A high-performance inference engine for AI models
kv cache clear fix
LuckyIYI91e1ef9cd310
diff --git a/crates/uzu/src/backends/metal/forward_pass/kv_cache.rs b/crates/uzu/src/backends/metal/forward_pass/kv_cache.rsindex 959e6d206..2a16a8381 100644--- a/crates/uzu/src/backends/metal/forward_pass/kv_cache.rs+++ b/crates/uzu/src/backends/metal/forward_pass/kv_cache.rs@@ -368,6 +368,7 @@ impl KVCache {prefix_len,} => {*prefix_len = 0;+ layer.prefix_token_positions.clear();},KVCacheLayerState::Windowed {ring_offset,@@ -376,9 +377,9 @@ impl KVCache {} => {*ring_offset = 0;*ring_length = 0;+ layer.prefix_token_positions.fill(INVALID_POSITION);},}- layer.prefix_token_positions.fill(INVALID_POSITION);}}diff --git a/crates/uzu/tests/text_session_test.rs b/crates/uzu/tests/text_session_test.rsindex 268c948c9..6fafb2aaa 100644--- a/crates/uzu/tests/text_session_test.rs+++ b/crates/uzu/tests/text_session_test.rs@@ -80,9 +80,6 @@ fn run_scenario(system_prompt: Option<String>,user_prompts: Vec<String>,) {- unsafe {- std::env::set_var("UZU_SKIP_RAM_CHECK", "1");- }let mut session =Session::new(build_model_path(), build_decoding_config()).unwrap();