fix(embedding): download embeddinggemma external-data sibling (.onnx_data)

[JeanBaptisteRenard]

Problem

MEMPALACE_EMBEDDING_MODEL=embeddinggemma crashes on first use on a clean cache:

onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL :
External data path validation failed for initializer: model.embed_tokens.weight_quantized.
External data path does not exist: .../onnx/model_quantized.onnx_data

The q8 export (onnx-community/embeddinggemma-300m-ONNX) keeps its weights in a sibling model_quantized.onnx_data file that the .onnx references by relative path. EmbeddinggemmaONNX._lazy_load() only hf_hub_downloads model_quantized.onnx, so the weights file is never fetched and onnxruntime can't resolve it.

Fix

Also fetch the .onnx_data sibling into the same snapshot dir (one extra hf_hub_download), wrapped in try/except so inlined-weight exports without a sibling stay non-fatal.

Repro / verification

• Before: clean HF cache → embeddinggemma → crash above (only model_quantized.onnx present in the snapshot onnx/ dir).
• After: the .onnx_data lands next to the .onnx; embeddinggemma loads and embeds. Verified the model works once both files are co-located — cross-lingual FR/EN cosine 0.89 (parallel) vs 0.40 (unrelated) on a 6-pair probe, matching the model card's multilingual claim.

No behavior change for minilm. Single-file additive fetch; honors the local-first principle (still HF-hosted weights, same as before — just the missing half).

[gemini-code-assist]

Code Review

This pull request updates mempalace/embedding.py to download the external-data sibling file for the quantized ONNX model if it exists, preventing runtime failures when weights are stored externally. The review feedback recommends catching a specific EntryNotFoundError from huggingface_hub.utils instead of a broad Exception to avoid swallowing unrelated errors, and suggests pinning the repository revision to ensure consistency between downloads.

[gemini-code-assist]

Specific Exception Handling & Revision Pinning

1. Avoid Broad Exception Catching: Catching Exception can swallow unrelated errors (such as network timeouts, DNS failures, permission issues, or disk full errors) and lead to confusing ONNX Runtime errors later. Instead, import and catch EntryNotFoundError from huggingface_hub.utils to specifically handle the case where the sibling file does not exist on the hub.
2. Revision Pinning Recommendation: Currently, the downloads do not specify a revision (defaulting to "main"). If a new commit is pushed to the repository between the download of the .onnx file and the .onnx_data file, they could be resolved to different commit hashes, leading to mismatched files or a missing sibling error. Pinning the revision to a specific commit hash (e.g., via a _EMBEDDINGGEMMA_REVISION constant) ensures consistency, prevents breaking changes from upstream updates, and allows huggingface_hub to load cached files instantly without making remote HEAD requests.

Suggested change

	try:
	hf_hub_download(
	_EMBEDDINGGEMMA_REPO,
	subfolder="onnx",
	filename=_EMBEDDINGGEMMA_ONNX + "_data",
	)
	except Exception:
	# Some exports inline their weights — a missing sibling is non-fatal.
	logger.debug("No external-data sibling for %s", _EMBEDDINGGEMMA_ONNX)
	try:
	from huggingface_hub.utils import EntryNotFoundError
	hf_hub_download(
	_EMBEDDINGGEMMA_REPO,
	subfolder="onnx",
	filename=_EMBEDDINGGEMMA_ONNX + "_data",
	)
	except EntryNotFoundError:
	# Some exports inline their weights — a missing sibling is non-fatal.
	logger.debug("No external-data sibling for %s", _EMBEDDINGGEMMA_ONNX)