rdy for merge

leejet · Sep 7, 2024 · f69fbfd · f69fbfd
1 parent e904b86
commit f69fbfd
Showing 1 changed file with 6 additions and 7 deletions.
diff --git a/ggml_extend.hpp b/ggml_extend.hpp
@@ -710,11 +710,10 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
 
     float scale = (1.0f / sqrt((float)d_head));
 
-    if (flash_attn) {
-        // TODO: remove before merge
-        LOG_DEBUG("attention_ext L_q:%d L_k:%d n_head:%d C:%d d_head:%d N:%d", L_q, L_k, n_head, C, d_head, N);
-    }
-    // is there anything oddly shaped??
+    //if (flash_attn) {
+    //    LOG_DEBUG("attention_ext L_q:%d L_k:%d n_head:%d C:%d d_head:%d N:%d", L_q, L_k, n_head, C, d_head, N);
+    //}
+    // is there anything oddly shaped?? ping Green-Sky if you can trip this assert
     GGML_ASSERT(((L_k % 256 == 0) && L_q == L_k) || !(L_k % 256 == 0));
 
     bool can_use_flash_attn = true;
@@ -725,12 +724,12 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
     can_use_flash_attn = can_use_flash_attn && d_head <= 256; // double check
 
     if (mask != nullptr) {
-        // TODO: figure out if we can bend t5 to work too
+        // TODO(Green-Sky): figure out if we can bend t5 to work too
         can_use_flash_attn = can_use_flash_attn && mask->ne[2] == 1;
         can_use_flash_attn = can_use_flash_attn && mask->ne[3] == 1;
     }
 
-    // TODO: more pad or disable for funny tensor shapes
+    // TODO(Green-Sky): more pad or disable for funny tensor shapes
 
     ggml_tensor* kqv = nullptr;
     //GGML_ASSERT((flash_attn && can_use_flash_attn) || !flash_attn);