cont : simplify SWA mask condition

ggml-ci

Author: ggerganov

src/llama-hparams.h

@@ -102,7 +102,7 @@ struct llama_hparams {
     // Sliding Window Attention (SWA)
     llama_swa_type swa_type = LLAMA_SWA_TYPE_STANDARD;
 
-    uint32_t n_swa = 0;         // sliding window attention (SWA)
+    uint32_t n_swa = 0;         // the size of the sliding window (0 - no SWA)
     uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention
 
     // for State Space Models

src/llama-kv-cache.cpp

@@ -687,10 +687,7 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
                                 {
                                     const llama_pos pos_chunk_start = (pos / hparams.n_swa) * hparams.n_swa;
 
-                                    // TODO: should this be simply:
-                                    //  if (cells[i].pos < pos_chunk_start) {
-                                    //
-                                    if (cells[i].pos < pos_chunk_start || pos < pos_chunk_start) {
+                                    if (cells[i].pos < pos_chunk_start) {
                                         f = -INFINITY;
                                     }
                                 } break;