer_model("llama") self.gguf_writer.add_tokenizer_pre("default") self.gguf_writer.add_token_list(tokens) @@ -2168,28 +2212,16 @@ class InternLM2Model(Model): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) old_eos = special_vocab.special_token_ids["eos"] - if "chat" in os.path.basename(self.dir_model.absolute()): + if chat_eos_token_id is not None: # For the chat model, we replace the eos with '<|im_end|>'. # TODO: this is a hack, should be fixed # https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048 - special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer) - logger.warning(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \ -in chat mode so that the conversation can end normally.") + special_vocab.special_token_ids["eos"] = chat_eos_token_id + logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}" + " in chat mode so that the conversation can end normally.") special_vocab.add_to_gguf(self.gguf_writer) - def _try_get_sft_eos(self, tokenizer): - unused_145_list = tokenizer.Encode('[UNUSED_TOKEN_145]') - im_end_list = tokenizer.Encode('<|im_end|>') - eos_token = None - assert (len(unused_145_list) == 1) ^ (len(im_end_list) == 1) - if len(unused_145_list) == 1: - eos_token = unused_145_list[0] - if len(im_end_list) == 1: - eos_token = im_end_list[0] - assert eos_token - return eos_token - def _hf_permute_qk(self, weights, n_head: int, n_head_kv: int): if n_head_kv is not None and n_head != n_head_kv: n_head = n_head_kv @@ -2208,6 +2240,10 @@ in chat mode so that the conversation can end normally.") 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) + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "linear": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: num_heads = self.hparams["num_attention_heads"] From a8be1e6f5995bec3b1d8474d48ce3a139553a8e1 Mon Sep 17 00:00:00 2001 From: fairydreaming <166155368+fairydreaming@users.noreply.github.com> Date: Wed, 10 Jul 2024 13:38:58 +0200 Subject: [PATCH 24/31] llama : add assert about missing llama_encode() call (#8400) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-authored-by: Stanisław Szymczyk --- src/llama.cpp | 2 ++ 1 file changed, 2 insertions(+) diff --git a/src/llama.cpp b/src/llama.cpp index 2b9ace285..80cc1da26 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -13200,6 +13200,8 @@ struct llm_build_context { LLM_NORM_RMS, cb, -1); cb(cur, "result_norm", -1); } else { + GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first"); + struct ggml_tensor * embd_enc = llm_build_inp_embd_enc(); struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true); From 7a80710d93ee0f4e0d335d65984ae747e62c0337 Mon Sep 17 00:00:00 2001 From: Borislav Stanimirov Date: Wed, 10 Jul 2024 14:40:53 +0300 Subject: [PATCH 25/31] msvc : silence codecvt c++17 deprecation warnings (#8395) --- common/common.cpp | 4 ++++ src/unicode.cpp | 4 ++++ 2 files changed, 8 insertions(+) diff --git a/common/common.cpp b/common/common.cpp index fc0f3b350..1e5fc30dd 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1,3 +1,7 @@ +#if defined(_MSC_VER) +#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING +#endif + #include "common.h" // Change JSON_ASSERT from assert() to GGML_ASSERT: #define JSON_ASSERT GGML_ASSERT diff --git a/src/unicode.cpp b/src/unicode.cpp index 51daa15af..e05fb9d17 100644 --- a/src/unicode.cpp +++ b/src/unicode.cpp @@ -1,3 +1,7 @@ +#if defined(_MSC_VER) +#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING +#endif + #include "unicode.h" #include "unicode-data.h" From cc61948b1f97165b46990f4c2d9699bf9bb64443 Mon Sep 17 00:00:00 2001 From: Borislav Stanimirov Date: Wed, 10 Jul 2024 14:45:44 +0300 Subject: [PATCH 26/31] llama : C++20 compatibility for u8 strings (#8408) --- src/llama.cpp | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/src/llama.cpp b/src/llama.cpp index 80cc1da26..98a242081 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -57,6 +57,12 @@ #include #endif +#if __cplusplus >= 202000L + #define LU8(x) (const char*)(u8##x) +#else + #define LU8(x) u8##x +#endif + #include #include #include @@ -21511,12 +21517,12 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>"; } - } else if (tmpl == "minicpm" || tmpl_contains(u8"<用户>")) { + } else if (tmpl == "minicpm" || tmpl_contains(LU8("<用户>"))) { // MiniCPM-3B-OpenHermes-2.5-v2-GGUF for (auto message : chat) { std::string role(message->role); if (role == "user") { - ss << u8"<用户>"; + ss << LU8("<用户>"); ss << trim(message->content); ss << ""; } else { @@ -21532,7 +21538,7 @@ static int32_t llama_chat_apply_template_internal( } else if (role == "user") { ss << "User: " << message->content << "\n\n"; } else if (role == "assistant") { - ss << "Assistant: " << message->content << u8"<｜end▁of▁sentence｜>"; + ss << "Assistant: " << message->content << LU8("<｜end▁of▁sentence｜>"); } } if (add_ass) { From 83321c6958acf20747d288031174cc1bf8816e33 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=2E=20Yusuf=20Sar=C4=B1g=C3=B6z?= Date: Wed, 10 Jul 2024 15:12:35 +0300 Subject: [PATCH 27/31] gguf-py rel pipeline (#8410) * Upd gguf-py/readme * Bump patch version for release --- gguf-py/README.md | 1 - gguf-py/pyproject.toml | 2 +- 2 files changed, 1 insertion(+), 2 deletions(-) diff --git a/gguf-py/README.md b/gguf-py/README.md index bc46d6e1d..9dd888f31 100644 --- a/gguf-py/README.md +++ b/gguf-py/README.md @@ -79,5 +79,4 @@ python -m twine upload dist/* ``` ## TODO -- [ ] Add tests - [ ] Include conversion scripts as command line entry points in this package. diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index 36e63ee3b..62129126b 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.9.0" +version = "0.9.1" description = "Read and write ML models in GGUF for GGML" authors = ["GGML "] packages = [ From 0f1a39f3439825acf7e3a1663566d410be152170 Mon Sep 17 00:00:00 2001 From: Dibakar Gope Date: Wed, 10 Jul 2024 07:14:51 -0500 Subject: [PATCH 28/31] ggml : add AArch64 optimized GEMV and GEMM Q4 kernels (#5780) * Arm AArch64: optimized GEMV and GEMM kernels for q4_0_q8_0, and q8_0_q8_0 quantization * Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions * Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions * Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions * Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions * Arm AArch64: add copyright claim only to ggml-aarch64.cpp and ggml-aarch64.h files * Arm AArch64: minor code refactoring for rebase * Arm AArch64: minor code refactoring for resolving a build issue with cmake * Arm AArch64: minor code refactoring to split the Q4_0_AARC64 type into three separate types: Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8 * Arm AArch64: minor code change for resolving a build issue with server-windows * retrigger checks * Arm AArch64: minor code changes for rebase * Arm AArch64: minor changes to skip the pr#7433 vec_dot code for arm cpus with SVE VL not equal to 256 bits * Arm AArch64: remove stale LLAMA_QKK_64 from CMakeLists.txt and delete build.zig * Arm AArch64: add reference scalar gemm and gemv, and avoid dynamic memory allocations during quantization for Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8 * Arm AArch64: add multithreaded quantization support for the new types: Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8 * Arm AArch64: minor code refactoring * Arm AArch64: simplify logic for calling gemm and gemv functions in ggml_compute_forward_mul_mat * Arm AArch64: minimize changes in ggml_compute_forward_mul_mat * Arm AArch64: minor code refactoring, and add reference scalar code to quantize routines for new quant types * Arm AArch64: minor code refactoring * Arm AArch64: minor code refactoring * Arm AArch64: minor code refactoring * rebase on the latest master commit 3fd62a6 and adapt to the new directory structure * Arm AArch64: remove a redundant comment * Arm AArch64: add pragma in ggml-aarch64.c to turn -Woverlength-strings warning off * Arm AArch64: use __aarch64__ check to guard 64-bit neon kernels * Arm AArch64: update docs/build.md README to include compile time flags for buiilding the Q4_0_4_4 quant type --- Makefile | 10 +- Package.swift | 1 + docs/build.md | 2 + examples/quantize/quantize.cpp | 3 + ggml/include/ggml.h | 17 + ggml/src/CMakeLists.txt | 1 + ggml/src/ggml-aarch64.c | 2187 ++++++++++++++++++++++++++++++++ ggml/src/ggml-aarch64.h | 39 + ggml/src/ggml-common.h | 24 + ggml/src/ggml-impl.h | 4 + ggml/src/ggml-quants.c | 122 +- ggml/src/ggml.c | 150 ++- include/llama.h | 3 + src/llama.cpp | 24 +- 14 files changed, 2534 insertions(+), 53 deletions(-) create mode 100644 ggml/src/ggml-aarch64.c create mode 100644 ggml/src/ggml-aarch64.h diff --git a/Makefile b/Makefile index 104120090..b70ebaed5 100644 --- a/Makefile +++ b/Makefile @@ -835,7 +835,8 @@ OBJ_GGML += \ ggml/src/ggml.o \ ggml/src/ggml-alloc.o \ ggml/src/ggml-backend.o \ - ggml/src/ggml-quants.o + ggml/src/ggml-quants.o \ + ggml/src/ggml-aarch64.o OBJ_LLAMA = \ src/llama.o \ @@ -969,6 +970,13 @@ ggml/src/ggml-quants.o: \ ggml/src/ggml-common.h $(CC) $(CFLAGS) -c $< -o $@ +ggml/src/ggml-aarch64.o: \ + ggml/src/ggml-aarch64.c \ + ggml/include/ggml.h \ + ggml/src/ggml-aarch64.h \ + ggml/src/ggml-common.h + $(CC) $(CFLAGS) -c $< -o $@ + ggml/src/ggml-blas.o: \ ggml/src/ggml-blas.cpp \ ggml/include/ggml-blas.h diff --git a/Package.swift b/Package.swift index 77fed86df..d40a48385 100644 --- a/Package.swift +++ b/Package.swift @@ -10,6 +10,7 @@ var sources = [ "ggml/src/ggml-alloc.c", "ggml/src/ggml-backend.c", "ggml/src/ggml-quants.c", + "ggml/src/ggml-aarch64.c", ] var resources: [Resource] = [] diff --git a/docs/build.md b/docs/build.md index bf41bfdf9..d70f72f4c 100644 --- a/docs/build.md +++ b/docs/build.md @@ -28,6 +28,7 @@ In order to build llama.cpp you have four different options. ``` - Notes: + - For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`. - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel. - For faster repeated compilation, install [ccache](https://ccache.dev/). - For debug builds, run `make LLAMA_DEBUG=1` @@ -41,6 +42,7 @@ In order to build llama.cpp you have four different options. **Notes**: + - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`. - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel. - For faster repeated compilation, install [ccache](https://ccache.dev/). - For debug builds, there are two cases: diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 76e2052d5..1578c4afb 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -46,6 +46,9 @@ static const std::vector QUANT_OPTIONS = { { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", }, { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", }, { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", }, + { "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, { "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", }, { "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", }, { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index d895c9acd..1e3677537 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -383,6 +383,9 @@ extern "C" { GGML_TYPE_F64 = 28, GGML_TYPE_IQ1_M = 29, GGML_TYPE_BF16 = 30, + GGML_TYPE_Q4_0_4_4 = 31, + GGML_TYPE_Q4_0_4_8 = 32, + GGML_TYPE_Q4_0_8_8 = 33, GGML_TYPE_COUNT, }; @@ -424,6 +427,9 @@ extern "C" { GGML_FTYPE_MOSTLY_IQ4_XS = 22, // except 1d tensors GGML_FTYPE_MOSTLY_IQ1_M = 23, // except 1d tensors GGML_FTYPE_MOSTLY_BF16 = 24, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors }; // available tensor operations: @@ -2406,6 +2412,12 @@ extern "C" { typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx, const void * GGML_RESTRICT y, size_t by, int nrc); + typedef void (*ggml_from_float_to_mat_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, + int64_t k, int64_t bx); + typedef void (*ggml_gemv_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, + const void * GGML_RESTRICT y, int nr, int nc); + typedef void (*ggml_gemm_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, + const void * GGML_RESTRICT y, int nr, int nc); typedef struct { const char * type_name; @@ -2418,6 +2430,11 @@ extern "C" { ggml_vec_dot_t vec_dot; enum ggml_type vec_dot_type; int64_t nrows; // number of rows to process simultaneously; + int64_t ncols; // number of columns to process simultaneously; + int64_t interleave_blcksize; // interleave elements in blocks of interleave_blcksize; + ggml_from_float_to_mat_t from_float_to_mat; + ggml_gemv_t gemv; + ggml_gemm_t gemm; } ggml_type_traits_t; GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type); diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index c6694df67..aae5b8e9f 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -1153,6 +1153,7 @@ add_library(ggml ${GGML_SOURCES_ROCM} ${GGML_HEADERS_ROCM} ${GGML_SOURCES_BLAS} ${GGML_HEADERS_BLAS} ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE} + ggml-aarch64.c ggml-aarch64.h ) if (EMSCRIPTEN) diff --git a/ggml/src/ggml-aarch64.c b/ggml/src/ggml-aarch64.c new file mode 100644 index 000000000..008718634 --- /dev/null +++ b/ggml/src/ggml-aarch64.c @@ -0,0 +1,2187 @@ +// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd. +#define GGML_COMMON_IMPL_C +#include "ggml-common.h" + +#include "ggml-quants.h" +#include "ggml-impl.h" + +#include +#include +#include +#include +#include // for qsort +#include // for GGML_ASSERT + +#include "ggml-aarch64.h" + +#pragma GCC diagnostic ignored "-Woverlength-strings" + +#define UNUSED GGML_UNUSED + +// Functions to create the interleaved data layout formats + +// interleave 4 block_q4_0s in blocks of interleave_blcksize +// returns an interleaved block_q4_0x4 +// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks +// first, then interleave quants from 4 block_q4_0s in blocks of interleave_blcksize +// +// - in : an array of block_q4_0 pointers +// - interleave_blcksize : the block_q4_0 quants bytes are interleaved in blocks of +// interleave_blcksize bytes +// - xor_mask : the mask to convert the nibbles in block_q4_0 quants bytes +// from bias offset form to pure sign form (this saves subtract +// operations durin unpacking) +// +static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int interleave_blcksize, unsigned int xor_mask) { + block_q4_0x4 out; + + for (int i = 0; i < 4; i++) { + out.d[i] = in[i].d; + } + + for (int i = 0; i < QK4_0 * 2; i++) { + int src_offset = (i / (4 * interleave_blcksize)) * interleave_blcksize; + int src_id = (i % (4 * interleave_blcksize)) / interleave_blcksize; + src_offset += (i % interleave_blcksize); + + out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask; + } + + return out; +} + +// interleave 8 block_q4_0s in blocks of interleave_blcksize +// returns an interleaved block_q4_0x8 +// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks +// first, then interleave quants from 8 block_q4_0s in blocks of interleave_blcksize +static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int interleave_blcksize, unsigned int xor_mask) { + block_q4_0x8 out; + + for (int i = 0; i < 8; i++) { + out.d[i] = in[i].d; + } + + for (int i = 0; i < QK4_0 * 4; i++) { + int src_offset = (i / (8 * interleave_blcksize)) * interleave_blcksize; + int src_id = (i % (8 * interleave_blcksize)) / interleave_blcksize; + src_offset += (i % interleave_blcksize); + + out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask; + } + + return out; +} + +void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) { + assert(QK8_0 == 32); + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + block_q8_0x4 * restrict y = (block_q8_0x4 *) vy; + +#if defined(__ARM_NEON) + float32x4_t srcv[4][8]; + float id[4]; + + for (int i = 0; i < nb; i++) { + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int row_iter = 0; row_iter < 4; row_iter++) { + for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]); + + for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]); + for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]); + for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + id[row_iter] = d ? 1.0f / d : 0.0f; + + y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + } + + for (int j = 0; j < 8; j++) { + float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]); + int32x4_t vi = vcvtnq_s32_f32(v); + y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3); + + v = vmulq_n_f32(srcv[1][j], id[1]); + vi = vcvtnq_s32_f32(v); + y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0); + y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1); + y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2); + y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3); + + v = vmulq_n_f32(srcv[2][j], id[2]); + vi = vcvtnq_s32_f32(v); + y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0); + y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1); + y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2); + y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3); + + v = vmulq_n_f32(srcv[3][j], id[3]); + vi = vcvtnq_s32_f32(v); + y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0); + y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1); + y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2); + y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3); + } + } +#else + // scalar + const int interleave_blcksize = 4; + float srcv[4][QK8_0]; + float id[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; + amax = MAX(amax, fabsf(srcv[row_iter][j])); + } + + const float d = amax / ((1 << 7) - 1); + id[row_iter] = d ? 1.0f / d : 0.0f; + + y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + } + + for (int j = 0; j < QK8_0 * 4; j++) { + int src_offset = (j / (4 * interleave_blcksize)) * interleave_blcksize; + int src_id = (j % (4 * interleave_blcksize)) / interleave_blcksize; + src_offset += (j % interleave_blcksize); + + float x0 = srcv[src_id][src_offset] * id[src_id]; + y[i].qs[j] = roundf(x0);; + } + } +#endif +} + +void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) { + assert(QK8_0 == 32); + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + block_q8_0x4 * restrict y = (block_q8_0x4 *) vy; + +#if defined(__ARM_NEON) + float32x4_t srcv[4][8]; + float id[4]; + + for (int i = 0; i < nb; i++) { + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int row_iter = 0; row_iter < 4; row_iter++) { + for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]); + + for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]); + for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]); + for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + id[row_iter] = d ? 1.0f / d : 0.0f; + + y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + } + + for (int j = 0; j < 4; j++) { + float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]); + int32x4_t vi = vcvtnq_s32_f32(v); + y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3); + v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]); + vi = vcvtnq_s32_f32(v); + y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0); + y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1); + y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2); + y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3); + + v = vmulq_n_f32(srcv[1][2 * j], id[1]); + vi = vcvtnq_s32_f32(v); + y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0); + y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1); + y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2); + y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3); + v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]); + vi = vcvtnq_s32_f32(v); + y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0); + y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1); + y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2); + y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3); + + v = vmulq_n_f32(srcv[2][2 * j], id[2]); + vi = vcvtnq_s32_f32(v); + y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0); + y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1); + y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2); + y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3); + v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]); + vi = vcvtnq_s32_f32(v); + y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0); + y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1); + y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2); + y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3); + + v = vmulq_n_f32(srcv[3][2 * j], id[3]); + vi = vcvtnq_s32_f32(v); + y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0); + y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1); + y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2); + y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3); + v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]); + vi = vcvtnq_s32_f32(v); + y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0); + y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1); + y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2); + y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3); + } + } +#else + // scalar + const int interleave_blcksize = 8; + float srcv[4][QK8_0]; + float id[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; + amax = MAX(amax, fabsf(srcv[row_iter][j])); + } + + const float d = amax / ((1 << 7) - 1); + id[row_iter] = d ? 1.0f / d : 0.0f; + + y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + } + + for (int j = 0; j < QK8_0 * 4; j++) { + int src_offset = (j / (4 * interleave_blcksize)) * interleave_blcksize; + int src_id = (j % (4 * interleave_blcksize)) / interleave_blcksize; + src_offset += (j % interleave_blcksize); + + float x0 = srcv[src_id][src_offset] * id[src_id]; + y[i].qs[j] = roundf(x0);; + } + } +#endif +} + +void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t interleave_blcksize) { + assert(nrow == 4); + UNUSED(nrow); + if (interleave_blcksize == 4) quantize_q8_0_4x4(x, vy, n_per_row); + else if (interleave_blcksize == 8) quantize_q8_0_4x8(x, vy, n_per_row); + else assert(false); +} + +static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int interleave_blcksize) { + assert(n_per_row % QK4_0 == 0); + const int nb = n_per_row / QK4_0; + + void * out_ptr = NULL; + if (nrows_interleaved == 8) { + out_ptr = (block_q4_0x8 *) dst; + } + else if (nrows_interleaved == 4) { + out_ptr = (block_q4_0x4 *) dst; + } + assert(nrows_interleaved <= 8); + block_q4_0 dst_tmp[8]; + + for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) { + + for (int64_t x = 0; x < nb; x++) { + + for (int i = 0; i < nrows_interleaved; i++ ) { + quantize_row_q4_0_reference(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0); + } + + if (nrows_interleaved == 8) { + *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, interleave_blcksize, 0x88); + out_ptr = (block_q4_0x8 *) out_ptr + 1; + } + else if (nrows_interleaved == 4) { + *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, interleave_blcksize, 0x88); + out_ptr = (block_q4_0x4 *) out_ptr + 1; + } + } + } + + return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0)); +} + +size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + if (!quant_weights) { + return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4); + } + else { + assert(false); + return 0; + } +} + +size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + if (!quant_weights) { + return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8); + } + else { + assert(false); + return 0; + } +} + +size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + if (!quant_weights) { + return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8); + } + else { + assert(false); + return 0; + } +} + +void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__ARM_FEATURE_SVE) + if (svcntw() == 8) { + GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) && + "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); + } +#endif +#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) && + "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance"); +#elif defined(__ARM_NEON) && defined(__aarch64__) + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + + __asm__ __volatile__( + "movi v31.16b, #0x4\n" + "movi v30.16b, #0xf0\n" + "add %x[b_ptr], %x[b_ptr], #0x8\n" + "1:" // Column loop + "add x22, %x[a_ptr], #0x2\n" + "movi v29.16b, #0x0\n" + "mov x21, %x[nb]\n" + "2:" // Block loop + "ldr q28, [%x[b_ptr], #0x0]\n" + "ldr q27, [x22, #0x0]\n" + "movi v26.4s, #0x0\n" + "sub x20, x22, #0x2\n" + "ldr q25, [x22, #0x10]\n" + "ldr q24, [%x[b_ptr], #0x10]\n" + "sub x21, x21, #0x1\n" + "add x22, x22, #0x22\n" + "ldr q23, [%x[b_ptr], #0x20]\n" + "ldr q22, [%x[b_ptr], #0x30]\n" + "ld1r { v21.8h }, [x20]\n" + "ldr q20, [%x[b_ptr], #-0x8]\n" + "sshl v16.16b, v28.16b, v31.16b\n" + "and v28.16b, v28.16b, v30.16b\n" + "sshl v19.16b, v24.16b, v31.16b\n" + "and v24.16b, v24.16b, v30.16b\n" + "add %x[b_ptr], %x[b_ptr], #0x48\n" + "sshl v18.16b, v23.16b, v31.16b\n" + "and v23.16b, v23.16b, v30.16b\n" + ".inst 0x4f9be21a // sdot v26.4s, v16.16b, v27.4b[0]\n" + "sshl v17.16b, v22.16b, v31.16b\n" + "and v22.16b, v22.16b, v30.16b\n" + "fcvtl v21.4s, v21.4h\n" + "fcvtl v16.4s, v20.4h\n" + ".inst 0x4f99e39a // sdot v26.4s, v28.16b, v25.4b[0]\n" + "fmul v16.4s, v16.4s, v21.4s\n" + ".inst 0x4fbbe27a // sdot v26.4s, v19.16b, v27.4b[1]\n" + ".inst 0x4fb9e31a // sdot v26.4s, v24.16b, v25.4b[1]\n" + ".inst 0x4f9bea5a // sdot v26.4s, v18.16b, v27.4b[2]\n" + ".inst 0x4f99eafa // sdot v26.4s, v23.16b, v25.4b[2]\n" + ".inst 0x4fbbea3a // sdot v26.4s, v17.16b, v27.4b[3]\n" + ".inst 0x4fb9eada // sdot v26.4s, v22.16b, v25.4b[3]\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "fmla v29.4s, v26.4s, v16.4s\n" + "cbnz x21, 2b\n" + "sub %x[nc], %x[nc], #0x4\n" + "str q29, [%x[res_ptr], #0x0]\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "cbnz %x[nc], 1b\n" + : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) + : [a_ptr] "r" (a_ptr), [nb] "r" (nb) + : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22" + ); +#else + float sumf[4]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +#endif +} + +void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 8; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__ARM_FEATURE_SVE) + if (svcntw() == 8) { + GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) && + "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); + } +#endif +#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + + __asm__ __volatile__( + "movi v2.16b, #0x4\n" + "movi v1.16b, #0xf0\n" + "add %x[b_ptr], %x[b_ptr], #0x8\n" + "1:" // Column loop + "add x23, %x[a_ptr], #0x2\n" + "movi v0.16b, #0x0\n" + "mov x22, %x[nb]\n" + "2:" // Block loop + "ldr q31, [%x[b_ptr], #0x0]\n" + "ldr q30, [%x[b_ptr], #0x10]\n" + "mov x21, x23\n" + "movi v29.4s, #0x0\n" + "ldr q28, [%x[b_ptr], #0x20]\n" + "ldr q27, [%x[b_ptr], #0x30]\n" + "movi v26.4s, #0x0\n" + "sub x20, x23, #0x2\n" + "ld1r { v25.8h }, [x20]\n" + "ldr q24, [%x[b_ptr], #-0x8]\n" + "sub x22, x22, #0x1\n" + "add x23, x23, #0x22\n" + "ld1r { v23.2d }, [x21], #0x8\n" + "sshl v22.16b, v31.16b, v2.16b\n" + "sshl v16.16b, v30.16b, v2.16b\n" + "add %x[b_ptr], %x[b_ptr], #0x48\n" + "ld1r { v21.2d }, [x21], #0x8\n" + "sshl v20.16b, v28.16b, v2.16b\n" + "sshl v19.16b, v27.16b, v2.16b\n" + "ld1r { v18.2d }, [x21], #0x8\n" + "ld1r { v17.2d }, [x21], #0x8\n" + "and v31.16b, v31.16b, v1.16b\n" + "and v30.16b, v30.16b, v1.16b\n" + ".inst 0x4e9796dd // sdot v29.4s, v22.16b, v23.16b\n" + ".inst 0x4e97961a // sdot v26.4s, v16.16b, v23.16b\n" + "and v28.16b, v28.16b, v1.16b\n" + "and v27.16b, v27.16b, v1.16b\n" + "fcvtl v25.4s, v25.4h\n" + "fcvtl v16.4s, v24.4h\n" + ".inst 0x4e95969d // sdot v29.4s, v20.16b, v21.16b\n" + ".inst 0x4e95967a // sdot v26.4s, v19.16b, v21.16b\n" + "fmul v16.4s, v16.4s, v25.4s\n" + ".inst 0x4e9297fd // sdot v29.4s, v31.16b, v18.16b\n" + ".inst 0x4e9297da // sdot v26.4s, v30.16b, v18.16b\n" + ".inst 0x4e91979d // sdot v29.4s, v28.16b, v17.16b\n" + ".inst 0x4e91977a // sdot v26.4s, v27.16b, v17.16b\n" + "addp v29.4s, v29.4s, v26.4s\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "fmla v0.4s, v29.4s, v16.4s\n" + "cbnz x22, 2b\n" + "sub %x[nc], %x[nc], #0x4\n" + "str q0, [%x[res_ptr], #0x0]\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "cbnz %x[nc], 1b\n" + : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) + : [a_ptr] "r" (a_ptr), [nb] "r" (nb) + : "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23" + ); +#elif defined(__ARM_NEON) && defined(__aarch64__) + GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && + "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " + "performance"); +#else + float sumf[4]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +#endif +} + +void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__ARM_FEATURE_SVE) + if (svcntw() == 8) { + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + + __asm__ __volatile__( + "ptrue p0.b\n" + "add %x[b_ptr], %x[b_ptr], #0x10\n" + "1:" // Column loop + "add x22, %x[a_ptr], #0x2\n" + "mov z31.b, #0x0\n" + "mov x21, %x[nb]\n" + "2:" // Block loop + "ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n" + "ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n" + "mov z28.s, #0x0\n" + "mov z27.s, #0x0\n" + "ld1rd { z26.d }, p0/Z, [x22]\n" + "ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n" + "sub x20, x22, #0x2\n" + "sub x21, x21, #0x1\n" + "ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n" + "ld1rd { z23.d }, p0/Z, [x22, #8]\n" + "lsl z22.b, z30.b, #0x4\n" + "lsl z16.b, z29.b, #0x4\n" + "and z30.b, z30.b, #0xf0\n" + "and z29.b, z29.b, #0xf0\n" + "ld1rd { z21.d }, p0/Z, [x22, #16]\n" + "ld1rd { z20.d }, p0/Z, [x22, #24]\n" + "lsl z19.b, z25.b, #0x4\n" + "and z25.b, z25.b, #0xf0\n" + "ld1rh { z17.h }, p0/Z, [x20]\n" + "ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n" + "sdot z28.s, z22.b, z26.b\n" + "sdot z27.s, z16.b, z26.b\n" + "lsl z16.b, z24.b, #0x4\n" + "add x22, x22, #0x22\n" + "and z24.b, z24.b, #0xf0\n" + "add %x[b_ptr], %x[b_ptr], #0x90\n" + "fcvt z17.s, p0/m, z17.h\n" + "fcvt z18.s, p0/m, z18.h\n" + "sdot z28.s, z19.b, z23.b\n" + "sdot z27.s, z16.b, z23.b\n" + "fmul z18.s, z18.s, z17.s\n" + "sdot z28.s, z30.b, z21.b\n" + "sdot z27.s, z29.b, z21.b\n" + "sdot z28.s, z25.b, z20.b\n" + "sdot z27.s, z24.b, z20.b\n" + "uzp1 z17.s, z28.s, z27.s\n" + "uzp2 z16.s, z28.s, z27.s\n" + "add z17.s, z17.s, z16.s\n" + "asr z17.s, z17.s, #0x4\n" + "scvtf z17.s, p0/m, z17.s\n" + "fmla z31.s, p0/M, z17.s, z18.s\n" + "cbnz x21, 2b\n" + "sub %x[nc], %x[nc], #0x8\n" + "st1w { z31.s }, p0, [%x[res_ptr]]\n" + "add %x[res_ptr], %x[res_ptr], #0x20\n" + "cbnz %x[nc], 1b\n" + : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) + : [a_ptr] "r" (a_ptr), [nb] "r" (nb) + : "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + ); + return; + } + else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { + GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) && + "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal " + "performance"); + } + else if (ggml_cpu_has_neon()) { + GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) && + "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 " + "quantization format for optimal performance"); + } +#endif +#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + GGML_ASSERT(ggml_cpu_has_sve() && + "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance"); +#elif defined(__ARM_NEON) && defined(__aarch64__) + GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && + "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " + "performance"); +#else + float sumf[8]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +#endif +} + +void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) + if (svcntw() == 8) { + GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) && + "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); + } +#endif +#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) && + "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance"); +#elif defined(__ARM_NEON) && defined(__aarch64__) + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + size_t res_stride = bs * sizeof(float); + + __asm__ __volatile__( + "mov x10, %x[nr]\n" + "mov x9, #0x88\n" + "cmp x10, #0x10\n" + "mul x9, %x[nb], x9\n" + "blt 4f\n" + "1:" // Row loop + "add x28, %x[b_ptr], #0x8\n" + "mov x27, %x[nc]\n" + "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" + "2:" // Column loop + "add x25, %x[a_ptr], #0x8\n" + "movi v15.16b, #0x0\n" + "movi v19.16b, #0x0\n" + "mov x24, %x[nb]\n" + "add x23, x25, x9\n" + "movi v18.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "add x22, x23, x9\n" + "movi v11.16b, #0x0\n" + "movi v13.16b, #0x0\n" + "add x21, x22, x9\n" + "movi v23.16b, #0x0\n" + "movi v16.16b, #0x0\n" + "movi v25.16b, #0x0\n" + "movi v7.16b, #0x0\n" + "movi v0.16b, #0x0\n" + "movi v4.16b, #0x0\n" + "movi v5.16b, #0x0\n" + "movi v21.16b, #0x0\n" + "movi v8.16b, #0x0\n" + "movi v1.16b, #0x0\n" + "3:" // Block loop + "ldr q3, [x28, #0x0]\n" + "ldr q31, [x25, #0x0]\n" + "movi v28.16b, #0x4\n" + "movi v10.4s, #0x0\n" + "ldr q22, [x28, #0x10]\n" + "ldr q6, [x25, #0x10]\n" + "movi v29.4s, #0x0\n" + "movi v9.4s, #0x0\n" + "ldr q27, [x28, #0x20]\n" + "ldr q30, [x28, #0x30]\n" + "movi v20.4s, #0x0\n" + "movi v24.16b, #0xf0\n" + "ldr d2, [x25, #-0x8]\n" + "ldr d26, [x23, #-0x8]\n" + "sshl v12.16b, v3.16b, v28.16b\n" + "sub x20, x28, #0x8\n" + "ldr d17, [x20, #0x0]\n" + "and v3.16b, v3.16b, v24.16b\n" + "subs x24, x24, #0x1\n" + "add x28, x28, #0x48\n" + ".inst 0x4f9fe18a // sdot v10.4s, v12.16b, v31.4b[0]\n" + ".inst 0x4fbfe19d // sdot v29.4s, v12.16b, v31.4b[1]\n" + ".inst 0x4f9fe989 // sdot v9.4s, v12.16b, v31.4b[2]\n" + ".inst 0x4fbfe994 // sdot v20.4s, v12.16b, v31.4b[3]\n" + "sshl v31.16b, v22.16b, v28.16b\n" + "and v22.16b, v22.16b, v24.16b\n" + "fcvtl v17.4s, v17.4h\n" + "fcvtl v2.4s, v2.4h\n" + "fcvtl v26.4s, v26.4h\n" + ".inst 0x4f86e3ea // sdot v10.4s, v31.16b, v6.4b[0]\n" + ".inst 0x4fa6e3fd // sdot v29.4s, v31.16b, v6.4b[1]\n" + ".inst 0x4f86ebe9 // sdot v9.4s, v31.16b, v6.4b[2]\n" + ".inst 0x4fa6ebf4 // sdot v20.4s, v31.16b, v6.4b[3]\n" + "sshl v6.16b, v27.16b, v28.16b\n" + "sshl v28.16b, v30.16b, v28.16b\n" + "and v27.16b, v27.16b, v24.16b\n" + "and v30.16b, v30.16b, v24.16b\n" + "ldr q24, [x25, #0x20]\n" + ".inst 0x4f98e0ca // sdot v10.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8c9 // sdot v9.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8d4 // sdot v20.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x30]\n" + ".inst 0x4f98e38a // sdot v10.4s, v28.16b, v24.4b[0]\n" + ".inst 0x4fb8e39d // sdot v29.4s, v28.16b, v24.4b[1]\n" + ".inst 0x4f98eb89 // sdot v9.4s, v28.16b, v24.4b[2]\n" + ".inst 0x4fb8eb94 // sdot v20.4s, v28.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x40]\n" + ".inst 0x4f98e06a // sdot v10.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e869 // sdot v9.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e874 // sdot v20.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x50]\n" + ".inst 0x4f98e2ca // sdot v10.4s, v22.16b, v24.4b[0]\n" + ".inst 0x4fb8e2dd // sdot v29.4s, v22.16b, v24.4b[1]\n" + ".inst 0x4f98eac9 // sdot v9.4s, v22.16b, v24.4b[2]\n" + ".inst 0x4fb8ead4 // sdot v20.4s, v22.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x60]\n" + ".inst 0x4f98e36a // sdot v10.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb69 // sdot v9.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb74 // sdot v20.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x70]\n" + "add x25, x25, #0x88\n" + ".inst 0x4f98e3ca // sdot v10.4s, v30.16b, v24.4b[0]\n" + ".inst 0x4fb8e3dd // sdot v29.4s, v30.16b, v24.4b[1]\n" + ".inst 0x4f98ebc9 // sdot v9.4s, v30.16b, v24.4b[2]\n" + ".inst 0x4fb8ebd4 // sdot v20.4s, v30.16b, v24.4b[3]\n" + "fmul v24.4s, v17.4s, v2.s[0]\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v15.4s, v10.4s, v24.4s\n" + "ldr q24, [x23, #0x0]\n" + "fmul v10.4s, v17.4s, v2.s[1]\n" + "fmla v19.4s, v29.4s, v10.4s\n" + "ldr q10, [x23, #0x10]\n" + "fmul v29.4s, v17.4s, v2.s[2]\n" + "fmul v2.4s, v17.4s, v2.s[3]\n" + "fmla v18.4s, v9.4s, v29.4s\n" + "movi v9.4s, #0x0\n" + "movi v29.4s, #0x0\n" + ".inst 0x4f98e189 // sdot v9.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e19d // sdot v29.4s, v12.16b, v24.4b[1]\n" + "fmla v14.4s, v20.4s, v2.4s\n" + "movi v20.4s, #0x0\n" + "movi v2.4s, #0x0\n" + ".inst 0x4f98e994 // sdot v20.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x20]\n" + ".inst 0x4f8ae3e9 // sdot v9.4s, v31.16b, v10.4b[0]\n" + ".inst 0x4faae3fd // sdot v29.4s, v31.16b, v10.4b[1]\n" + ".inst 0x4f8aebf4 // sdot v20.4s, v31.16b, v10.4b[2]\n" + ".inst 0x4faaebe2 // sdot v2.4s, v31.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x30]\n" + ".inst 0x4f98e0c9 // sdot v9.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8d4 // sdot v20.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x40]\n" + ".inst 0x4f8ae389 // sdot v9.4s, v28.16b, v10.4b[0]\n" + ".inst 0x4faae39d // sdot v29.4s, v28.16b, v10.4b[1]\n" + ".inst 0x4f8aeb94 // sdot v20.4s, v28.16b, v10.4b[2]\n" + ".inst 0x4faaeb82 // sdot v2.4s, v28.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x50]\n" + ".inst 0x4f98e069 // sdot v9.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e874 // sdot v20.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x60]\n" + ".inst 0x4f8ae2c9 // sdot v9.4s, v22.16b, v10.4b[0]\n" + ".inst 0x4faae2dd // sdot v29.4s, v22.16b, v10.4b[1]\n" + ".inst 0x4f8aead4 // sdot v20.4s, v22.16b, v10.4b[2]\n" + ".inst 0x4faaeac2 // sdot v2.4s, v22.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x70]\n" + "add x23, x23, #0x88\n" + ".inst 0x4f98e369 // sdot v9.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb74 // sdot v20.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x0]\n" + ".inst 0x4f8ae3c9 // sdot v9.4s, v30.16b, v10.4b[0]\n" + ".inst 0x4faae3dd // sdot v29.4s, v30.16b, v10.4b[1]\n" + ".inst 0x4f8aebd4 // sdot v20.4s, v30.16b, v10.4b[2]\n" + ".inst 0x4faaebc2 // sdot v2.4s, v30.16b, v10.4b[3]\n" + "fmul v10.4s, v17.4s, v26.s[0]\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "fmla v11.4s, v9.4s, v10.4s\n" + "ldr q9, [x22, #0x10]\n" + "fmul v10.4s, v17.4s, v26.s[1]\n" + "fmla v13.4s, v29.4s, v10.4s\n" + "ldr d29, [x22, #-0x8]\n" + "fmul v10.4s, v17.4s, v26.s[2]\n" + "fmul v26.4s, v17.4s, v26.s[3]\n" + "fcvtl v29.4s, v29.4h\n" + "fmla v23.4s, v20.4s, v10.4s\n" + "movi v20.4s, #0x0\n" + "movi v10.4s, #0x0\n" + "fmla v16.4s, v2.4s, v26.4s\n" + "movi v26.4s, #0x0\n" + "movi v2.4s, #0x0\n" + ".inst 0x4f98e194 // sdot v20.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" + ".inst 0x4f98e99a // sdot v26.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x20]\n" + ".inst 0x4f89e3f4 // sdot v20.4s, v31.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" + ".inst 0x4f89ebfa // sdot v26.4s, v31.16b, v9.4b[2]\n" + ".inst 0x4fa9ebe2 // sdot v2.4s, v31.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x30]\n" + ".inst 0x4f98e0d4 // sdot v20.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0ca // sdot v10.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8da // sdot v26.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x40]\n" + ".inst 0x4f89e394 // sdot v20.4s, v28.16b, v9.4b[0]\n" + ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" + ".inst 0x4f89eb9a // sdot v26.4s, v28.16b, v9.4b[2]\n" + ".inst 0x4fa9eb82 // sdot v2.4s, v28.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x50]\n" + ".inst 0x4f98e074 // sdot v20.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e06a // sdot v10.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e87a // sdot v26.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x60]\n" + ".inst 0x4f89e2d4 // sdot v20.4s, v22.16b, v9.4b[0]\n" + ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" + ".inst 0x4f89eada // sdot v26.4s, v22.16b, v9.4b[2]\n" + ".inst 0x4fa9eac2 // sdot v2.4s, v22.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x70]\n" + "add x22, x22, #0x88\n" + ".inst 0x4f98e374 // sdot v20.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e36a // sdot v10.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb7a // sdot v26.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x21, #0x0]\n" + ".inst 0x4f89e3d4 // sdot v20.4s, v30.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ca // sdot v10.4s, v30.16b, v9.4b[1]\n" + ".inst 0x4f89ebda // sdot v26.4s, v30.16b, v9.4b[2]\n" + ".inst 0x4fa9ebc2 // sdot v2.4s, v30.16b, v9.4b[3]\n" + "fmul v9.4s, v17.4s, v29.s[0]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "fmla v25.4s, v20.4s, v9.4s\n" + "ldr q9, [x21, #0x10]\n" + "fmul v20.4s, v17.4s, v29.s[1]\n" + "fmla v7.4s, v10.4s, v20.4s\n" + "ldr d20, [x21, #-0x8]\n" + "fmul v10.4s, v17.4s, v29.s[2]\n" + "fmul v29.4s, v17.4s, v29.s[3]\n" + "fcvtl v20.4s, v20.4h\n" + "fmla v0.4s, v26.4s, v10.4s\n" + "movi v26.4s, #0x0\n" + "movi v10.4s, #0x0\n" + "fmla v4.4s, v2.4s, v29.4s\n" + "movi v2.4s, #0x0\n" + "movi v29.4s, #0x0\n" + ".inst 0x4f98e19a // sdot v26.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" + ".inst 0x4f98e982 // sdot v2.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e99d // sdot v29.4s, v12.16b, v24.4b[3]\n" + "ldr q12, [x21, #0x20]\n" + "fmul v24.4s, v17.4s, v20.s[0]\n" + ".inst 0x4f89e3fa // sdot v26.4s, v31.