From 8917ea0432161ca4f310faf397d3737d07ab2fa5 Mon Sep 17 00:00:00 2001 From: Pascal Date: Wed, 23 Sep 2026 17:29:00 +0200 Subject: [PATCH] metal: add the missing f32 x bf16 mul_mv variants (llama/28741) ggml_conv_1d_dw builds its im2col as f32 when the kernel is bf16, then multiplies the two, so a depthwise convolution over bf16 weights asks for kernel_mul_mv_f32_bf16, which was never instantiated. The base, the _4 and the _short families are filled in next to their bf16 neighbours, inside the same runtime guard, so a device without bf16 support is unaffected. --- ggml/src/ggml-metal/kernels/mul_mv.metal | 3 +++ 1 file changed, 3 insertions(+) diff --git a/ggml/src/ggml-metal/kernels/mul_mv.metal b/ggml/src/ggml-metal/kernels/mul_mv.metal index 8e2df2765..0d1069f2c 100644 --- a/ggml/src/ggml-metal/kernels/mul_mv.metal +++ b/ggml/src/ggml-metal/kernels/mul_mv.metal @@ -1043,6 +1043,7 @@ template [[host_name("kernel_mul_mv_f16_f16")]] kernel mul_mv_t_t kernel_mul_m #if defined(GGML_METAL_HAS_BF16) template [[host_name("kernel_mul_mv_bf16_f32")]] kernel mul_mv_t_t kernel_mul_mv_t_t; template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t_t kernel_mul_mv_t_t; +template [[host_name("kernel_mul_mv_f32_bf16")]] kernel mul_mv_t_t kernel_mul_mv_t_t; #endif template @@ -1167,6 +1168,7 @@ template [[host_name("kernel_mul_mv_f16_f16_4")]] kernel mul_mv_t_t_4 kernel_m #if defined(GGML_METAL_HAS_BF16) template [[host_name("kernel_mul_mv_bf16_f32_4")]] kernel mul_mv_t_t_4 kernel_mul_mv_t_t_4; template [[host_name("kernel_mul_mv_bf16_bf16_4")]] kernel mul_mv_t_t_4 kernel_mul_mv_t_t_4; +template [[host_name("kernel_mul_mv_f32_bf16_4")]] kernel mul_mv_t_t_4 kernel_mul_mv_t_t_4; #endif template @@ -1232,6 +1234,7 @@ template [[host_name("kernel_mul_mv_f16_f16_short")]] kernel mul_mv_t_t_short_t #if defined(GGML_METAL_HAS_BF16) template [[host_name("kernel_mul_mv_bf16_f32_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short; template [[host_name("kernel_mul_mv_bf16_bf16_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short; +template [[host_name("kernel_mul_mv_f32_bf16_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short; #endif template