Commit 2e53a0bab for llama.cpp
commit 2e53a0bab751a288d7e48e642827cb4773b7b9c9
Author: thelittlefireman <5165783+thelittlefireman@users.noreply.github.com>
Date: Tue Nov 17 23:37:14 2026 +0100
ggml-cuda: hip: Tune stream_k algo and mmq configs to use stream_k on AMD GCN (#30022)
* ggml-cuda: hip: enable stream_k on GCN
* ggml-cuda: Test on CDNA done no improvement, remove TODO
* ggml-cuda: HIP: GCN cleanup
diff --git a/ggml/src/ggml-cuda/mmq-config-gcn.cuh b/ggml/src/ggml-cuda/mmq-config-gcn.cuh
index 2aea107f5..4cf57f40a 100644
--- a/ggml/src/ggml-cuda/mmq-config-gcn.cuh
+++ b/ggml/src/ggml-cuda/mmq-config-gcn.cuh
@@ -1,326 +1,326 @@
static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_gcn(ggml_type type, int J, bool fallback) {
- CASE(GGML_TYPE_Q1_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q1_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q1_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q1_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q1_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q1_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q1_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q1_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
CASE(GGML_TYPE_Q1_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q2_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q2_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q2_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q2_0, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q2_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_0, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q2_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_0, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_Q2_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_Q4_0, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q4_0, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_1, 512, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_1, 512, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_1, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_Q5_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q5_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_0, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q5_0, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q5_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_0, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_Q5_0, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_0, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q5_0, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_Q5_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_1, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_1, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q5_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_1, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_1, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q5_1, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q5_1, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_1, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_1, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_1, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q5_1, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_1, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_1, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_1, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_Q5_1, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_1, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_1, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q5_1, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_1, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q8_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q8_0, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q8_0, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q8_0, 512, 2, 64, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q8_0, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 64, 80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q8_0, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
// ---------------------------------------------------------------------------------------------
- CASE(GGML_TYPE_Q2_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q2_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q2_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q2_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q2_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q2_K, 512, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q2_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q2_K, 512, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_Q3_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q3_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q3_K, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q3_K, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q3_K, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q3_K, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q3_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q3_K, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q3_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_Q3_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
CASE(GGML_TYPE_Q3_K, 256, 2, 128, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q3_K, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q3_K, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_Q3_K, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q3_K, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q3_K, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_Q4_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q4_K, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q4_K, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
CASE(GGML_TYPE_Q4_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q4_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q4_K, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_Q5_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_K, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q5_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_K, 256, 3, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q5_K, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_K, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_K, 256, 3, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q5_K, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_Q6_K, 256, 2, 64, 56, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_Q6_K, 256, 2, 64, 56, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_Q6_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_Q6_K, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
// ---------------------------------------------------------------------------------------------
- CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ1_S, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ1_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ1_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
// CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ1_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_IQ1_S, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ1_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ1_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
// CASE(GGML_TYPE_IQ1_S, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ2_XXS, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XXS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ2_XXS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
// CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_IQ2_XXS, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XXS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ2_XXS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
// CASE(GGML_TYPE_IQ2_XXS, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ2_XS, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ2_XS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
// CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ2_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_IQ2_XS, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ2_XS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
// CASE(GGML_TYPE_IQ2_XS, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
// CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ2_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
// CASE(GGML_TYPE_IQ2_S, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ3_XXS, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_XXS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ3_XXS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
// CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_IQ3_XXS, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_XXS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ3_XXS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
// CASE(GGML_TYPE_IQ3_XXS, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ3_S, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ3_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
// CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ3_S, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_IQ3_S, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ3_S, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
// CASE(GGML_TYPE_IQ3_S, 512, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ4_XS, 512, 2, 64, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ4_XS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ4_XS, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ4_XS, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ4_XS, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ4_NL, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ4_NL, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_IQ4_NL, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_IQ4_NL, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
// ---------------------------------------------------------------------------------------------
- CASE(GGML_TYPE_MXFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_MXFP4, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_MXFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_MXFP4, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_MXFP4, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_MXFP4, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_MXFP4, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_MXFP4, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_MXFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_MXFP4, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_MXFP4, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_MXFP4, 512, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_MXFP4, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_MXFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_MXFP4, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
CASE(GGML_TYPE_MXFP4, 512, 2, 128, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_MXFP4, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_MXFP4, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_MXFP4, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_MXFP4, 512, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_MXFP4, 512, 2, 128, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_MXFP4, 512, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_NVFP4, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_NVFP4, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_NVFP4, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
CASE(GGML_TYPE_NVFP4, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_NVFP4, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_NVFP4, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
- CASE(GGML_TYPE_NVFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_NVFP4, 512, 2, 128, 16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 128, 40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
+ CASE(GGML_TYPE_NVFP4, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 64, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
}
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 219b567dd..7fb09a353 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -1017,8 +1017,9 @@ static __device__ __forceinline__ void mul_mat_q_process_tile(
// The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
+// force_tiling lets the host use the native tiled path for configs that normally use Stream-K.
