ommlds 0.0.0.dev467__py3-none-any.whl → 0.0.0.dev469__py3-none-any.whl

ommlds/.omlish-manifests.json

@@ -170,6 +170,21 @@
     "attr": null,
     "file": "ommlds/minichain/backends/impls/llamacpp/chat.py",
     "line": 33,
+    "value": {
+      "!.minichain.backends.strings.manifests.BackendStringsManifest": {
+        "service_cls_names": [
+          "ChatChoicesService"
+        ],
+        "backend_name": "llamacpp",
+        "model_names": null
+      }
+    }
+  },
+  {
+    "module": ".minichain.backends.impls.llamacpp.chat",
+    "attr": null,
+    "file": "ommlds/minichain/backends/impls/llamacpp/chat.py",
+    "line": 42,
     "value": {
       "!.minichain.registries.manifests.RegistryManifest": {
         "module": "ommlds.minichain.backends.impls.llamacpp.chat",
@@ -200,6 +215,21 @@
     "attr": null,
     "file": "ommlds/minichain/backends/impls/llamacpp/stream.py",
     "line": 32,
+    "value": {
+      "!.minichain.backends.strings.manifests.BackendStringsManifest": {
+        "service_cls_names": [
+          "ChatChoicesStreamService"
+        ],
+        "backend_name": "llamacpp",
+        "model_names": null
+      }
+    }
+  },
+  {
+    "module": ".minichain.backends.impls.llamacpp.stream",
+    "attr": null,
+    "file": "ommlds/minichain/backends/impls/llamacpp/stream.py",
+    "line": 41,
     "value": {
       "!.minichain.registries.manifests.RegistryManifest": {
         "module": "ommlds.minichain.backends.impls.llamacpp.stream",
@@ -271,6 +301,52 @@
       }
     }
   },
+  {
+    "module": ".minichain.backends.impls.ollama.chat",
+    "attr": null,
+    "file": "ommlds/minichain/backends/impls/ollama/chat.py",
+    "line": 38,
+    "value": {
+      "!.minichain.backends.strings.manifests.BackendStringsManifest": {
+        "service_cls_names": [
+          "ChatChoicesService",
+          "ChatChoicesStreamService"
+        ],
+        "backend_name": "ollama",
+        "model_names": null
+      }
+    }
+  },
+  {
+    "module": ".minichain.backends.impls.ollama.chat",
+    "attr": null,
+    "file": "ommlds/minichain/backends/impls/ollama/chat.py",
+    "line": 96,
+    "value": {
+      "!.minichain.registries.manifests.RegistryManifest": {
+        "module": "ommlds.minichain.backends.impls.ollama.chat",
+        "attr": "OllamaChatChoicesService",
+        "name": "ollama",
+        "aliases": null,
+        "type": "ChatChoicesService"
+      }
+    }
+  },
+  {
+    "module": ".minichain.backends.impls.ollama.chat",
+    "attr": null,
+    "file": "ommlds/minichain/backends/impls/ollama/chat.py",
+    "line": 143,
+    "value": {
+      "!.minichain.registries.manifests.RegistryManifest": {
+        "module": "ommlds.minichain.backends.impls.ollama.chat",
+        "attr": "OllamaChatChoicesStreamService",
+        "name": "ollama",
+        "aliases": null,
+        "type": "ChatChoicesStreamService"
+      }
+    }
+  },
   {
     "module": ".minichain.backends.impls.openai.chat",
     "attr": null,
@@ -450,7 +526,23 @@
     "module": ".minichain.backends.impls.transformers.transformers",
     "attr": null,
     "file": "ommlds/minichain/backends/impls/transformers/transformers.py",
-    "line": 43,
+    "line": 50,
+    "value": {
+      "!.minichain.backends.strings.manifests.BackendStringsManifest": {
+        "service_cls_names": [
+          "ChatChoicesService",
+          "ChatChoicesStreamService"
+        ],
+        "backend_name": "transformers",
+        "model_names": null
+      }
+    }
+  },
+  {
+    "module": ".minichain.backends.impls.transformers.transformers",
+    "attr": null,
+    "file": "ommlds/minichain/backends/impls/transformers/transformers.py",
+    "line": 66,
     "value": {
       "!.minichain.registries.manifests.RegistryManifest": {
         "module": "ommlds.minichain.backends.impls.transformers.transformers",
@@ -467,7 +559,7 @@
     "module": ".minichain.backends.impls.transformers.transformers",
     "attr": null,
     "file": "ommlds/minichain/backends/impls/transformers/transformers.py",
-    "line": 131,
+    "line": 197,
     "value": {
       "!.minichain.registries.manifests.RegistryManifest": {
         "module": "ommlds.minichain.backends.impls.transformers.transformers",
@@ -480,6 +572,21 @@
       }
     }
   },
+  {
+    "module": ".minichain.backends.impls.transformers.transformers",
+    "attr": null,
+    "file": "ommlds/minichain/backends/impls/transformers/transformers.py",
+    "line": 227,
+    "value": {
+      "!.minichain.registries.manifests.RegistryManifest": {
+        "module": "ommlds.minichain.backends.impls.transformers.transformers",
+        "attr": "TransformersChatChoicesStreamService",
+        "name": "transformers",
+        "aliases": null,
+        "type": "ChatChoicesStreamService"
+      }
+    }
+  },
   {
     "module": ".minichain.chat.choices.services",
     "attr": null,

