From 2aef9767d982beb7ab61b3ae359366f627eb74dd Mon Sep 17 00:00:00 2001
From: John <78893154+cmp-nct@users.noreply.github.com>
Date: Sat, 31 Jan 2026 05:15:49 +0100
Subject: [PATCH] Switch to base cache for position tracking

Updated position tracking to use base cache instead of SWA cache to avoid kv cache token losses.
---
 src/llama-kv-cache-iswa.cpp | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/src/llama-kv-cache-iswa.cpp b/src/llama-kv-cache-iswa.cpp
index 3a34102a23..3033e9891f 100644
--- a/src/llama-kv-cache-iswa.cpp
+++ b/src/llama-kv-cache-iswa.cpp
@@ -107,12 +107,14 @@ void llama_kv_cache_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p
 }
 
 llama_pos llama_kv_cache_iswa::seq_pos_min(llama_seq_id seq_id) const {
-    // the base cache is a superset of the SWA cache, so we can just check the SWA cache
-    return kv_swa->seq_pos_min(seq_id);
+    // base cache preserves all positions (no SWA cell reuse), making it the ground truth for position tracking
+    // SWA cache can lose older positions when cells are reused for new tokens outside the SWA window
+    return kv_base->seq_pos_min(seq_id);
 }
 
 llama_pos llama_kv_cache_iswa::seq_pos_max(llama_seq_id seq_id) const {
-    return kv_swa->seq_pos_max(seq_id);
+    // base cache preserves all positions (no SWA cell reuse), making it the ground truth for position tracking
+    return kv_base->seq_pos_max(seq_id);
 }
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_iswa::memory_breakdown() const {