-template <ggml_type type, int J, bool fallback, ggml_prec prec_src1, bool has_fusion>
+template <ggml_type type, int J, bool fallback, ggml_prec prec_src1, bool has_fusion, bool force_tiling = false>
__launch_bounds__(ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1), ggml_cuda_mmq_get_occupancy(type, J, fallback, prec_src1))
static __global__ void mul_mat_q(
const char * __restrict__ x, const char * __restrict__ x_gate, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
@@ -1072,7 +1073,7 @@ static __global__ void mul_mat_q(
__syncthreads();
// The fused kernel needs the whole K range of a tile in one block, so it always uses tiling.
- if constexpr (has_fusion || !ggml_cuda_mmq_get_stream_k(type, J, fallback, prec_src1)) {
+ if constexpr (has_fusion || force_tiling || !ggml_cuda_mmq_get_stream_k(type, J, fallback, prec_src1)) {
// The fused kernel writes the GLU of I/2 up rows and I/2 gate rows per tile.
constexpr int I_dst = has_fusion ? I/2 : I;
@@ -1307,14 +1308,25 @@ static __global__ void mul_mat_q(
}
template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
-__launch_bounds__(ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1)/2, 1)
+static constexpr __device__ int mul_mat_q_stream_k_fixup_nthreads() {
+#ifdef GCN
+ // With NVIDIA warp32, the half-sized fixup launch has nthreads/64 warps.
+ // GCN wave64 needs the full MMQ thread count to keep the same nthreads/64 wavefronts.
+ return ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1);
+#else
+ return ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1) / 2;
+#endif // GCN
+}
+
+template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
+__launch_bounds__((mul_mat_q_stream_k_fixup_nthreads<type, J, fallback, prec_src1>()), 1)
static __global__ void mul_mat_q_stream_k_fixup(
const int32_t * __restrict__ ids_dst, const ggml_cuda_mm_id_tile_spec * __restrict__ tile_spec, float * __restrict__ dst,
float * __restrict__ tmp_last_tile, const uint3 blocks_per_ne00, const int nrows_x, const int ncols_dst,
const int stride_col_dst, const uint3 nchannels_y, const int stride_channel_dst, const uint3 nsamples_y,
const int stride_sample_dst, const uint3 ntj) {
constexpr int warp_size = ggml_cuda_get_physical_warp_size();
- constexpr int nwarps = (ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1) / 2) / warp_size;
+ constexpr int nwarps = mul_mat_q_stream_k_fixup_nthreads<type, J, fallback, prec_src1>() / warp_size;
constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback, prec_src1);
constexpr int qk = ggml_cuda_type_traits<type>::qk;
constexpr int ITER_K = ggml_cuda_mmq_get_K_vram(type, J, fallback, prec_src1);
@@ -1520,16 +1532,42 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
return;
}
- // For the stream-k kernel it is possible to run it with tiling by setting the number of CUDA blocks equal to the number of tiles.
