Update modeling_mistral.py
Browse files- modeling_mistral.py +1 -35
modeling_mistral.py
CHANGED
@@ -34,7 +34,7 @@ from transformers.utils import (
|
|
34 |
)
|
35 |
from .configuration_mistral import MistralConfig
|
36 |
from transformers.models.mistral.modeling_mistral import MistralRMSNorm
|
37 |
-
|
38 |
|
39 |
logger = logging.get_logger(__name__)
|
40 |
|
@@ -58,40 +58,6 @@ class MistralMLP(nn.Module):
|
|
58 |
return down_proj
|
59 |
|
60 |
|
61 |
-
def rotate_half(x):
|
62 |
-
"""Rotates half the hidden dims of the input."""
|
63 |
-
x1 = x[..., : x.shape[-1] // 2]
|
64 |
-
x2 = x[..., x.shape[-1] // 2 :]
|
65 |
-
return torch.cat((-x2, x1), dim=-1)
|
66 |
-
|
67 |
-
|
68 |
-
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
69 |
-
"""Applies Rotary Position Embedding to the query and key tensors.
|
70 |
-
|
71 |
-
Args:
|
72 |
-
q (`torch.Tensor`): The query tensor.
|
73 |
-
k (`torch.Tensor`): The key tensor.
|
74 |
-
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
75 |
-
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
76 |
-
position_ids (`torch.Tensor`, *optional*):
|
77 |
-
Deprecated and unused.
|
78 |
-
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
79 |
-
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
80 |
-
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
81 |
-
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
82 |
-
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
83 |
-
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
84 |
-
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
85 |
-
Returns:
|
86 |
-
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
87 |
-
"""
|
88 |
-
cos = cos.unsqueeze(unsqueeze_dim)
|
89 |
-
sin = sin.unsqueeze(unsqueeze_dim)
|
90 |
-
q_embed = (q * cos) + (rotate_half(q) * sin)
|
91 |
-
k_embed = (k * cos) + (rotate_half(k) * sin)
|
92 |
-
return q_embed, k_embed
|
93 |
-
|
94 |
-
|
95 |
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
96 |
"""
|
97 |
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
|
|
34 |
)
|
35 |
from .configuration_mistral import MistralConfig
|
36 |
from transformers.models.mistral.modeling_mistral import MistralRMSNorm
|
37 |
+
from transformers.models.mistral.modeling_mistral import apply_rotary_pos_emb
|
38 |
|
39 |
logger = logging.get_logger(__name__)
|
40 |
|
|
|
58 |
return down_proj
|
59 |
|
60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
62 |
"""
|
63 |
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|