ommlds/__about__.py

@@ -37,8 +37,8 @@ class Project(ProjectBase):
         ],
 
         'huggingface': [
-            'huggingface-hub ~= 0.35',
-            'datasets ~= 4.2',
+            'huggingface-hub ~= 0.36',
+            'datasets ~= 4.3',
         ],
 
         'numpy': [

ommlds/_hacks/__init__.py

@@ -0,0 +1,4 @@
+"""
+Currently, and ideally, the only thing this codebase will ever interact with requiring these kinds of awful hacks is ML
+code - thus these 'helpers' are kept here, not in the core library.
+"""

ommlds/_hacks/funcs.py

@@ -0,0 +1,110 @@
+import dataclasses as dc
+import linecache
+import textwrap
+import threading
+import types
+import typing as ta
+import uuid
+import warnings
+
+from omlish import check
+from omlish import lang
+
+from .names import NamespaceBuilder
+from .params import render_param_spec_call
+from .params import render_param_spec_def
+
+
+##
+
+
+@dc.dataclass()
+class _ReservedFilenameEntry:
+    unique_id: str
+    seq: int = 0
+
+
+_RESERVED_FILENAME_UUID_TLS = threading.local()
+
+
+def reserve_linecache_filename(prefix: str) -> str:
+    try:
+        e = _RESERVED_FILENAME_UUID_TLS.unique_id
+    except AttributeError:
+        e = _RESERVED_FILENAME_UUID_TLS.unique_id = _ReservedFilenameEntry(str(uuid.uuid4()))
+
+    while True:
+        unique_filename = f'<generated:{prefix}:{e.seq}>'
+        cache_line = (1, None, (e.unique_id,), unique_filename)
+        e.seq += 1
+        if linecache.cache.setdefault(unique_filename, cache_line) == cache_line:  # type: ignore
+            return unique_filename
+
+
+##
+
+
+def create_function(
+        name: str,
+        params: lang.CanParamSpec,
+        body: str,
+        *,
+        globals: ta.Mapping[str, ta.Any] | None = None,  # noqa
+        locals: ta.Mapping[str, ta.Any] | None = None,  # noqa
+        indent: str = '    ',
+) -> types.FunctionType:
+    params = lang.ParamSpec.of(params)
+    check.isinstance(body, str)
+    locals = dict(locals or {})  # noqa
+
+    nsb = NamespaceBuilder(reserved_names=set(locals) | set(globals or []))
+    sig = render_param_spec_def(params, nsb)
+    for k, v in nsb.items():
+        check.not_in(k, locals)
+        locals[k] = v
+
+    body_txt = '\n'.join([
+        f'def {name}{sig}:',
+        textwrap.indent(textwrap.dedent(body.strip()), indent),
+    ])
+
+    exec_txt = '\n'.join([
+        f'def __create_fn__({", ".join(locals.keys())}):',
+        textwrap.indent(body_txt, indent),
+        f'{indent}return {name}',
+    ])
+
+    ns: dict = {}
+    filename = reserve_linecache_filename(name)
+    bytecode = compile(exec_txt, filename, 'exec')
+    eval(bytecode, globals or {}, ns)  # type: ignore  # noqa
+
+    fn = ns['__create_fn__'](**locals)
+    fn.__source__ = body_txt
+    linecache.cache[filename] = (len(exec_txt), None, exec_txt.splitlines(True), filename)
+    return fn
+
+
+##
+
+
+def create_detour(
+        params: lang.CanParamSpec,
+        target: ta.Callable,
+        *,
+        as_kwargs: bool = False,
+) -> types.CodeType:
+    params = lang.ParamSpec.of(params)
+    check.callable(target)
+
+    with warnings.catch_warnings():
+        warnings.filterwarnings('ignore', category=SyntaxWarning)
+
+        gfn = create_function(
+            '_',
+            params,
+            f'return 1{render_param_spec_call(params, as_kwargs=as_kwargs)}',
+        )
+
+    check.state(gfn.__code__.co_consts[:2] == (None, 1))
+    return gfn.__code__.replace(co_consts=(None, target, *gfn.__code__.co_consts[2:]))

