Minor fixes

Author: matanbt

.gitignore

@@ -24,3 +24,10 @@ tropt/recipe_hub/PGD.py
 tropt/recipe_hub/PGD__geisler2024.py
 tropt/optimizer/pgd_optimizer.py
 
+# Local artifacts & dev scratch — never tracked or published
+wandb/
+logs/
+claude_stuff/
+.claude/
+*.egg-info/
+

CONTRIBUTING.md

@@ -58,7 +58,7 @@ Open an issue with `[feature]` or `[recipe]` before writing code, especially for
 A contribution that touches a component must conform to its interface contract. Read in this order:
 
 - [`DESIGN.md`](DESIGN.md) — full design philosophy (modularity, backend vs. frontend, why the contracts are the way they are).
-- [`docs/guides/`](docs/guides/) — step-by-step per-component walkthroughs (source of truth for each contribution type).[TODO link to the actual docs]
+- [`docs/guides/`](docs/guides/) — step-by-step per-component walkthroughs (source of truth for each contribution type).
 - **Agent files.**
     - [`CLAUDE.md`](CLAUDE.md) — fast-path orientation: the four components, mixin contract, `ModelInput`/`ModelOutput`/`Targets`, setup-then-compute.
     - [`skills/tropt/SKILL.md`](skills/tropt/SKILL.md) — task-routed pointers and cross-cutting pitfalls (mixin mismatches, thinking-model alignment, multi-model OOM, etc.).
@@ -75,7 +75,7 @@ A contribution that touches a component must conform to its interface contract.
 
 For every contribution:
 
-1. **Mirror the source layout in `tests/`.** Cover output shape, sign conventions, mixin requirements, and known input/output pairs. See [`TESTING.md`](TESTING.md) for tolerances and per-component checklists.
+1. **Add tests** following the conventions and per-component checklists in [`TESTING.md`](TESTING.md); match the layout of the existing tests under [`tests/`](tests/). Cover output shape, sign conventions, mixin requirements, and known input/output pairs. Note that the end-to-end suite runs tiny CPU models with few steps, so tests assert shapes and finiteness — not loss decrease or trigger content, which are flaky at that scale.
 2. **Regenerate the compatibility matrix** if you touched a model, optimizer, mixin, or loss `require_*` flag: `python docs/scripts/generate_compat_matrix.py`. Commit the regenerated [`docs/guides/compatibility_matrix.md`](docs/guides/compatibility_matrix.md).
 3. **Update the relevant README row** — recipes need an entry in [`tropt/recipe_hub/README.md`](tropt/recipe_hub/README.md) (key, description, target model, required access, paper, file).
 4. **Smoke-test end-to-end** on a small model (e.g. `google/gemma-3-270m-it`) before submitting.

DESIGN.md

@@ -287,8 +287,6 @@ The repository exposes two interfaces for managing recipes:
 
 <!-- * **Config Runner [`runner/main.py`].** A flexible runner that constructs a recipe from a YAML configuration file. The runner uses [Hydra](https://hydra.cc/) to manage configurations, allowing users to specify the model, loss, optimizer, and their parameters in a structured way without writing new code. -->
 
-* **Full evaluations [WIP].** [TODO]
-
 ## Summary
 
 

docs/_static/custom.css

@@ -406,7 +406,17 @@ html[data-theme="dark"] .tropt-cta-secondary:hover {
 /* (Dot-grid layer is now `.tropt-hero-dots` defined above with the hero.) */
 
 @media (max-width: 720px) {
-  .tropt-optbar { gap: 6px; padding: 8px 10px; }
+  /* The slot-machine row is wider than a phone viewport when forced onto one
+     line, so it gets clipped at the screen edge (the trailing "!" disappears)
+     and reads as a broken banner. Let it wrap onto multiple lines and never
+     exceed the viewport. */
+  .tropt-optbar {
+    gap: 6px;
+    padding: 8px 10px;
+    flex-wrap: wrap;
+    max-width: 100%;
+    box-sizing: border-box;
+  }
   .tropt-tok { font-size: 12.5px; padding: 7px 10px; }
   .tropt-tok-opt { min-width: 100px; }
 }

pyproject.toml

@@ -12,6 +12,28 @@ license = {text = "MIT"}
 authors = [
   {name = "Matan Ben-Tov"}
 ]
+keywords = [
+  "adversarial-attacks",
+  "jailbreak",
+  "red-teaming",
+  "trigger-optimization",
+  "discrete-optimization",
+  "prompt-optimization",
+  "llm",
+  "nlp",
+  "robustness",
+  "ai-safety",
+]
+classifiers = [
+  "Development Status :: 3 - Alpha",
+  "Intended Audience :: Science/Research",
+  "License :: OSI Approved :: MIT License",
+  "Operating System :: OS Independent",
+  "Programming Language :: Python :: 3.10",
+  "Programming Language :: Python :: 3.11",
+  "Programming Language :: Python :: 3.12",
+  "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
 
 # Core dependencies - minimal for basic HuggingFace model optimization
 dependencies = [
@@ -37,8 +59,9 @@ dependencies = [
 ]
 
 [project.urls]
-Repository = "https://github.com/matanbt/tropt"
-Issues = "https://github.com/matanbt/tropt/issues"
+Homepage = "https://matanbt.github.io/TROPT"
+Repository = "https://github.com/matanbt/TROPT"
+Issues = "https://github.com/matanbt/TROPT/issues"
 
 [project.optional-dependencies]
 # API model integrations (install separately as needed)
@@ -134,6 +157,15 @@ allow-direct-references = true
 [tool.hatch.build.targets.wheel]
 packages = ["tropt"]
 
