| 71 | |
| 72 | |
| 73 | class LlamaAttentionFused(nn.Module): |
| 74 | def __init__(self, args): |
| 75 | super().__init__() |
| 76 | self.args = args |
| 77 | self.n_local_heads = args.num_attention_heads |
| 78 | self.hidden_size = args.hidden_size |
| 79 | self.num_heads = args.num_attention_heads |
| 80 | self.head_dim = self.hidden_size // self.num_heads |
| 81 | |
| 82 | self.num_key_value_heads = args.num_key_value_heads |
| 83 | self.num_key_value_groups = self.num_heads // self.num_key_value_heads |
| 84 | self.max_position_embeddings = args.max_position_embeddings |
| 85 | # self.rope_theta = args.rope_theta |
| 86 | |
| 87 | kv_max_seq_len = min(max_seq_len, self.max_position_embeddings) |
| 88 | |
| 89 | self.q_proj = nn.Linear( |
| 90 | self.hidden_size, |
| 91 | self.num_heads * self.head_dim, |
| 92 | bias=False, |
| 93 | ) |
| 94 | self.k_proj = nn.Linear( |
| 95 | self.hidden_size, |
| 96 | self.num_key_value_heads * self.head_dim, |
| 97 | bias=False, |
| 98 | ) |
| 99 | self.v_proj = nn.Linear( |
| 100 | self.hidden_size, |
| 101 | self.num_key_value_heads * self.head_dim, |
| 102 | bias=False, |
| 103 | ) |
| 104 | self.o_proj = nn.Linear( |
| 105 | self.num_heads * self.head_dim, |
| 106 | self.hidden_size, |
| 107 | bias=False, |
| 108 | ) |
| 109 | |
| 110 | # following fastertransformer definition |
| 111 | self.cache_v = ( |
| 112 | torch.zeros( |
| 113 | ( |
| 114 | max_batch_size, |
| 115 | self.num_key_value_heads, |
| 116 | # args.max_position_embeddings, |
| 117 | kv_max_seq_len, |
| 118 | self.head_dim, |
| 119 | ) |
| 120 | ) |
| 121 | .cuda() |
| 122 | .half() |
| 123 | ) # added to half |
| 124 | # 8: pack 8 fp16 in FT, if fp32 then use 4 |
| 125 | self.cache_k = ( |
| 126 | torch.zeros( |
| 127 | ( |
| 128 | max_batch_size, |
| 129 | self.num_key_value_heads, |
| 130 | self.head_dim // 8, |