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" + ".inst 0x4f89ebe2 // sdot v2.4s, v31.16b, v9.4b[2]\n" + ".inst 0x4fa9ebfd // sdot v29.4s, v31.16b, v9.4b[3]\n" + "ldr q9, [x21, #0x30]\n" + "fmul v31.4s, v17.4s, v20.s[1]\n" + ".inst 0x4f8ce0da // sdot v26.4s, v6.16b, v12.4b[0]\n" + ".inst 0x4face0ca // sdot v10.4s, v6.16b, v12.4b[1]\n" + ".inst 0x4f8ce8c2 // sdot v2.4s, v6.16b, v12.4b[2]\n" + ".inst 0x4face8dd // sdot v29.4s, v6.16b, v12.4b[3]\n" + "ldr q12, [x21, #0x40]\n" + "fmul v6.4s, v17.4s, v20.s[2]\n" + "fmul v20.4s, v17.4s, v20.s[3]\n" + ".inst 0x4f89e39a // sdot v26.4s, v28.16b, v9.4b[0]\n" + ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" + ".inst 0x4f89eb82 // sdot v2.4s, v28.16b, v9.4b[2]\n" + ".inst 0x4fa9eb9d // sdot v29.4s, v28.16b, v9.4b[3]\n" + "ldr q9, [x21, #0x50]\n" + ".inst 0x4f8ce07a // sdot v26.4s, v3.16b, v12.4b[0]\n" + ".inst 0x4face06a // sdot v10.4s, v3.16b, v12.4b[1]\n" + ".inst 0x4f8ce862 // sdot v2.4s, v3.16b, v12.4b[2]\n" + ".inst 0x4face87d // sdot v29.4s, v3.16b, v12.4b[3]\n" + "ldr q12, [x21, #0x60]\n" + ".inst 0x4f89e2da // sdot v26.4s, v22.16b, v9.4b[0]\n" + ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" + ".inst 0x4f89eac2 // sdot v2.4s, v22.16b, v9.4b[2]\n" + ".inst 0x4fa9eadd // sdot v29.4s, v22.16b, v9.4b[3]\n" + "ldr q17, [x21, #0x70]\n" + "add x21, x21, #0x88\n" + ".inst 0x4f8ce37a // sdot v26.4s, v27.16b, v12.4b[0]\n" + ".inst 0x4face36a // sdot v10.4s, v27.16b, v12.4b[1]\n" + ".inst 0x4f8ceb62 // sdot v2.4s, v27.16b, v12.4b[2]\n" + ".inst 0x4faceb7d // sdot v29.4s, v27.16b, v12.4b[3]\n" + ".inst 0x4f91e3da // sdot v26.4s, v30.16b, v17.4b[0]\n" + ".inst 0x4fb1e3ca // sdot v10.4s, v30.16b, v17.4b[1]\n" + ".inst 0x4f91ebc2 // sdot v2.4s, v30.16b, v17.4b[2]\n" + ".inst 0x4fb1ebdd // sdot v29.4s, v30.16b, v17.4b[3]\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "fmla v5.4s, v26.4s, v24.4s\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "fmla v21.4s, v10.4s, v31.4s\n" + "fmla v8.4s, v2.4s, v6.4s\n" + "fmla v1.4s, v29.4s, v20.4s\n" + "bgt 3b\n" + "mov x20, %x[res_ptr]\n" + "subs x27, x27, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "str q15, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q19, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q18, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q14, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q11, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q13, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q23, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q16, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q25, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q7, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q0, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q4, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q5, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q21, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q8, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q1, [x20, #0x0]\n" + "bne 2b\n" + "mov x20, #0x4\n" + "sub x10, x10, #0x10\n" + "cmp x10, #0x10\n" + "mov %x[res_ptr], x26\n" + "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" + "bge 1b\n" + "4:" // Row loop skip + "cbz x10, 9f\n" + "5:" // Row tail: Row loop + "add x24, %x[b_ptr], #0x8\n" + "mov x23, %x[nc]\n" + "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" + "6:" // Row tail: Column loop + "movi v15.16b, #0x0\n" + "movi v19.16b, #0x0\n" + "add x25, %x[a_ptr], #0x8\n" + "mov x21, %x[nb]\n" + "movi v18.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "7:" // Row tail: Block loop + "ldr q7, [x24, #0x0]\n" + "ldr q5, [x25, #0x0]\n" + "movi v9.16b, #0x4\n" + "movi v4.4s, #0x0\n" + "ldr q3, [x24, #0x10]\n" + "ldr q2, [x25, #0x10]\n" + "movi v1.4s, #0x0\n" + "movi v0.4s, #0x0\n" + "ldr q13, [x24, #0x20]\n" + "ldr q31, [x25, #0x20]\n" + "movi v30.4s, #0x0\n" + "movi v29.16b, #0xf0\n" + "ldr q28, [x24, #0x30]\n" + "ldr q27, [x25, #0x30]\n" + "sshl v20.16b, v7.16b, v9.16b\n" + "sub x20, x24, #0x8\n" + "ldr q26, [x25, #0x40]\n" + "ldr q25, [x25, #0x50]\n" + "sshl v17.16b, v3.16b, v9.16b\n" + "and v7.16b, v7.16b, v29.16b\n" + "ldr q24, [x25, #0x60]\n" + "ldr q16, [x25, #0x70]\n" + "sshl v22.16b, v13.16b, v9.16b\n" + "and v3.16b, v3.16b, v29.16b\n" + "ldr d21, [x20, #0x0]\n" + "ldr d12, [x25, #-0x8]\n" + ".inst 0x4f85e284 // sdot v4.4s, v20.16b, v5.4b[0]\n" + ".inst 0x4fa5e281 // sdot v1.4s, v20.16b, v5.4b[1]\n" + ".inst 0x4f85ea80 // sdot v0.4s, v20.16b, v5.4b[2]\n" + ".inst 0x4fa5ea9e // sdot v30.4s, v20.16b, v5.4b[3]\n" + "sshl v9.16b, v28.16b, v9.16b\n" + "subs x21, x21, #0x1\n" + "and v13.16b, v13.16b, v29.16b\n" + "and v28.16b, v28.16b, v29.16b\n" + "add x25, x25, #0x88\n" + "add x24, x24, #0x48\n" + "fcvtl v21.4s, v21.4h\n" + "fcvtl v12.4s, v12.4h\n" + ".inst 0x4f82e224 // sdot v4.4s, v17.16b, v2.4b[0]\n" + ".inst 0x4fa2e221 // sdot v1.4s, v17.16b, v2.4b[1]\n" + ".inst 0x4f82ea20 // sdot v0.4s, v17.16b, v2.4b[2]\n" + ".inst 0x4fa2ea3e // sdot v30.4s, v17.16b, v2.4b[3]\n" + "fmul v11.4s, v21.4s, v12.s[0]\n" + "fmul v23.4s, v21.4s, v12.s[1]\n" + "fmul v17.4s, v21.4s, v12.s[2]\n" + ".inst 0x4f9fe2c4 // sdot v4.4s, v22.16b, v31.4b[0]\n" + "fmul v6.4s, v21.4s, v12.s[3]\n" + ".inst 0x4fbfe2c1 // sdot v1.4s, v22.16b, v31.4b[1]\n" + ".inst 0x4f9feac0 // sdot v0.4s, v22.16b, v31.4b[2]\n" + ".inst 0x4fbfeade // sdot v30.4s, v22.16b, v31.4b[3]\n" + ".inst 0x4f9be124 // sdot v4.4s, v9.16b, v27.4b[0]\n" + ".inst 0x4fbbe121 // sdot v1.4s, v9.16b, v27.4b[1]\n" + ".inst 0x4f9be920 // sdot v0.4s, v9.16b, v27.4b[2]\n" + ".inst 0x4fbbe93e // sdot v30.4s, v9.16b, v27.4b[3]\n" + ".inst 0x4f9ae0e4 // sdot v4.4s, v7.16b, v26.4b[0]\n" + ".inst 0x4fbae0e1 // sdot v1.4s, v7.16b, v26.4b[1]\n" + ".inst 0x4f9ae8e0 // sdot v0.4s, v7.16b, v26.4b[2]\n" + ".inst 0x4fbae8fe // sdot v30.4s, v7.16b, v26.4b[3]\n" + ".inst 0x4f99e064 // sdot v4.4s, v3.16b, v25.4b[0]\n" + ".inst 0x4fb9e061 // sdot v1.4s, v3.16b, v25.4b[1]\n" + ".inst 0x4f99e860 // sdot v0.4s, v3.16b, v25.4b[2]\n" + ".inst 0x4fb9e87e // sdot v30.4s, v3.16b, v25.4b[3]\n" + ".inst 0x4f98e1a4 // sdot v4.4s, v13.16b, v24.4b[0]\n" + ".inst 0x4fb8e1a1 // sdot v1.4s, v13.16b, v24.4b[1]\n" + ".inst 0x4f98e9a0 // sdot v0.4s, v13.16b, v24.4b[2]\n" + ".inst 0x4fb8e9be // sdot v30.4s, v13.16b, v24.4b[3]\n" + ".inst 0x4f90e384 // sdot v4.4s, v28.16b, v16.4b[0]\n" + ".inst 0x4fb0e381 // sdot v1.4s, v28.16b, v16.4b[1]\n" + ".inst 0x4f90eb80 // sdot v0.4s, v28.16b, v16.4b[2]\n" + ".inst 0x4fb0eb9e // sdot v30.4s, v28.16b, v16.4b[3]\n" + "scvtf v4.4s, v4.4s, #0x4\n" + "scvtf v1.4s, v1.4s, #0x4\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "fmla v15.4s, v4.4s, v11.4s\n" + "scvtf v30.4s, v30.4s, #0x4\n" + "fmla v19.4s, v1.4s, v23.4s\n" + "fmla v18.4s, v0.4s, v17.4s\n" + "fmla v14.4s, v30.4s, v6.4s\n" + "bgt 7b\n" + "mov x20, %x[res_ptr]\n" + "cmp x10, #0x1\n" + "str q15, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x2\n" + "str q19, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x3\n" + "str q18, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "str q14, [x20, #0x0]\n" + "8:" // Row tail: Accumulator store skip + "subs x23, x23, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "bne 6b\n" + "subs x10, x10, #0x4\n" + "add %x[a_ptr], %x[a_ptr], x9\n" + "mov %x[res_ptr], x22\n" + "bgt 5b\n" + "9:" // Row tail: Row loop skip + : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) + : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + ); +#else + float sumf[4][4]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } +#endif +} + +void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 8; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) + if (svcntw() == 8) { + GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) && + "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); + } +#endif +#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + size_t res_stride = bs * sizeof(float); + + __asm__ __volatile__( + "mov x10, %x[nr]\n" + "mov x9, #0x88\n" + "cmp x10, #0x10\n" + "mul x9, %x[nb], x9\n" + "blt 4f\n" + "1:" // Row loop + "add x28, %x[b_ptr], #0x8\n" + "mov x27, %x[nc]\n" + "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" + "2:" // Column loop + "add x25, %x[a_ptr], #0x8\n" + "movi v2.16b, #0x0\n" + "movi v10.16b, #0x0\n" + "mov x24, %x[nb]\n" + "add x23, x25, x9\n" + "movi v12.16b, #0x0\n" + "movi v28.16b, #0x0\n" + "add x22, x23, x9\n" + "movi v11.16b, #0x0\n" + "movi v13.16b, #0x0\n" + "add x21, x22, x9\n" + "movi v22.16b, #0x0\n" + "movi v23.16b, #0x0\n" + "movi v25.16b, #0x0\n" + "movi v5.16b, #0x0\n" + "movi v7.16b, #0x0\n" + "movi v4.16b, #0x0\n" + "movi v6.16b, #0x0\n" + "movi v30.16b, #0x0\n" + "movi v24.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "3:" // Block loop + "ldr q21, [x28, #0x0]\n" + "ldr q16, [x28, #0x10]\n" + "movi v1.16b, #0x4\n" + "movi v19.4s, #0x0\n" + "ldr q27, [x25, #0x0]\n" + "ldr q15, [x25, #0x10]\n" + "movi v26.4s, #0x0\n" + "movi v18.4s, #0x0\n" + "ldr q29, [x28, #0x20]\n" + "ldr q3, [x28, #0x30]\n" + "movi v17.4s, #0x0\n" + "movi v0.16b, #0xf0\n" + "ldr d20, [x25, #-0x8]\n" + "ldr d9, [x23, #-0x8]\n" + "sshl v8.16b, v21.16b, v1.16b\n" + "sshl v31.16b, v16.16b, v1.16b\n" + "and v21.16b, v21.16b, v0.16b\n" + "and v16.16b, v16.16b, v0.16b\n" + "sub x20, x28, #0x8\n" + "subs x24, x24, #0x1\n" + "add x28, x28, #0x48\n" + ".inst 0x4e88a773 // smmla v19.4s, v27.16b, v8.16b\n" + ".inst 0x4e9fa77a // smmla v26.4s, v27.16b, v31.16b\n" + "ldr q27, [x25, #0x20]\n" + ".inst 0x4e88a5f2 // smmla v18.4s, v15.16b, v8.16b\n" + ".inst 0x4e9fa5f1 // smmla v17.4s, v15.16b, v31.16b\n" + "sshl v15.16b, v29.16b, v1.16b\n" + "sshl v1.16b, v3.16b, v1.16b\n" + "and v29.16b, v29.16b, v0.16b\n" + "and v3.16b, v3.16b, v0.16b\n" + "ldr q0, [x25, #0x30]\n" + "fcvtl v20.4s, v20.4h\n" + ".inst 0x4e8fa773 // smmla v19.4s, v27.16b, v15.16b\n" + "fcvtl v9.4s, v9.4h\n" + ".inst 0x4e81a77a // smmla v26.4s, v27.16b, v1.16b\n" + "ldr q27, [x25, #0x40]\n" + ".inst 0x4e8fa412 // smmla v18.4s, v0.16b, v15.16b\n" + ".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n" + "ldr q0, [x25, #0x50]\n" + ".inst 0x4e95a773 // smmla v19.4s, v27.16b, v21.16b\n" + ".inst 0x4e90a77a // smmla v26.4s, v27.16b, v16.16b\n" + "ldr q27, [x25, #0x60]\n" + ".inst 0x4e95a412 // smmla v18.4s, v0.16b, v21.16b\n" + ".inst 0x4e90a411 // smmla v17.4s, v0.16b, v16.16b\n" + "ldr q0, [x25, #0x70]\n" + "add x25, x25, #0x88\n" + ".inst 0x4e9da773 // smmla v19.4s, v27.16b, v29.16b\n" + ".inst 0x4e83a77a // smmla v26.4s, v27.16b, v3.16b\n" + "ldr d27, [x20, #0x0]\n" + ".inst 0x4e9da412 // smmla v18.4s, v0.16b, v29.16b\n" + ".inst 0x4e83a411 // smmla v17.4s, v0.16b, v3.16b\n" + "fcvtl v27.4s, v27.4h\n" + "uzp1 v0.2d, v19.2d, v26.2d\n" + "uzp2 v26.2d, v19.2d, v26.2d\n" + "fmul v19.4s, v27.4s, v20.s[0]\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "fmla v2.4s, v0.4s, v19.4s\n" + "ldr q19, [x23, #0x0]\n" + "uzp1 v0.2d, v18.2d, v17.2d\n" + "uzp2 v18.2d, v18.2d, v17.2d\n" + "fmul v17.4s, v27.4s, v20.s[1]\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "fmla v10.4s, v26.4s, v17.4s\n" + "ldr q17, [x23, #0x10]\n" + "fmul v26.4s, v27.4s, v20.s[2]\n" + "fmul v20.4s, v27.4s, v20.s[3]\n" + "fmla v12.4s, v0.4s, v26.4s\n" + "ldr d0, [x22, #-0x8]\n" + "ldr d26, [x21, #-0x8]\n" + "fcvtl v0.4s, v0.4h\n" + "fmla v28.4s, v18.4s, v20.4s\n" + "movi v20.4s, #0x0\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" + ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" + "ldr q19, [x23, #0x20]\n" + "fcvtl v26.4s, v26.4h\n" + ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" + ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" + "ldr q19, [x23, #0x40]\n" + ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" + ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" + "ldr q19, [x23, #0x60]\n" + ".inst 0x4e9da674 // smmla v20.4s, v19.16b, v29.16b\n" + ".inst 0x4e83a672 // smmla v18.4s, v19.16b, v3.16b\n" + "uzp1 v19.2d, v20.2d, v18.2d\n" + "scvtf v19.4s, v19.4s, #0x4\n" + "uzp2 v20.2d, v20.2d, v18.2d\n" + "fmul v18.4s, v27.4s, v9.s[0]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v11.4s, v19.4s, v18.4s\n" + "ldr q18, [x22, #0x0]\n" + "fmul v19.4s, v27.4s, v9.s[1]\n" + "fmla v13.4s, v20.4s, v19.4s\n" + "movi v19.4s, #0x0\n" + "movi v20.4s, #0x0\n" + ".inst 0x4e88a633 // smmla v19.4s, v17.16b, v8.16b\n" + ".inst 0x4e9fa634 // smmla v20.4s, v17.16b, v31.16b\n" + "ldr q17, [x23, #0x30]\n" + ".inst 0x4e8fa633 // smmla v19.4s, v17.16b, v15.16b\n" + ".inst 0x4e81a634 // smmla v20.4s, v17.16b, v1.16b\n" + "ldr q17, [x23, #0x50]\n" + ".inst 0x4e95a633 // smmla v19.4s, v17.16b, v21.16b\n" + ".inst 0x4e90a634 // smmla v20.4s, v17.16b, v16.16b\n" + "ldr q17, [x23, #0x70]\n" + "add x23, x23, #0x88\n" + ".inst 0x4e9da633 // smmla v19.4s, v17.16b, v29.16b\n" + ".inst 0x4e83a634 // smmla v20.4s, v17.16b, v3.16b\n" + "uzp1 v17.2d, v19.2d, v20.2d\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "uzp2 v20.2d, v19.2d, v20.2d\n" + "fmul v19.4s, v27.4s, v9.s[2]\n" + "fmul v9.4s, v27.4s, v9.s[3]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v22.4s, v17.4s, v19.4s\n" + "ldr q17, [x22, #0x10]\n" + "movi v19.4s, #0x0\n" + ".inst 0x4e88a653 // smmla v19.4s, v18.16b, v8.16b\n" + "fmla v23.4s, v20.4s, v9.4s\n" + "movi v20.4s, #0x0\n" + "movi v9.4s, #0x0\n" + ".inst 0x4e9fa654 // smmla v20.4s, v18.16b, v31.16b\n" + "ldr q18, [x22, #0x20]\n" + ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" + ".inst 0x4e8fa653 // smmla v19.4s, v18.16b, v15.16b\n" + ".inst 0x4e81a654 // smmla v20.4s, v18.16b, v1.16b\n" + "ldr q18, [x22, #0x40]\n" + ".inst 0x4e95a653 // smmla v19.4s, v18.16b, v21.16b\n" + ".inst 0x4e90a654 // smmla v20.4s, v18.16b, v16.16b\n" + "ldr q18, [x22, #0x60]\n" + ".inst 0x4e9da653 // smmla v19.4s, v18.16b, v29.16b\n" + ".inst 0x4e83a654 // smmla v20.4s, v18.16b, v3.16b\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e9fa632 // smmla v18.4s, v17.16b, v31.16b\n" + "ldr q17, [x22, #0x30]\n" + ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" + ".inst 0x4e81a632 // smmla v18.4s, v17.16b, v1.16b\n" + "ldr q17, [x22, #0x50]\n" + ".inst 0x4e95a629 // smmla v9.4s, v17.16b, v21.16b\n" + ".inst 0x4e90a632 // smmla v18.4s, v17.16b, v16.16b\n" + "ldr q17, [x22, #0x70]\n" + "add x22, x22, #0x88\n" + ".inst 0x4e9da629 // smmla v9.4s, v17.16b, v29.16b\n" + ".inst 0x4e83a632 // smmla v18.4s, v17.16b, v3.16b\n" + "uzp1 v17.2d, v19.2d, v20.2d\n" + "uzp2 v20.2d, v19.2d, v20.2d\n" + "fmul v19.4s, v27.4s, v0.s[0]\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v25.4s, v17.4s, v19.4s\n" + "ldr q19, [x21, #0x0]\n" + "fmul v17.4s, v27.4s, v0.s[1]\n" + "fmla v5.4s, v20.4s, v17.4s\n" + "ldr q17, [x21, #0x10]\n" + "uzp1 v20.2d, v9.2d, v18.2d\n" + "uzp2 v9.2d, v9.2d, v18.2d\n" + "fmul v18.4s, v27.4s, v0.s[2]\n" + "fmul v0.4s, v27.4s, v0.s[3]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "fmla v7.4s, v20.4s, v18.4s\n" + "movi v20.4s, #0x0\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" + ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" + "ldr q19, [x21, #0x20]\n" + "fmla v4.4s, v9.4s, v0.4s\n" + "movi v9.4s, #0x0\n" + "movi v0.4s, #0x0\n" + ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" + "fmul v8.4s, v27.4s, v26.s[0]\n" + ".inst 0x4e9fa620 // smmla v0.4s, v17.16b, v31.16b\n" + "ldr q17, [x21, #0x30]\n" + ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" + "fmul v31.4s, v27.4s, v26.s[1]\n" + ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" + "ldr q19, [x21, #0x40]\n" + ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" + "fmul v15.4s, v27.4s, v26.s[2]\n" + "fmul v27.4s, v27.4s, v26.s[3]\n" + ".inst 0x4e81a620 // smmla v0.4s, v17.16b, v1.16b\n" + "ldr q1, [x21, #0x50]\n" + ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" + ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" + "ldr q26, [x21, #0x60]\n" + ".inst 0x4e95a429 // smmla v9.4s, v1.16b, v21.16b\n" + ".inst 0x4e90a420 // smmla v0.4s, v1.16b, v16.16b\n" + "ldr q21, [x21, #0x70]\n" + "add x21, x21, #0x88\n" + ".inst 0x4e9da754 // smmla v20.4s, v26.16b, v29.16b\n" + ".inst 0x4e83a752 // smmla v18.4s, v26.16b, v3.16b\n" + ".inst 0x4e9da6a9 // smmla v9.4s, v21.16b, v29.16b\n" + ".inst 0x4e83a6a0 // smmla v0.4s, v21.16b, v3.16b\n" + "uzp1 v29.2d, v20.2d, v18.2d\n" + "uzp2 v21.2d, v20.2d, v18.2d\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "uzp1 v18.2d, v9.2d, v0.2d\n" + "uzp2 v16.2d, v9.2d, v0.2d\n" + "scvtf v21.4s, v21.4s, #0x4\n" + "fmla v6.4s, v29.4s, v8.4s\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "scvtf v16.4s, v16.4s, #0x4\n" + "fmla v30.4s, v21.4s, v31.4s\n" + "fmla v24.4s, v18.4s, v15.4s\n" + "fmla v14.4s, v16.4s, v27.4s\n" + "bgt 3b\n" + "mov x20, %x[res_ptr]\n" + "subs x27, x27, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "str q2, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q10, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q12, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q28, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q11, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q13, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q22, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q23, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q25, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q5, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q7, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q4, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q6, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q30, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q24, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q14, [x20, #0x0]\n" + "bne 2b\n" + "mov x20, #0x4\n" + "sub x10, x10, #0x10\n" + "cmp x10, #0x10\n" + "mov %x[res_ptr], x26\n" + "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" + "bge 1b\n" + "4:" // Row loop skip + "cbz x10, 9f\n" + "5:" // Row tail: Row loop + "add x24, %x[b_ptr], #0x8\n" + "mov x23, %x[nc]\n" + "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" + "6:" // Row tail: Column loop + "movi v2.16b, #0x0\n" + "movi v10.16b, #0x0\n" + "add x25, %x[a_ptr], #0x8\n" + "mov x21, %x[nb]\n" + "movi v12.16b, #0x0\n" + "movi v28.16b, #0x0\n" + "7:" // Row tail: Block loop + "ldr q6, [x24, #0x0]\n" + "ldr q5, [x24, #0x10]\n" + "movi v17.16b, #0x4\n" + "movi v8.4s, #0x0\n" + "ldr q4, [x25, #0x0]\n" + "ldr q13, [x25, #0x10]\n" + "movi v27.4s, #0x0\n" + "movi v0.4s, #0x0\n" + "ldr q31, [x24, #0x20]\n" + "ldr q14, [x24, #0x30]\n" + "movi v29.4s, #0x0\n" + "movi v22.16b, #0xf0\n" + "ldr q11, [x25, #0x20]\n" + "ldr q23, [x25, #0x30]\n" + "sshl v21.16b, v6.16b, v17.16b\n" + "sshl v16.16b, v5.16b, v17.16b\n" + "ldr q20, [x25, #0x40]\n" + "ldr q26, [x25, #0x50]\n" + "and v6.16b, v6.16b, v22.16b\n" + "and v5.16b, v5.16b, v22.16b\n" + "ldr q25, [x25, #0x60]\n" + "ldr q3, [x25, #0x70]\n" + "sshl v19.16b, v31.16b, v17.16b\n" + "sshl v18.16b, v14.16b, v17.16b\n" + "ldr d17, [x25, #-0x8]\n" + ".inst 0x4e95a488 // smmla v8.4s, v4.16b, v21.16b\n" + ".inst 0x4e90a49b // smmla v27.4s, v4.16b, v16.16b\n" + "and v31.16b, v31.16b, v22.16b\n" + ".inst 0x4e95a5a0 // smmla v0.4s, v13.16b, v21.16b\n" + ".inst 0x4e90a5bd // smmla v29.4s, v13.16b, v16.16b\n" + "and v14.16b, v14.16b, v22.16b\n" + "sub x20, x24, #0x8\n" + "ldr d16, [x20, #0x0]\n" + "subs x21, x21, #0x1\n" + "add x25, x25, #0x88\n" + "fcvtl v17.4s, v17.4h\n" + "add x24, x24, #0x48\n" + ".inst 0x4e93a568 // smmla v8.4s, v11.16b, v19.16b\n" + ".inst 0x4e92a57b // smmla v27.4s, v11.16b, v18.16b\n" + ".inst 0x4e93a6e0 // smmla v0.4s, v23.16b, v19.16b\n" + ".inst 0x4e92a6fd // smmla v29.4s, v23.16b, v18.16b\n" + "fcvtl v16.4s, v16.4h\n" + ".inst 0x4e86a688 // smmla v8.4s, v20.16b, v6.16b\n" + ".inst 0x4e85a69b // smmla v27.4s, v20.16b, v5.16b\n" + "fmul v23.4s, v16.4s, v17.s[0]\n" + "fmul v21.4s, v16.4s, v17.s[1]\n" + "fmul v1.4s, v16.4s, v17.s[2]\n" + "fmul v20.4s, v16.4s, v17.s[3]\n" + ".inst 0x4e86a740 // smmla v0.4s, v26.16b, v6.16b\n" + ".inst 0x4e85a75d // smmla v29.4s, v26.16b, v5.16b\n" + ".inst 0x4e9fa728 // smmla v8.4s, v25.16b, v31.16b\n" + ".inst 0x4e8ea73b // smmla v27.4s, v25.16b, v14.16b\n" + ".inst 0x4e9fa460 // smmla v0.4s, v3.16b, v31.16b\n" + ".inst 0x4e8ea47d // smmla v29.4s, v3.16b, v14.16b\n" + "uzp1 v19.2d, v8.2d, v27.2d\n" + "uzp2 v18.2d, v8.2d, v27.2d\n" + "scvtf v19.4s, v19.4s, #0x4\n" + "uzp1 v17.2d, v0.2d, v29.2d\n" + "uzp2 v16.2d, v0.2d, v29.2d\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "fmla v2.4s, v19.4s, v23.4s\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "scvtf v16.4s, v16.4s, #0x4\n" + "fmla v10.4s, v18.4s, v21.4s\n" + "fmla v12.4s, v17.4s, v1.4s\n" + "fmla v28.4s, v16.4s, v20.4s\n" + "bgt 7b\n" + "mov x20, %x[res_ptr]\n" + "cmp x10, #0x1\n" + "str q2, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x2\n" + "str q10, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x3\n" + "str q12, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "str q28, [x20, #0x0]\n" + "8:" // Row tail: Accumulator store skip + "subs x23, x23, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "bne 6b\n" + "subs x10, x10, #0x4\n" + "add %x[a_ptr], %x[a_ptr], x9\n" + "mov %x[res_ptr], x22\n" + "bgt 5b\n" + "9:" // Row tail: Row loop skip + : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) + : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + ); +#elif defined(__ARM_NEON) && defined(__aarch64__) + GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && + "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " + "performance"); +#else + float sumf[4][4]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } +#endif +} + +void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) + if (svcntw() == 8) { + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + size_t res_stride = bs * sizeof(float); + + __asm__ __volatile__( + "mov x20, #0x4\n" + "mov x13, %x[nr]\n" + "mov z28.s, #-0x4\n" + "mov x12, #0x88\n" + "ptrue p1.b\n" + "whilelt p0.s, XZR, x20\n" + "cmp x13, #0x10\n" + "mul x12, %x[nb], x12\n" + "blt 4f\n" + "1:" // Row loop + "add x11, %x[b_ptr], #0x10\n" + "mov x10, %x[nc]\n" + "add x9, %x[res_ptr], %x[res_stride], LSL #4\n" + "2:" // Column loop + "add x28, %x[a_ptr], #0x8\n" + "mov z24.b, #0x0\n" + "mov z15.b, #0x0\n" + "mov x27, %x[nb]\n" + "add x26, x28, x12\n" + "mov z12.b, #0x0\n" + "mov z0.b, #0x0\n" + "add x25, x26, x12\n" + "mov z13.b, #0x0\n" + "mov z1.b, #0x0\n" + "add x24, x25, x12\n" + "mov z20.b, #0x0\n" + "mov z25.b, #0x0\n" + "mov z11.b, #0x0\n" + "mov z16.b, #0x0\n" + "mov z19.b, #0x0\n" + "mov z26.b, #0x0\n" + "mov z8.b, #0x0\n" + "mov z29.b, #0x0\n" + "mov z27.b, #0x0\n" + "mov z10.b, #0x0\n" + "3:" // Block loop + "ld1b { z30.b }, p1/Z, [x11]\n" + "ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n" + "mov z18.s, #0x0\n" + "mov z7.s, #0x0\n" + "ld1rqb { z3.b }, p1/Z, [x28]\n" + "ld1rqb { z5.b }, p1/Z, [x28, #16]\n" + "mov z9.s, #0x0\n" + "mov z22.s, #0x0\n" + "ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n" + "ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n" + "sub x20, x11, #0x10\n" + "sub x23, x28, #0x8\n" + "lsl z31.b, z30.b, #0x4\n" + "lsl z6.b, z21.b, #0x4\n" + "ld1h { z23.s }, p1/Z, [x20]\n" + "sub x22, x26, #0x8\n" + "and z30.b, z30.b, #0xf0\n" + "and z21.b, z21.b, #0xf0\n" + "sub x21, x25, #0x8\n" + "sub x20, x24, #0x8\n" + "lsl z14.b, z4.b, #0x4\n" + "lsl z2.b, z17.b, #0x4\n" + "subs x27, x27, #0x1\n" + "add x11, x11, #0x90\n" + ".inst 0x451f9872 // smmla z18.s, z3.b, z31.b\n" + ".inst 0x45069867 // smmla z7.s, z3.b, z6.b\n" + "ld1rqb { z3.b }, p1/Z, [x28, #32]\n" + "and z4.b, z4.b, #0xf0\n" + ".inst 0x451f98a9 // smmla z9.s, z5.b, z31.b\n" + ".inst 0x450698b6 // smmla z22.s, z5.b, z6.b\n" + "ld1rqb { z5.b }, p1/Z, [x28, #48]\n" + "and z17.b, z17.b, #0xf0\n" + "fcvt z23.s, p1/m, z23.h\n" + ".inst 0x450e9872 // smmla z18.s, z3.b, z14.b\n" + ".inst 0x45029867 // smmla z7.s, z3.b, z2.b\n" + "ld1rqb { z3.b }, p1/Z, [x28, #64]\n" + ".inst 0x450e98a9 // smmla z9.s, z5.b, z14.b\n" + ".inst 0x450298b6 // smmla z22.s, z5.b, z2.b\n" + "ld1rqb { z5.b }, p1/Z, [x28, #80]\n" + "fscale z23.s, p1/m, z23.s, z28.s\n" + ".inst 0x451e9872 // smmla z18.s, z3.b, z30.b\n" + ".inst 0x45159867 // smmla z7.s, z3.b, z21.b\n" + "ld1rqb { z3.b }, p1/Z, [x28, #96]\n" + ".inst 0x451e98a9 // smmla z9.s, z5.b, z30.b\n" + ".inst 0x451598b6 // smmla z22.s, z5.b, z21.b\n" + "ld1rqb { z5.b }, p1/Z, [x28, #112]\n" + "add x28, x28, #0x88\n" + ".inst 0x45049872 // smmla z18.s, z3.b, z4.b\n" + ".inst 0x45119867 // smmla z7.s, z3.b, z17.b\n" + "ld1h { z3.s }, p0/Z, [x23]\n" + ".inst 0x450498a9 // smmla z9.s, z5.b, z4.b\n" + ".inst 0x451198b6 // smmla z22.s, z5.b, z17.b\n" + "fcvt z3.s, p1/m, z3.h\n" + "uzp1 z5.d, z18.d, z7.d\n" + "uzp2 z18.d, z18.d, z7.d\n" + "mov z3.q, z3.q[0]\n" + "uzp1 z7.d, z9.d, z22.d\n" + "uzp2 z22.d, z9.d, z22.d\n" + "fmul z9.s, z23.s, z3.s[0]\n" + "scvtf z5.s, p1/m, z5.s\n" + "scvtf z18.s, p1/m, z18.s\n" + "scvtf z7.s, p1/m, z7.s\n" + "scvtf z22.s, p1/m, z22.s\n" + "fmla z24.s, p1/M, z5.s, z9.s\n" + "ld1rqb { z5.b }, p1/Z, [x26]\n" + "fmul z9.s, z23.s, z3.s[1]\n" + "fmla z15.s, p1/M, z18.s, z9.s\n" + "ld1rqb { z18.b }, p1/Z, [x26, #16]\n" + "fmul z9.s, z23.s, z3.s[2]\n" + "fmul z3.s, z23.s, z3.s[3]\n" + "fmla z12.s, p1/M, z7.s, z9.s\n" + "mov z9.s, #0x0\n" + "ld1h { z7.s }, p0/Z, [x22]\n" + ".inst 0x451f98a9 // smmla z9.s, z5.b, z31.b\n" + "fmla z0.s, p1/M, z22.s, z3.s\n" + "mov z22.s, #0x0\n" + "ld1h { z3.s }, p0/Z, [x21]\n" + ".inst 0x450698b6 // smmla z22.s, z5.b, z6.b\n" + "ld1rqb { z5.b }, p1/Z, [x26, #32]\n" + "fcvt z7.s, p1/m, z7.h\n" + "fcvt z3.s, p1/m, z3.h\n" + ".inst 0x450e98a9 // smmla z9.s, z5.b, z14.b\n" + ".inst 0x450298b6 // smmla z22.s, z5.b, z2.b\n" + "ld1rqb { z5.b }, p1/Z, [x26, #64]\n" + "mov z7.q, z7.q[0]\n" + "mov z3.q, z3.q[0]\n" + ".inst 0x451e98a9 // smmla z9.s, z5.b, z30.b\n" + ".inst 0x451598b6 // smmla z22.s, z5.b, z21.b\n" + "ld1rqb { z5.b }, p1/Z, [x26, #96]\n" + ".inst 0x450498a9 // smmla z9.s, z5.b, z4.b\n" + ".inst 0x451198b6 // smmla z22.s, z5.b, z17.b\n" + "uzp1 z5.d, z9.d, z22.d\n" + "scvtf z5.s, p1/m, z5.s\n" + "uzp2 z22.d, z9.d, z22.d\n" + "fmul z9.s, z23.s, z7.s[0]\n" + "scvtf z22.s, p1/m, z22.s\n" + "fmla z13.s, p1/M, z5.s, z9.s\n" + "ld1rqb { z9.b }, p1/Z, [x25]\n" + "fmul z5.s, z23.s, z7.s[1]\n" + "fmla z1.s, p1/M, z22.s, z5.s\n" + "mov z5.s, #0x0\n" + "mov z22.s, #0x0\n" + ".inst 0x451f9a45 // smmla z5.s, z18.b, z31.b\n" + ".inst 0x45069a56 // smmla z22.s, z18.b, z6.b\n" + "ld1rqb { z18.b }, p1/Z, [x26, #48]\n" + ".inst 0x450e9a45 // smmla z5.s, z18.b, z14.b\n" + ".inst 0x45029a56 // smmla z22.s, z18.b, z2.b\n" + "ld1rqb { z18.b }, p1/Z, [x26, #80]\n" + ".inst 0x451e9a45 // smmla z5.s, z18.b, z30.b\n" + ".inst 0x45159a56 // smmla z22.s, z18.b, z21.b\n" + "ld1rqb { z18.b }, p1/Z, [x26, #112]\n" + "add x26, x26, #0x88\n" + ".inst 0x45049a45 // smmla z5.s, z18.b, z4.b\n" + ".inst 0x45119a56 // smmla z22.s, z18.b, z17.b\n" + "uzp1 z18.d, z5.d, z22.d\n" + "scvtf z18.s, p1/m, z18.s\n" + "uzp2 z22.d, z5.d, z22.d\n" + "fmul z5.s, z23.s, z7.s[2]\n" + "fmul z7.s, z23.s, z7.s[3]\n" + "scvtf z22.s, p1/m, z22.s\n" + "fmla z20.s, p1/M, z18.s, z5.s\n" + "ld1rqb { z18.b }, p1/Z, [x25, #16]\n" + "ld1h { z5.s }, p0/Z, [x20]\n" + "fcvt z5.s, p1/m, z5.h\n" + "fmla z25.s, p1/M, z22.s, z7.s\n" + "mov z22.s, #0x0\n" + "mov z7.s, #0x0\n" + ".inst 0x451f9936 // smmla z22.s, z9.b, z31.b\n" + ".inst 0x45069927 // smmla z7.s, z9.b, z6.b\n" + "ld1rqb { z9.b }, p1/Z, [x25, #32]\n" + "mov z5.q, z5.q[0]\n" + ".inst 0x450e9936 // smmla z22.s, z9.b, z14.b\n" + ".inst 0x45029927 // smmla z7.s, z9.b, z2.b\n" + "ld1rqb { z9.b }, p1/Z, [x25, #64]\n" + ".inst 0x451e9936 // smmla z22.s, z9.b, z30.b\n" + ".inst 0x45159927 // smmla z7.s, z9.b, z21.b\n" + "ld1rqb { z9.b }, p1/Z, [x25, #96]\n" + ".inst 0x45049936 // smmla z22.s, z9.b, z4.b\n" + ".inst 0x45119927 // smmla z7.s, z9.b, z17.b\n" + "uzp1 z9.d, z22.d, z7.d\n" + "scvtf z9.s, p1/m, z9.s\n" + "uzp2 z22.d, z22.d, z7.d\n" + "fmul z7.s, z23.s, z3.s[0]\n" + "scvtf z22.s, p1/m, z22.s\n" + "fmla z11.s, p1/M, z9.s, z7.s\n" + "ld1rqb { z9.b }, p1/Z, [x24]\n" + "fmul z7.s, z23.s, z3.s[1]\n" + "fmla z16.s, p1/M, z22.s, z7.s\n" + "mov z22.s, #0x0\n" + "mov z7.s, #0x0\n" + ".inst 0x451f9a56 // smmla z22.s, z18.b, z31.b\n" + ".inst 0x45069a47 // smmla z7.s, z18.b, z6.b\n" + "ld1rqb { z18.b }, p1/Z, [x25, #48]\n" + ".inst 0x450e9a56 // smmla z22.s, z18.b, z14.b\n" + ".inst 0x45029a47 // smmla z7.s, z18.b, z2.b\n" + "ld1rqb { z18.b }, p1/Z, [x25, #80]\n" + ".inst 0x451e9a56 // smmla z22.s, z18.b, z30.b\n" + ".inst 0x45159a47 // smmla z7.s, z18.b, z21.b\n" + "ld1rqb { z18.b }, p1/Z, [x25, #112]\n" + "add x25, x25, #0x88\n" + ".inst 0x45049a56 // smmla z22.s, z18.b, z4.b\n" + ".inst 0x45119a47 // smmla z7.s, z18.b, z17.b\n" + "uzp1 z18.d, z22.d, z7.d\n" + "scvtf z18.s, p1/m, z18.s\n" + "uzp2 z7.d, z22.d, z7.d\n" + "fmul z22.s, z23.s, z3.s[2]\n" + "fmul z3.s, z23.s, z3.s[3]\n" + "scvtf z7.s, p1/m, z7.s\n" + "fmla z19.s, p1/M, z18.s, z22.s\n" + "ld1rqb { z18.b }, p1/Z, [x24, #16]\n" + "fmul z22.s, z23.s, z5.s[0]\n" + "fmla z26.s, p1/M, z7.s, z3.s\n" + "mov z3.s, #0x0\n" + "mov z7.s, #0x0\n" + ".inst 0x451f9923 // smmla z3.s, z9.b, z31.b\n" + ".inst 0x45069927 // smmla z7.s, z9.b, z6.b\n" + "ld1rqb { z9.b }, p1/Z, [x24, #32]\n" + ".inst 0x450e9923 // smmla z3.s, z9.b, z14.b\n" + ".inst 0x45029927 // smmla z7.s, z9.b, z2.b\n" + "mov z9.s, #0x0\n" + ".inst 0x451f9a49 // smmla z9.s, z18.b, z31.b\n" + "mov z31.s, #0x0\n" + ".inst 0x45069a5f // smmla z31.s, z18.b, z6.b\n" + "ld1rqb { z6.b }, p1/Z, [x24, #48]\n" + "ld1rqb { z18.b }, p1/Z, [x24, #64]\n" + ".inst 0x450e98c9 // smmla z9.s, z6.b, z14.b\n" + "fmul z14.s, z23.s, z5.s[1]\n" + ".inst 0x450298df // smmla z31.s, z6.b, z2.b\n" + "ld1rqb { z6.b }, p1/Z, [x24, #80]\n" + "fmul z2.s, z23.s, z5.s[2]\n" + "fmul z23.s, z23.s, z5.s[3]\n" + ".inst 0x451e9a43 // smmla z3.s, z18.b, z30.b\n" + ".inst 0x45159a47 // smmla z7.s, z18.b, z21.b\n" + "ld1rqb { z5.b }, p1/Z, [x24, #96]\n" + ".inst 0x451e98c9 // smmla z9.s, z6.b, z30.b\n" + ".inst 0x451598df // smmla z31.s, z6.b, z21.b\n" + "ld1rqb { z18.b }, p1/Z, [x24, #112]\n" + "add x24, x24, #0x88\n" + ".inst 0x450498a3 // smmla z3.s, z5.b, z4.b\n" + ".inst 0x451198a7 // smmla z7.s, z5.b, z17.b\n" + ".inst 0x45049a49 // smmla z9.s, z18.b, z4.b\n" + ".inst 0x45119a5f // smmla z31.s, z18.b, z17.b\n" + "uzp1 z18.d, z3.d, z7.d\n" + "uzp2 z5.d, z3.d, z7.d\n" + "scvtf z18.s, p1/m, z18.s\n" + "uzp1 z6.d, z9.d, z31.d\n" + "uzp2 z9.d, z9.d, z31.d\n" + "scvtf z5.s, p1/m, z5.s\n" + "fmla z8.s, p1/M, z18.s, z22.s\n" + "scvtf z6.s, p1/m, z6.s\n" + "scvtf z9.s, p1/m, z9.s\n" + "fmla z29.s, p1/M, z5.s, z14.s\n" + "fmla z27.s, p1/M, z6.s, z2.s\n" + "fmla z10.s, p1/M, z9.s, z23.s\n" + "bgt 3b\n" + "mov x20, %x[res_ptr]\n" + "subs x10, x10, #0x8\n" + "add %x[res_ptr], %x[res_ptr], #0x20\n" + "st1w { z24.