From: Georgi Gerganov <redacted>
Date: Mon, 10 Nov 2025 19:33:35 +0000 (+0200)
Subject: metal : cap threadgroups size of set_rows (llama/17146)
X-Git-Tag: upstream/0.9.4.395~202
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=8832578bc54cbc668c033de9a688d0778a405029;p=pkg%2Fggml%2Fsources%2Fggml

metal : cap threadgroups size of set_rows (llama/17146)
---

diff --git a/src/ggml-metal/ggml-metal-ops.cpp b/src/ggml-metal/ggml-metal-ops.cpp
index 7a85edbd..5a8f150a 100644
--- a/src/ggml-metal/ggml-metal-ops.cpp
+++ b/src/ggml-metal/ggml-metal-ops.cpp
@@ -1036,6 +1036,11 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
 
     nth = std::min(nth, nk0);
 
+    if (nth*nrptg > ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
+        nrptg = 1;
+    }
+
     ggml_metal_kargs_set_rows args = {
         /*.nk0  =*/ nk0,
         /*.ne01 =*/ ne01,