fix(embed): mark all tokens as output to suppress llama.cpp 'overriding' INFO (#2208) (#2212)

Author: Anai-Guo

CHANGELOG.md

@@ -10,6 +10,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - feat: Update llama.cpp to ggerganov/llama.cpp@5d6f18a63 and sync Python bindings
 - fix: Correct batched embedding outputs for multi-sequence `embed()` calls by @Anai-Guo in #2205
 - fix: Configure embedding contexts with enough sequence slots for batched `embed()` calls
+- fix: Mark all embedding input tokens as outputs to avoid llama.cpp override warnings by @Anai-Guo in #2212
 
 ## [0.3.22]
 

llama_cpp/llama.py

@@ -1040,7 +1040,13 @@ def embed(
 
         # get pooling information
         pooling_type = self.pooling_type()
-        logits_all = pooling_type == llama_cpp.LLAMA_POOLING_TYPE_NONE
+        # In embedding mode every input token must be marked as an output, regardless of
+        # pooling type. llama.cpp would otherwise override per-token `logits[i]` and emit
+        # "embeddings required but some input tokens were not marked as outputs ->
+        # overriding" once per input. Pooling NONE vs MEAN/CLS only changes how the
+        # per-token outputs are read back (see decode_batch below), not whether they are
+        # produced. See abetlen/llama-cpp-python#2208.
+        logits_all = True
 
         if self.context_params.embeddings is False:
             raise RuntimeError(