pyrox.dev
rocmPackages.composable_kernel: make more parts big-parallel
All parts except pool are too slow without the higher core limit on big-parallel builders.
Infra channel discussion: https://matrix.to/#/!RROtHmAaQIkiJzJZZE:nixos.org/
+8
pkgs/development/rocm-modules/6/composable_kernel/default.nix
+8
pkgs/development/rocm-modules/6/composable_kernel/default.nix
···
36
36
"device_grouped_conv2d_fwd_instance"
37
37
"device_grouped_conv2d_fwd_dynamic_op_instance"
38
38
];
39
+
requiredSystemFeatures = [ "big-parallel" ];
39
40
};
40
41
grouped_conv_bwd_3d = {
41
42
targets = [
···
46
47
"device_grouped_conv3d_bwd_weight_bilinear_instance"
47
48
"device_grouped_conv3d_bwd_weight_scale_instance"
48
49
];
50
+
requiredSystemFeatures = [ "big-parallel" ];
49
51
};
50
52
grouped_conv_fwd_3d = {
51
53
targets = [
···
60
62
"device_grouped_conv3d_fwd_scaleadd_ab_instance"
61
63
"device_grouped_conv3d_fwd_scaleadd_scaleadd_relu_instance"
62
64
];
65
+
requiredSystemFeatures = [ "big-parallel" ];
63
66
};
64
67
batched_gemm = {
65
68
targets = [
···
77
80
"device_grouped_gemm_fixed_nk_multi_abd_instance"
78
81
"device_grouped_gemm_tile_loop_instance"
79
82
];
83
+
requiredSystemFeatures = [ "big-parallel" ];
80
84
};
81
85
gemm_universal = {
82
86
targets = [
···
108
112
"device_gemm_splitk_instance"
109
113
"device_gemm_streamk_instance"
110
114
];
115
+
requiredSystemFeatures = [ "big-parallel" ];
111
116
};
112
117
conv = {
113
118
targets = [
···
118
123
"device_conv2d_fwd_bias_relu_add_instance"
119
124
"device_conv3d_bwd_data_instance"
120
125
];
126
+
requiredSystemFeatures = [ "big-parallel" ];
121
127
};
122
128
pool = {
123
129
targets = [
···
139
145
"device_normalization_bwd_gamma_beta_instance"
140
146
"device_normalization_fwd_instance"
141
147
];
148
+
requiredSystemFeatures = [ "big-parallel" ];
142
149
};
143
150
other2 = {
144
151
targets = [
···
150
157
"device_softmax_instance"
151
158
"device_transpose_instance"
152
159
];
160
+
requiredSystemFeatures = [ "big-parallel" ];
153
161
};
154
162
};
155
163
tensorOpBuilder =