ommlds/_hacks/names.py

@@ -0,0 +1,158 @@
+"""
+TODO:
+ - releaseable names
+"""
+import string
+import typing as ta
+
+from omlish import check
+from omlish import collections as col
+
+
+##
+
+
+class NameGenerator(ta.Protocol):
+    def __call__(self, prefix: str = '') -> str: ...
+
+
+#
+
+
+class NameGeneratorImpl:
+    DEFAULT_PREFIX: ta.ClassVar[str] = '_'
+
+    def __init__(
+            self,
+            *,
+            reserved_names: ta.Iterable[str] | None = None,
+            global_prefix: str | None = None,
+            use_global_prefix_if_present: bool = False,
+            add_global_prefix_before_number: bool = False,
+    ) -> None:
+        super().__init__()
+
+        check.arg(not isinstance(reserved_names, str))
+        self._reserved_names = {check.isinstance(n, str) for n in (reserved_names or [])}
+        self._global_prefix = global_prefix if global_prefix is not None else self.DEFAULT_PREFIX
+        self._use_global_prefix_if_present = bool(use_global_prefix_if_present)
+        self._add_global_prefix_before_number = bool(add_global_prefix_before_number)
+
+        self._name_counts: dict[str, int] = {}
+
+    def __call__(self, prefix: str = '') -> str:
+        if self._use_global_prefix_if_present and prefix.startswith(self._global_prefix):
+            base_name = prefix
+        else:
+            base_name = self._global_prefix + prefix
+
+        base_count = -1
+        if base_name[-1] in string.digits:
+            i = len(base_name) - 2
+            while i >= 0 and base_name[i] in string.digits:
+                i -= 1
+            i += 1
+            base_count = int(base_name[i:])
+            base_name = base_name[:i]
+
+        if self._add_global_prefix_before_number:
+            if not (self._use_global_prefix_if_present and base_name.endswith(self._global_prefix)):
+                base_name += self._global_prefix
+
+        if base_count >= 0:
+            count = self._name_counts.setdefault(base_name, 0)
+            if base_count > count:
+                self._name_counts[base_name] = base_count
+
+        while True:
+            count = self._name_counts.get(base_name, 0)
+            self._name_counts[base_name] = count + 1
+            name = base_name + str(count)
+            if name not in self._reserved_names:
+                return name
+
+
+name_generator = NameGeneratorImpl
+
+
+##
+
+
+class NamespaceBuilder(ta.Mapping[str, ta.Any]):
+    def __init__(
+            self,
+            *,
+            reserved_names: ta.Iterable[str] | None = None,
+            name_generator: NameGenerator | None = None,  # noqa
+    ) -> None:
+        super().__init__()
+
+        self._reserved_names = {
+            check.isinstance(n, str)
+            for n in (check.not_isinstance(reserved_names, str) or [])
+        }
+        self._name_generator = check.callable(name_generator) if name_generator is not None else \
+            NameGeneratorImpl(reserved_names=self._reserved_names, use_global_prefix_if_present=True)
+
+        self._dct: ta.MutableMapping[str, ta.Any] = {}
+        self._dedupe_dct: ta.MutableMapping[ta.Any, str] = col.IdentityKeyDict()
+
+    @property
+    def reserved_names(self) -> ta.AbstractSet[str]:
+        return self._reserved_names
+
+    @property
+    def name_generator(self) -> NameGenerator:
+        return self._name_generator
+
+    def __getitem__(self, k: str) -> ta.Any:
+        return self._dct[k]
+
+    def __len__(self):
+        return len(self._dct)
+
+    def __iter__(self) -> ta.Iterator[str]:
+        return iter(self._dct)
+
+    def items(self) -> ta.ItemsView[str, ta.Any]:
+        return self._dct.items()
+
+    def put(
+            self,
+            value: ta.Any,
+            name: str | None = None,
+            *,
+            exact: bool = False,
+            dedupe: bool = False,
+    ) -> str:
+        check.arg(not (name is None and exact))
+        if name is not None:
+            check.isinstance(name, str)
+
+        if dedupe:
+            try:
+                return self._dedupe_dct[value]
+            except KeyError:
+                pass
+
+        if name is not None:
+            if name not in self._reserved_names:
+                try:
+                    existing = self._dct[name]
+                except KeyError:
+                    self._dct[name] = value
+                    if dedupe:
+                        self._dedupe_dct[value] = name
+                    return name
+                else:
+                    if existing is value:
+                        return name
+            if exact:
+                raise KeyError(name)
+
+        gen_name = self._name_generator(name or '')
+        check.not_in(gen_name, self._dct)
+        self._dct[gen_name] = value
+        if dedupe:
+            self._dedupe_dct[value] = gen_name
+        return gen_name

