|
19 | 19 |
|
20 | 20 | name2qmap = {} |
21 | 21 |
|
22 | | -if lib and lib.compiled_with_cuda: |
23 | | - """C FUNCTIONS FOR OPTIMIZERS""" |
24 | | - str2optimizer32bit = { |
25 | | - "adam": ( |
26 | | - lib.cadam32bit_grad_fp32, |
27 | | - lib.cadam32bit_grad_fp16, |
28 | | - lib.cadam32bit_grad_bf16, |
29 | | - ), |
30 | | - "momentum": ( |
31 | | - lib.cmomentum32bit_grad_32, |
32 | | - lib.cmomentum32bit_grad_16, |
33 | | - ), |
34 | | - "rmsprop": ( |
35 | | - lib.crmsprop32bit_grad_32, |
36 | | - lib.crmsprop32bit_grad_16, |
37 | | - ), |
38 | | - "lion": ( |
39 | | - lib.clion32bit_grad_fp32, |
40 | | - lib.clion32bit_grad_fp16, |
41 | | - lib.clion32bit_grad_bf16, |
42 | | - ), |
43 | | - "adagrad": ( |
44 | | - lib.cadagrad32bit_grad_32, |
45 | | - lib.cadagrad32bit_grad_16, |
46 | | - ), |
47 | | - "lamb": ( |
48 | | - lib.cadam32bit_grad_fp32, |
49 | | - lib.cadam32bit_grad_fp16, |
50 | | - lib.cadam32bit_grad_bf16, |
51 | | - ), |
52 | | - "ademamix": ( |
53 | | - lib.cademamix32bit_grad_fp32, |
54 | | - lib.cademamix32bit_grad_fp16, |
55 | | - lib.cademamix32bit_grad_bf16, |
56 | | - ), |
57 | | - } |
58 | | - |
59 | | - str2optimizer8bit = { |
60 | | - "adam": ( |
61 | | - lib.cadam_static_8bit_grad_32, |
62 | | - lib.cadam_static_8bit_grad_16, |
63 | | - ), |
64 | | - "momentum": ( |
65 | | - lib.cmomentum_static_8bit_grad_32, |
66 | | - lib.cmomentum_static_8bit_grad_16, |
67 | | - ), |
68 | | - "rmsprop": ( |
69 | | - lib.crmsprop_static_8bit_grad_32, |
70 | | - lib.crmsprop_static_8bit_grad_16, |
71 | | - ), |
72 | | - "lion": ( |
73 | | - lib.clion_static_8bit_grad_32, |
74 | | - lib.clion_static_8bit_grad_16, |
75 | | - ), |
76 | | - "lamb": ( |
77 | | - lib.cadam_static_8bit_grad_32, |
78 | | - lib.cadam_static_8bit_grad_16, |
79 | | - ), |
80 | | - "lars": ( |
81 | | - lib.cmomentum_static_8bit_grad_32, |
82 | | - lib.cmomentum_static_8bit_grad_16, |
83 | | - ), |
84 | | - } |
85 | | - |
86 | | - str2optimizer8bit_blockwise = { |
87 | | - "adam": ( |
88 | | - lib.cadam_8bit_blockwise_grad_fp32, |
89 | | - lib.cadam_8bit_blockwise_grad_fp16, |
90 | | - lib.cadam_8bit_blockwise_grad_bf16, |
91 | | - ), |
92 | | - "momentum": ( |
93 | | - lib.cmomentum_8bit_blockwise_grad_fp32, |
94 | | - lib.cmomentum_8bit_blockwise_grad_fp16, |
95 | | - lib.cmomentum_8bit_blockwise_grad_bf16, |
96 | | - ), |
97 | | - "rmsprop": ( |
98 | | - lib.crmsprop_8bit_blockwise_grad_fp32, |
99 | | - lib.crmsprop_8bit_blockwise_grad_fp16, |
100 | | - lib.crmsprop_8bit_blockwise_grad_bf16, |
101 | | - ), |
102 | | - "lion": ( |
103 | | - lib.clion_8bit_blockwise_grad_fp32, |
104 | | - lib.clion_8bit_blockwise_grad_fp16, |
105 | | - lib.clion_8bit_blockwise_grad_bf16, |
106 | | - ), |
107 | | - "adagrad": ( |
108 | | - lib.cadagrad_8bit_blockwise_grad_fp32, |
109 | | - lib.cadagrad_8bit_blockwise_grad_fp16, |
110 | | - lib.cadagrad_8bit_blockwise_grad_bf16, |
111 | | - ), |
112 | | - "ademamix": ( |
113 | | - lib.cademamix_8bit_blockwise_grad_fp32, |
114 | | - lib.cademamix_8bit_blockwise_grad_fp16, |
115 | | - lib.cademamix_8bit_blockwise_grad_bf16, |
116 | | - ), |
