Bugfix/move demo

.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,25 @@
+---
+name: Bug Report
+about: Create a report to help us improve
+title: ""
+labels: ["bug"]
+assignees: []
+---
+
+<!-- Rest of the template body here -->
+
+
+## What's the problem?
+...
+
+## Steps to Reproduce
+...
+
+## Screenshot
+If applicable.
+
+## Proposed Solution
+...
+
+## Notes
+...

.github/ISSUE_TEMPLATE/feature_request.md

@@ -0,0 +1,16 @@
+---
+name: Enhancement Request
+about: Ask for a change.
+title: ""
+labels: ["enhancement"]
+assignees: []
+---
+
+## Desired Change
+...
+
+## Rationale
+...
+
+## Notes
+...

.github/pull_request_template.md

@@ -0,0 +1,10 @@
+## Description
+...
+
+## Summary of Changes
+...
+
+Use ```Closes #X``` to automatically close an issue addressed by this PR.
+
+# Proof it Works
+...

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "llm"  # The pip install <name>.
-version = "0.4.0"
+version = "0.4.1"
 description = "Library for easy use of LLMs."
 readme = "README.md"
 authors = [
@@ -48,7 +48,7 @@ explicit = true  # Only used for torch packages explicitly mapped below.
 [tool.uv.sources]
 torch = { index = "pytorch-cu130" }
 torchvision = { index = "pytorch-cu130" }
-llm-conversation = { git = "https://github.com/EricApgar/llm-conversation", rev = "v0.2.0" }
+llm-conversation = { git = "https://github.com/EricApgar/llm-conversation" }
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]

src/llm/other/demo.py → scripts/demo.py

@@ -9,15 +9,15 @@
 
 if __name__ == '__main__':
 
-    # model = GptOss20b()
     model = GptOss20b()
+    model_cache_dir = input(prompt='Enter model cache dir: ')
     model.load(location=r'/home/eric/Repos/model_cache')
 
     c = Conversation()
-    c.set_overall_prompt(text='Your name is Seamus OFinnegan. Respond as the assistant in character.')
-    c.add_context(text='Youre a real salt of the earth Irish coal miner.')
-    c.add_context(text='You get straight to the point and dont waste words on small talk.')
-    c.add_context(text='You are married to your wife of 10 years, Gurdy. Your favorite hobby is fighting.')
+    c.set_overall_prompt(text='Your name is Samson McTavish. Respond in character.')
+    c.add_context(text="You're a 3rd year student at a school for magic.")
+    c.add_context(text='Your favorite class is potions.')
+    c.add_context(text='Your favorite spell is "Mimble Wimble".')
 
     # Loop:
     '''
@@ -35,6 +35,6 @@
 
         system_response = model.ask(prompt=c)
 
-        print(f'[Seamus]: {system_response}\n')
+        print(f'[Samson]: {system_response}\n')
 
         c.add_response(role='assistant', text=system_response)

src/llm/__init__.py

@@ -1,27 +1,10 @@
-'''
-Desired interface:
-
-import llm
-
-model = llm.model(
-    name='openai/gpt-oss-20b',
-    hf_token=<hf token>)
-model.load(
-    location=<path to save dir>,
-    remote=true, 
-    commit=<git commit>,
-    quantization='4-bit')
-response = model.ask(prompt='Tell me a joke.')
-'''
-
 from __future__ import annotations
 from typing import TYPE_CHECKING, Any
 
 
 __all__ = (
     "model",
-    "embedding",
-    'Conversation')
+    "embedding")
 
 
 if TYPE_CHECKING:
@@ -40,11 +23,6 @@ def __getattr__(name: str) -> Any:
         from .models.embed import EmbeddingModel
         globals()[name] = EmbeddingModel
         return EmbeddingModel
-
-    if name == 'Conversation':
-        from .other.conversations import Conversation
-        globals()[name] = Conversation
-        return Conversation
 
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
 

src/llm/models/embed.py

@@ -1,14 +1,21 @@
+import os
+
 from sentence_transformers import util, SentenceTransformer
 import numpy as np
 import torch
 
+from llm.models.selection import SUPPORTED_EMBEDDING_MODELS
+
 
 class EmbeddingModel:
 
-    def __init__(self, hf_token: str|None=None):
+    def __init__(self, name: str='all-mpnet-base-v2', hf_token: str=None):
 
