Improve DeepSeek batched processing speed (#282)

* Improve DeepSeek batched processing speed

* Revert the commented out section in iqk_mul_mat.cpp

It does have some benefit at long contexts.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

Author: ikawrakow

ggml/src/iqk/iqk_mul_mat.cpp

@@ -17265,13 +17265,25 @@ template <int step_k, typename KHelper, typename VHelper>
 inline void iqk_deepseek_helper(KHelper& kh, VHelper& vh,
                         int nq1, int nk1, int stride_q, int stride_m, int stride_qkv,
                         const float * q, const char * mask, float scale, float softcap, float * qkv, float * M, float * S) {
-    if (nq1 % 8 == 0) {
+    if (nq1 >= 8) {
         FlashAttn<576, 512, 8, step_k> fa(scale, softcap);
         fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S);
-    } else {
+    }
+    else if (nq1 >= 4) {
+        FlashAttn<576, 512, 4, step_k> fa(scale, softcap);
+        fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S);
+    }
+    else {
         FlashAttn<576, 512, 1, step_k> fa(scale, softcap);
         fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S);
     }
+    //if (nq1 % 8 == 0) {
+    //    FlashAttn<576, 512, 8, step_k> fa(scale, softcap);
+    //    fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S);
+    //} else {
+    //    FlashAttn<576, 512, 1, step_k> fa(scale, softcap);
+    //    fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S);
+    //}
 }
 
 template <int step_k>

src/llama.cpp

@@ -13896,7 +13896,7 @@ struct llm_build_context {
 
         // whether to use n_tokens as the matrix dimension during multiplication or n_head
         // n_tokens is higher during prompt processing, this allows to optimize for this case
-        bool pp_opt = n_tokens > n_head;
+        bool pp_opt = n_tokens >= 128; // Is it a fixed constant or is it somehow relared to n_head? original: n_tokens > n_head;
 
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * inpSA = inpL;