upload custom code
Browse files- attention.py +771 -0
- blocks.py +120 -0
- configuration.py +207 -0
- modeling_mpt.py +837 -0
- utils.py +17 -0
attention.py
ADDED
|
@@ -0,0 +1,771 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Adapted from https://github.com/mosaicml/llm-foundry
|
| 2 |
+
# Classes changed: MultiheadAttention
|
| 3 |
+
# Functions changed: scaled_multihead_dot_product_attention, build_alibi_bias, build_attn_bias
|
| 4 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 5 |
+
|
| 6 |
+
"""Attention layers."""
|
| 7 |
+
import math
|
| 8 |
+
import warnings
|
| 9 |
+
from typing import Optional
|
| 10 |
+
import torch
|
| 11 |
+
import torch.nn as nn
|
| 12 |
+
from einops import rearrange
|
| 13 |
+
from packaging import version
|
| 14 |
+
from torch import nn
|
| 15 |
+
from torch.linalg import vector_norm
|
| 16 |
+
from llmfoundry.models.layers.norm import LPLayerNorm
|
| 17 |
+
from torch.nn import functional as F
|
| 18 |
+
|
| 19 |
+
def _reset_is_causal(num_query_tokens: int, num_key_tokens: int,
|
| 20 |
+
original_is_causal: bool):
|
| 21 |
+
# disable causal when it is not needed
|
| 22 |
+
# necessary for flash & triton for generation with kv_cache
|
| 23 |
+
if original_is_causal and num_query_tokens != num_key_tokens:
|
| 24 |
+
if num_query_tokens != 1:
|
| 25 |
+
raise NotImplementedError(
|
| 26 |
+
'MPT does not support query and key with different number of tokens, unless number of query tokens is 1.'
|
| 27 |
+
)
|
| 28 |
+
else:
|
| 29 |
+
return False
|
| 30 |
+
return original_is_causal
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def scaled_multihead_dot_product_attention(
|
| 34 |
+
query,
|
| 35 |
+
key,
|
| 36 |
+
value,
|
| 37 |
+
n_heads,
|
| 38 |
+
past_key_value=None,
|
| 39 |
+
long_range_past_key_value=None,
|
| 40 |
+
softmax_scale=None,
|
| 41 |
+
attn_bias=None,
|
| 42 |
+
attn_bias_ae=None,
|
| 43 |
+
key_padding_mask=None,
|
| 44 |
+
is_causal=False,
|
| 45 |
+
dropout_p=0.0,
|
| 46 |
+
training=False,
|
| 47 |
+
needs_weights=False,
|
| 48 |
+
multiquery=False,
|
| 49 |
+
topk=None,
|
| 50 |
+
faiss_indexes=None,
|
| 51 |
+
n_layers=None,
|
| 52 |
+
current_layer=None,
|
| 53 |
+
mask_by_sim=False,
|
| 54 |
+
sim_threshold=0.0
|
| 55 |
+
):
|
| 56 |
+
q = rearrange(query, 'b s (h d) -> b h s d', h=n_heads)
|
| 57 |
+
kv_n_heads = 1 if multiquery else n_heads
|
| 58 |
+
k = rearrange(key, 'b s (h d) -> b h d s', h=kv_n_heads)
|
| 59 |
+
v = rearrange(value, 'b s (h d) -> b h s d', h=kv_n_heads)
|
| 60 |
+
|
| 61 |
+
had_kv=False
|
| 62 |
+
if past_key_value is not None:
|
| 63 |
+
# attn_impl: flash & triton use kernels which expect input shape [b, s, h, d_head].
|
| 64 |
+
# kv_cache is therefore stored using that shape.
|
| 65 |
+
# attn_impl: torch stores the kv_cache in the ordering which is most advantageous
|
| 66 |
+
# for its attn computation ie
|
| 67 |
+
# keys are stored as tensors with shape [b, h, d_head, s] and
|
| 68 |
+
# values are stored as tensors with shape [b, h, s, d_head]
|
| 69 |
+
if len(past_key_value) != 0:
|
| 70 |
+
k = torch.cat([past_key_value[0], k], dim=3)
|
| 71 |
+
v = torch.cat([past_key_value[1], v], dim=2)
|
| 72 |
+
had_kv=True
|
| 73 |
+
|
| 74 |
+
past_key_value = (k, v)
|
| 75 |
+
|
| 76 |
+
b, h, s_q, d = q.shape
|
| 77 |
+
s_k = k.size(-1)
|
| 78 |
+
|
| 79 |
+
if softmax_scale is None:
|
| 80 |
+
softmax_scale = 1 / math.sqrt(d)
|
| 81 |
+
|
| 82 |
+
attn_weight = q.matmul(k) * softmax_scale
|
| 83 |
+
|
| 84 |
+
if attn_bias is not None:
|
| 85 |
+
# clamp to 0 necessary for torch 2.0 compile()
|
| 86 |
+
_s_q = max(0, attn_bias.size(2) - s_q)
|
| 87 |
+
_s_k = max(0, attn_bias.size(3) - s_k)
|
| 88 |
+
attn_bias = attn_bias[:, :, _s_q:, _s_k:]
|
| 89 |
+
|
| 90 |
+
if (attn_bias.size(-1) != 1 and
|
| 91 |
+
attn_bias.size(-1) != s_k) or (attn_bias.size(-2) != 1 and
|
| 92 |
+
attn_bias.size(-2) != s_q):
|
| 93 |
+
raise RuntimeError(
|
| 94 |
+
f'attn_bias (shape: {attn_bias.shape}) is expected to broadcast to shape: {attn_weight.shape}.'
|
| 95 |
+
)
|
| 96 |
+
attn_weight = attn_weight + attn_bias
|
| 97 |
+
|
| 98 |
+
if needs_weights:
|
| 99 |
+
reshaped_idx = None
|
| 100 |
+
if long_range_past_key_value is not None or faiss_indexes is not None:
|
| 101 |
+
if long_range_past_key_value is not None: #manual external memories
|
| 102 |
+
|
| 103 |
+
k_cache, v_cache = long_range_past_key_value
|
| 104 |
+
s_cache = k_cache.size(-1)
|
| 105 |
+
|
| 106 |
+
k_cache = k_cache.to(k.device)
|
| 107 |
+
v_cache = v_cache.to(k.device)
|
| 108 |
+
|
| 109 |
+
q_n = q/vector_norm(q, ord=2, dim=-1, keepdim=True)
|
| 110 |
+
k_n = k_cache/vector_norm(k_cache, ord=2, dim=-2, keepdim=True)
|
| 111 |
+
|
| 112 |
+
sim = q_n.matmul(k_n)
|
| 113 |
+
if s_cache<topk:
|
| 114 |
+
topk = s_cache #number of tokens in cache < topk
|
| 115 |
+
val, idx = torch.topk(sim, k=topk, dim=-1)
|
| 116 |
+
|
| 117 |
+
reshaped_idx = idx.reshape(b, h, s_q * topk)
|
| 118 |
+
|
| 119 |
+
selected_k = k_cache.gather(dim=-1, index=reshaped_idx.unsqueeze(-2).expand(-1, -1, d, -1))
|
| 120 |
+
selected_v = v_cache.gather(dim=-2, index=reshaped_idx.unsqueeze(-1).expand(-1, -1, -1, d))
|
| 121 |
+
|
| 122 |
+
sim_mask = rearrange(~ (val > sim_threshold).bool(), 'b h s i -> b h (s i)').unsqueeze(-2).expand(-1, -1, s_q, -1)
|
| 123 |
+
min_val = torch.finfo(selected_k.dtype).min
|
| 124 |
+
|
| 125 |
+
elif faiss_indexes is not None: #faiss indexes
|
| 126 |
+
|
| 127 |
+
kn_index, kv_index = faiss_indexes
|
| 128 |
+
q_n = q/vector_norm(q, ord=2, dim=-1, keepdim=True)
|
| 129 |
+
|
| 130 |
+
one_hot_encodings = F.one_hot(torch.arange(0, n_heads*n_layers, device=q.device))*10
|
| 131 |
+
q_n = torch.concat([rearrange(q_n, 'b h s d -> b (h s) d', h=n_heads), one_hot_encodings[n_heads*current_layer:n_heads*(current_layer+1)].unsqueeze(0).repeat_interleave(repeats=q.size(-2), dim=-2)], dim=-1).squeeze()
|
| 132 |
+
|
| 133 |
+
D, I = kn_index.search(q_n.to('cpu').numpy(), k=topk)
|
| 134 |
+
|
| 135 |
+
selected_k=rearrange(torch.tensor(kv_index.reconstruct_batch(I.flatten()))[:,:d], '(h s) d -> 1 h d s', h=32).to(q.device)
|
| 136 |
+
selected_v=rearrange(torch.tensor(kv_index.reconstruct_batch(I.flatten()))[:,d:], '(h s) d -> 1 h s d', h=32).to(q.device)
|
| 137 |
+
|
| 138 |
+
s_k_ae = selected_k.size(-1)
|
| 139 |
+
s_k += s_k_ae
|
| 140 |
+
attn_weight_cache = q.matmul(selected_k) * softmax_scale
|
| 141 |
+
if mask_by_sim:
|
| 142 |
+
attn_weight_cache = attn_weight_cache.masked_fill(sim_mask, min_val)
|
| 143 |
+
|
| 144 |
+
if attn_bias_ae is not None:
|
| 145 |
+
# clamp to 0 necessary for torch 2.0 compile()
|
| 146 |
+
_s_q = max(0, attn_bias_ae.size(2) - s_q)
|
| 147 |
+
_s_k = max(0, attn_bias_ae.size(3) - s_k_ae)
|
| 148 |
+
attn_bias_ae = attn_bias_ae[:, :, _s_q:, _s_k:]
|
| 149 |
+
|
| 150 |
+
if (attn_bias_ae.size(-1) != 1 and
|
| 151 |
+
attn_bias_ae.size(-1) != s_k_ae) or (attn_bias_ae.size(-2) != 1 and
|
| 152 |
+
attn_bias_ae.size(-2) != s_q):
|
| 153 |
+
raise RuntimeError(
|
| 154 |
+
f'attn_bias (shape: {attn_bias_ae.shape}) is expected to broadcast to shape: {attn_weight_cache.shape}.'
|
| 155 |
+
)
|
| 156 |
+
attn_weight_cache = attn_weight_cache + attn_bias_ae
|
| 157 |
+
|
| 158 |
+
attn_weight = torch.cat([attn_weight_cache, attn_weight], dim=-1)
|
| 159 |
+
v = torch.cat([selected_v, v], dim=-2)
|
| 160 |
+
|
| 161 |
+
min_val = torch.finfo(q.dtype).min
|
| 162 |
+
|
| 163 |
+
if key_padding_mask is not None:
|
| 164 |
+
if attn_bias is not None:
|
| 165 |
+
warnings.warn(
|
| 166 |
+
'Propogating key_padding_mask to the attention module ' +\
|
| 167 |
+
'and applying it within the attention module can cause ' +\
|
| 168 |
+
'unneccessary computation/memory usage. Consider integrating ' +\
|
| 169 |
+
'into attn_bias once and passing that to each attention ' +\
|
| 170 |
+
'module instead.'
|
| 171 |
+
)
|
| 172 |
+
attn_weight = attn_weight.masked_fill(
|
| 173 |
+
~key_padding_mask.view((b, 1, 1, s_k)), min_val)
|
| 174 |
+
|
| 175 |
+
def _create_active_externalism_mask(k, s_q, device):
|
| 176 |
+
mask = torch.zeros(s_q, s_q * k, device=device, dtype=torch.bool)
|
| 177 |
+
for i in range(s_q):
|
| 178 |
+
mask[i, i * k : (i + 1) * k] = 1
|
| 179 |
+
return ~mask
|
| 180 |
+
|
| 181 |
+
if is_causal and (not q.size(2) == 1):
|
| 182 |
+
s = max(s_q, s_k)
|
| 183 |
+
causal_mask = attn_weight.new_ones(s, s, dtype=torch.float16)
|
| 184 |
+
causal_mask = causal_mask.tril()
|
| 185 |
+
causal_mask = causal_mask.to(torch.bool)
|
| 186 |
+
causal_mask = ~causal_mask
|
| 187 |
+
causal_mask = causal_mask[-s_q:, -s_k:]
|
| 188 |
+
|
| 189 |
+
if long_range_past_key_value is not None:
|
| 190 |
+
mask = _create_active_externalism_mask(k=topk,s_q=s_q, device=attn_weight.device)
|
| 191 |
+
s=s_q
|
| 192 |
+
if had_kv:
|
| 193 |
+
s += (past_key_value[0][0].size(-1) -s_q)
|
| 194 |
+
causal_mask = torch.cat([mask, causal_mask[:,-s:]], dim=1)
|
| 195 |
+
|
| 196 |
+
attn_weight = attn_weight.masked_fill(causal_mask.view(1, 1, s_q, s_k),
|
| 197 |
+
min_val)
|
| 198 |
+
|
| 199 |
+
attn_weight = torch.softmax(attn_weight, dim=-1)
|
| 200 |
+
|
| 201 |
+
if dropout_p:
|
| 202 |
+
attn_weight = torch.nn.functional.dropout(attn_weight,
|
| 203 |
+
p=dropout_p,
|
| 204 |
+
training=training,
|
| 205 |
+
inplace=True)
|
| 206 |
+
|
| 207 |
+
out = attn_weight.to(v.dtype).matmul(v)
|
| 208 |
+
out = rearrange(out, 'b h s d -> b s (h d)')
|
| 209 |
+
|
| 210 |
+
if needs_weights:
|
| 211 |
+
return out, attn_weight, past_key_value, reshaped_idx
|
| 212 |
+
return out, None, past_key_value, None
|
| 213 |
+
|
| 214 |
+
|
| 215 |
+
def check_valid_inputs(*tensors, valid_dtypes=[torch.float16, torch.bfloat16]):
|
| 216 |
+
for tensor in tensors:
|
| 217 |
+
if tensor.dtype not in valid_dtypes:
|
| 218 |
+
raise TypeError(f'{tensor.dtype=} must be in {valid_dtypes=}.')
|
| 219 |
+
if not tensor.is_cuda:
|
| 220 |
+
raise TypeError(f'Inputs must be cuda tensors ({tensor.is_cuda=}).')
|
| 221 |
+
|
| 222 |
+
|
| 223 |
+
def flash_attn_fn(
|
| 224 |
+
query,
|
| 225 |
+
key,
|
| 226 |
+
value,
|
| 227 |
+
n_heads,
|
| 228 |
+
past_key_value=None,
|
| 229 |
+
softmax_scale=None,
|
| 230 |
+
attn_bias=None,
|
| 231 |
+
key_padding_mask=None,
|
| 232 |
+
is_causal=False,
|
| 233 |
+
dropout_p=0.0,
|
| 234 |
+
training=False,
|
| 235 |
+
needs_weights=False,
|
| 236 |
+
multiquery=False,
|
| 237 |
+
):
|
| 238 |
+
try:
|
| 239 |
+
from flash_attn import bert_padding, flash_attn_interface # type: ignore # yapf: disable # isort: skip
|
| 240 |
+
except:
|
| 241 |
+
raise RuntimeError('Please install flash-attn==1.0.3.post0')
|
| 242 |
+
|
| 243 |
+
check_valid_inputs(query, key, value)
|
| 244 |
+
|
| 245 |
+
if past_key_value is not None:
|
| 246 |
+
if len(past_key_value) != 0:
|
| 247 |
+
key = torch.cat([past_key_value[0], key], dim=1)
|
| 248 |
+
value = torch.cat([past_key_value[1], value], dim=1)
|
| 249 |
+
|
| 250 |
+
past_key_value = (key, value)
|
| 251 |
+
|
| 252 |
+
if attn_bias is not None:
|
| 253 |
+
# clamp to 0 necessary for torch 2.0 compile()
|
| 254 |
+
_s_q = max(0, attn_bias.size(2) - query.size(1))
|
| 255 |
+
_s_k = max(0, attn_bias.size(3) - key.size(1))
|
| 256 |
+
attn_bias = attn_bias[:, :, _s_q:, _s_k:]
|
| 257 |
+
|
| 258 |
+
if attn_bias is not None:
|
| 259 |
+
raise NotImplementedError(f'attn_bias not implemented for flash attn.')
|
| 260 |
+
|
| 261 |
+
batch_size, seqlen = query.shape[:2]
|
| 262 |
+
|
| 263 |
+
if key_padding_mask is None:
|
| 264 |
+
key_padding_mask = torch.ones_like(key[:, :, 0], dtype=torch.bool)
|
| 265 |
+
query_padding_mask = key_padding_mask[:, -query.size(1):]
|
| 266 |
+
|
| 267 |
+
query_unpad, indices_q, cu_seqlens_q, max_seqlen_q = bert_padding.unpad_input(
|
| 268 |
+
query, query_padding_mask)
|
| 269 |
+
query_unpad = rearrange(query_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
|
| 270 |
+
|
| 271 |
+
key_unpad, _, cu_seqlens_k, max_seqlen_k = bert_padding.unpad_input(
|
| 272 |
+
key, key_padding_mask)
|
| 273 |
+
key_unpad = rearrange(key_unpad,
|
| 274 |
+
'nnz (h d) -> nnz h d',
|
| 275 |
+
h=1 if multiquery else n_heads)
|
| 276 |
+
|
| 277 |
+
value_unpad, _, _, _ = bert_padding.unpad_input(value, key_padding_mask)
|
| 278 |
+
value_unpad = rearrange(value_unpad,
|
| 279 |
+
'nnz (h d) -> nnz h d',
|
| 280 |
+
h=1 if multiquery else n_heads)
|
| 281 |
+
|
| 282 |
+
if multiquery:
|
| 283 |
+
# Expanding a tensor does not allocate new memory, but only creates a new
|
| 284 |
+
# view on the existing tensor where a dimension of size one is expanded
|
| 285 |
+
# to a larger size by setting the stride to 0.
|
| 286 |
+
# - pytorch docs
|
| 287 |
+
#
|
| 288 |
+
# hopefully the kernels can utilize this and we're jot just wasting BW here
|
| 289 |
+
key_unpad = key_unpad.expand(key_unpad.size(0), n_heads,
|
| 290 |
+
key_unpad.size(-1))
|
| 291 |
+
value_unpad = value_unpad.expand(value_unpad.size(0), n_heads,
|
| 292 |
+
value_unpad.size(-1))
|
| 293 |
+
|
| 294 |
+
dropout_p = dropout_p if training else 0.0
|
| 295 |
+
|
| 296 |
+
reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
|
| 297 |
+
|
| 298 |
+
output_unpad = flash_attn_interface.flash_attn_unpadded_func(
|
| 299 |
+
query_unpad,
|
| 300 |
+
key_unpad,
|
| 301 |
+
value_unpad,
|
| 302 |
+
cu_seqlens_q,
|
| 303 |
+
cu_seqlens_k,
|
| 304 |
+
max_seqlen_q,
|
| 305 |
+
max_seqlen_k,
|
| 306 |
+
dropout_p,
|
| 307 |
+
softmax_scale=softmax_scale,
|
| 308 |
+
causal=reset_is_causal,
|
| 309 |
+
return_attn_probs=needs_weights)
|
| 310 |
+
|
| 311 |
+
output = bert_padding.pad_input(
|
| 312 |
+
rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices_q, batch_size,
|
| 313 |
+
seqlen)
|
| 314 |
+
return output, None, past_key_value
|
| 315 |
+
|
| 316 |
+
|
| 317 |
+
def triton_flash_attn_fn(
|
| 318 |
+
query,
|
| 319 |
+
key,
|
| 320 |
+
value,
|
| 321 |
+
n_heads,
|
| 322 |
+
past_key_value=None,
|
| 323 |
+
softmax_scale=None,
|
| 324 |
+
attn_bias=None,
|
| 325 |
+
key_padding_mask=None,
|
| 326 |
+
is_causal=False,
|
| 327 |
+
dropout_p=0.0,
|
| 328 |
+
training=False,
|
| 329 |
+
needs_weights=False,
|
| 330 |
+
multiquery=False,
|
| 331 |
+
):
|
| 332 |
+
try:
|
| 333 |
+
from llmfoundry.models.layers.flash_attn_triton import flash_attn_func
|
| 334 |
+
except:
|
| 335 |
+
_installed = False
|
| 336 |
+
if version.parse(torch.__version__) < version.parse('2.0.0'):
|
| 337 |
+
_installed = True
|
| 338 |
+
# if torch1.13.1 revert to using triton flash attn from HazyResearch
|
| 339 |
+
# with flash-attn==1.0.3.post0 and triton==2.0.0.dev20221202
|
| 340 |
+
try:
|
| 341 |
+
from flash_attn.flash_attn_triton import flash_attn_func
|
| 342 |
+
except:
|
| 343 |
+
_installed = False
|
| 344 |
+
if not _installed:
|
| 345 |
+
# installing triton-pre-mlir works for both torch1.13.1 and torch2.0+
|
| 346 |
+
# default recommendation is to install this variant
|
| 347 |
+
raise RuntimeError(
|
| 348 |
+
'Requirements for `attn_impl: triton` not installed. Either (1) have a CUDA-compatible GPU '
|
| 349 |
+
'and `pip install .[gpu]` if installing from llm-foundry source or '
|
| 350 |
+
'`pip install triton-pre-mlir@git+https://github.com/vchiley/triton.git@triton_pre_mlir#subdirectory=python` '
|
| 351 |
+
'if installing from pypi, or (2) use torch attn model.attn_config.attn_impl=torch (torch attn_impl will be slow). '
|
| 352 |
+
'Note: (1) requires you have CMake and PyTorch already installed.'
