fixed needle rope bug, moved repeated rope table instantiations from …

[kar-m]

…components into weights

[kar-m]

@jakmro if you look at current model bundles for qwen, lfm, and gemma, they are pretty big outside of the weights because rope tables were being stored in the component folders. Some of these tables were being stored in several copies (gemma being the worst offender, storing 1.2gb worth of rope tables). Changed it so unique rope tables are stored in weights and just reused between components

[Copilot]

Pull request overview

This PR targets Needle long-context / RoPE stability and reduces duplicated RoPE table materialization across component graphs by externalizing shared tables into weight files, plus adds an engine-level Needle tool-calling test.

Changes:

• Update RoPE table precompute matching logic in optimize_graph.py (remove constant-ancestor tracing for inv_freq).
• Pad Needle input_ids up to a fixed context length (1024) when building component specs.
• Deduplicate/externalize c_rope_table_* bound constants into shared .weights files during bundle writing, and add a new Needle tool-call engine test.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
python/cactus/transpile/optimize_graph.py	Adjusts how RoPE precompute identifies the inv_freq constant input.
python/cactus/transpile/model_adapters.py	Pads Needle inputs to a fixed context length and builds attention masks.
python/cactus/transpile/hf_model.py	Externalizes/deduplicates shared RoPE tables into standalone weight files during bundle creation.
cactus-engine/tests/test_needle.cpp	Adds a Needle-specific engine regression test for forced tool calling.

Comments suppressed due to low confidence (1)

python/cactus/transpile/model_adapters.py:5342

• When pad_token_id is None and input_ids are padded up to NEEDLE_CONTEXT_LENGTH, the code sets attention_mask = torch.ones_like(...), which marks the newly padded tokens as real tokens. This changes the encoder inputs vs. the original shorter sequence and can break Needle models that rely on masking rather than a dedicated pad token. Build the mask using current_len so padded positions are 0 even when pad_token_id is unknown.

    if pad_token_id is None:
        attention_mask = torch.ones_like(input_ids, dtype=torch.int64)
    else:
        attention_mask = (input_ids != int(pad_token_id)).to(dtype=torch.int64)

---

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