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z15.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z12.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z0.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z13.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z1.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z20.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z25.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z11.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z16.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z19.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z26.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z8.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z29.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z27.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "st1w { z10.s }, p1, [x20]\n" + "bne 2b\n" + "mov x20, #0x4\n" + "sub x13, x13, #0x10\n" + "cmp x13, #0x10\n" + "mov %x[res_ptr], x9\n" + "madd %x[a_ptr], x20, x12, %x[a_ptr]\n" + "bge 1b\n" + "4:" // Row loop skip + "cbz x13, 9f\n" + "5:" // Row tail: Row loop + "add x25, %x[b_ptr], #0x10\n" + "mov x24, %x[nc]\n" + "add x23, %x[res_ptr], %x[res_stride], LSL #2\n" + "6:" // Row tail: Column loop + "mov z24.b, #0x0\n" + "mov z15.b, #0x0\n" + "add x28, %x[a_ptr], #0x8\n" + "mov x22, %x[nb]\n" + "mov z12.b, #0x0\n" + "mov z0.b, #0x0\n" + "7:" // Row tail: Block loop + "ld1b { z3.b }, p1/Z, [x25]\n" + "ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n" + "mov z2.s, #0x0\n" + "mov z25.s, #0x0\n" + "ld1rqb { z26.b }, p1/Z, [x28]\n" + "ld1rqb { z21.b }, p1/Z, [x28, #16]\n" + "mov z27.s, #0x0\n" + "mov z19.s, #0x0\n" + "ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n" + "ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n" + "sub x21, x25, #0x10\n" + "sub x20, x28, #0x8\n" + "lsl z20.b, z3.b, #0x4\n" + "lsl z4.b, z6.b, #0x4\n" + "ld1rqb { z10.b }, p1/Z, [x28, #32]\n" + "ld1rqb { z23.b }, p1/Z, [x28, #48]\n" + "and z3.b, z3.b, #0xf0\n" + "and z6.b, z6.b, #0xf0\n" + "ld1rqb { z11.b }, p1/Z, [x28, #64]\n" + "ld1rqb { z7.b }, p1/Z, [x28, #80]\n" + "lsl z8.b, z29.b, #0x4\n" + "lsl z14.b, z16.b, #0x4\n" + "ld1rqb { z18.b }, p1/Z, [x28, #96]\n" + "ld1rqb { z30.b }, p1/Z, [x28, #112]\n" + ".inst 0x45149b42 // smmla z2.s, z26.b, z20.b\n" + ".inst 0x45049b59 // smmla z25.s, z26.b, z4.b\n" + "and z29.b, z29.b, #0xf0\n" + "ld1h { z17.s }, p1/Z, [x21]\n" + ".inst 0x45149abb // smmla z27.s, z21.b, z20.b\n" + ".inst 0x45049ab3 // smmla z19.s, z21.b, z4.b\n" + "and z16.b, z16.b, #0xf0\n" + "ld1h { z4.s }, p0/Z, [x20]\n" + "subs x22, x22, #0x1\n" + "add x28, x28, #0x88\n" + "fcvt z17.s, p1/m, z17.h\n" + "add x25, x25, #0x90\n" + ".inst 0x45089942 // smmla z2.s, z10.b, z8.b\n" + ".inst 0x450e9959 // smmla z25.s, z10.b, z14.b\n" + "fcvt z4.s, p1/m, z4.h\n" + ".inst 0x45089afb // smmla z27.s, z23.b, z8.b\n" + ".inst 0x450e9af3 // smmla z19.s, z23.b, z14.b\n" + "fscale z17.s, p1/m, z17.s, z28.s\n" + "mov z4.q, z4.q[0]\n" + ".inst 0x45039962 // smmla z2.s, z11.b, z3.b\n" + ".inst 0x45069979 // smmla z25.s, z11.b, z6.b\n" + "fmul z23.s, z17.s, z4.s[0]\n" + "fmul z9.s, z17.s, z4.s[1]\n" + "fmul z21.s, z17.s, z4.s[2]\n" + "fmul z4.s, z17.s, z4.s[3]\n" + ".inst 0x450398fb // smmla z27.s, z7.b, z3.b\n" + ".inst 0x450698f3 // smmla z19.s, z7.b, z6.b\n" + ".inst 0x451d9a42 // smmla z2.s, z18.b, z29.b\n" + ".inst 0x45109a59 // smmla z25.s, z18.b, z16.b\n" + ".inst 0x451d9bdb // smmla z27.s, z30.b, z29.b\n" + ".inst 0x45109bd3 // smmla z19.s, z30.b, z16.b\n" + "uzp1 z31.d, z2.d, z25.d\n" + "uzp2 z13.d, z2.d, z25.d\n" + "scvtf z31.s, p1/m, z31.s\n" + "uzp1 z17.d, z27.d, z19.d\n" + "uzp2 z18.d, z27.d, z19.d\n" + "scvtf z13.s, p1/m, z13.s\n" + "fmla z24.s, p1/M, z31.s, z23.s\n" + "scvtf z17.s, p1/m, z17.s\n" + "scvtf z18.s, p1/m, z18.s\n" + "fmla z15.s, p1/M, z13.s, z9.s\n" + "fmla z12.s, p1/M, z17.s, z21.s\n" + "fmla z0.s, p1/M, z18.s, z4.s\n" + "bgt 7b\n" + "mov x20, %x[res_ptr]\n" + "cmp x13, #0x1\n" + "st1w { z24.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x13, #0x2\n" + "st1w { z15.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x13, #0x3\n" + "st1w { z12.s }, p1, [x20]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "st1w { z0.s }, p1, [x20]\n" + "8:" // Row tail: Accumulator store skip + "subs x24, x24, #0x8\n" + "add %x[res_ptr], %x[res_ptr], #0x20\n" + "bne 6b\n" + "subs x13, x13, #0x4\n" + "add %x[a_ptr], %x[a_ptr], x12\n" + "mov %x[res_ptr], x23\n" + "bgt 5b\n" + "9:" // Row tail: Row loop skip + : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) + : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) + : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + ); + return; + } + else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { + GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) && + "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal " + "performance"); + } + else if (ggml_cpu_has_neon()) { + GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) && + "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 " + "quantization format for optimal performance"); + } +#endif +#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + GGML_ASSERT(ggml_cpu_has_sve() && + "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance"); +#elif defined(__ARM_NEON) && defined(__aarch64__) + GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && + "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " + "performance"); +#else + float sumf[4][8]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } +#endif +} diff --git a/ggml/src/ggml-aarch64.h b/ggml/src/ggml-aarch64.h new file mode 100644 index 000000000..65ead1efe --- /dev/null +++ b/ggml/src/ggml-aarch64.h @@ -0,0 +1,39 @@ +// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd. +#pragma once + +#define GGML_COMMON_DECL_C +#include "ggml-common.h" + +#include "ggml.h" + +// GGML internal header + +#ifdef __cplusplus +extern "C" { +#endif + +// Quantization +void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); + +void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t interleave_blcksize); + +// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization") +size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); + +// GEMV +void ggml_gemv_q4_0_4x4_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q4_0_4x8_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q4_0_8x8_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); + +// GEMM +void ggml_gemm_q4_0_4x4_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q4_0_4x8_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q4_0_8x8_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); + +#ifdef __cplusplus +} +#endif + diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index c74060cc4..fafd5fa7a 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -199,6 +199,30 @@ typedef struct { } block_q8_1; static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding"); +typedef struct { + ggml_half d[4]; // deltas for 4 q4_0 blocks + uint8_t qs[QK4_0 * 2]; // nibbles / quants for 4 q4_0 blocks +} block_q4_0x4; +static_assert(sizeof(block_q4_0x4) == 4 * sizeof(ggml_half) + QK4_0 * 2, "wrong q4_0x4 block size/padding"); + +typedef struct { + ggml_half d[8]; // deltas for 8 q4_0 blocks + uint8_t qs[QK4_0 * 4]; // nibbles / quants for 8 q4_0 blocks +} block_q4_0x8; +static_assert(sizeof(block_q4_0x8) == 8 * sizeof(ggml_half) + QK4_0 * 4, "wrong q4_0x8 block size/padding"); + +typedef struct { + ggml_half d[4]; // deltas for 4 q8_0 blocks + int8_t qs[QK8_0 * 4]; // quants for 4 q8_0 blocks +} block_q8_0x4; +static_assert(sizeof(block_q8_0x4) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong q8_0x4 block size/padding"); + +typedef struct { + ggml_half d[8]; // deltas for 8 q8_0 blocks + int8_t qs[QK8_0 * 8]; // quants for 8 q8_0 blocks +} block_q8_0x8; +static_assert(sizeof(block_q8_0x8) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong q8_0x8 block size/padding"); + // // Super-block quantization structures // diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h index 1d2336190..a2c8dbec0 100644 --- a/ggml/src/ggml-impl.h +++ b/ggml/src/ggml-impl.h @@ -609,6 +609,10 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { #endif // defined(__ARM_NEON) && (!defined(__MSC_VER) +#ifdef __ARM_FEATURE_SVE +#include +#endif // __ARM_FEATURE_SVE + // precomputed f32 table for f16 (256 KB) // defined in ggml.c, initialized in ggml_init() extern float ggml_table_f32_f16[1 << 16]; diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 0eb52e485..cbe377cf5 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -3814,43 +3814,47 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r } #endif #if defined(__ARM_FEATURE_SVE) - const svbool_t ptrueh = svptrue_pat_b8(SV_VL16); - const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh); + if (svcntb() == QK8_0) { + const svbool_t ptrueh = svptrue_pat_b8(SV_VL16); + const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh); - svfloat32_t sumv0 = svdup_n_f32(0.0f); - svfloat32_t sumv1 = svdup_n_f32(0.0f); + svfloat32_t sumv0 = svdup_n_f32(0.0f); + svfloat32_t sumv1 = svdup_n_f32(0.0f); - assert(nb % 2 == 0); // TODO: handle odd nb + assert(nb % 2 == 0); // TODO: handle odd nb - for (int i = 0; i < nb; i += 2) { - const block_q4_0 * restrict x0 = &x[i + 0]; - const block_q4_0 * restrict x1 = &x[i + 1]; - const block_q8_0 * restrict y0 = &y[i + 0]; - const block_q8_0 * restrict y1 = &y[i + 1]; + for (int i = 0; i < nb; i += 2) { + const block_q4_0 * restrict x0 = &x[i + 0]; + const block_q4_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; - // load x - const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs); - const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs); + // load x + const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs); + const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs); - // 4-bit -> 8-bit - const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04)); - const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04)); + // 4-bit -> 8-bit + const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04)); + const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04)); - // sub 8 - const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8); - const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8); + // sub 8 + const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8); + const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8); - // load y - const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); - const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); + // load y + const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); + const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); - // dot product - sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + // dot product + sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); + return; } - - *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); -#elif defined(__ARM_NEON) +#endif +#if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -5422,31 +5426,35 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r } #endif #if defined(__ARM_FEATURE_SVE) - svfloat32_t sumv0 = svdup_n_f32(0.0f); - svfloat32_t sumv1 = svdup_n_f32(0.0f); + if (svcntb() == QK8_0) { + svfloat32_t sumv0 = svdup_n_f32(0.0f); + svfloat32_t sumv1 = svdup_n_f32(0.0f); - assert(nb % 2 == 0); // TODO: handle odd nb + assert(nb % 2 == 0); // TODO: handle odd nb - for (int i = 0; i < nb; i += 2) { - const block_q8_0 * restrict x0 = &x[i + 0]; - const block_q8_0 * restrict x1 = &x[i + 1]; - const block_q8_0 * restrict y0 = &y[i + 0]; - const block_q8_0 * restrict y1 = &y[i + 1]; + for (int i = 0; i < nb; i += 2) { + const block_q8_0 * restrict x0 = &x[i + 0]; + const block_q8_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; - // load x - const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs); - const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs); + // load x + const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs); + const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs); - // load y - const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); - const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); + // load y + const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); + const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); - sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); + return; } - - *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); -#elif defined(__ARM_NEON) +#endif +#if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -14760,6 +14768,16 @@ static bool validate_fp16(ggml_fp16_t f, size_t i) { } \ } +#define VALIDATE_ROW_DATA_DVEC_F16_IMPL(type, data, nb, nr) \ + const type * q = (const type *) (data); \ + for (size_t i = 0; i < (nb); ++i) { \ + for (size_t j = 0; j < (nr); ++j) { \ + if (!validate_fp16(q[i].d[j], i)) { \ + return false; \ + } \ + } \ + } + bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes) { if (type < 0 || type >= GGML_TYPE_COUNT) { fprintf(stderr, "%s: invalid type %d\n", __func__, type); @@ -14977,6 +14995,16 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte { VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb); } break; + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + { + VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x4, data, nbytes / sizeof(block_q4_0x4), 4); + } break; + case GGML_TYPE_Q4_0_8_8: + { + VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x8, data, nbytes / sizeof(block_q4_0x8), 8); + } break; + case GGML_TYPE_I8: case GGML_TYPE_I16: case GGML_TYPE_I32: diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index bc91ac3a7..c0aced3d2 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -4,6 +4,7 @@ #include "ggml-impl.h" #include "ggml-quants.h" #include "ggml.h" +#include "ggml-aarch64.h" #if defined(_MSC_VER) || defined(__MINGW32__) @@ -37,7 +38,7 @@ #include #endif -#ifdef __ARM_FEATURE_MATMUL_INT8 +#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8) #undef GGML_USE_LLAMAFILE #endif @@ -692,6 +693,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { #else .nrows = 1, #endif + .from_float_to_mat = quantize_mat_q8_0, }, [GGML_TYPE_Q8_1] = { .type_name = "q8_1", @@ -889,6 +891,54 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .vec_dot = (ggml_vec_dot_t) ggml_vec_dot_bf16, .vec_dot_type = GGML_TYPE_BF16, .nrows = 1, + }, + [GGML_TYPE_Q4_0_4_4] = { + .type_name = "q4_0_4x4", + .blck_size = QK4_0, + .type_size = sizeof(block_q4_0), + .is_quantized = true, + .to_float = NULL, + .from_float = NULL, + .from_float_reference = NULL, + .vec_dot = NULL, + .vec_dot_type = GGML_TYPE_Q8_0, + .nrows = 1, + .ncols = 4, + .interleave_blcksize = 4, + .gemv = ggml_gemv_q4_0_4x4_q8_0, + .gemm = ggml_gemm_q4_0_4x4_q8_0, + }, + [GGML_TYPE_Q4_0_4_8] = { + .type_name = "q4_0_4x8", + .blck_size = QK4_0, + .type_size = sizeof(block_q4_0), + .is_quantized = true, + .to_float = NULL, + .from_float = NULL, + .from_float_reference = NULL, + .vec_dot = NULL, + .vec_dot_type = GGML_TYPE_Q8_0, + .nrows = 1, + .ncols = 4, + .interleave_blcksize = 8, + .gemv = ggml_gemv_q4_0_4x8_q8_0, + .gemm = ggml_gemm_q4_0_4x8_q8_0, + }, + [GGML_TYPE_Q4_0_8_8] = { + .type_name = "q4_0_8x8", + .blck_size = QK4_0, + .type_size = sizeof(block_q4_0), + .is_quantized = true, + .to_float = NULL, + .from_float = NULL, + .from_float_reference = NULL, + .vec_dot = NULL, + .vec_dot_type = GGML_TYPE_Q8_0, + .nrows = 1, + .ncols = 8, + .interleave_blcksize = 8, + .gemv = ggml_gemv_q4_0_8x8_q8_0, + .gemm = ggml_gemm_q4_0_8x8_q8_0, } }; @@ -3188,6 +3238,9 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ4_XS: wtype = GGML_TYPE_IQ4_XS; break; case GGML_FTYPE_MOSTLY_IQ3_S: wtype = GGML_TYPE_IQ3_S; break; case GGML_FTYPE_MOSTLY_IQ2_S: wtype = GGML_TYPE_IQ2_S; break; + case GGML_FTYPE_MOSTLY_Q4_0_4_4: wtype = GGML_TYPE_Q4_0_4_4; break; + case GGML_FTYPE_MOSTLY_Q4_0_4_8: wtype = GGML_TYPE_Q4_0_4_8; break; + case GGML_FTYPE_MOSTLY_Q4_0_8_8: wtype = GGML_TYPE_Q4_0_8_8; break; case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break; case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break; } @@ -9432,6 +9485,9 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: { ggml_compute_forward_add_q_f32(params, dst); } break; @@ -9807,6 +9863,9 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: { ggml_compute_forward_add1_q_f32(params, dst); } break; @@ -9932,6 +9991,9 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: default: { GGML_ASSERT(false); @@ -12134,6 +12196,12 @@ static void ggml_compute_forward_mul_mat( enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type; ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float; int64_t const vec_dot_num_rows = type_traits[type].nrows; + int64_t const matmul_num_cols = type_traits[type].ncols; + int64_t const interleave_blcksize = type_traits[type].interleave_blcksize; + ggml_from_float_to_mat_t const from_float_to_mat + = type_traits[vec_dot_type].from_float_to_mat; + ggml_gemv_t const gemv = type_traits[type].gemv; + ggml_gemm_t const gemm = type_traits[type].gemm; GGML_ASSERT(ne0 == ne01); GGML_ASSERT(ne1 == ne11); @@ -12192,7 +12260,16 @@ UseGgmlGemm1:; for (int64_t i13 = 0; i13 < ne13; ++i13) { for (int64_t i12 = 0; i12 < ne12; ++i12) { - for (int64_t i11 = ith; i11 < ne11; i11 += nth) { + int64_t i11_processed = 0; + if ((ggml_n_dims(src1) == 2) && from_float_to_mat && gemm) { + for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { + from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), + (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), + 4, ne10, interleave_blcksize); + } + i11_processed = ne11 - ne11 % 4; + } + for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) { from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), ne10); @@ -12273,6 +12350,28 @@ UseGgmlGemm2:; const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0; const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1; + if ((ggml_n_dims(src0) == 2) && gemv) { + const void * src1_wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; + const size_t src1_col_stride = ggml_is_contiguous(src1) || src1->type != vec_dot_type ? ggml_row_size(vec_dot_type, ne10) : nb11; + int64_t src0_start = (ith * ne01) / nth; + int64_t src0_end = ((ith + 1) * ne01) / nth; + src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start; + src0_end = (src0_end % matmul_num_cols) ? src0_end + matmul_num_cols - (src0_end % matmul_num_cols): src0_end; + if (src0_start >= src0_end) return; + + // If there are more than three rows in src1, use gemm; otherwise, use gemv. + if (gemm && (ne11 > 3)) { + gemm(ne00, (float *)((char *) dst->data) + src0_start, ne01, (const char *) src0->data + src0_start * nb01, + (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start); + } + for (int iter = gemm ? ne11 - ne11 % 4 : 0; iter < ne11; iter++) { + gemv(ne00, (float *)((char *) dst->data + (iter * nb1)) + src0_start, ne01, + (const char *) src0->data + src0_start * nb01, (const char *) src1_wdata + (src1_col_stride * iter), 1, + src0_end - src0_start); + } + return; + } + // The first chunk comes from our thread_id, the rest will get auto-assigned. int current_chunk = ith; @@ -12318,6 +12417,8 @@ static void ggml_compute_forward_mul_mat_id( ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot; enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type; ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float; + int64_t const matmul_num_cols = type_traits[type].ncols; + ggml_gemv_t const gemv = type_traits[type].gemv; // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == ggml_type_size(type)); @@ -12403,6 +12504,34 @@ static void ggml_compute_forward_mul_mat_id( const int64_t nr0 = ne01; // src0 rows const int64_t nr1 = cne1; // src1 rows + if (((ggml_n_dims(src0) - 1) == 2) && gemv) { + int64_t src0_cur_start = (ith * ne01) / nth; + int64_t src0_cur_end = ((ith + 1) * ne01) / nth; + src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start; + src0_cur_end = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end; + if (src0_cur_start >= src0_cur_end) return; + + for (int ir1 = 0; ir1 < nr1; ir1++) { + struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1); + const int id = row_mapping.i1; // selected expert index + + const int64_t i11 = id % ne11; + const int64_t i12 = row_mapping.i2; // row index in src1 + + const int64_t i1 = id; // selected expert index + const int64_t i2 = i12; // row + + const char * src1_col = (const char *) wdata + + (src1_cont || src1->type != vec_dot_type + ? (i11 + i12 * ne11) * row_size + : (i11 * nb11 + i12 * nb12)); + + gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01, + (const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start); + } + continue; + } + // distribute the thread work across the inner or outer loop based on which one is larger const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows @@ -12704,6 +12833,9 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: { ggml_compute_forward_out_prod_q_f32(params, dst); } break; @@ -12889,6 +13021,9 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: default: { GGML_ASSERT(false); @@ -13148,6 +13283,9 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: { ggml_compute_forward_get_rows_q(params, dst); } break; @@ -13734,6 +13872,9 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q8_K: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: case GGML_TYPE_I8: case GGML_TYPE_I16: case GGML_TYPE_I32: @@ -20457,6 +20598,9 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ1_M: result = quantize_iq1_m (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_NL: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_F16: { size_t elemsize = sizeof(ggml_fp16_t); @@ -21759,8 +21903,6 @@ int ggml_cpu_has_neon(void) { int ggml_cpu_has_sve(void) { #if defined(__ARM_FEATURE_SVE) - // TODO: Currently, SVE 256 bit is only supported. - GGML_ASSERT(svcntb() == QK8_0); return 1; #else return 0; diff --git a/include/llama.h b/include/llama.h index bb4b05ba6..3970c3aeb 100644 --- a/include/llama.h +++ b/include/llama.h @@ -162,6 +162,9 @@ extern "C" { LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors LLAMA_FTYPE_MOSTLY_BF16 = 32, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // except 1d tensors LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file }; diff --git a/src/llama.cpp b/src/llama.cpp index 98a242081..b19d786e2 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3788,6 +3788,9 @@ struct llama_model_loader { case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break; case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break; case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break; + case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break; + case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break; + case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break; default: { LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); @@ -4481,6 +4484,9 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw"; + case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4"; + case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8"; + case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8"; default: return "unknown, may not work"; } @@ -17768,6 +17774,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { new_type = GGML_TYPE_IQ3_S; } + else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || + new_type == GGML_TYPE_Q4_0_8_8) { + new_type = GGML_TYPE_Q4_0; + } } } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) { @@ -18080,6 +18090,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break; case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break; case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break; + case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break; + case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break; + case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break; default: throw std::runtime_error(format("invalid output file type %d\n", ftype)); } @@ -18390,6 +18403,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s f32_data = (float *) f32_conv_buf.data(); } + int chunk_size_multiplier = 1; + if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) { + if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0; + else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0; + if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8; + else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4; + } + LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type)); fflush(stdout); @@ -18402,7 +18423,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s const int64_t nrows = tensor->ne[1]; static const int64_t min_chunk_size = 32 * 512; - const int64_t chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row); + const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) * + chunk_size_multiplier; const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1]; const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size; From 6b2a849d1f43d46b82d2f9c08c3275137b528784 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 10 Jul 2024 15:23:29 +0300 Subject: [PATCH 29/31] ggml : move sgemm sources to llamafile subfolder (#8394) ggml-ci --- Makefile | 8 ++++---- ggml/CMakeLists.txt | 2 +- ggml/src/CMakeLists.txt | 6 +++--- ggml/src/ggml.c | 3 +-- ggml/src/{ => llamafile}/sgemm.cpp | 0 ggml/src/{ => llamafile}/sgemm.h | 0 6 files changed, 9 insertions(+), 10 deletions(-) rename ggml/src/{ => llamafile}/sgemm.cpp (100%) rename ggml/src/{ => llamafile}/sgemm.h (100%) diff --git a/Makefile b/Makefile index b70ebaed5..68197fef8 100644 --- a/Makefile +++ b/Makefile @@ -554,7 +554,7 @@ endif # GGML_BLIS ifndef GGML_NO_LLAMAFILE MK_CPPFLAGS += -DGGML_USE_LLAMAFILE - OBJ_GGML += ggml/src/sgemm.o + OBJ_GGML += ggml/src/llamafile/sgemm.o endif ifdef GGML_RPC @@ -983,9 +983,9 @@ ggml/src/ggml-blas.o: \ $(CXX) $(CXXFLAGS) -c $< -o $@ ifndef GGML_NO_LLAMAFILE -ggml/src/sgemm.o: \ - ggml/src/sgemm.cpp \ - ggml/src/sgemm.h \ +ggml/src/llamafile/sgemm.o: \ + ggml/src/llamafile/sgemm.cpp \ + ggml/src/llamafile/sgemm.h \ ggml/include/ggml.h $(CXX) $(CXXFLAGS) -c $< -o $@ endif # GGML_NO_LLAMAFILE diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 0d0d52d57..649ac3dcc 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -104,7 +104,7 @@ option(GGML_ACCELERATE "ggml: enable Accelerate framework" option(GGML_BLAS "ggml: use BLAS" ${GGML_BLAS_DEFAULT}) set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING "ggml: BLAS library