+        if name not in SUPPORTED_EMBEDDING_MODELS:
+            raise ValueError(f'Invalid embedding model "{self.name}". Supported: {SUPPORTED_EMBEDDING_MODELS}.')
+
+        self.name: str = name
         self.hf_token: str = hf_token
-        self.name: str = None
 
         self.location: str = None
         self.remote: bool = False
@@ -20,23 +27,19 @@ def __init__(self, hf_token: str|None=None):
 
     def load(self,
         location: str,
-        name: str='all-mpnet-base-v2',
         remote: bool=False,
         commit: str=None) -> None:
 
-        VALID_MODELS = ['all-mpnet-base-v2', 'all-MiniLM-L6-v2']
+        if (not remote) and (not os.path.isdir(location)):
+            raise ValueError(f'Nonexistant location ({location}) - fix or set remote=True.')
 
-        if name not in VALID_MODELS:
-            raise ValueError(f'Invalid embedding model "{name}". Supported: {VALID_MODELS}')
-
         self.location = location
-        self.name = name
         self.remote = remote
         self.commit = commit
 
         self.model = SentenceTransformer(
             cache_folder=location,
-            model_name_or_path=name,
+            model_name_or_path=self.name,
             local_files_only=not self.remote,
             trust_remote_code=self.remote,
             token=self.hf_token)
@@ -72,10 +75,19 @@ def get_similarity(self,
             similarity = self._torch_dot_similarity(v1=v1, v2=v2)
 
         return similarity
-
+
+
     @staticmethod
     def _dot_similarity(v1: np.array, v2: np.array) -> float:
 
+        result = torch.dot(v1, v2)
+
+        return result
+
+
+    @staticmethod
+    def _cosine_similarity(v1: np.array, v2: np.array) -> float:
+
         dot_product = torch.dot(v1, v2)
 
         norm_vector_1 = torch.sqrt(torch.sum(v1**2))
@@ -95,12 +107,13 @@ def _torch_dot_similarity(v1: np.array, v2: np.array) -> float:
 
 
 if __name__ == '__main__':
-    model = EmbeddingModel()
-    model.load(location=r'<path to model cache>')
-    e1 = model.embed(text='What shape is best?')
-    e2 = model.embed(text='Hexagons are the bestagons.')
-    similarity = model.get_similarity(v1=e1, v2=e2)
 
-    print(similarity)
+    # model = EmbeddingModel()
+    # model.load(location=<path to model cache>)  # NOTE: set <path to model cache>.
+    # e1 = model.embed(text='What shape is best?')
+    # e2 = model.embed(text='Hexagons are the bestagons.')
+    # similarity = model.get_similarity(v1=e1, v2=e2)
+
+    # print(similarity)
 
     pass

src/llm/models/gpt_oss_20b.py

@@ -117,10 +117,10 @@ def _format_prompt(prompt: str | Conversation, reasoning_level: str=None) -> lis
 
 if __name__ == '__main__':
 
-    model = GptOss20b()
-    model.load(location=r'/home/eric/Repos/model_cache')
-    response = model.ask(prompt='Name a primary color.')
+    # model = GptOss20b()
+    # model.load(location=<path to model cache>)  # NOTE: set <path to model cache>.
+    # response = model.ask(prompt='Name a primary color.')
 
-    print(response)
+    # print(response)
 
     pass

src/llm/models/gpt_oss_20b_dev.py

@@ -204,10 +204,10 @@ def get_good_token_start(token_list: list[str]) -> int:
 
 if __name__ == '__main__':
 
-    model = GptOss20bDev()
-    model.load(location=r'/home/eric/Repos/model_cache')
-    response = model.ask(prompt='Name a primary color.')
+    # model = GptOss20bDev()
+    # model.load(location=<path to model cache>)  # NOTE: set <path to model cache>.
+    # response = model.ask(prompt='Name a primary color.')
 
-    print(response)
+    # print(response)
 
     pass

src/llm/models/phi4_multimodal_instruct.py

@@ -198,23 +198,23 @@ def get_usable_length(self, new_seq_length: int, layer_idx: int=0) -> int:
 
 if __name__ == '__main__':
 