- // This is worthwhile if the efficiency of tiling is high and skipping the fixup kernel is more important.
+ // Stream-K adds scheduling and fixup overhead when regular tiling already provides enough parallel work.
const int ntiles_dst = args.ntj * nti * ntzw;
const int tiles_nwaves = (ntiles_dst + nsm - 1) / nsm;
const int tiles_efficiency_percent = 100 * ntiles_dst / (nsm*tiles_nwaves);
+ const int stream_k_blocks = GGML_CUDA_CC_IS_GCN(cc) ? nsm * config.occupancy : nsm;
+
+ const int k_chunks = (args.ncols_x + config.K_vram - 1) / config.K_vram;
+ // Use tiling when Stream-K is unlikely to improve performance:
+ // With at least 2 full waves of tiles, tiling already saturates the device and Stream-K only adds overhead.
+ // With few K chunks per tile, the extra parallelism of Stream-K is not needed.
+ const bool force_tiling = GGML_CUDA_CC_IS_GCN(cc) &&
+ (tiles_efficiency_percent >= 90 || ntiles_dst >= 2*stream_k_blocks ||
+ k_chunks <= config.occupancy * tiles_nwaves);
+
+ if (force_tiling) {
+ GGML_ASSERT(nti <= 65535);
+ CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, J, fallback, prec_src1, has_fusion, true>), nbytes_shared);
+ mul_mat_q<type, J, fallback, prec_src1, has_fusion, true><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
+ (args.x, (const char *) args.fusion.gate, args.y, args.ids_dst, args.tile_spec, args.dst, nullptr,
+ args.y_scale, args.fusion.glu_op, args.fusion.glu_limit, (const float *) args.fusion.x_scale, (const float *) args.fusion.gate_scale,
+ blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
+ channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+ sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+ ntj_fd);
+ return;
+ }
GGML_ASSERT(ntiles_dst * blocks_per_ne00_fd.z < (1 << 30)); // Assert that variable kbc will not overflow.
- const int nwork = ntiles_dst * ((args.ncols_x + config.K_vram - 1) / config.K_vram); // Maximum number of CUDA blocks that can do useful work.
- const dim3 block_nums_stream_k(GGML_CUDA_CC_IS_NVIDIA(cc) && tiles_efficiency_percent >= 90 ? ntiles_dst : std::min(nsm, nwork), 1, 1);
+ const int nwork = ntiles_dst * k_chunks; // Maximum number of CUDA blocks that can do useful work.
+ int nblocks_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && tiles_efficiency_percent >= 90 ? ntiles_dst : std::min(stream_k_blocks, nwork);
+ if (GGML_CUDA_CC_IS_GCN(cc) && ntiles_dst < nsm) {
+ // Use a 512-thread budget per CU when the tiles do not fill the device.
+ nblocks_stream_k = std::min(nwork, nsm * std::max(1, 512 / config.nthreads));
+ }
+ const dim3 block_nums_stream_k(nblocks_stream_k, 1, 1);
const bool fixup_needed = ntiles_dst % block_nums_stream_k.x != 0;
@@ -1540,7 +1578,10 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
}
const dim3 block_nums_fixup(block_nums_stream_k.x, config.I/warp_size, 1);
- const dim3 block_dims_fixup(block_dims.x, block_dims.y/2, block_dims.z);
+
+ // GCN wave64 already halves the warp count relative to NVIDIA warp32, so halving nwarps again would halve fixup parallelism.
+ const int nwarps_fixup = GGML_CUDA_CC_IS_GCN(cc) ? nwarps : nwarps/2;
+ const dim3 block_dims_fixup(block_dims.x, nwarps_fixup, block_dims.z);
mul_mat_q<type, J, fallback, prec_src1, has_fusion><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
(args.x, (const char *) args.fusion.gate, args.y, args.ids_dst, args.tile_spec, args.dst, tmp_fixup.ptr,