ommlds/_hacks/params.py

@@ -0,0 +1,73 @@
+import typing as ta
+
+from omlish import lang
+
+from .names import NamespaceBuilder
+
+
+##
+
+
+def render_param_spec_call(
+        params: lang.ParamSpec,
+        *,
+        as_kwargs: bool = False,
+) -> str:
+    src = ['(']
+
+    for i, p in enumerate(params):
+        if isinstance(p, lang.ParamSeparator):
+            continue
+
+        if i:
+            src.append(', ')
+
+        if as_kwargs:
+            if isinstance(p, lang.Param):
+                src.append(f'{p.name}={p.name}')
+            else:
+                raise TypeError(p)
+
+        else:
+            if isinstance(p, lang.ArgsParam):
+                src.append(f'*{p.name}')
+            elif isinstance(p, lang.KwargsParam):
+                src.append(f'**{p.name}')
+            elif isinstance(p, lang.PosOnlyParam):
+                src.append(p.name)
+            elif isinstance(p, lang.KwOnlyParam):
+                src.append(f'{p.name}={p.name}')
+            elif isinstance(p, lang.ValParam):
+                src.append(p.name)
+            else:
+                raise TypeError(p)
+
+    src.append(')')
+
+    return ''.join(src)
+
+
+def render_param_spec_def(
+        params: lang.ParamSpec,
+        nsb: NamespaceBuilder,
+        *,
+        return_ann: lang.Maybe[ta.Any] = lang.empty(),
+) -> str:
+    src = ['(']
+
+    for i, p in enumerate(params):
+        if i:
+            src.append(', ')
+
+        src.append(lang.param_render(
+            p,
+            render_annotation=lambda ann: nsb.put(ann, f'ann_{p.name}'),  # noqa
+            render_default=lambda dfl: nsb.put(dfl, f'dfl_{p.name}'),  # noqa
+        ))
+
+    src.append(')')
+
+    if return_ann.present:
+        src.append(f' -> {nsb.put(return_ann.must(), "ann_return")}')
+
+    return ''.join(src)

ommlds/_hacks/patches.py

@@ -1,7 +1,4 @@
 """
-Currently, and ideally, the only thing this codebase will ever interact with requiring these kinds of awful hacks is ML
-code - thus these 'helpers' are kept here, not in the core library.
-
 TODO:
  - patch lock
  - thread / context local gating