|
| 353 |
+
)
|
| 354 |
+
|
| 355 |
+
check_valid_inputs(query, key, value)
|
| 356 |
+
|
| 357 |
+
if past_key_value is not None:
|
| 358 |
+
if len(past_key_value) != 0:
|
| 359 |
+
key = torch.cat([past_key_value[0], key], dim=1)
|
| 360 |
+
value = torch.cat([past_key_value[1], value], dim=1)
|
| 361 |
+
|
| 362 |
+
past_key_value = (key, value)
|
| 363 |
+
|
| 364 |
+
if attn_bias is not None:
|
| 365 |
+
# clamp to 0 necessary for torch 2.0 compile()
|
| 366 |
+
_s_q = max(0, attn_bias.size(2) - query.size(1))
|
| 367 |
+
_s_k = max(0, attn_bias.size(3) - key.size(1))
|
| 368 |
+
attn_bias = attn_bias[:, :, _s_q:, _s_k:]
|
| 369 |
+
|
| 370 |
+
if dropout_p:
|
| 371 |
+
raise NotImplementedError(
|
| 372 |
+
f'Dropout not implemented for attn_impl: triton.')
|
| 373 |
+
|
| 374 |
+
if needs_weights:
|
| 375 |
+
raise NotImplementedError(
|
| 376 |
+
f'attn_impl: triton cannot return attn weights.')
|
| 377 |
+
|
| 378 |
+
if key_padding_mask is not None:
|
| 379 |
+
warnings.warn(
|
| 380 |
+
'Propagating key_padding_mask to the attention module ' +\
|
| 381 |
+
'and applying it within the attention module can cause ' +\
|
| 382 |
+
'unnecessary computation/memory usage. Consider integrating ' +\
|
| 383 |
+
'into attn_bias once and passing that to each attention ' +\
|
| 384 |
+
'module instead.'
|
| 385 |
+
)
|
| 386 |
+
b_size, s_k = key_padding_mask.shape[:2]
|
| 387 |
+
|
| 388 |
+
if attn_bias is None:
|
| 389 |
+
attn_bias = query.new_zeros(b_size, 1, 1, s_k)
|
| 390 |
+
|
| 391 |
+
attn_bias = attn_bias.masked_fill(
|
| 392 |
+
~key_padding_mask.view((b_size, 1, 1, s_k)),
|
| 393 |
+
torch.finfo(query.dtype).min)
|
| 394 |
+
|
| 395 |
+
query = rearrange(query, 'b s (h d) -> b s h d', h=n_heads)
|
| 396 |
+
key = rearrange(key, 'b s (h d) -> b s h d', h=1 if multiquery else n_heads)
|
| 397 |
+
value = rearrange(value,
|
| 398 |
+
'b s (h d) -> b s h d',
|
| 399 |
+
h=1 if multiquery else n_heads)
|
| 400 |
+
|
| 401 |
+
if multiquery:
|
| 402 |
+
# Expanding a tensor does not allocate new memory, but only creates a new
|
| 403 |
+
# view on the existing tensor where a dimension of size one is expanded
|
| 404 |
+
# to a larger size by setting the stride to 0.
|
| 405 |
+
# - pytorch docs
|
| 406 |
+
#
|
| 407 |
+
# hopefully the kernels can utilize this and we're jot just wasting BW here
|
| 408 |
+
key = key.expand(*key.shape[:2], n_heads, key.size(-1))
|
| 409 |
+
value = value.expand(*value.shape[:2], n_heads, value.size(-1))
|
| 410 |
+
|
| 411 |
+
reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
|
| 412 |
+
attn_output = flash_attn_func(query, key, value, attn_bias, reset_is_causal,
|
| 413 |
+
softmax_scale)
|
| 414 |
+
|
| 415 |
+
output = attn_output.view(*attn_output.shape[:2], -1)
|
| 416 |
+
|
| 417 |
+
return output, None, past_key_value
|
| 418 |
+
|
| 419 |
+
|
| 420 |
+
class MultiheadAttention(nn.Module):
|
| 421 |
+
"""Multi-head self attention.
|
| 422 |
+
|
| 423 |
+
Using torch or triton attention implemetation enables user to also use
|
| 424 |
+
additive bias.
|
| 425 |
+
"""
|
| 426 |
+
|
| 427 |
+
def __init__(
|
| 428 |
+
self,
|
| 429 |
+
d_model: int,
|
| 430 |
+
n_heads: int,
|
| 431 |
+
attn_impl: str = 'triton',
|
| 432 |
+
clip_qkv: Optional[float] = None,
|
| 433 |
+
qk_ln: bool = False,
|
| 434 |
+
softmax_scale: Optional[float] = None,
|
| 435 |
+
attn_pdrop: float = 0.0,
|
| 436 |
+
low_precision_layernorm: bool = False,
|
| 437 |
+
verbose: int = 0,
|
| 438 |
+
device: Optional[str] = None,
|
| 439 |
+
):
|
| 440 |
+
super().__init__()
|
| 441 |
+
|
| 442 |
+
self.attn_impl = attn_impl
|
| 443 |
+
self.clip_qkv = clip_qkv
|
| 444 |
+
self.qk_ln = qk_ln
|
| 445 |
+
|
| 446 |
+
self.d_model = d_model
|
| 447 |
+
self.n_heads = n_heads
|
| 448 |
+
self.softmax_scale = softmax_scale
|
| 449 |
+
if self.softmax_scale is None:
|
| 450 |
+
self.softmax_scale = 1 / math.sqrt(self.d_model / self.n_heads)
|
| 451 |
+
self.attn_dropout_p = attn_pdrop
|
| 452 |
+
|
| 453 |
+
self.Wqkv = nn.Linear(self.d_model, 3 * self.d_model, device=device)
|
| 454 |
+
# for param init fn; enables shape based init of fused layers
|
| 455 |
+
fuse_splits = (d_model, 2 * d_model)
|
| 456 |
+
self.Wqkv._fused = (0, fuse_splits) # type: ignore
|
| 457 |
+
|
| 458 |
+
if self.qk_ln:
|
| 459 |
+
layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
|
| 460 |
+
self.q_ln = layernorm_class(self.d_model, device=device)
|
| 461 |
+
self.k_ln = layernorm_class(self.d_model, device=device)
|
| 462 |
+
|
| 463 |
+
if self.attn_impl == 'flash':
|
| 464 |
+
self.attn_fn = flash_attn_fn
|
| 465 |
+
elif self.attn_impl == 'triton':
|
| 466 |
+
self.attn_fn = triton_flash_attn_fn
|
| 467 |
+
if verbose:
|
| 468 |
+
warnings.warn(
|
| 469 |
+
'While `attn_impl: triton` can be faster than `attn_impl: flash` ' +\
|
| 470 |
+
'it uses more memory. When training larger models this can trigger ' +\
|
| 471 |
+
'alloc retries which hurts performance. If encountered, we recommend ' +\
|
| 472 |
+
'using `attn_impl: flash` if your model does not use `alibi` or `prefix_lm`.'
|
| 473 |
+
)
|
| 474 |
+
elif self.attn_impl == 'torch':
|
| 475 |
+
self.attn_fn = scaled_multihead_dot_product_attention
|
| 476 |
+
if torch.cuda.is_available() and verbose:
|
| 477 |
+
warnings.warn(
|
| 478 |
+
'Using `attn_impl: torch`. If your model does not use `alibi` or ' +\
|
| 479 |
+
'`prefix_lm` we recommend using `attn_impl: flash` otherwise ' +\
|
| 480 |
+
'we recommend using `attn_impl: triton`.'
|
| 481 |
+
)
|
| 482 |
+
else:
|
| 483 |
+
raise ValueError(f'{attn_impl=} is an invalid setting.')
|
| 484 |
+
|
| 485 |
+
self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
|
| 486 |
+
self.out_proj._is_residual = True # type: ignore
|
| 487 |
+
|
| 488 |
+
def forward(
|
| 489 |
+
self,
|
| 490 |
+
x,
|
| 491 |
+
past_key_value=None,
|
| 492 |
+
long_range_past_key_value=None,
|
| 493 |
+
attn_bias=None,
|
| 494 |
+
attn_bias_ae=None,
|
| 495 |
+
attention_mask=None,
|
| 496 |
+
is_causal=True,
|
| 497 |
+
needs_weights=False,
|
| 498 |
+
topk=None,
|
| 499 |
+
faiss_indexes=None,
|
| 500 |
+
n_layers=None,
|
| 501 |
+
current_layer=None,
|
| 502 |
+
mask_by_sim=None,
|
| 503 |
+
sim_threshold=None
|
| 504 |
+
):
|
| 505 |
+
qkv = self.Wqkv(x)
|
| 506 |
+
|
| 507 |
+
if self.clip_qkv:
|
| 508 |
+
qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
|
| 509 |
+
|
| 510 |
+
query, key, value = qkv.chunk(3, dim=2)
|
| 511 |
+
|
| 512 |
+
key_padding_mask = attention_mask
|
| 513 |
+
|
| 514 |
+
if self.qk_ln:
|
| 515 |
+
# Applying layernorm to qk
|
| 516 |
+
dtype = query.dtype
|
| 517 |
+
query = self.q_ln(query).to(dtype)
|
| 518 |
+
key = self.k_ln(key).to(dtype)
|
| 519 |
+
|
| 520 |
+
context, attn_weights, past_key_value, reshaped_idx = self.attn_fn(
|
| 521 |
+
query,
|
| 522 |
+
key,
|
| 523 |
+
value,
|
| 524 |
+
self.n_heads,
|
| 525 |
+
past_key_value=past_key_value,
|
| 526 |
+
long_range_past_key_value=long_range_past_key_value,
|
| 527 |
+
softmax_scale=self.softmax_scale,
|
| 528 |
+
attn_bias=attn_bias,
|
| 529 |
+
attn_bias_ae=attn_bias_ae,
|
| 530 |
+
key_padding_mask=key_padding_mask,
|
| 531 |
+
is_causal=is_causal,
|
| 532 |
+
dropout_p=self.attn_dropout_p,
|
| 533 |
+
training=self.training,
|
| 534 |
+
needs_weights=needs_weights,
|
| 535 |
+
topk=topk,
|
| 536 |
+
faiss_indexes=faiss_indexes,
|
| 537 |
+
n_layers=n_layers,
|
| 538 |
+
current_layer=current_layer,
|
| 539 |
+
mask_by_sim=mask_by_sim,
|
| 540 |
+
sim_threshold=sim_threshold
|
| 541 |
+
)
|
| 542 |
+
|
| 543 |
+
return self.out_proj(context), attn_weights, past_key_value, reshaped_idx
|
| 544 |
+
|
| 545 |
+
|
| 546 |
+
class MultiQueryAttention(nn.Module):
|
| 547 |
+
"""Multi-Query self attention.