117 | | - } |
| 22 | +"""C FUNCTIONS FOR OPTIMIZERS""" |
| 23 | +str2optimizer32bit = { |
| 24 | + "adam": ( |
| 25 | + lib.cadam32bit_grad_fp32, |
| 26 | + lib.cadam32bit_grad_fp16, |
| 27 | + lib.cadam32bit_grad_bf16, |
| 28 | + ), |
| 29 | + "momentum": ( |
| 30 | + lib.cmomentum32bit_grad_32, |
| 31 | + lib.cmomentum32bit_grad_16, |
| 32 | + ), |
| 33 | + "rmsprop": ( |
| 34 | + lib.crmsprop32bit_grad_32, |
| 35 | + lib.crmsprop32bit_grad_16, |
| 36 | + ), |
| 37 | + "lion": ( |
| 38 | + lib.clion32bit_grad_fp32, |
| 39 | + lib.clion32bit_grad_fp16, |
| 40 | + lib.clion32bit_grad_bf16, |
| 41 | + ), |
| 42 | + "adagrad": ( |
| 43 | + lib.cadagrad32bit_grad_32, |
| 44 | + lib.cadagrad32bit_grad_16, |
| 45 | + ), |
| 46 | + "lamb": ( |
| 47 | + lib.cadam32bit_grad_fp32, |
| 48 | + lib.cadam32bit_grad_fp16, |
| 49 | + lib.cadam32bit_grad_bf16, |
| 50 | + ), |
| 51 | + "ademamix": ( |
| 52 | + lib.cademamix32bit_grad_fp32, |
| 53 | + lib.cademamix32bit_grad_fp16, |
| 54 | + lib.cademamix32bit_grad_bf16, |
| 55 | + ), |
| 56 | +} |
| 57 | + |
| 58 | +str2optimizer8bit = { |
| 59 | + "adam": ( |
| 60 | + lib.cadam_static_8bit_grad_32, |
| 61 | + lib.cadam_static_8bit_grad_16, |
| 62 | + ), |
| 63 | + "momentum": ( |
| 64 | + lib.cmomentum_static_8bit_grad_32, |
| 65 | + lib.cmomentum_static_8bit_grad_16, |
| 66 | + ), |
| 67 | + "rmsprop": ( |
| 68 | + lib.crmsprop_static_8bit_grad_32, |
| 69 | + lib.crmsprop_static_8bit_grad_16, |
| 70 | + ), |
| 71 | + "lion": ( |
| 72 | + lib.clion_static_8bit_grad_32, |
| 73 | + lib.clion_static_8bit_grad_16, |
| 74 | + ), |
| 75 | + "lamb": ( |
| 76 | + lib.cadam_static_8bit_grad_32, |
| 77 | + lib.cadam_static_8bit_grad_16, |
| 78 | + ), |
| 79 | + "lars": ( |
| 80 | + lib.cmomentum_static_8bit_grad_32, |
| 81 | + lib.cmomentum_static_8bit_grad_16, |
| 82 | + ), |
| 83 | +} |
| 84 | + |
| 85 | +str2optimizer8bit_blockwise = { |
| 86 | + "adam": ( |
| 87 | + lib.cadam_8bit_blockwise_grad_fp32, |
| 88 | + lib.cadam_8bit_blockwise_grad_fp16, |
| 89 | + lib.cadam_8bit_blockwise_grad_bf16, |
| 90 | + ), |
| 91 | + "momentum": ( |
| 92 | + lib.cmomentum_8bit_blockwise_grad_fp32, |
| 93 | + lib.cmomentum_8bit_blockwise_grad_fp16, |
| 94 | + lib.cmomentum_8bit_blockwise_grad_bf16, |
| 95 | + ), |
| 96 | + "rmsprop": ( |
| 97 | + lib.crmsprop_8bit_blockwise_grad_fp32, |
| 98 | + lib.crmsprop_8bit_blockwise_grad_fp16, |
| 99 | + lib.crmsprop_8bit_blockwise_grad_bf16, |
| 100 | + ), |
| 101 | + "lion": ( |
| 102 | + lib.clion_8bit_blockwise_grad_fp32, |
| 103 | + lib.clion_8bit_blockwise_grad_fp16, |
| 104 | + lib.clion_8bit_blockwise_grad_bf16, |
| 105 | + ), |
| 106 | + "adagrad": ( |
| 107 | + lib.cadagrad_8bit_blockwise_grad_fp32, |
| 108 | + lib.cadagrad_8bit_blockwise_grad_fp16, |
| 109 | + lib.cadagrad_8bit_blockwise_grad_bf16, |
| 110 | + ), |
| 111 | + "ademamix": ( |
| 112 | + lib.cademamix_8bit_blockwise_grad_fp32, |
| 113 | + lib.cademamix_8bit_blockwise_grad_fp16, |
| 114 | + lib.cademamix_8bit_blockwise_grad_bf16, |
| 115 | + ), |
| 116 | +} |
118 | 117 |
|
119 | 118 |
|
120 | 119 | class GlobalPageManager: |
|
0 commit comments