hey! this is a static git repo viewer. commit diffs, file history, older commits and other revisions won't load without javascript.

you are still able to view source files on branch/tag tips, but the experience is degraded.

char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Piotr Wilkin (ilintar)Revert "CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#24546)" (#28551)e71b805102026-09-07

1static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_cdna(ggml_type type, int J, bool fallback) {
2    CASE(GGML_TYPE_Q1_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
3    CASE(GGML_TYPE_Q1_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
4    CASE(GGML_TYPE_Q1_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
5    CASE(GGML_TYPE_Q1_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
6    CASE(GGML_TYPE_Q1_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
7    CASE(GGML_TYPE_Q1_0, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
8    CASE(GGML_TYPE_Q1_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
9
10    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
11    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
12    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
13    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
14    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
15    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
16    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
17
18    CASE(GGML_TYPE_Q4_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
19    CASE(GGML_TYPE_Q4_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
20    CASE(GGML_TYPE_Q4_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
21    CASE(GGML_TYPE_Q4_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
22    CASE(GGML_TYPE_Q4_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
23    CASE(GGML_TYPE_Q4_0, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
24    CASE(GGML_TYPE_Q4_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
25
26    CASE(GGML_TYPE_Q4_1, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
27    CASE(GGML_TYPE_Q4_1, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
28    CASE(GGML_TYPE_Q4_1, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
29    CASE(GGML_TYPE_Q4_1, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
30    CASE(GGML_TYPE_Q4_1, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
31    CASE(GGML_TYPE_Q4_1, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
32    CASE(GGML_TYPE_Q4_1, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
33
34    CASE(GGML_TYPE_Q5_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
35    CASE(GGML_TYPE_Q5_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
36    CASE(GGML_TYPE_Q5_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
37    CASE(GGML_TYPE_Q5_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
38    CASE(GGML_TYPE_Q5_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
39    CASE(GGML_TYPE_Q5_0, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
40    CASE(GGML_TYPE_Q5_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
41
42    CASE(GGML_TYPE_Q5_1, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
43    CASE(GGML_TYPE_Q5_1, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
44    CASE(GGML_TYPE_Q5_1, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
45    CASE(GGML_TYPE_Q5_1, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
46    CASE(GGML_TYPE_Q5_1, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
47    CASE(GGML_TYPE_Q5_1, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
48    CASE(GGML_TYPE_Q5_1, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
49
50    CASE(GGML_TYPE_Q8_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
51    CASE(GGML_TYPE_Q8_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
52    CASE(GGML_TYPE_Q8_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
53    CASE(GGML_TYPE_Q8_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
54    CASE(GGML_TYPE_Q8_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
55    CASE(GGML_TYPE_Q8_0, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
56    CASE(GGML_TYPE_Q8_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
57
58// ---------------------------------------------------------------------------------------------
59
60    CASE(GGML_TYPE_Q2_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
61    CASE(GGML_TYPE_Q2_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
62    CASE(GGML_TYPE_Q2_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, true);
63    CASE(GGML_TYPE_Q2_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
64    CASE(GGML_TYPE_Q2_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
65    CASE(GGML_TYPE_Q2_K, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
66    CASE(GGML_TYPE_Q2_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, true, false);
67
68    CASE(GGML_TYPE_Q3_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
69    CASE(GGML_TYPE_Q3_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
70    CASE(GGML_TYPE_Q3_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
71    CASE(GGML_TYPE_Q3_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
72    CASE(GGML_TYPE_Q3_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
73    CASE(GGML_TYPE_Q3_K, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
74    CASE(GGML_TYPE_Q3_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
75
76    CASE(GGML_TYPE_Q4_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
77    CASE(GGML_TYPE_Q4_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
78    CASE(GGML_TYPE_Q4_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
79    CASE(GGML_TYPE_Q4_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
80    CASE(GGML_TYPE_Q4_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
81    CASE(GGML_TYPE_Q4_K, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
82    CASE(GGML_TYPE_Q4_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
83
84    CASE(GGML_TYPE_Q5_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
85    CASE(GGML_TYPE_Q5_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
86    CASE(GGML_TYPE_Q5_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
87    CASE(GGML_TYPE_Q5_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
88    CASE(GGML_TYPE_Q5_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
89    CASE(GGML_TYPE_Q5_K, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