|
| 548 |
+
|
| 549 |
+
Using torch or triton attention implemetation enables user to also use
|
| 550 |
+
additive bias.
|
| 551 |
+
"""
|
| 552 |
+
|
| 553 |
+
def __init__(
|
| 554 |
+
self,
|
| 555 |
+
d_model: int,
|
| 556 |
+
n_heads: int,
|
| 557 |
+
attn_impl: str = 'triton',
|
| 558 |
+
clip_qkv: Optional[float] = None,
|
| 559 |
+
qk_ln: bool = False,
|
| 560 |
+
softmax_scale: Optional[float] = None,
|
| 561 |
+
attn_pdrop: float = 0.0,
|
| 562 |
+
low_precision_layernorm: bool = False,
|
| 563 |
+
verbose: int = 0,
|
| 564 |
+
device: Optional[str] = None,
|
| 565 |
+
):
|
| 566 |
+
super().__init__()
|
| 567 |
+
|
| 568 |
+
self.attn_impl = attn_impl
|
| 569 |
+
self.clip_qkv = clip_qkv
|
| 570 |
+
self.qk_ln = qk_ln
|
| 571 |
+
|
| 572 |
+
self.d_model = d_model
|
| 573 |
+
self.n_heads = n_heads
|
| 574 |
+
self.head_dim = d_model // n_heads
|
| 575 |
+
self.softmax_scale = softmax_scale
|
| 576 |
+
if self.softmax_scale is None:
|
| 577 |
+
self.softmax_scale = 1 / math.sqrt(self.head_dim)
|
| 578 |
+
self.attn_dropout_p = attn_pdrop
|
| 579 |
+
|
| 580 |
+
# NOTE: if we ever want to make attn TensorParallel, I'm pretty sure we'll
|
| 581 |
+
# want to split Wqkv into Wq and Wkv where Wq can be TensorParallel but
|
| 582 |
+
# Wkv shouldn't be TensorParallel
|
| 583 |
+
# - vchiley
|
| 584 |
+
self.Wqkv = nn.Linear(
|
| 585 |
+
d_model,
|
| 586 |
+
d_model + 2 * self.head_dim,
|
| 587 |
+
device=device,
|
| 588 |
+
)
|
| 589 |
+
# for param init fn; enables shape based init of fused layers
|
| 590 |
+
fuse_splits = (d_model, d_model + self.head_dim)
|
| 591 |
+
self.Wqkv._fused = (0, fuse_splits) # type: ignore
|
| 592 |
+
|
| 593 |
+
if self.qk_ln:
|
| 594 |
+
layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
|
| 595 |
+
self.q_ln = layernorm_class(d_model, device=device)
|
| 596 |
+
self.k_ln = layernorm_class(self.head_dim, device=device)
|
| 597 |
+
|
| 598 |
+
if self.attn_impl == 'flash':
|
| 599 |
+
self.attn_fn = flash_attn_fn
|
| 600 |
+
elif self.attn_impl == 'triton':
|
| 601 |
+
self.attn_fn = triton_flash_attn_fn
|
| 602 |
+
if verbose:
|
| 603 |
+
warnings.warn(
|
| 604 |
+
'While `attn_impl: triton` can be faster than `attn_impl: flash` ' +\
|
| 605 |
+
'it uses more memory. When training larger models this can trigger ' +\
|
| 606 |
+
'alloc retries which hurts performance. If encountered, we recommend ' +\
|
| 607 |
+
'using `attn_impl: flash` if your model does not use `alibi` or `prefix_lm`.'
|
| 608 |
+
)
|
| 609 |
+
elif self.attn_impl == 'torch':
|
| 610 |
+
self.attn_fn = scaled_multihead_dot_product_attention
|
| 611 |
+
if torch.cuda.is_available() and verbose:
|
| 612 |
+
warnings.warn(
|
| 613 |
+
'Using `attn_impl: torch`. If your model does not use `alibi` or ' +\
|
| 614 |
+
'`prefix_lm` we recommend using `attn_impl: flash` otherwise ' +\
|
| 615 |
+
'we recommend using `attn_impl: triton`.'
|
| 616 |
+
)
|
| 617 |
+
else:
|
| 618 |
+
raise ValueError(f'{attn_impl=} is an invalid setting.')
|
| 619 |
+
|
| 620 |
+
self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
|
| 621 |
+
self.out_proj._is_residual = True # type: ignore
|
| 622 |
+
|
| 623 |
+
def forward(
|
| 624 |
+
self,
|
| 625 |
+
x,
|
| 626 |
+
past_key_value=None,
|
| 627 |
+
attn_bias=None,
|
| 628 |
+
attention_mask=None,
|
| 629 |
+
is_causal=True,
|
| 630 |
+
needs_weights=False,
|
| 631 |
+
):
|
| 632 |
+
qkv = self.Wqkv(x)
|
| 633 |
+
|
| 634 |
+
if self.clip_qkv:
|
| 635 |
+
qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
|
| 636 |
+
|
| 637 |
+
query, key, value = qkv.split(
|
| 638 |
+
[self.d_model, self.head_dim, self.head_dim], dim=2)
|
| 639 |
+
|
| 640 |
+
key_padding_mask = attention_mask
|
| 641 |
+
|
| 642 |
+
if self.qk_ln:
|
| 643 |
+
# Applying layernorm to qk
|
| 644 |
+
dtype = query.dtype
|
| 645 |
+
query = self.q_ln(query).to(dtype)
|
| 646 |
+
key = self.k_ln(key).to(dtype)
|
| 647 |
+
|
| 648 |
+
context, attn_weights, past_key_value = self.attn_fn(
|
| 649 |
+
query,
|
| 650 |
+
key,
|
| 651 |
+
value,
|
| 652 |
+
self.n_heads,
|
| 653 |
+
past_key_value=past_key_value,
|
| 654 |
+
softmax_scale=self.softmax_scale,
|
| 655 |
+
attn_bias=attn_bias,
|
| 656 |
+
key_padding_mask=key_padding_mask,
|
| 657 |
+
is_causal=is_causal,
|
| 658 |
+
dropout_p=self.attn_dropout_p,
|
| 659 |
+
training=self.training,
|
| 660 |
+
needs_weights=needs_weights,
|
| 661 |
+
multiquery=True,
|
| 662 |
+
)
|
| 663 |
+
|
| 664 |
+
return self.out_proj(context), attn_weights, past_key_value
|
| 665 |
+
|
| 666 |
+
|
| 667 |
+
def attn_bias_shape(attn_impl, n_heads, seq_len, alibi, prefix_lm, causal,
|
| 668 |
+
use_sequence_id):
|
| 669 |
+
if attn_impl == 'flash':
|
| 670 |
+
return None
|
| 671 |
+
elif attn_impl in ['torch', 'triton']:
|
| 672 |
+
if alibi:
|
| 673 |
+
if (prefix_lm or not causal) or use_sequence_id:
|
| 674 |
+
return (1, n_heads, seq_len, seq_len)
|
| 675 |
+
return (1, n_heads, 1, seq_len)
|
| 676 |
+
elif prefix_lm or use_sequence_id:
|
| 677 |
+
return (1, 1, seq_len, seq_len)
|
| 678 |
+
return None
|
| 679 |
+
else:
|
| 680 |
+
raise ValueError(f'{attn_impl=} is an invalid setting.')
|
| 681 |
+
|
| 682 |
+
|
| 683 |
+
def build_attn_bias(
|
| 684 |
+
attn_impl,
|
| 685 |
+
n_heads,
|
| 686 |
+
seq_len,
|
| 687 |
+
attn_bias=None,
|
| 688 |
+
causal=False,
|
| 689 |
+
alibi=False,
|
| 690 |
+
alibi_bias_max=8,
|
| 691 |
+
for_ae=False,
|
| 692 |
+
topk=0,
|
| 693 |
+
device=None,
|
| 694 |
+
dtype=None
|
| 695 |
+
):
|
| 696 |
+
if attn_impl == 'flash':
|
| 697 |
+
return None
|
| 698 |
+
elif attn_impl in ['torch', 'triton']:
|
| 699 |
+
if alibi:
|
| 700 |
+
# in place add alibi to attn bias
|
| 701 |
+
if attn_bias is not None:
|
| 702 |
+
attn_bias = attn_bias.add(
|
| 703 |
+
build_alibi_bias(
|
| 704 |
+
n_heads,
|
| 705 |
+
seq_len,
|
| 706 |
+
full=not causal,
|
| 707 |
+
alibi_bias_max=alibi_bias_max,
|
| 708 |
+
device=device,
|
| 709 |
+
dtype=dtype,
|
| 710 |
+
for_ae=for_ae,
|
| 711 |
+
topk=topk
|
| 712 |
+
))
|
| 713 |
+
else:
|
| 714 |
+
attn_bias = build_alibi_bias(
|
| 715 |
+
n_heads,
|
| 716 |
+
seq_len,
|
| 717 |
+
full=not causal,
|
| 718 |
+
alibi_bias_max=alibi_bias_max,
|
| 719 |
+
for_ae=for_ae,
|
| 720 |
+
topk=topk)
|
| 721 |
+
return attn_bias
|
| 722 |
+
|
| 723 |
+
|
| 724 |
+
def gen_slopes(n_heads, alibi_bias_max=8, device=None):
|
| 725 |
+
_n_heads = 2**math.ceil(math.log2(n_heads))
|
| 726 |
+
m = torch.arange(1, _n_heads + 1, dtype=torch.float32, device=device)
|
| 727 |
+
m = m.mul(alibi_bias_max / _n_heads)
|
| 728 |
+
slopes = (1. / torch.pow(2, m))
|
| 729 |
+
|
| 730 |
+
if _n_heads != n_heads:
|
| 731 |
+
# if n_heads is not a power of two,
|
| 732 |
+
# Huggingface and FasterTransformer calculate slopes normally,
|
| 733 |
+
# then return this strided concatenation of slopes
|
| 734 |
+
slopes = torch.concat([slopes[1::2], slopes[::2]])[:n_heads]
|
| 735 |
+
|
| 736 |
+
return slopes.view(1, n_heads, 1, 1)
|
| 737 |
+
|
| 738 |
+
|
| 739 |
+
def build_alibi_bias(
|
| 740 |
+
n_heads,
|
| 741 |
+
seq_len,
|
| 742 |
+
full=False,
|
| 743 |
+
alibi_bias_max=8,
|
| 744 |
+
device=None,
|
| 745 |
+
dtype=None,
|
| 746 |
+
for_ae=False,
|
| 747 |
+
topk=0
|
| 748 |
+
):
|
| 749 |
+
if not for_ae:
|
| 750 |
+
alibi_bias = torch.arange(1 - seq_len, 1, dtype=torch.int32,
|
| 751 |
+
device=device).view(1, 1, 1, seq_len)
|
| 752 |
+
else:
|
| 753 |
+
alibi_bias = torch.tensor(-seq_len, dtype=torch.int32,
|
| 754 |
+
device=device).repeat(seq_len*topk).view(1, 1, 1, seq_len*(topk))
|
| 755 |
+
if full:
|
| 756 |
+
# generate 1 x Heads x SeqLen x SeqLen alibi bias mask
|
| 757 |
+
# otherwise the mask is 1 x Heads x 1 x SeqLen (which is broadcast to the appropriate size)
|
| 758 |
+
alibi_bias = alibi_bias - torch.arange(
|
| 759 |
+
1 - seq_len, 1, dtype=torch.int32, device=device).view(
|
| 760 |
+
1, 1, seq_len, 1)
|
| 761 |
+
alibi_bias = alibi_bias.abs().mul(-1)
|
| 762 |
+
|
| 763 |
+
slopes = gen_slopes(n_heads, alibi_bias_max, device=device)
|
| 764 |
+
alibi_bias = alibi_bias * slopes
|
| 765 |
+
return alibi_bias.to(dtype=dtype)
|
| 766 |
+
|
| 767 |
+
|
| 768 |
+
ATTN_CLASS_REGISTRY = {
|
| 769 |
+
'multihead_attention': MultiheadAttention,
|
| 770 |
+
'multiquery_attention': MultiQueryAttention,
|
| 771 |
+
}
|
blocks.py
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Adapted from https://github.com/mosaicml/llm-foundry
|
| 2 |
+
# Classes changed: MPTBlock
|
| 3 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 4 |
+
|
| 5 |
+
"""GPT Blocks used for the GPT Model."""