vendor") -option(GGML_LLAMAFILE "ggml: use ggml SGEMM" OFF) +option(GGML_LLAMAFILE "ggml: use LLAMAFILE" OFF) option(GGML_CUDA "ggml: use CUDA" OFF) option(GGML_CUDA_FORCE_DMMV "ggml: use dmmv instead of mmvq CUDA kernels" OFF) diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index aae5b8e9f..c5ee7e425 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -238,12 +238,12 @@ if (GGML_BLAS) endif() if (GGML_LLAMAFILE) - message(STATUS "Using ggml SGEMM") + message(STATUS "Using llamafile") add_compile_definitions(GGML_USE_LLAMAFILE) - set(GGML_HEADERS_LLAMAFILE sgemm.h) - set(GGML_SOURCES_LLAMAFILE sgemm.cpp) + set(GGML_HEADERS_LLAMAFILE llamafile/sgemm.h) + set(GGML_SOURCES_LLAMAFILE llamafile/sgemm.cpp) endif() if (GGML_CUDA) diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index c0aced3d2..1bb731e16 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -6,7 +6,6 @@ #include "ggml.h" #include "ggml-aarch64.h" - #if defined(_MSC_VER) || defined(__MINGW32__) #include // using malloc.h with MSC/MINGW #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__) @@ -43,7 +42,7 @@ #endif #ifdef GGML_USE_LLAMAFILE -#include "sgemm.h" +#include #endif #if defined(_MSC_VER) diff --git a/ggml/src/sgemm.cpp b/ggml/src/llamafile/sgemm.cpp similarity index 100% rename from ggml/src/sgemm.cpp rename to ggml/src/llamafile/sgemm.cpp diff --git a/ggml/src/sgemm.h b/ggml/src/llamafile/sgemm.h similarity index 100% rename from ggml/src/sgemm.h rename to ggml/src/llamafile/sgemm.h From f4444d992c16b6b9442f4770c7c3a10b19a08343 Mon Sep 17 00:00:00 2001 From: AidanBeltonS Date: Wed, 10 Jul 2024 16:10:49 +0100 Subject: [PATCH 30/31] [SYCL] Use multi_ptr to clean up deprecated warnings (#8256) --- ggml/src/ggml-sycl/common.hpp | 6 ++ ggml/src/ggml-sycl/convert.cpp | 2 +- ggml/src/ggml-sycl/mmq.cpp | 184 ++++++++++++++++----------------- ggml/src/ggml-sycl/norm.cpp | 6 +- ggml/src/ggml-sycl/softmax.cpp | 2 +- 5 files changed, 103 insertions(+), 97 deletions(-) diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index 9a1c161b6..68d41411b 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -346,4 +346,10 @@ inline sycl::vec vec_aligned_load(const Tp* aligned_ptr) { return *reinterpret_cast*>(aligned_ptr); } +// Helper for accessing pointers with no warnings +template +static __dpct_inline__ Tp* get_pointer(sycl::local_accessor acc) { + return acc.template get_multi_ptr().get(); +} + #endif // GGML_SYCL_COMMON_HPP diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp index a15271b51..39c28753c 100644 --- a/ggml/src/ggml-sycl/convert.cpp +++ b/ggml/src/ggml-sycl/convert.cpp @@ -158,7 +158,7 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k, sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_K(vx, y, scale_local_acc.get_pointer(), item_ct1); + dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1); }); }); } diff --git a/ggml/src/ggml-sycl/mmq.cpp b/ggml/src/ggml-sycl/mmq.cpp index b514f0040..3107ba919 100644 --- a/ggml/src/ggml-sycl/mmq.cpp +++ b/ggml/src/ggml-sycl/mmq.cpp @@ -1835,10 +1835,10 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, mul_mat_q4_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_qs_q4_0_acc_ct1.get_pointer(), - tile_x_d_q4_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_qs_q4_0_acc_ct1), + get_pointer(tile_x_d_q4_0_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -1870,10 +1870,10 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, mul_mat_q4_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_qs_q4_0_acc_ct1.get_pointer(), - tile_x_d_q4_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_qs_q4_0_acc_ct1), + get_pointer(tile_x_d_q4_0_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -1950,10 +1950,10 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, mul_mat_q4_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_qs_q4_1_acc_ct1.get_pointer(), - tile_x_dm_q4_1_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_qs_q4_1_acc_ct1), + get_pointer(tile_x_dm_q4_1_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -1985,10 +1985,10 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, mul_mat_q4_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_qs_q4_1_acc_ct1.get_pointer(), - tile_x_dm_q4_1_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_qs_q4_1_acc_ct1), + get_pointer(tile_x_dm_q4_1_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2065,10 +2065,10 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, mul_mat_q5_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q5_0_acc_ct1.get_pointer(), - tile_x_d_q5_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q5_0_acc_ct1), + get_pointer(tile_x_d_q5_0_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2100,10 +2100,10 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, mul_mat_q5_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q5_0_acc_ct1.get_pointer(), - tile_x_d_q5_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q5_0_acc_ct1), + get_pointer(tile_x_d_q5_0_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2180,10 +2180,10 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, mul_mat_q5_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q5_1_acc_ct1.get_pointer(), - tile_x_dm_q5_1_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q5_1_acc_ct1), + get_pointer(tile_x_dm_q5_1_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2215,10 +2215,10 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, mul_mat_q5_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q5_1_acc_ct1.get_pointer(), - tile_x_dm_q5_1_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q5_1_acc_ct1), + get_pointer(tile_x_dm_q5_1_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2295,10 +2295,10 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, mul_mat_q8_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_qs_q8_0_acc_ct1.get_pointer(), - tile_x_d_q8_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_qs_q8_0_acc_ct1), + get_pointer(tile_x_d_q8_0_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2330,10 +2330,10 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, mul_mat_q8_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_qs_q8_0_acc_ct1.get_pointer(), - tile_x_d_q8_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_qs_q8_0_acc_ct1), + get_pointer(tile_x_d_q8_0_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2412,11 +2412,11 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q2_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q2_K_acc_ct1.get_pointer(), - tile_x_dm_q2_K_acc_ct1.get_pointer(), - tile_x_sc_q2_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q2_K_acc_ct1), + get_pointer(tile_x_dm_q2_K_acc_ct1), + get_pointer(tile_x_sc_q2_K_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2450,11 +2450,11 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q2_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q2_K_acc_ct1.get_pointer(), - tile_x_dm_q2_K_acc_ct1.get_pointer(), - tile_x_sc_q2_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q2_K_acc_ct1), + get_pointer(tile_x_dm_q2_K_acc_ct1), + get_pointer(tile_x_sc_q2_K_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2537,12 +2537,12 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q3_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q3_K_acc_ct1.get_pointer(), - tile_x_dm_q3_K_acc_ct1.get_pointer(), - tile_x_qh_q3_K_acc_ct1.get_pointer(), - tile_x_sc_q3_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q3_K_acc_ct1), + get_pointer(tile_x_dm_q3_K_acc_ct1), + get_pointer(tile_x_qh_q3_K_acc_ct1), + get_pointer(tile_x_sc_q3_K_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2578,12 +2578,12 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q3_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q3_K_acc_ct1.get_pointer(), - tile_x_dm_q3_K_acc_ct1.get_pointer(), - tile_x_qh_q3_K_acc_ct1.get_pointer(), - tile_x_sc_q3_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q3_K_acc_ct1), + get_pointer(tile_x_dm_q3_K_acc_ct1), + get_pointer(tile_x_qh_q3_K_acc_ct1), + get_pointer(tile_x_sc_q3_K_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2663,11 +2663,11 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q4_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q4_K_acc_ct1.get_pointer(), - tile_x_dm_q4_K_acc_ct1.get_pointer(), - tile_x_sc_q4_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q4_K_acc_ct1), + get_pointer(tile_x_dm_q4_K_acc_ct1), + get_pointer(tile_x_sc_q4_K_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2701,11 +2701,11 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q4_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q4_K_acc_ct1.get_pointer(), - tile_x_dm_q4_K_acc_ct1.get_pointer(), - tile_x_sc_q4_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q4_K_acc_ct1), + get_pointer(tile_x_dm_q4_K_acc_ct1), + get_pointer(tile_x_sc_q4_K_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2784,11 +2784,11 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q5_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q5_K_acc_ct1.get_pointer(), - tile_x_dm_q5_K_acc_ct1.get_pointer(), - tile_x_sc_q5_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q5_K_acc_ct1), + get_pointer(tile_x_dm_q5_K_acc_ct1), + get_pointer(tile_x_sc_q5_K_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2822,11 +2822,11 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q5_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_q5_K_acc_ct1.get_pointer(), - tile_x_dm_q5_K_acc_ct1.get_pointer(), - tile_x_sc_q5_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_q5_K_acc_ct1), + get_pointer(tile_x_dm_q5_K_acc_ct1), + get_pointer(tile_x_sc_q5_K_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2905,11 +2905,11 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q6_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_acc_ct1.get_pointer(), - tile_x_dm_acc_ct1.get_pointer(), - tile_x_sc_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_acc_ct1), + get_pointer(tile_x_dm_acc_ct1), + get_pointer(tile_x_sc_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } @@ -2943,11 +2943,11 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, mul_mat_q6_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, - tile_x_ql_acc_ct1.get_pointer(), - tile_x_dm_acc_ct1.get_pointer(), - tile_x_sc_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + get_pointer(tile_x_ql_acc_ct1), + get_pointer(tile_x_dm_acc_ct1), + get_pointer(tile_x_sc_acc_ct1), + get_pointer(tile_y_qs_acc_ct1), + get_pointer(tile_y_ds_acc_ct1)); }); }); } diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp index e0c5dfeca..cccf87d06 100644 --- a/ggml/src/ggml-sycl/norm.cpp +++ b/ggml/src/ggml-sycl/norm.cpp @@ -218,7 +218,7 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols, [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); + get_pointer(s_sum_acc_ct1), work_group_size); }); }); } @@ -265,7 +265,7 @@ static void group_norm_f32_sycl(const float* x, float* dst, [[intel::reqd_sub_group_size(WARP_SIZE)]] { group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); + get_pointer(s_sum_acc_ct1), work_group_size); }); }); } @@ -306,7 +306,7 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { rms_norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); + get_pointer(s_sum_acc_ct1), work_group_size); }); }); } diff --git a/ggml/src/ggml-sycl/softmax.cpp b/ggml/src/ggml-sycl/softmax.cpp index e624b6ba3..c5d9a837e 100644 --- a/ggml/src/ggml-sycl/softmax.cpp +++ b/ggml/src/ggml-sycl/softmax.cpp @@ -136,7 +136,7 @@ static void soft_max_f32_submitter(const float * x, const float * mask, float * soft_max_f32(x, mask, dst, ncols_par, nrows_y, scale, max_bias, m0, m1, n_head_log2, item_ct1, - local_buf_acc.get_pointer()); + get_pointer(local_buf_acc)); }); }); } From dd07a123b79f9bd9e8a4ba0447427b3083e9347a Mon Sep 17 00:00:00 2001 From: Clint Herron Date: Wed, 10 Jul 2024 12:35:18 -0400 Subject: [PATCH 31/31] Name Migration: Build the deprecation-warning 'main' binary every time (#8404) * Modify the deprecation-warning 'main' binary to build every time, instead of only when a legacy binary is present. This is to help users of tutorials and other instruction sets from knowing what to do when the 'main' binary is missing and they are trying to follow instructions. * Adjusting 'server' name-deprecation binary to build all the time, similar to the 'main' legacy name binary. --- Makefile | 24 ++++++++---------------- 1 file changed, 8 insertions(+), 16 deletions(-) diff --git a/Makefile b/Makefile index 68197fef8..668b38b99 100644 --- a/Makefile +++ b/Makefile @@ -1513,15 +1513,17 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \ # Mark legacy binary targets as .PHONY so that they are always checked. .PHONY: main quantize perplexity embedding server finetune +# NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate. +# Eventually we will want to remove these target from building all the time. main: examples/deprecation-warning/deprecation-warning.cpp -ifneq (,$(wildcard main)) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - @echo "#########" - @echo "WARNING: The 'main' binary is deprecated. Please use 'llama-cli' instead." - @echo " Remove the 'main' binary to remove this warning." - @echo "#########" -endif + @echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead." + +server: examples/deprecation-warning/deprecation-warning.cpp + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead." quantize: examples/deprecation-warning/deprecation-warning.cpp ifneq (,$(wildcard quantize)) @@ -1553,16 +1555,6 @@ ifneq (,$(wildcard embedding)) @echo "#########" endif -server: examples/deprecation-warning/deprecation-warning.cpp -ifneq (,$(wildcard server)) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - @echo "#########" - @echo "WARNING: The 'server' binary is deprecated. Please use 'llama-server' instead." - @echo " Remove the 'server' binary to remove this warning." - @echo "#########" -endif - finetune: examples/deprecation-warning/deprecation-warning.cpp ifneq (,$(wildcard finetune)) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)