From: Georgi Gerganov <redacted>
Date: Tue, 16 Jan 2024 17:34:54 +0000 (+0200)
Subject: perplexity : fix kv cache handling for hellaswag (#4981)
X-Git-Tag: upstream/0.0.4488~2598
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=959ef0c0df725c013c7f712eaa7790b8e38a8e20;p=pkg%2Fggml%2Fsources%2Fllama.cpp

perplexity : fix kv cache handling for hellaswag (#4981)

ggml-ci
---

diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
index 9a77beca..b4fedf80 100644
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -428,6 +428,7 @@ static std::vector<float> hellaswag_evaluate_tokens(
     for (size_t i_chunk = 0; i_chunk < n_chunk; ++i_chunk) {
         size_t n_tokens = tokens.size() - i_chunk * n_batch;
         n_tokens = std::min(n_tokens, size_t(n_batch));
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
         if (llama_decode(ctx, llama_batch_get_one(tokens.data() + i_chunk * n_batch, n_tokens, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return {};