|
| 6 |
+
|
| 7 |
+
from typing import Dict, Optional, Tuple
|
| 8 |
+
import torch
|
| 9 |
+
import torch.nn as nn
|
| 10 |
+
from .attention import ATTN_CLASS_REGISTRY
|
| 11 |
+
from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
|
| 12 |
+
|
| 13 |
+
class MPTMLP(nn.Module):
|
| 14 |
+
|
| 15 |
+
def __init__(self,
|
| 16 |
+
d_model: int,
|
| 17 |
+
expansion_ratio: int,
|
| 18 |
+
device: Optional[str] = None):
|
| 19 |
+
super().__init__()
|
| 20 |
+
self.up_proj = nn.Linear(d_model,
|
| 21 |
+
expansion_ratio * d_model,
|
| 22 |
+
device=device)
|
| 23 |
+
self.act = nn.GELU(approximate='none')
|
| 24 |
+
self.down_proj = nn.Linear(expansion_ratio * d_model,
|
| 25 |
+
d_model,
|
| 26 |
+
device=device)
|
| 27 |
+
self.down_proj._is_residual = True # type: ignore
|
| 28 |
+
|
| 29 |
+
def forward(self, x):
|
| 30 |
+
return self.down_proj(self.act(self.up_proj(x)))
|
| 31 |
+
|
| 32 |
+
class MPTBlock(nn.Module):
|
| 33 |
+
def __init__(
|
| 34 |
+
self,
|
| 35 |
+
d_model: int,
|
| 36 |
+
n_heads: int,
|
| 37 |
+
expansion_ratio: int,
|
| 38 |
+
attn_config: Dict = {
|
| 39 |
+
'attn_type': 'multihead_attention',
|
| 40 |
+
'attn_pdrop': 0.0,
|
| 41 |
+
'attn_impl': 'triton',
|
| 42 |
+
'qk_ln': False,
|
| 43 |
+
'clip_qkv': None,
|
| 44 |
+
'softmax_scale': None,
|
| 45 |
+
'prefix_lm': False,
|
| 46 |
+
'attn_uses_sequence_id': False,
|
| 47 |
+
'alibi': False,
|
| 48 |
+
'alibi_bias_max': 8,
|
| 49 |
+
},
|
| 50 |
+
resid_pdrop: float = 0.0,
|
| 51 |
+
norm_type: str = 'low_precision_layernorm',
|
| 52 |
+
verbose: int = 0,
|
| 53 |
+
device: Optional[str] = None,
|
| 54 |
+
**kwargs):
|
| 55 |
+
del kwargs # unused, just to capture any extra args from the config
|
| 56 |
+
super().__init__()
|
| 57 |
+
|
| 58 |
+
norm_class = NORM_CLASS_REGISTRY[norm_type.lower()]
|
| 59 |
+
attn_class = ATTN_CLASS_REGISTRY[attn_config['attn_type']]
|
| 60 |
+
|
| 61 |
+
self.norm_1 = norm_class(d_model, device=device)
|
| 62 |
+
self.attn = attn_class(
|
| 63 |
+
attn_impl=attn_config['attn_impl'],
|
| 64 |
+
clip_qkv=attn_config['clip_qkv'],
|
| 65 |
+
qk_ln=attn_config['qk_ln'],
|
| 66 |
+
softmax_scale=attn_config['softmax_scale'],
|
| 67 |
+
attn_pdrop=attn_config['attn_pdrop'],
|
| 68 |
+
d_model=d_model,
|
| 69 |
+
n_heads=n_heads,
|
| 70 |
+
verbose=verbose,
|
| 71 |
+
device=device,
|
| 72 |
+
)
|
| 73 |
+
self.norm_2 = norm_class(d_model, device=device)
|
| 74 |
+
self.ffn = MPTMLP(
|
| 75 |
+
d_model=d_model,
|
| 76 |
+
expansion_ratio=expansion_ratio,
|
| 77 |
+
device=device,
|
| 78 |
+
)
|
| 79 |
+
self.resid_attn_dropout = nn.Dropout(resid_pdrop)
|
| 80 |
+
self.resid_ffn_dropout = nn.Dropout(resid_pdrop)
|
| 81 |
+
|
| 82 |
+
def forward(
|
| 83 |
+
self,
|
| 84 |
+
x: torch.Tensor,
|
| 85 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
| 86 |
+
long_range_past_key_value:Optional[Tuple[torch.Tensor]] = None,
|
| 87 |
+
attn_bias: Optional[torch.Tensor] = None,
|
| 88 |
+
attn_bias_ae: Optional[torch.Tensor] = None,
|
| 89 |
+
attention_mask: Optional[torch.ByteTensor] = None,
|
| 90 |
+
is_causal: bool = True,
|
| 91 |
+
topk:int=None,
|
| 92 |
+
needs_weights:bool=None,
|
| 93 |
+
faiss_indexes:Tuple=None,
|
| 94 |
+
n_layers:int=None,
|
| 95 |
+
current_layer:int=None,
|
| 96 |
+
mask_by_sim:bool=False,
|
| 97 |
+
sim_threshold:float=None
|
| 98 |
+
) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
|
| 99 |
+
a = self.norm_1(x)
|
| 100 |
+
b, attn_weights, past_key_value, reshaped_idx = self.attn(
|
| 101 |
+
a,
|
| 102 |
+
past_key_value=past_key_value,
|
| 103 |
+
long_range_past_key_value=long_range_past_key_value,
|
| 104 |
+
attn_bias=attn_bias,
|
| 105 |
+
attn_bias_ae=attn_bias_ae,
|
| 106 |
+
attention_mask=attention_mask,
|
| 107 |
+
is_causal=is_causal,
|
| 108 |
+
topk=topk,
|
| 109 |
+
needs_weights=needs_weights,
|
| 110 |
+
faiss_indexes=faiss_indexes,
|
| 111 |
+
n_layers=n_layers,
|
| 112 |
+
current_layer=current_layer,
|
| 113 |
+
mask_by_sim=mask_by_sim,
|
| 114 |
+
sim_threshold=sim_threshold
|
| 115 |
+
)
|
| 116 |
+
x = x + self.resid_attn_dropout(b)
|
| 117 |
+
m = self.norm_2(x)
|
| 118 |
+
n = self.ffn(m)
|
| 119 |
+
x = x + self.resid_ffn_dropout(n)
|
| 120 |
+
return x, attn_weights, past_key_value, reshaped_idx
|
configuration.py
ADDED
|
@@ -0,0 +1,207 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Adapted from https://github.com/mosaicml/llm-foundry
|
| 2 |
+
# Classes changed: MPTConfig
|
| 3 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
"""A HuggingFace-style model configuration."""
|
| 7 |
+
|
| 8 |
+
from typing import Dict, List, Optional, Union
|
| 9 |
+
from transformers import PretrainedConfig
|
| 10 |
+
|
| 11 |
+
attn_config_defaults: Dict = {
|
| 12 |
+
'attn_type': 'multihead_attention',
|
| 13 |
+
'attn_pdrop': 0.0,
|
| 14 |
+
'attn_impl': 'torch',
|
| 15 |
+
'qk_ln': False,
|
| 16 |
+
'clip_qkv': None,
|
| 17 |
+
'softmax_scale': None,
|
| 18 |
+
'prefix_lm': False,
|
| 19 |
+
'attn_uses_sequence_id': False,
|
| 20 |
+
'alibi': True,
|
| 21 |
+
'alibi_bias_max': 8,
|
| 22 |
+
"topk": 10,
|
| 23 |
+
'mask_by_sim':True,
|
| 24 |
+
'sim_threshold':0.25,
|
| 25 |
+
'use_active_externalism':True,
|
| 26 |
+
'memory_type':'manual'
|
| 27 |
+
}
|
| 28 |
+
|
| 29 |
+
init_config_defaults: Dict = {
|
| 30 |
+
'name': 'kaiming_normal_',
|
| 31 |
+
'fan_mode': 'fan_in',
|
| 32 |
+
'init_nonlinearity': 'relu',
|
| 33 |
+
'init_div_is_residual': True,
|
| 34 |
+
'emb_init_std': None,
|
| 35 |
+
'emb_init_uniform_lim': None,
|
| 36 |
+
'init_std': None,
|
| 37 |
+
'init_gain': 0.0,
|
| 38 |
+
}
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
class ExtendedMPTConfig(PretrainedConfig):
|
| 42 |
+
model_type = 'extended-mpt'
|
| 43 |
+
|
| 44 |
+
def __init__(
|
| 45 |
+
self,
|
| 46 |
+
d_model: int = 4096,
|
| 47 |
+
n_heads: int = 32,
|
| 48 |
+
n_layers: int = 32,
|
| 49 |
+
expansion_ratio: int = 4,
|
| 50 |
+
max_seq_len: int = 2048,
|
| 51 |
+
vocab_size: int = 50432,
|
| 52 |
+
resid_pdrop: float = 0.0,
|
| 53 |
+
emb_pdrop: float = 0.0,
|
| 54 |
+
learned_pos_emb: bool = True,
|
| 55 |
+
attn_config: Dict = attn_config_defaults,
|
| 56 |
+
init_device: str = 'cpu',
|
| 57 |
+
logit_scale: Optional[Union[float, str]] = None,
|
| 58 |
+
no_bias: bool = True,
|
| 59 |
+
verbose: int = 0,
|
| 60 |
+
embedding_fraction: float = 1.0,
|
| 61 |
+
norm_type: str = 'low_precision_layernorm',
|
| 62 |
+
use_cache: bool = False,
|
| 63 |
+
init_config: Dict = init_config_defaults,
|
| 64 |
+
use_active_externalism_by_layer: List[bool] = [True for _ in range(32)],
|
| 65 |
+
memory_device:str = 'cpu',
|
| 66 |
+
**kwargs,
|
| 67 |
+
):
|
| 68 |
+
"""The MPT configuration class.
