ggml-cpu: stabilise nnpa fp32<->fp16

Signed-off-by: Aaron Teo <[email protected]>

Author: taronaeo

common/common.h

@@ -312,7 +312,13 @@ struct common_params {
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
+    #if defined(GGML_NNPA) || defined(__NNPA__)
+    // disable Flash Attention on NNPA
+    // see: https://github.com/ggml-org/llama.cpp/issues/15721
+    enum llama_flash_attn_type   flash_attn_type   = LLAMA_FLASH_ATTN_TYPE_DISABLED;
+    #else
     enum llama_flash_attn_type   flash_attn_type   = LLAMA_FLASH_ATTN_TYPE_AUTO; // whether to use Flash Attention
+    #endif
 
     struct common_params_sampling    sampling;
     struct common_params_speculative speculative;

ggml/src/ggml-cpu/ggml-cpu.c

@@ -3219,7 +3219,10 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
         float32x4_t v_zero = vec_splats(0.0f);
         uint16x8_t v_yd = vec_round_from_fp32(v_x, v_zero, 0);
         uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
-        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+        y[i + 0] = vec_extract(v_y, 0);
+        y[i + 1] = vec_extract(v_y, 1);
+        y[i + 2] = vec_extract(v_y, 2);
+        y[i + 3] = vec_extract(v_y, 3);
     }
 #endif
     for (; i < n; ++i) {