0029-ggml-cuda-skip-large-batches.patch

From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Michael Yang <git@mxy.ng>
Date: Tue, 18 Nov 2025 11:13:04 -0800
Subject: [PATCH] ggml-cuda: skip large batches

cuda panics on batches larger than 1024 so mark it as unsupported to
fallback to cpu
---
 ggml/src/ggml-cuda/ggml-cuda.cu | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index c1bfadb3e..16c166a08 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -4570,6 +4570,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
                     return false;
                 }
+                if (op->op == GGML_OP_MUL_MAT && b->ne[2] * b->ne[3] > 1024) {
+                    return false;
+                }
 #ifdef GGML_USE_MUSA
                 const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
                 if (b->ne[2]*b->ne[3] > 1 && !ggml_is_transposed(a) && !ggml_is_transposed(b)) {