reproducibility/imatrix-gemma-embedding.patch

diff --git a/tools/imatrix/imatrix.cpp b/tools/imatrix/imatrix.cpp
index 5e9f911e0..d7c531fd8 100644
--- a/tools/imatrix/imatrix.cpp
+++ b/tools/imatrix/imatrix.cpp
@@ -1426,7 +1426,9 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params, c
 
     const bool add_bos = llama_vocab_get_add_bos(vocab);
 
-    if (llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_LAST) {
+    // Embedding models may append EOS while using mean pooling. This is valid
+    // for activation collection, which does not require next-token PPL.
+    if (params.compute_ppl && llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_LAST) {
         GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
     }