|
| 69 |
+
|
| 70 |
+
Args:
|
| 71 |
+
d_model (int): The size of the embedding dimension of the model.
|
| 72 |
+
n_heads (int): The number of attention heads.
|
| 73 |
+
n_layers (int): The number of layers in the model.
|
| 74 |
+
expansion_ratio (int): The ratio of the up/down scale in the MLP.
|
| 75 |
+
max_seq_len (int): The maximum sequence length of the model.
|
| 76 |
+
vocab_size (int): The size of the vocabulary.
|
| 77 |
+
resid_pdrop (float): The dropout probability applied to the attention output before combining with residual.
|
| 78 |
+
emb_pdrop (float): The dropout probability for the embedding layer.
|
| 79 |
+
learned_pos_emb (bool): Whether to use learned positional embeddings
|
| 80 |
+
attn_config (Dict): A dictionary used to configure the model's attention module:
|
| 81 |
+
attn_type (str): type of attention to use. Options: multihead_attention, multiquery_attention
|
| 82 |
+
attn_pdrop (float): The dropout probability for the attention layers.
|
| 83 |
+
attn_impl (str): The attention implementation to use. One of 'torch', 'flash', or 'triton'.
|
| 84 |
+
qk_ln (bool): Whether to apply layer normalization to the queries and keys in the attention layer.
|
| 85 |
+
clip_qkv (Optional[float]): If not None, clip the queries, keys, and values in the attention layer to
|
| 86 |
+
this value.
|
| 87 |
+
softmax_scale (Optional[float]): If not None, scale the softmax in the attention layer by this value. If None,
|
| 88 |
+
use the default scale of ``1/sqrt(d_keys)``.
|
| 89 |
+
prefix_lm (Optional[bool]): Whether the model should operate as a Prefix LM. This requires passing an
|
| 90 |
+
extra `prefix_mask` argument which indicates which tokens belong to the prefix. Tokens in the prefix
|
| 91 |
+
can attend to one another bi-directionally. Tokens outside the prefix use causal attention.
|
| 92 |
+
attn_uses_sequence_id (Optional[bool]): Whether to restrict attention to tokens that have the same sequence_id.
|
| 93 |
+
When the model is in `train` mode, this requires passing an extra `sequence_id` argument which indicates
|
| 94 |
+
which sub-sequence each token belongs to.
|
| 95 |
+
Defaults to ``False`` meaning any provided `sequence_id` will be ignored.
|
| 96 |
+
alibi (bool): Whether to use the alibi bias instead of position embeddings.
|
| 97 |
+
alibi_bias_max (int): The maximum value of the alibi bias.
|
| 98 |
+
init_device (str): The device to use for parameter initialization.
|
| 99 |
+
logit_scale (Optional[Union[float, str]]): If not None, scale the logits by this value.
|
| 100 |
+
no_bias (bool): Whether to use bias in all layers.
|
| 101 |
+
verbose (int): The verbosity level. 0 is silent.
|
| 102 |
+
embedding_fraction (float): The fraction to scale the gradients of the embedding layer by.
|
| 103 |
+
norm_type (str): choose type of norm to use
|
| 104 |
+
multiquery_attention (bool): Whether to use multiquery attention implementation.
|
| 105 |
+
use_cache (bool): Whether or not the model should return the last key/values attentions
|
| 106 |
+
init_config (Dict): A dictionary used to configure the model initialization:
|
| 107 |
+
init_config.name: The parameter initialization scheme to use. Options: 'default_', 'baseline_',
|
| 108 |
+
'kaiming_uniform_', 'kaiming_normal_', 'neox_init_', 'small_init_', 'xavier_uniform_', or
|
| 109 |
+
'xavier_normal_'. These mimic the parameter initialization methods in PyTorch.
|
| 110 |
+
init_div_is_residual (Union[int, float, str, bool]): Value to divide initial weights by if ``module._is_residual`` is True.
|
| 111 |
+
emb_init_std (Optional[float]): The standard deviation of the normal distribution used to initialize the embedding layer.
|
| 112 |
+
emb_init_uniform_lim (Optional[Union[Tuple[float, float], float]]): The lower and upper limits of the uniform distribution
|
| 113 |
+
used to initialize the embedding layer. Mutually exclusive with ``emb_init_std``.
|
| 114 |
+
init_std (float): The standard deviation of the normal distribution used to initialize the model,
|
| 115 |
+
if using the baseline_ parameter initialization scheme.
|
| 116 |
+
init_gain (float): The gain to use for parameter initialization with kaiming or xavier initialization schemes.
|
| 117 |
+
fan_mode (str): The fan mode to use for parameter initialization with kaiming initialization schemes.
|
| 118 |
+
init_nonlinearity (str): The nonlinearity to use for parameter initialization with kaiming initialization schemes.
|
| 119 |
+
---
|
| 120 |
+
See llmfoundry.models.utils.param_init_fns.py for info on other param init config options
|
| 121 |
+
"""
|
| 122 |
+
self.d_model = d_model
|
| 123 |
+
self.n_heads = n_heads
|
| 124 |
+
self.n_layers = n_layers
|
| 125 |
+
self.expansion_ratio = expansion_ratio
|
| 126 |
+
self.max_seq_len = max_seq_len
|
| 127 |
+
self.vocab_size = vocab_size
|
| 128 |
+
self.resid_pdrop = resid_pdrop
|
| 129 |
+
self.emb_pdrop = emb_pdrop
|
| 130 |
+
self.learned_pos_emb = learned_pos_emb
|
| 131 |
+
self.attn_config = attn_config
|
| 132 |
+
self.init_device = init_device
|
| 133 |
+
self.logit_scale = logit_scale
|
| 134 |
+
self.no_bias = no_bias
|
| 135 |
+
self.verbose = verbose
|
| 136 |
+
self.embedding_fraction = embedding_fraction
|
| 137 |
+
self.norm_type = norm_type
|
| 138 |
+
self.use_cache = use_cache
|
| 139 |
+
self.init_config = init_config
|
| 140 |
+
self.use_active_externalism_by_layer = use_active_externalism_by_layer
|
| 141 |
+
self.memory_device = memory_device
|
| 142 |
+
if 'name' in kwargs:
|
| 143 |
+
del kwargs['name']
|
| 144 |
+
if 'loss_fn' in kwargs:
|
| 145 |
+
del kwargs['loss_fn']
|
| 146 |
+
super().__init__(**kwargs)
|
| 147 |
+
|
| 148 |
+
self._validate_config()
|
| 149 |
+
|
| 150 |
+
def _set_config_defaults(self, config, config_defaults):
|
| 151 |
+
# set config defaults
|
| 152 |
+
for k, v in config_defaults.items():
|
| 153 |
+
if k not in config:
|
| 154 |
+
config[k] = v
|
| 155 |
+
return config
|
| 156 |
+
|
| 157 |
+
def _validate_config(self):
|
| 158 |
+
# set config defaults
|
| 159 |
+
self.attn_config = self._set_config_defaults(
|
| 160 |
+
self.attn_config,
|
| 161 |
+
attn_config_defaults,
|
| 162 |
+
)
|
| 163 |
+
self.init_config = self._set_config_defaults(
|
| 164 |
+
self.init_config,
|
| 165 |
+
init_config_defaults,
|
| 166 |
+
)
|
| 167 |
+
|
| 168 |
+
if self.d_model % self.n_heads != 0:
|
| 169 |
+
raise ValueError('d_model must be divisible by n_heads')
|
| 170 |
+
if any(
|
| 171 |
+
prob < 0 or prob > 1 for prob in
|
| 172 |
+
[self.attn_config['attn_pdrop'], self.resid_pdrop, self.emb_pdrop]):
|
| 173 |
+
raise ValueError(
|
| 174 |
+
"self.attn_config['attn_pdrop'], resid_pdrop, emb_pdrop are probabilities and must be between 0 and 1"
|
| 175 |
+
)
|
| 176 |
+
if self.attn_config['attn_impl'] not in ['torch', 'flash', 'triton']:
|
| 177 |
+
raise ValueError(
|
| 178 |
+
f"Unknown attn_impl={self.attn_config['attn_impl']}")
|
| 179 |
+
if self.attn_config['prefix_lm'] and self.attn_config[
|
| 180 |
+
'attn_impl'] not in ['torch', 'triton']:
|
| 181 |
+
raise NotImplementedError(
|
| 182 |
+
'prefix_lm only implemented with torch and triton attention.')
|
| 183 |
+
if self.attn_config['alibi'] and self.attn_config['attn_impl'] not in [
|
| 184 |
+
'torch', 'triton'
|
| 185 |
+
]:
|
| 186 |
+
raise NotImplementedError(
|
| 187 |
+
'alibi only implemented with torch and triton attention.')
|
| 188 |
+
if self.attn_config['attn_uses_sequence_id'] and self.attn_config[
|
| 189 |
+
'attn_impl'] not in ['torch', 'triton']:
|
| 190 |
+
raise NotImplementedError(
|
| 191 |
+
'attn_uses_sequence_id only implemented with torch and triton attention.'
|
| 192 |
+
)
|
| 193 |
+
if self.embedding_fraction > 1 or self.embedding_fraction <= 0:
|
| 194 |
+
raise ValueError(
|
| 195 |
+
'model.embedding_fraction must be between 0 (exclusive) and 1 (inclusive)!'
|
| 196 |
+
)
|
| 197 |
+
if isinstance(self.logit_scale,
|
| 198 |
+
str) and self.logit_scale != 'inv_sqrt_d_model':
|
| 199 |
+
raise ValueError(
|
| 200 |
+
f"{self.logit_scale=} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'."
|
| 201 |
+
)
|
| 202 |
+
if self.init_config.get('name', None) is None:
|
| 203 |
+
raise ValueError(f"{self.init_config=} 'name' needs to be set.")