ommlds/backends/ollama/protocol.py

@@ -0,0 +1,170 @@
+"""
+https://docs.ollama.com/api
+"""
+import typing as ta
+
+from omlish import dataclasses as dc
+from omlish import lang
+
+
+##
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@dc.extra_class_params(default_repr_fn=dc.opt_repr)
+class Options:
+    # loading
+    numa: bool | None = None
+    num_ctx: int | None = None
+    num_batch: int | None = None
+    num_gpu: int | None = None
+    main_gpu: int | None = None
+    low_vram: bool | None = None
+    f16_kv: bool | None = None
+    logits_all: bool | None = None
+    vocab_only: bool | None = None
+    use_mmap: bool | None = None
+    use_mlock: bool | None = None
+    embedding_only: bool | None = None
+    num_thread: int | None = None
+
+    # querying
+    num_keep: int | None = None
+    seed: int | None = None
+    num_predict: int | None = None
+    top_k: int | None = None
+    top_p: float | None = None
+    tfs_z: float | None = None
+    typical_p: float | None = None
+    repeat_last_n: int | None = None
+    temperature: float | None = None
+    repeat_penalty: float | None = None
+    presence_penalty: float | None = None
+    frequency_penalty: float | None = None
+    mirostat: int | None = None
+    mirostat_tau: float | None = None
+    mirostat_eta: float | None = None
+    penalize_newline: bool | None = None
+    stop: ta.Sequence[str] | None = None
+
+
+##
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+class BaseRequest(lang.Abstract):
+    model: str
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+class BaseStreamableRequest(BaseRequest, lang.Abstract):
+    stream: bool | None = None
+
+
+##
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+class BaseGenerateRequest(BaseStreamableRequest, lang.Abstract):
+    options: Options | None = None
+    format: ta.Literal['', 'json'] | None = None  # TODO: jsonschema
+    keep_alive: float | str | None = None
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@dc.extra_class_params(default_repr_fn=dc.opt_repr)
+class GenerateRequest(BaseGenerateRequest):
+    prompt: str | None = None
+    suffix: str | None = None
+    system: str | None = None
+    template: str | None = None
+    context: ta.Sequence[int] | None = None
+    raw: bool | None = None
+    images: ta.Sequence[bytes] | None = None
+    think: bool | ta.Literal['low', 'medium', 'high'] | None = None
+
+
+#
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+class BaseGenerateResponse(lang.Abstract):
+    model: str | None = None
+    created_at: str | None = None
+    done: bool | None = None
+    done_reason: str | None = None
+    total_duration: int | None = None
+    load_duration: int | None = None
+    prompt_eval_count: int | None = None
+    prompt_eval_duration: int | None = None
+    eval_count: int | None = None
+    eval_duration: int | None = None
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@dc.extra_class_params(default_repr_fn=dc.opt_repr)
+class GenerateResponse(BaseGenerateResponse):
+    response: str
+    thinking: str | None = None
+    context: ta.Sequence[int] | None = None
+
+
+##
+
+
+Role: ta.TypeAlias = ta.Literal[
+    'system',
+    'user',
+    'assistant',
+    'tool',
+]
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@dc.extra_class_params(default_repr_fn=dc.opt_repr)
+class Message:
+    role: Role
+    content: str | None = None
+    thinking: str | None = None
+    images: ta.Sequence[bytes] | None = None
+    tool_name: str | None = None
+
+    @dc.dataclass(frozen=True, kw_only=True)
+    class ToolCall:
+        @dc.dataclass(frozen=True, kw_only=True)
+        class Function:
+            name: str
+            arguments: ta.Mapping[str, ta.Any]
+
+        function: Function
+
+    tool_calls: ta.Sequence[ToolCall] | None = None
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@dc.extra_class_params(default_repr_fn=dc.opt_repr)
+class Tool:
+    type: str | None = 'function'
+
+    @dc.dataclass(frozen=True, kw_only=True)
+    @dc.extra_class_params(default_repr_fn=dc.opt_repr)
+    class Function:
+        name: str | None = None
+        description: str | None = None
+        parameters: ta.Any | None = None
+
+    function: Function | None = None
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@dc.extra_class_params(default_repr_fn=dc.opt_repr)
+class ChatRequest(BaseGenerateRequest):
+    messages: ta.Sequence[Message] | None = None
+    tools: ta.Sequence[Tool] | None = None
+    think: bool | ta.Literal['low', 'medium', 'high'] | None = None
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@dc.extra_class_params(default_repr_fn=dc.opt_repr)
+class ChatResponse(BaseGenerateResponse):
+    message: Message