iswa : for now, always enable shifts (experiment)

ggerganov · ggerganov · commit e7bef6bd5692 · 2025-05-15T08:06:11.000+03:00
ggml-ci
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
@@ -1715,7 +1715,10 @@ 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
`@@ -1715,7 +1715,10 @@ llama_pos llama_kv_cache_unified_iswa::get_pos_max() const {`
`1715`	`1715`	`}`
`1716`	`1716`
`1717`	`1717`	`bool llama_kv_cache_unified_iswa::get_can_shift() const {`
`1718`		`- return kv_base->get_size() == kv_swa->get_size();`
	`1718`	`+ // TODO: for now allow this, eventhough it's not mathematically correct`
	`1719`	`+ // but some initial tests indicate that the results are not bad`
	`1720`	`+ return true;`
	`1721`	`+ //return kv_base->get_size() == kv_swa->get_size();`
`1719`	`1722`	`}`
`1720`	`1723`
`1721`	`1724`	`void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {`