vulkan: For coopmat2 FA, use fp16 accumulators for the final result (#19376)

The cpu and cuda backends use fp16 for the VKQ accumulator type, this change
does the same for vulkan. This helps particularly with large head sizes which
are very register-limited.

I tried this for the coopmat1 path and it slowed down a bit. I didn't try for
scalar.

I applied the softmax bias that the cuda backend uses to avoid overflow,
although I was not able to reproduce the original bug without it.

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl

@@ -240,3 +240,7 @@ void init_indices()
     // and breaking the alignment detection.
     m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
 }
+
+// Bias applied to softmax to stay in fp16 range.
+// Based on ggml-cuda issue https://github.com/ggml-org/llama.cpp/issues/18606
+const float FATTN_KQ_MAX_OFFSET = 3.0f*0.6931f;

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -117,7 +117,7 @@ void main() {
     Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -223,6 +223,8 @@ void main() {
 
         coopMatReduceNV(rowmax, S, gl_CooperativeMatrixReduceRowNV, maxReduce);
 
+        rowmax += coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(FATTN_KQ_MAX_OFFSET);
+
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> Mold = M;
 
         // M = max(rowmax, Mold)
@@ -265,11 +267,8 @@ void main() {
         // resize eM by using smear/reduce
         coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
-        // multiply with fp16 accumulation, then add to O.
+        O *= coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(eMdiag);
-        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
+        O = coopMatMulAdd(P_A, V, O);
-        PV = coopMatMulAdd(P_A, V, PV);
-
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(PV);
     }
 
     // If there is split_k, then the split_k resolve shader does the final
@@ -311,7 +310,7 @@ void main() {
             if (sink > Mr[i]) {
                 ms = exp(Mr[i] - sink);
 
-                O[i] *= ms;
+                O[i] *= float16_t(ms);
             } else {
                 vs = exp(sink - Mr[i]);
             }
@@ -325,15 +324,16 @@ void main() {
         Ldiag[k] = (Ldiag[k] == 0.0) ? ACC_TYPE(0.0) : (ACC_TYPE(1.0) / Ldiag[k]);
     }
 
-    O = Ldiag*O;
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
+
+    O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(Ldiag)*O_D;
 
 #if defined(ACC_TYPE_MAX)
-    [[unroll]] for (uint i = 0; i < O.length(); ++i) { O[i] = clamp(O[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
+    [[unroll]] for (uint i = 0; i < O_D.length(); ++i) { O_D[i] = clamp(O_D[i], D_TYPE(-ACC_TYPE_MAX), D_TYPE(ACC_TYPE_MAX)); }
 #endif
 
     uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
 
-    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
     } else {