(self, args)
| 94 | |
| 95 | class MPTAttentionFused(nn.Module): |
| 96 | def __init__(self, args): |
| 97 | super().__init__() |
| 98 | self.args = args |
| 99 | self.n_local_heads = args.n_heads |
| 100 | self.head_dim = args.d_model // args.n_heads |
| 101 | args.max_seq_len = min(args.max_seq_len, global_max_seq_len) |
| 102 | |
| 103 | self.Wqkv = nn.Linear( |
| 104 | args.d_model, |
| 105 | args.n_heads * self.head_dim * 3, |
| 106 | bias=False, |
| 107 | ) |
| 108 | |
| 109 | self.out_proj = nn.Linear( |
| 110 | args.n_heads * self.head_dim, |
| 111 | args.d_model, |
| 112 | bias=False, |
| 113 | ) |
| 114 | |
| 115 | # following fastertransformer definition |
| 116 | |
| 117 | self.cache_v = ( |
| 118 | torch.zeros( |
| 119 | ( |
| 120 | max_batch_size, |
| 121 | self.n_local_heads, |
| 122 | args.max_seq_len, |
| 123 | self.head_dim, |
| 124 | ) |
| 125 | ) |
| 126 | .cuda() |
| 127 | .half() |
| 128 | ) # added to half |
| 129 | # 8: pack 8 fp16 in FT, if fp32 then use 4 |
| 130 | self.cache_k = ( |
| 131 | torch.zeros( |
| 132 | ( |
| 133 | max_batch_size, |
| 134 | self.n_local_heads, |
| 135 | self.head_dim // 8, |
| 136 | args.max_seq_len, |
| 137 | 8, |
| 138 | ) |
| 139 | ) |
| 140 | .cuda() |
| 141 | .half() |
| 142 | ) # added to half |
| 143 | |
| 144 | alibi_slopes, alibi_bias = build_alibi_bias( |
| 145 | self.n_local_heads, args.max_seq_len |
| 146 | ) |
| 147 | # TODO (Haotian): fix device |
| 148 | self.alibi_slopes = alibi_slopes.float().to("cuda:0") |
| 149 | self.alibi_bias = alibi_bias.to("cuda:0") |
| 150 | |
| 151 | def forward( |
| 152 | self, |
nothing calls this directly
no test coverage detected