+# Explicit allowlist so the sdist ships only the package + metadata.
+[tool.hatch.build.targets.sdist]
+include = [
+  "tropt/",
+  "README.md",
+  "LICENSE",
+  "pyproject.toml",
+]
+
 [tool.ruff]
 include = ["tropt/**/*.py", "tests/**/*.py"]
 extend-exclude = ["tests/_legacy"]

tropt/__init__.py

@@ -0,0 +1,9 @@
+# Expose tropt.__version__, read from install metadata (pyproject) with a source-tree fallback.
+import importlib.metadata as _metadata
+
+try:
+    __version__ = _metadata.version("tropt")
+except _metadata.PackageNotFoundError:  # running from a source tree without install
+    __version__ = "0.0.0+unknown"
+
+del _metadata

tropt/common.py

@@ -231,7 +231,7 @@ def to_device(self, device: torch.device | str) -> "Targets":
         for k, v in self:
             if isinstance(v, Tensor):
                 updates[k] = v.to(device)
-            if isinstance(v, list) and isinstance(v[0], Tensor):
+            elif isinstance(v, list) and v and isinstance(v[0], Tensor):
                 updates[k] = [t.to(device) for t in v]
         return self.model_copy(update=updates)
 

tropt/model/google/encoder.py

@@ -58,31 +58,55 @@ class EncoderGeminiModel(EncoderBaseModel, LossTextAccessMixin):
     """
 
     def __init__(
-        self, model_name="gemini-embedding-001", d_model: int = 3072, **kwargs
+        self,
+        model_name="gemini-embedding-001",
+        d_model: int = 3072,
+
+        # Vertex configuration:
+        use_vertex: bool = False,
+        project: Optional[str] = None,
+        location: str = "us-central1",
+        default_text_type: Optional[str] = None,
+        **kwargs,
     ):
         """
         Initializes the Gemini Encoder Model wrapper.
 
         Args:
             model_name: The name of the Gemini embedding model to use.
             d_model: The dimensionality of the embeddings (e.g., 768, 3072).
+            use_vertex: Embed via the Vertex AI backend (ADC) instead of AI Studio.
+                Needed for Vertex-only models such as ``text-embedding-005``.
+            project: Vertex project (only used when ``use_vertex``; falls back to the
+                ``GOOGLE_CLOUD_PROJECT`` env var when None). ``location`` is the region.
+            default_text_type: Fallback ``text_type`` ("document"/"query") used when a
+                caller doesn't pass one — e.g. the optimizer's
+                ``compute_loss_from_texts``, so candidates embed as documents.
 
         Note:
-        Requires `os.environ["GOOGLE_API_KEY"]` to be set externally.
-
+            AI Studio backend requires ``os.environ["GOOGLE_API_KEY"]``; Vertex backend
+            requires Application Default Credentials.
         """
         # Import google.genai only when instantiating (optional dependency)
         from google import genai
 
-        self._client = genai.Client(
-            http_options=genai.types.HttpOptions(timeout=_REQUEST_TIMEOUT_MS),
-        )
+        http_options = genai.types.HttpOptions(timeout=_REQUEST_TIMEOUT_MS)
+        if use_vertex:
+            client_kwargs = {"vertexai": True, "location": location,
+                             "http_options": http_options}
+            if project:
+                client_kwargs["project"] = project
+            self._client = genai.Client(**client_kwargs)
+        else:
+            self._client = genai.Client(http_options=http_options)
         self.model_name = model_name
         self._d_model = d_model  # for gemini-embedding-001: could be 768, 1536, or 3072
+        self._default_text_type = default_text_type
         self._text_to_task_type = {
             "document": "RETRIEVAL_DOCUMENT",
             "query": "RETRIEVAL_QUERY",
         }
+        self._max_batch = 100
 
     @property
     def d_model(self) -> int:
@@ -104,6 +128,8 @@ def invoke_from_texts(
         Returns:
             A ModelOutput containing the generated embeddings.
         """
+        if text_type is None:
+            text_type = self._default_text_type
         assert text_type in (
             None,
             "document",
@@ -113,8 +139,6 @@ def invoke_from_texts(
 
         import google.genai as genai
 
-        # Gemini's BatchEmbedContents caps at 100 requests per call; chunk.
-        MAX_BATCH = 100
         all_embeddings = []
         total_tokens = 0
         cfg = genai.types.EmbedContentConfig(
@@ -134,8 +158,8 @@ def _embed_chunk(chunk):
                 contents=chunk, model=self.model_name, config=cfg,
             )
 
-        for start in range(0, len(input_texts), MAX_BATCH):
-            chunk = input_texts[start : start + MAX_BATCH]
+        for start in range(0, len(input_texts), self._max_batch):
+            chunk = input_texts[start : start + self._max_batch]
             response = _embed_chunk(chunk)
             assert response is not None and response.embeddings is not None, (
                 "embed_content returned no embeddings"

tropt/model/huggingface/lm.py

@@ -531,6 +531,7 @@ def invoke_from_texts(
         ).to(self.device)
         padded_seq_len = inputs.input_ids.shape[1]
         n_prompt_tokens = inputs.input_ids.numel()
+        # n_prompt_tokens = int(inputs.attention_mask.sum().item())  # [Disabled] an alternative that excludes padding tokens
 
         # ---- Shared forward pass (prefill logits and/or first-token logprobs) ----
         prefill_response_logits = None

tropt/optimizer/gcg_optimizer.py

@@ -70,6 +70,9 @@ def __init__(
         self.token_constraints = token_constraints
         self.use_retokenize = use_retokenize
 
+        # validations:
+        assert self.loss_func.is_differentiable, "GCGOptimizer requires a differentiable loss function."
+
     def optimize_trigger(
         self,
         templates: TextTemplates,