From: Andreas Kieslinger <redacted>
Date: Fri, 29 May 2026 04:46:10 +0000 (+0200)
Subject: cuda : disables launch_fattn PDL enrollment due to compiler bug (#23825)
X-Git-Tag: upstream/0.0.10438~1034
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=241cbd41d21fca058c5ce900502e5d6d8d5b041b;p=pkg%2Fggml%2Fsources%2Fllama.cpp

cuda : disables launch_fattn PDL enrollment due to compiler bug (#23825)
---

diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index debcb6e54..d650b5fbd 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -1153,8 +1153,8 @@ void launch_fattn(
 
     GGML_ASSERT(block_dim.x % warp_size == 0);
 
-    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num, block_dim, nbytes_shared, main_stream);
-    ggml_cuda_kernel_launch(fattn_kernel, launch_params,
+        // disabled PDL enrollment for now due to a compiler bug.
+        fattn_kernel<<<blocks_num, block_dim, nbytes_shared, main_stream>>>(
         (const char *) Q->data,
         K_data,
         V_data,