fix(embed): mark all tokens for output to suppress llama.cpp 'overriding' warning (#2208)

[Anai-Guo]

Summary

Fixes #2208.

When Llama.embed() is called with a model whose pooling type is not NONE, every input token's logits flag is set to False except the last (in LlamaBatch.add_sequence). When llama.cpp later runs the embedding pass it requires every embedded token to have its output flag enabled, so it emits one info line per input:

init: embeddings required but some input tokens were not marked as outputs -> overriding

then forces all tokens on internally. The output is correct but the log is noisy — see also the matching ollama issue ollama/ollama#12381 referenced in the bug report.

Fix

Set logits_all = True unconditionally inside embed() so the Python side marks every token, matching what llama.cpp does internally. No behavioural change for LLAMA_POOLING_TYPE_NONE (already True); for the other pooling modes the override-warning loop is suppressed.

Test plan

• model.embed(texts) with a pooled embedding model (e.g. nomic-embed-text-v1.5) — confirm the overriding lines are gone and returned vectors are unchanged.
• model.embed(text, normalize=True) single-string call — output identical.
• model.embed(texts) with a LLAMA_POOLING_TYPE_NONE model — per-token embeddings unchanged.

🤖 Generated with Claude Code