Add vocab sphere with attn layers

[klei22]

No description provided.

[Copilot]

Pull request overview

This PR extends the hf-vocab-sphere tool with model-aware vector arithmetic (attention/MLP helper functions and norm helpers), adds vector “participation ratio”/effective-dimension metadata to projection points, and updates the frontend to expose these new controls and label components.

Changes:

• Expanded vector_math’s safe expression evaluator to support model-dispatched functions (norm, invnorm, wqk/wkq, attn_layer, mlp_layer, etc.) and finite SLERP extrapolation.
• Added effective-dimension metrics to API schemas and projection responses, and surfaced them in the UI as optional label components + hotkeys.
• Implemented auxiliary tensor loading/caching to support attention/MLP computations from safetensors.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
report/threejs/hf-vocab-sphere/tests/test_vector_math.py	Adds coverage for nested model functions, SLERP extrapolation, and effective-dimension metrics.
report/threejs/hf-vocab-sphere/tests/test_frontend_contract.py	Extends contract checks for new UI controls, settings keys, and hotkeys.
report/threejs/hf-vocab-sphere/app/vector_math.py	Adds vector metrics + model-function dispatch support to the safe evaluator.
report/threejs/hf-vocab-sphere/app/templates/index.html	Adds UI controls for layer-helper insertion, SLERP extrapolation input, and label toggles.
report/threejs/hf-vocab-sphere/app/static/app.js	Implements new label components, hotkeys, settings wiring, and expression insertion helpers.
report/threejs/hf-vocab-sphere/app/schemas.py	Extends API models with effective-dimension fields.
report/threejs/hf-vocab-sphere/app/model_store.py	Adds safetensors scanning/caching and model-vector helper implementations for attention/MLP ops.
report/threejs/hf-vocab-sphere/app/main.py	Wires model-function dispatch into arithmetic evaluation and adds effective-dimension metrics to points.

Comments suppressed due to low confidence (4)

report/threejs/hf-vocab-sphere/app/vector_math.py:185

• String literals are now supported as function selectors, but unary +/- still blindly multiplies operand.value by a float. For expressions like -'attn' this will raise a TypeError instead of a controlled ValueError (and can surface as a 500). Reject unary operators for string values explicitly.

        if isinstance(node, ast.UnaryOp) and isinstance(node.op, (ast.UAdd, ast.USub)):
            operand = self._visit(node.operand)
            sign = 1.0 if isinstance(node.op, ast.UAdd) else -1.0
            return _Value(operand.kind, operand.value * sign)

report/threejs/hf-vocab-sphere/app/model_store.py:1036

• Cache population for _AUX_TENSOR_CACHE is also unsynchronized. Even with a locked read, writes should be guarded too to keep cache state consistent and make behavior predictable under concurrency.

                        handle.get_tensor(str(name)).detach().cpu().to(dtype=torch.float32).contiguous()
                    )
    _AUX_TENSOR_CACHE[cache_key] = tensors
    return tensors

report/threejs/hf-vocab-sphere/app/vector_math.py:180

• String literals should only be usable as selectors inside model-function calls (e.g. norm(A, 0, 'attn', 'input')). As written, a string literal can appear anywhere in an expression and later hit arithmetic nodes, causing TypeError. Gate string-literal parsing behind an internal flag that is only enabled while visiting model-function arguments.

        if isinstance(node, ast.Constant) and isinstance(node.value, str):
            text = node.value.strip().casefold()
            if not text or len(text) > 40:
                raise ValueError("String arguments must be short, non-blank selectors.")
            return _Value("string", text)

report/threejs/hf-vocab-sphere/app/vector_math.py:213

• After restricting string literals to model-function argument lists, the model-function argument evaluation needs to temporarily allow strings while visiting node.args, then disable it again so strings cannot escape into the surrounding expression.

            if function_name in {"norm", "invnorm", "inverse_norm", "wov", "wqk", "wkq", "attn_layer", "mlp_layer"}:
                if self.model_function is None:
                    raise ValueError(f"{node.func.id} requires model auxiliary tensors loaded by the projection API.")
                values = [self._visit(argument) for argument in node.args]
                args: list[float | str | np.ndarray] = [
                    np.asarray(value.value)
                    if value.kind == "vector"
                    else (str(value.value) if value.kind == "string" else float(value.value))
                    for value in values
                ]
                return _Value("vector", self.model_function(function_name, args))

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.