90    CASE(GGML_TYPE_Q5_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
91
92    CASE(GGML_TYPE_Q6_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
93    CASE(GGML_TYPE_Q6_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
94    CASE(GGML_TYPE_Q6_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, true);
95    CASE(GGML_TYPE_Q6_K, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
96    CASE(GGML_TYPE_Q6_K, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
97    CASE(GGML_TYPE_Q6_K, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
98    CASE(GGML_TYPE_Q6_K, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, true, false);
99
100// ---------------------------------------------------------------------------------------------
101
102    CASE(GGML_TYPE_IQ1_S, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
103    CASE(GGML_TYPE_IQ1_S, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
104    CASE(GGML_TYPE_IQ1_S, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
105    CASE(GGML_TYPE_IQ1_S, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
106    CASE(GGML_TYPE_IQ1_S, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
107    CASE(GGML_TYPE_IQ1_S, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
108    CASE(GGML_TYPE_IQ1_S, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
109
110    CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
111    CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
112    CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
113    CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
114    CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
115    CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
116    CASE(GGML_TYPE_IQ2_XXS, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
117
118    CASE(GGML_TYPE_IQ2_XS, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
119    CASE(GGML_TYPE_IQ2_XS, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
120    CASE(GGML_TYPE_IQ2_XS, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
121    CASE(GGML_TYPE_IQ2_XS, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
122    CASE(GGML_TYPE_IQ2_XS, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
123    CASE(GGML_TYPE_IQ2_XS, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
124    CASE(GGML_TYPE_IQ2_XS, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
125
126    CASE(GGML_TYPE_IQ2_S, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
127    CASE(GGML_TYPE_IQ2_S, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
128    CASE(GGML_TYPE_IQ2_S, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, true);
129    CASE(GGML_TYPE_IQ2_S, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
130    CASE(GGML_TYPE_IQ2_S, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
131    CASE(GGML_TYPE_IQ2_S, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
132    CASE(GGML_TYPE_IQ2_S, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, true, false);
133
134    CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
135    CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
136    CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
137    CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
138    CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
139    CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
140    CASE(GGML_TYPE_IQ3_XXS, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
141
142    CASE(GGML_TYPE_IQ3_S, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
143    CASE(GGML_TYPE_IQ3_S, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
144    CASE(GGML_TYPE_IQ3_S, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
145    CASE(GGML_TYPE_IQ3_S, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
146    CASE(GGML_TYPE_IQ3_S, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
147    CASE(GGML_TYPE_IQ3_S, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
148    CASE(GGML_TYPE_IQ3_S, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
149
150    CASE(GGML_TYPE_IQ4_XS, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
151    CASE(GGML_TYPE_IQ4_XS, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
152    CASE(GGML_TYPE_IQ4_XS, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
153    CASE(GGML_TYPE_IQ4_XS, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
154    CASE(GGML_TYPE_IQ4_XS, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
155    CASE(GGML_TYPE_IQ4_XS, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
156    CASE(GGML_TYPE_IQ4_XS, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
157
158    CASE(GGML_TYPE_IQ4_NL, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
159    CASE(GGML_TYPE_IQ4_NL, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
160    CASE(GGML_TYPE_IQ4_NL, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
161    CASE(GGML_TYPE_IQ4_NL, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
162    CASE(GGML_TYPE_IQ4_NL, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
163    CASE(GGML_TYPE_IQ4_NL, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
164    CASE(GGML_TYPE_IQ4_NL, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
165
166// ---------------------------------------------------------------------------------------------
167
168    CASE(GGML_TYPE_MXFP4, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
169    CASE(GGML_TYPE_MXFP4, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
170    CASE(GGML_TYPE_MXFP4, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, true);
171    CASE(GGML_TYPE_MXFP4, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
172    CASE(GGML_TYPE_MXFP4, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
173    CASE(GGML_TYPE_MXFP4, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
174    CASE(GGML_TYPE_MXFP4, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, true, false);
175
176    CASE(GGML_TYPE_NVFP4, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
177    CASE(GGML_TYPE_NVFP4, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
178    CASE(GGML_TYPE_NVFP4, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, true);
179    CASE(GGML_TYPE_NVFP4, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
180    CASE(GGML_TYPE_NVFP4, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
181    CASE(GGML_TYPE_NVFP4, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
182    CASE(GGML_TYPE_NVFP4, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
183
184    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
185}