Llama.embed() calls LlamaBatch.add_sequence with old 3-arg signature; missing logits_array

L’issue si trova in llama_cpp/llama.py intorno alla riga 1678, dove Llama.embed() chiama _batch.add_sequence con tre argomenti. Confrontala con la chiamata corretta a quattro argomenti in llama_cpp/llama_embedding.py intorno alla riga 262. Aggiorna la chiamata in modo che corrisponda alla firma, usando l’array dei token, l’array delle posizioni, gli ID delle sequenze e l’array dei logits. Esegui il test eseguendo lo script di riproduzione fornito con un modello che supporta gli embeddings.

Prerequisites

• I am running the latest code. Development is very rapid so there are no tagged versions as of now.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new bug or useful enhancement to share.

Expected Behavior

Llama.embed() should successfully compute embeddings when called on a model constructed with embeddings=True.

Current Behavior

Llama.embed() raises a TypeError immediately, before any embedding is computed:

TypeError: LlamaBatch.add_sequence() missing 1 required positional argument: 'logits_array'

The cause: Llama.embed() in llama_cpp/llama.py (around line 1678) calls add_sequence with three positional arguments:

self._batch.add_sequence(tokens, p_batch, logits_all)

But LlamaBatch.add_sequence in llama_cpp/_internals.py (around line 1013) requires four:

def add_sequence(
    self,
    token_array: Sequence[int],
    pos_array: Sequence[int],
    seq_ids: Sequence[Sequence[int]],
    logits_array: Sequence[bool]
)

llama_cpp/llama_embedding.py (around line 262) already calls add_sequence correctly with the four-arg shape — the call site in Llama.embed() was apparently missed during the LlamaBatch.add_sequence refactor.

Environment and Context

• Hardware: x86_64, NVIDIA GeForce RTX 4090
• OS: Windows 10 22H2
• Python 3.12.9
• llama-cpp-python 0.3.36 (CUDA 12.8 prebuilt wheel)

$ python --version
Python 3.12.9

$ pip show llama-cpp-python | findstr Version
Version: 0.3.36

Failure Information (for bugs)

This is a clean regression — LlamaBatch.add_sequence was refactored from a 3-arg signature to a 4-arg one, and the call sites were updated everywhere except in Llama.embed(). llama_embedding.py shows what the new shape should look like for the embedding code path.

Steps to Reproduce

from llama_cpp import Llama

m = Llama(model_path="path/to/model.gguf", embeddings=True)
m.embed("hello")

Result:

TypeError: LlamaBatch.add_sequence() missing 1 required positional argument: 'logits_array'

Failure Logs

Traceback (most recent call last):
  File "...\Lib\site-packages\llama_cpp\llama.py", line 1678, in embed
    self._batch.add_sequence(tokens, p_batch, logits_all)
TypeError: LlamaBatch.add_sequence() missing 1 required positional argument: 'logits_array'

Suggested fix

Mirror the call shape already used in llama_cpp/llama_embedding.py:

# In llama.py Llama.embed(), replace:
self._batch.add_sequence(tokens, p_batch, logits_all)

# With something like:
self._batch.add_sequence(
    token_array=tokens,
    pos_array=list(range(len(tokens))),
    seq_ids=[p_batch],
    logits_array=[True] * len(tokens) if logits_all else [False] * (len(tokens) - 1) + [True],
)

Workaround

Monkey-patching LlamaBatch.add_sequence to detect 3-arg legacy calls and synthesize the missing pos_array works as a stopgap. Hit while running Tencent's HY-Motion text-to-motion model, whose text encoder uses Llama.embed() against GGUF Qwen3 weights.