From: Georgi Gerganov <redacted>
Date: Fri, 14 Jun 2024 14:14:09 +0000 (+0300)
Subject: metal : utilize max shared memory for mul_mat_id (#7935)
X-Git-Tag: upstream/0.0.4488~1338
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=66ef1ceedf983773c8ceb4d925285d41d4e50e2a;p=pkg%2Fggml%2Fsources%2Fllama.cpp

metal : utilize max shared memory for mul_mat_id (#7935)
---

diff --git a/ggml-metal.m b/ggml-metal.m
index ec9e9530..f894274c 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -1862,9 +1862,10 @@ static enum ggml_status ggml_metal_graph_compute(
                         // ne21 = n_rows
                         const int dst_rows = ne20*ne21;
                         const int dst_rows_min = n_as;
+                        const int dst_rows_max = (ctx->device.maxThreadgroupMemoryLength - 32 - 8192)/4;
 
                         // max size of the rowids array in the kernel shared buffer
-                        GGML_ASSERT(dst_rows <= 2048);
+                        GGML_ASSERT(dst_rows <= dst_rows_max);
 
                         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel