iswa : for now, always enable shifts (experiment)

ggerganov · ggerganov · commit 8d8c8c4689a9 · 2025-05-11T19:02:18.000+03:00
ggml-ci
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
@@ -1707,7 +1707,10 @@ llama_pos llama_kv_cache_unified_iswa::get_pos_max() const {
 }
 
 bool llama_kv_cache_unified_iswa::get_can_shift() const {
-    return kv_base->get_size() == kv_swa->get_size();
+    // TODO: for now allow this, eventhough it's not mathematically correct
+    //       but some initial tests indicate that the results are not bad
+    return true;
+    //return kv_base->get_size() == kv_swa->get_size();
 }
 
 void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {

Original file line number	Diff line number	Diff line change
`@@ -1707,7 +1707,10 @@ llama_pos llama_kv_cache_unified_iswa::get_pos_max() const {`
`1707`	`1707`	`}`
`1708`	`1708`
`1709`	`1709`	`bool llama_kv_cache_unified_iswa::get_can_shift() const {`
`1710`		`- return kv_base->get_size() == kv_swa->get_size();`
	`1710`	`+ // TODO: for now allow this, eventhough it's not mathematically correct`
	`1711`	`+ // but some initial tests indicate that the results are not bad`
	`1712`	`+ return true;`
	`1713`	`+ //return kv_base->get_size() == kv_swa->get_size();`
`1711`	`1714`	`}`
`1712`	`1715`
`1713`	`1716`	`void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {`