`pip install "python>=99.0"` isn't a command anyone can run.

Greedy packing keeps each chunk's per_item_token * width * count under the budget, isolates a long outlier in its own chunk, and never emits an em

CachedMultipleNegativesRankingLoss composes an inner MultipleNegativesRankingLoss, but the documented attributes and ``get_config_dict`` keys must

The ``gather_across_devices`` path has no other test. Simulate world=2/rank=1 by having ``all_gather_with_grad`` return ``[local; local]``: the ga

The delegating properties must keep supporting assignment, as the plain attributes did before the rebase. That includes assigning an ``nn.Module``

``MatryoshkaLoss(CachedMultipleNegativesRankingLoss(...))``, the combination the visual document retrieval example ships with, must match ``Matryo

With ``mini_batch_num_tokens``, the embedding passes pack by token count, but the loss and gradient must still exactly match MultipleNegativesRank

KD scores are each query's own document group, the query-major block diagonal of the full matrix. Split out of the bfloat16 test below so it keeps

Every colbert scorer takes and forwards chunk_elements and length_normalize, so one kwargs dict works across the family and the default training p

The losses' default similarity_fct surface must keep maxsim's float32 scores, also under a future fused reimplementation.

The collator assigns ``task`` by column POSITION (column 0 = query, the rest = document) to match the losses, which score positionally. Column nam

A fresh MVE has ``document_length=None``. Before the fix the collator forced ``padding="max_length"`` which silently padded every batch to ``token

``router_mapping`` overrides the positional default per column.

The collator stamps the task each column was tokenized with as ``{column}_task`` so the losses re-run the model under the same task (instead of re

Within each column every row shares ``T`` (the loss reshapes per-column on ``dim=0``), but columns are independently padded to their own batch-lon

Test specifying custom column names.

A model with tuple modality ("image", "text") builds multimodal dicts from text+image columns.

A Kosmos-like model with only ("image", "text") still builds combined dicts.

Test case where all points form a single community due to a low threshold.

Test case for GPU support (if available).

Test case with a large dataset and batching.

Test case where communities are filtered based on minimum size.

Test case where no communities are found due to a high threshold.

Test case where input is a numpy array instead of a torch tensor.

Test case with overlapping communities (resolved by the function).

Members whose similarity equals the threshold exactly must not be truncated. The candidate window grew only while the smallest candidate was stri

Test case with two clear communities.

A dict with multiple modalities (e.g. text+image) should be expanded into separate content items, not wrapped as a single compound-typed item.

``.config`` and ``.model`` must degrade to ``None`` when the module stack has no underlying HuggingFace Transformer rather than raising AttributeE

`model.config` should return the underlying transformers PretrainedConfig so integrations like Deepspeed and transformers.Trainer can read `hidden

Assigning ``model.config`` should not raise and should round-trip through the underlying transformers model (fixes optimum ONNX export regression

If the model has no underlying transformers PreTrainedModel (e.g. StaticEmbedding-only), `model.config` returns `None` instead of raising.

Assigning ``model.config`` on a StaticEmbedding-only model (no underlying transformers model) should be silently ignored, i.e. ``model.config`` ke

https://github.com/huggingface/sentence-transformers/issues/3874

The same encoded source produces the same hash.

Non-string, non-list content is treated as non-text.

The SentenceTransformer branch reuses the saved modules and appends a SparseAutoEncoder (CSR), so the source's prompts and similarity stay meaning

Converting a SparseEncoder (or CrossEncoder) save rebuilds the default MVE modules, so the source's prompts and default_prompt_name are not inheri

Converting a CrossEncoder save rebuilds this family's default modules, so the source's prompts and default_prompt_name no longer describe the load

Converting a SentenceTransformer save appends a fresh classification head, so the source's embedding prompts and default_prompt_name are not inher

Converting a MultiVectorEncoder save rebuilds this family's default modules, so the source's prompts, default_prompt_name and similarity_fn_name a

A SparseEncoder save carries similarity_fn_name="dot", chosen for sparse embeddings. The converted dense model scores with the dense default inste

A dense SentenceTransformer is converted to a MultiVectorEncoder on load. Its saved ``similarity_fn_name`` ("cosine" / "dot" can't score ragged pe

Converting a dense SentenceTransformer WITH a Dense head (LaBSE-shape) redirects the head to token level: the conversion drops the Pooling, so sen

rank no longer declares them, so they are dropped as no-ops and the warning names rank.

predict still declares both, so they must pass through untouched and unwarned.

Variable-length embeddings cannot stack, so unlike every other model type there is no `convert_to_tensor` here. Copied-over calls are common enoug

Negative strides were never convertible and still are not. `torch.from_numpy` rejects them with the same error `torch.tensor` raised before the vi

Read-only buffers (memmaps, broadcast views) take the copying path instead of torch's non-writable-tensor warning.

A list of arrays (what encoding with `convert_to_numpy=True` returns) went through `torch.tensor`, which reads it one element at a time. Stacking

A numpy corpus is viewed, not copied: the copy costs as much as the scoring it feeds.

Test cosine similarity between sparse and dense representations.

The exponent is s(k,l) - s(i,j): the score of the pair with the *lower* expected similarity minus the score of the pair with the higher one, so ra

Non-text pairs reach pair_to_messages by a different route, but take the same roles.

Leading padding is left in place: trimming it would shift the positions a model derives from the sequence length.

ColBERT expansion positions carry attention_mask 0 but are still scored, so the trailing trim must not drop them: doing so silently shrinks a 32-t

A padded mini-batch of short sequences must be embedded at its own width, not the whole batch's padded width, or mini_batch_num_tokens would not b

Test using a cross-encoder for rescoring.

Test using a cross-encoder with different parameters.

Multi-label: shape includes num_labels, no rank section.

Multimodal pairs use _format_snippet_value for non-text elements and skip rank section.

run_usage_snippet with multimodal pairs gracefully falls back when predict fails.

Single-label: shape is (n,), rank section present.

data: URIs bypass the URL space check and should still be detected as images.