|
| 204 |
+
if not self.learned_pos_emb and not self.attn_config['alibi']:
|
| 205 |
+
raise ValueError(
|
| 206 |
+
f'Positional information must be provided to the model using either learned_pos_emb or alibi.'
|
| 207 |
+
)
|
modeling_mpt.py
ADDED
|
@@ -0,0 +1,837 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Adapted from https://github.com/mosaicml/llm-foundry
|
| 2 |
+
# Classes changed: MPTModel, MPTForCausalLM
|
| 3 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 4 |
+
|
| 5 |
+
"""A simple, flexible implementation of a GPT model.
|
| 6 |
+
|
| 7 |
+
Inspired by https://github.com/karpathy/minGPT/blob/master/mingpt/model.py
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
import math
|
| 11 |
+
import warnings
|
| 12 |
+
from typing import List, Optional, Tuple, Union
|
| 13 |
+
import torch
|
| 14 |
+
import torch.nn as nn
|
| 15 |
+
import torch.nn.functional as F
|
| 16 |
+
from torch.linalg import vector_norm
|
| 17 |
+
import faiss
|
| 18 |
+
from einops import rearrange
|
| 19 |
+
from composer.utils import dist
|
| 20 |
+
from omegaconf import DictConfig
|
| 21 |
+
|
| 22 |
+
from transformers import (PreTrainedModel, PreTrainedTokenizer,
|
| 23 |
+
PreTrainedTokenizerFast)
|
| 24 |
+
from transformers.modeling_outputs import (BaseModelOutputWithPast,
|
| 25 |
+
CausalLMOutputWithPast)
|
| 26 |
+
from llmfoundry.models.layers.custom_embedding import SharedEmbedding
|
| 27 |
+
from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
|
| 28 |
+
from llmfoundry.models.utils.param_init_fns import MODEL_INIT_REGISTRY
|
| 29 |
+
|
| 30 |
+
from .configuration import ExtendedMPTConfig
|
| 31 |
+
from .attention import attn_bias_shape, build_attn_bias
|
| 32 |
+
from .blocks import MPTBlock
|
| 33 |
+
from .utils import instantiate_from_config
|
| 34 |
+
|
| 35 |
+
Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
|
| 36 |
+
|
| 37 |
+
class MPTPreTrainedModel(PreTrainedModel):
|
| 38 |
+
config_class = ExtendedMPTConfig
|
| 39 |
+
base_model_prefix = 'model'
|
| 40 |
+
_no_split_modules = ['MPTBlock']
|
| 41 |
+
|
| 42 |
+
class ExtendedMPTModel(MPTPreTrainedModel):
|
| 43 |
+
|
| 44 |
+
def __init__(self, config: ExtendedMPTConfig):
|
| 45 |
+
config._validate_config()
|
| 46 |
+
super().__init__(config)
|
| 47 |
+
|
| 48 |
+
self.attn_impl = config.attn_config['attn_impl']
|
| 49 |
+
self.prefix_lm = config.attn_config['prefix_lm']
|
| 50 |
+
self.attn_uses_sequence_id = config.attn_config['attn_uses_sequence_id']
|
| 51 |
+
self.alibi = config.attn_config['alibi']
|
| 52 |
+
self.alibi_bias_max = config.attn_config['alibi_bias_max']
|
| 53 |
+
|
| 54 |
+
self.mask_by_sim = config.attn_config['mask_by_sim']
|
| 55 |
+
self.sim_threshold = config.attn_config['sim_threshold']
|
| 56 |
+
self.topk = config.attn_config['topk']
|
| 57 |
+
self.use_active_externalism = config.attn_config['use_active_externalism']
|
| 58 |
+
|
| 59 |
+
self.use_active_externalism_by_layer = config.use_active_externalism_by_layer
|
| 60 |
+
|
| 61 |
+
if config.init_device == 'mixed':
|
| 62 |
+
if dist.get_local_rank() == 0:
|
| 63 |
+
config.init_device = 'cpu'
|
| 64 |
+
else:
|
| 65 |
+
config.init_device = 'meta'
|
| 66 |
+
|
| 67 |
+
if config.norm_type.lower() not in NORM_CLASS_REGISTRY.keys():
|
| 68 |
+
norm_options = ' | '.join(NORM_CLASS_REGISTRY.keys())
|
| 69 |
+
raise NotImplementedError(
|
| 70 |
+
f'Requested norm type ({config.norm_type}) is not implemented within this repo (Options: {norm_options}).'
|
| 71 |
+
)
|
| 72 |
+
norm_class = NORM_CLASS_REGISTRY[config.norm_type.lower()]
|
| 73 |
+
|
| 74 |
+
# CogView (https://arxiv.org/abs/2105.13290) and GLM-130B (https://arxiv.org/abs/2210.02414)
|
| 75 |
+
# both report this helping with stabilizing training
|
| 76 |
+
self.embedding_fraction = config.embedding_fraction
|
| 77 |
+
|
| 78 |
+
self.wte = SharedEmbedding(config.vocab_size,
|
| 79 |
+
config.d_model,
|
| 80 |
+
device=config.init_device)
|
| 81 |
+
if not self.alibi:
|
| 82 |
+
self.wpe = torch.nn.Embedding(config.max_seq_len,
|
| 83 |
+
config.d_model,
|
| 84 |
+
device=config.init_device)
|
| 85 |
+
self.emb_drop = nn.Dropout(config.emb_pdrop)
|
| 86 |
+
self.blocks = nn.ModuleList([
|
| 87 |
+
MPTBlock(
|
| 88 |
+
device=config.init_device,
|
| 89 |
+
**config.to_dict(),
|
| 90 |
+
) for _ in range(config.n_layers)
|
| 91 |
+
])
|
| 92 |
+
self.norm_f = norm_class(config.d_model, device=config.init_device)
|
| 93 |
+
|
| 94 |
+
if config.init_device != 'meta':
|
| 95 |
+
print(
|
| 96 |
+
f'You are using {config.init_device=}, but you can also use config.init_device="meta" with Composer + FSDP for fast initialization.'
|
| 97 |
+
)
|
| 98 |
+
self.apply(self.param_init_fn)
|
| 99 |
+
|
| 100 |
+
self.is_causal = not self.prefix_lm
|
| 101 |
+
|
| 102 |
+
# define attn mask
|
| 103 |
+
self._attn_bias_initialized = False
|
| 104 |
+
self.attn_bias = None
|
| 105 |
+
self.attn_bias_shape = attn_bias_shape(
|
| 106 |
+
self.attn_impl,
|
| 107 |
+
config.n_heads,
|
| 108 |
+
config.max_seq_len,
|
| 109 |
+
self.alibi,
|
| 110 |
+
prefix_lm=self.prefix_lm,
|
| 111 |
+
causal=self.is_causal,
|
| 112 |
+
use_sequence_id=self.attn_uses_sequence_id,
|
| 113 |
+
)
|
| 114 |
+
self._attn_bias_ae_initialized = False
|
| 115 |
+
self.attn_bias_ae = None
|
| 116 |
+
|
| 117 |
+
if self.config.no_bias:
|
| 118 |
+
for module in self.modules():
|
| 119 |
+
if hasattr(module, 'bias') and isinstance(
|
| 120 |
+
module.bias, nn.Parameter):
|
| 121 |
+
if self.config.verbose:
|
| 122 |
+
warnings.warn(
|
| 123 |
+
f'Removing bias ({module.bias}) from {module}.')
|
| 124 |
+
module.register_parameter('bias', None)
|
| 125 |
+
|
| 126 |
+
# Print verbose info
|
| 127 |
+
if config.verbose and config.verbose > 2:
|
| 128 |
+
print(self)
|
| 129 |
+
if 'verbose' not in self.config.init_config:
|
| 130 |
+
self.config.init_config['verbose'] = self.config.verbose
|
| 131 |
+
if self.config.init_config['verbose'] > 1:
|
| 132 |
+
init_fn_name = self.config.init_config['name']
|
| 133 |
+
warnings.warn(f'Using {init_fn_name} initialization.')
|
| 134 |
+
|
| 135 |
+
def get_input_embeddings(self):
|
| 136 |
+
return self.wte
|
| 137 |
+
|
| 138 |
+
def set_input_embeddings(self, value: nn.Embedding):
|
| 139 |
+
self.wte = value
|
| 140 |
+
|
| 141 |
+
@torch.no_grad()
|
| 142 |
+
def _attn_bias(
|
| 143 |
+
self,
|
| 144 |
+
device,
|
| 145 |
+
dtype,
|
| 146 |
+
attention_mask: Optional[torch.ByteTensor] = None,
|
| 147 |
+
prefix_mask: Optional[torch.ByteTensor] = None,
|
| 148 |
+
sequence_id: Optional[torch.LongTensor] = None,
|
| 149 |
+
seq_len: Optional[int] = None,
|
| 150 |
+
use_active_externalism:bool=None,
|
| 151 |
+
topk=None,
|
| 152 |
+
):
|
| 153 |
+
if not self._attn_bias_initialized:
|
| 154 |
+
if self.attn_bias_shape:
|
| 155 |
+
self.attn_bias = torch.zeros(self.attn_bias_shape,
|
| 156 |
+
device=device,
|
| 157 |
+
dtype=dtype)
|
| 158 |
+
self.attn_bias = build_attn_bias(
|
| 159 |
+
self.attn_impl,
|
| 160 |
+
self.config.n_heads,
|
| 161 |
+
self.config.max_seq_len,
|
| 162 |
+
device=device,
|
| 163 |
+
dtype=dtype,
|
| 164 |
+
attn_bias = self.attn_bias,
|
| 165 |
+
causal=self.is_causal,
|
| 166 |
+
alibi=self.alibi,
|
| 167 |
+
alibi_bias_max=self.alibi_bias_max
|
| 168 |
+
)
|
| 169 |
+
self._attn_bias_initialized = True
|
| 170 |
+
|
| 171 |
+
if use_active_externalism:
|
| 172 |
+
self.attn_bias_ae = build_attn_bias(
|
| 173 |
+
self.attn_impl,
|
| 174 |
+
self.config.n_heads,
|
| 175 |
+
seq_len,
|
| 176 |
+
device=device,
|
| 177 |
+
dtype=dtype,
|
| 178 |
+
causal=self.is_causal,
|
| 179 |
+
alibi=self.alibi,
|
| 180 |
+
alibi_bias_max=self.alibi_bias_max,
|
| 181 |
+
for_ae=use_active_externalism,
|
| 182 |
+
topk=topk
|
| 183 |
+
)
|
| 184 |
+
|
| 185 |
+
self._attn_bias_ae_initialized = True
|
| 186 |
+
|
| 187 |
+
# flash does not support prefix_lm and will incorporate any
|
| 188 |
+
# attention_mask inside the attention module
|
| 189 |
+
if self.attn_impl == 'flash':
|
| 190 |
+
return self.attn_bias, attention_mask
|
| 191 |
+
|
| 192 |
+
if self.attn_bias is not None:
|
| 193 |
+
# .to(*args, **kwargs) is a no-op if tensor is already on
|
| 194 |
+
# specified device or of specificed dtype
|
| 195 |
+
self.attn_bias = self.attn_bias.to(dtype=dtype, device=device)
|
| 196 |
+
|
| 197 |
+
attn_bias = self.attn_bias
|
| 198 |
+
|
| 199 |
+
if self.attn_bias_ae is not None:
|
| 200 |
+
self.attn_bias_ae = self.attn_bias_ae.to(dtype=dtype, device=device)
|
| 201 |
+
attn_bias_ae = self.attn_bias_ae
|
| 202 |
+
|
| 203 |
+
# If using torch or triton, we incorporate the prefix_mask (if appropriate)
|
| 204 |
+
if self.prefix_lm:
|
| 205 |
+
assert isinstance(attn_bias, torch.Tensor) # pyright
|
| 206 |
+
assert isinstance(prefix_mask, torch.Tensor) # pyright
|
| 207 |
+
attn_bias = self._apply_prefix_mask(attn_bias, prefix_mask)
|
| 208 |
+
|
| 209 |
+
# If using torch or triton, we incorporate sequence_id (if appropriate)
|
| 210 |
+
if self.attn_uses_sequence_id and sequence_id is not None:
|
| 211 |
+
assert isinstance(attn_bias, torch.Tensor) # pyright
|
| 212 |
+
attn_bias = self._apply_sequence_id(attn_bias, sequence_id)