-    model = Phi4MultimodalInstruct()
-    model.load(location=r'/home/eric/Repos/model_cache')  # NOTE: set <path to model cache>.
-
-    response = model.ask(prompt='Name a primary color. Be brief.', max_tokens=256)
-    print(f'{response}\n')
-
-    convo = Conversation()
-    convo.set_overall_prompt(text='You are a helpful assistant.')
-    convo.add_context(text='Your favorite color is red.')
-    convo.add_context(text='Your favorite shape is the hexagon.')
-    convo.add_response(role='user', text='What is your favorite color-shape combination?')
-    response = model.ask(prompt=convo, max_tokens=256)
-    print(f'{response}\n')
-
-    from PIL import Image as PillowImage
-    image = PillowImage.open(r'/home/eric/Desktop/monkey.png')  # NOTE: Point to existing image.
-    response = model.ask(prompt='Describe the image.', images=[image], max_tokens=256)
-    print(f'{response}\n')
+    # model = Phi4MultimodalInstruct()
+    # model.load(location=<path to model cache>)  # NOTE: set <path to model cache>.
+
+    # response = model.ask(prompt='Name a primary color. Be brief.', max_tokens=256)
+    # print(f'{response}\n')
+
+    # convo = Conversation()
+    # convo.set_overall_prompt(text='You are a helpful assistant.')
+    # convo.add_context(text='Your favorite color is red.')
+    # convo.add_context(text='Your favorite shape is the hexagon.')
+    # convo.add_response(role='user', text='What is your favorite color-shape combination?')
+    # response = model.ask(prompt=convo, max_tokens=256)
+    # print(f'{response}\n')
+
+    # from PIL import Image as PillowImage
+    # image = PillowImage.open(r'/home/eric/Desktop/monkey.png')  # NOTE: Point to existing image.
+    # response = model.ask(prompt='Describe the image.', images=[image], max_tokens=256)
+    # print(f'{response}\n')
 
     pass

src/llm/models/selection.py

@@ -1,16 +1,32 @@
-'''
-Import the specific model named by the user.
-'''
+from typing import Literal
+
+from .gpt_oss_20b import GptOss20b
+from .phi4_multimodal_instruct import Phi4MultimodalInstruct
+
+
+SUPPORTED_LLMS = {
+    GptOss20b().name: GptOss20b,
+    Phi4MultimodalInstruct().name: Phi4MultimodalInstruct
+}
+
+SUPPORTED_EMBEDDING_MODELS = ['all-MiniLM-L6-v2', 'all-mpnet-base-v2']
+
 
 def model(name: str, hf_token: str=None):
 
-    if name == 'openai/gpt-oss-20b':
-        from .gpt_oss_20b import GptOss20b
-        return GptOss20b(hf_token=hf_token)
+    if name not in SUPPORTED_LLMS:
+        raise ValueError(f'Model "{name}" not supported! Call "list_models()" for options.')
+
+    return SUPPORTED_LLMS[name](hf_token=hf_token)
+
 
-    elif name == 'microsoft/Phi-4-multimodal-instruct':
-        from .phi4_multimodal_instruct import Phi4MultimodalInstruct
-        return Phi4MultimodalInstruct(hf_token=hf_token)
+def list_models(kind: Literal['llm', 'embed']):
 
+    if kind == 'llm':
+        supported_models = list(SUPPORTED_LLMS.keys())
+    elif kind == 'embed':
+        supported_models = SUPPORTED_EMBEDDING_MODELS
     else:
-        raise ValueError(f'Model "{name}" not supported!')
+        raise ValueError(f'Unsupported "kind" ({kind}). Must be "llm" or "embed".')
+
+    return supported_models

tests/test_embed.py

@@ -0,0 +1,43 @@
+"""
+Tests for EmbeddingModel.
+
+These tests load a real model and call embed() and get_similarity(), so they
+require local model weights. Sync the environment first, then pass
+LLM_MODEL_CACHE inline when invoking pytest so it only exists for that one
+command and does not persist in your shell:
+
+    uv sync --extra openai
+    LLM_MODEL_CACHE=<path to model cache dir> pytest
+
+Tests are skipped automatically if LLM_MODEL_CACHE is not set.
+
+Make sure the virtual environment is active.
+"""
+
+import os
+
+import pytest
+import torch
+
+from llm.models.embed import EmbeddingModel
+
+
+MODEL_CACHE = os.environ.get('LLM_MODEL_CACHE')
+
+
+@pytest.fixture(scope='module')
+def model():
+    if not MODEL_CACHE:
+        pytest.skip('LLM_MODEL_CACHE environment variable not set.')
+    m = EmbeddingModel()
+    m.load(location=MODEL_CACHE)
+    yield m
+    del m
+    torch.cuda.empty_cache()
+
+
+def test_get_similarity(model):
+    e1 = model.embed(text='What shape is best?')
+    e2 = model.embed(text='Hexagons are the bestagons.')
+    similarity = model.get_similarity(v1=e1, v2=e2)
+    assert isinstance(similarity, (int, float))