ggml: add MLA flash attention config for GLM-4.7-flash

GLM-4.7-flash has gqa_ratio=20, which requires ncols=4 support for the
576x512 (MLA) flash attention kernel. The kernel selection logic and
extern declarations exist but the CONFIG_CASE entries were missing,
causing performance regression.

This adds ncols=4 CONFIG_CASE entries for Ampere, Turing, Volta, and
RDNA architectures.

Fixes #14045

llama/patches/0032-ggml-add-MLA-flash-attention-config-for-GLM-4.7-flas.patch

@@ -0,0 +1,64 @@
+From d19e2dbe8e759d1341e3993601b6d9a1e6c5fe51 Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Tue, 3 Feb 2026 12:45:29 -0800
+Subject: [PATCH] ggml: add MLA flash attention config for GLM-4.7-flash
+
+GLM-4.7-flash has gqa_ratio=20, which requires ncols=4 support for the
+576x512 (MLA) flash attention kernel. The kernel selection logic and
+extern declarations exist but the CONFIG_CASE entries were missing.
+
+This adds ncols=4 CONFIG_CASE entries for:
+- Ampere
+- Turing
+- Volta
+- RDNA (AMD)
+---
+ ggml/src/ggml-cuda/fattn-mma-f16.cuh | 11 ++++++++---
+ 1 file changed, 8 insertions(+), 3 deletions(-)
+
+diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+index 0b8ef9079..79a40da5a 100644
+--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
++++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+@@ -66,7 +66,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
+ 
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
+@@ -80,7 +81,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
+ 
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32,  96,  64, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
+@@ -89,7 +91,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+ }
+ 
+ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_volta(const int DKQ, const int DV, const int ncols) {
+-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256,  64, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256,  64, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128,  64, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128,  64, 1, false);
+@@ -103,6 +106,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
+ 
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32,  96,  64, 128, 1, false);
++    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, true);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
+     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
+-- 
+2.50.1 (Apple Git-155)
+

ml/backend/ggml/ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -66,7 +66,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
 
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
@@ -80,7 +81,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
 
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
@@ -89,7 +91,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
 }
 
 static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_volta(const int DKQ, const int DV, const int ncols) {
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32, 288, 256,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256,  64, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128,  64, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128,  64, 1, false);
@@ -103,6 +106,8 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
 
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  4,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);