From 3c26dd3353bced6fa88c2e7e6c7d921a1e09dfcd Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sun, 6 Apr 2025 04:03:47 -0500
Subject: [PATCH] vulkan: fix NaN issue in flash attention shader (llama/12776)

Use -FLT_MAX/2 rather than -inf as the initial value for computing the maximum.
---
 ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
index eedbc6f8..8ddadb8a 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -227,8 +227,11 @@ void main() {
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
+    // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M.
+    const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
+
     L = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
-    M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(-1.0/0.0);
+    M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(NEG_FLT_MAX_OVER_2);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> slopeMat = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(1.0);
 
@@ -278,7 +281,7 @@ void main() {
             uint R = ((i + 1) * Br >  N) ?  (N % Br) : Br;
             uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc;
 
-            coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(-1.0/0.0), R, C);
+            coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(NEG_FLT_MAX_OVER_2), R, C);
         }
 
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> rowmax, P, rowsum, eM;