|
| 213 |
+
|
| 214 |
+
# If using torch or triton, we incorporate attention_mask. This will output
|
| 215 |
+
# None in place of attention_mask since it will not be further needed in the
|
| 216 |
+
# attention modules.
|
| 217 |
+
if attention_mask is not None:
|
| 218 |
+
s_k = attention_mask.shape[-1]
|
| 219 |
+
if attn_bias is None:
|
| 220 |
+
attn_bias = torch.zeros((1, 1, 1, s_k),
|
| 221 |
+
device=device,
|
| 222 |
+
dtype=dtype)
|
| 223 |
+
else:
|
| 224 |
+
# clamp to 0 necessary for torch 2.0 compile()
|
| 225 |
+
_s_k = max(0, attn_bias.size(-1) - s_k)
|
| 226 |
+
attn_bias = attn_bias[:, :, :, _s_k:]
|
| 227 |
+
if prefix_mask is not None and (attention_mask.shape !=
|
| 228 |
+
prefix_mask.shape):
|
| 229 |
+
raise ValueError(
|
| 230 |
+
f'attention_mask shape={attention_mask.shape} ' +
|
| 231 |
+
f'and prefix_mask shape={prefix_mask.shape} are not equal.')
|
| 232 |
+
min_val = torch.finfo(attn_bias.dtype).min
|
| 233 |
+
attn_bias = attn_bias.masked_fill(
|
| 234 |
+
~attention_mask.view(-1, 1, 1, s_k), min_val)
|
| 235 |
+
|
| 236 |
+
return attn_bias, attn_bias_ae, None
|
| 237 |
+
|
| 238 |
+
def _apply_prefix_mask(self, attn_bias: torch.Tensor,
|
| 239 |
+
prefix_mask: torch.Tensor):
|
| 240 |
+
s_k, s_q = attn_bias.shape[-2:]
|
| 241 |
+
if (s_k != self.config.max_seq_len) or (s_q != self.config.max_seq_len):
|
| 242 |
+
raise ValueError(
|
| 243 |
+
'attn_bias does not match the expected shape. ' +
|
| 244 |
+
f'The last two dimensions should both be {self.config.max_length} '
|
| 245 |
+
+ f'but are {s_k} and {s_q}.')
|
| 246 |
+
seq_len = prefix_mask.shape[-1]
|
| 247 |
+
if seq_len > self.config.max_seq_len:
|
| 248 |
+
raise ValueError(
|
| 249 |
+
f'prefix_mask sequence length cannot exceed max_seq_len={self.config.max_seq_len}'
|
| 250 |
+
)
|
| 251 |
+
|
| 252 |
+
# select seq_len subset of attn mask
|
| 253 |
+
attn_bias = attn_bias[..., :seq_len, :seq_len]
|
| 254 |
+
|
| 255 |
+
# Mix the causal max and the bidirectional mask to get the full
|
| 256 |
+
# allowable attention (i.e. full = not accounting for padding yet)
|
| 257 |
+
causal = torch.tril(
|
| 258 |
+
torch.ones((seq_len, seq_len),
|
| 259 |
+
dtype=torch.bool,
|
| 260 |
+
device=prefix_mask.device)).view(1, 1, seq_len, seq_len)
|
| 261 |
+
prefix = prefix_mask.view(-1, 1, 1, seq_len)
|
| 262 |
+
cannot_attend = ~torch.logical_or(causal, prefix.bool())
|
| 263 |
+
|
| 264 |
+
min_val = torch.finfo(attn_bias.dtype).min
|
| 265 |
+
attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
|
| 266 |
+
|
| 267 |
+
return attn_bias
|
| 268 |
+
|
| 269 |
+
def _apply_sequence_id(self, attn_bias: torch.Tensor,
|
| 270 |
+
sequence_id: torch.LongTensor):
|
| 271 |
+
seq_len = sequence_id.shape[-1]
|
| 272 |
+
if seq_len > self.config.max_seq_len:
|
| 273 |
+
raise ValueError(
|
| 274 |
+
f'sequence_id sequence length cannot exceed max_seq_len={self.config.max_seq_len}'
|
| 275 |
+
)
|
| 276 |
+
|
| 277 |
+
# select seq_len subset of attn mask
|
| 278 |
+
attn_bias = attn_bias[..., :seq_len, :seq_len]
|
| 279 |
+
|
| 280 |
+
# Restrict attention to tokens that share the same value
|
| 281 |
+
# in sequence_id
|
| 282 |
+
cannot_attend = torch.logical_not(
|
| 283 |
+
torch.eq(
|
| 284 |
+
sequence_id.view(-1, seq_len, 1),
|
| 285 |
+
sequence_id.view(-1, 1, seq_len),
|
| 286 |
+
)).unsqueeze(1)
|
| 287 |
+
min_val = torch.finfo(attn_bias.dtype).min
|
| 288 |
+
attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
|
| 289 |
+
|
| 290 |
+
return attn_bias
|
| 291 |
+
|
| 292 |
+
def forward(
|
| 293 |
+
self,
|
| 294 |
+
input_ids: torch.LongTensor,
|
| 295 |
+
past_key_values: Optional[List[Tuple[torch.FloatTensor]]] = None,
|
| 296 |
+
attention_mask: Optional[torch.ByteTensor] = None,
|
| 297 |
+
prefix_mask: Optional[torch.ByteTensor] = None,
|
| 298 |
+
sequence_id: Optional[torch.LongTensor] = None,
|
| 299 |
+
return_dict: Optional[bool] = None,
|
| 300 |
+
output_attentions: Optional[bool] = None,
|
| 301 |
+
output_hidden_states: Optional[bool] = None,
|
| 302 |
+
use_cache: Optional[bool] = None,
|
| 303 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 304 |
+
use_active_externalism:Optional[bool]=None,
|
| 305 |
+
long_range_past_key_values:Optional[List[Tuple[torch.FloatTensor]]] = None,
|
| 306 |
+
faiss_indexes:Tuple=None,
|
| 307 |
+
topk:int=None,
|
| 308 |
+
):
|
| 309 |
+
return_dict = (return_dict
|
| 310 |
+
if return_dict is not None else self.config.return_dict)
|
| 311 |
+
use_cache = (use_cache
|
| 312 |
+
if use_cache is not None else self.config.use_cache)
|
| 313 |
+
use_active_externalism = (use_active_externalism
|
| 314 |
+
if use_active_externalism is not None else self.use_active_externalism)
|
| 315 |
+
topk = (topk if topk is not None else self.topk)
|
| 316 |
+
|
| 317 |
+
if attention_mask is not None:
|
| 318 |
+
attention_mask = attention_mask.bool()
|
| 319 |
+
|
| 320 |
+
if prefix_mask is not None:
|
| 321 |
+
prefix_mask = prefix_mask.bool()
|
| 322 |
+
|
| 323 |
+
# These args are passed in by keyword in huggingface's generate function
|
| 324 |
+
# https://github.com/huggingface/transformers/blob/68287689f2f0d8b7063c400230b3766987abf18d/src/transformers/generation/utils.py#L2201-L2206
|
| 325 |
+
# but have not yet been fully implemented in MPTModel
|
| 326 |
+
if not return_dict:
|
| 327 |
+
raise NotImplementedError(
|
| 328 |
+
'return_dict False is not implemented yet for MPT')
|
| 329 |
+
if output_attentions:
|
| 330 |
+
if self.attn_impl != 'torch':
|
| 331 |
+
raise NotImplementedError(
|
| 332 |
+
'output_attentions is not implemented for MPT when using attn_impl `flash` or `triton`.'
|
| 333 |
+
)
|
| 334 |
+
|
| 335 |
+
if (attention_mask is not None and
|
| 336 |
+
attention_mask[:, 0].sum() != attention_mask.shape[0] and
|
| 337 |
+
self.training):
|
| 338 |
+
raise NotImplementedError(
|
| 339 |
+
'MPT does not support training with left padding.')
|
| 340 |
+
|
| 341 |
+
if self.prefix_lm and prefix_mask is None:
|
| 342 |
+
raise ValueError(
|
| 343 |
+
'prefix_mask is a required argument when MPT is configured with prefix_lm=True.'
|
| 344 |
+
)
|
| 345 |
+
|
| 346 |
+
# Raise a not implemented error if input_embeds is not None (this is an arg in huggingface transformers and we need to support it for PEFT)
|
| 347 |
+
if inputs_embeds is not None:
|
| 348 |
+
raise NotImplementedError(
|
| 349 |
+
'inputs_embeds is not implemented for MPT.')
|
| 350 |
+
|
| 351 |
+
if self.training:
|
| 352 |
+
if self.attn_uses_sequence_id and sequence_id is None:
|
| 353 |
+
raise ValueError(
|
| 354 |
+
'sequence_id is a required argument when MPT is configured with attn_uses_sequence_id=True '
|
| 355 |
+
+ 'and the model is in train mode.')
|
| 356 |
+
elif (self.attn_uses_sequence_id is False) and (sequence_id
|
| 357 |
+
is not None):
|
| 358 |
+
warnings.warn(
|
| 359 |
+
'MPT received non-None input for `sequence_id` but is configured with attn_uses_sequence_id=False. '
|
| 360 |
+
+
|
| 361 |
+
'This input will be ignored. If you want the model to use `sequence_id`, set attn_uses_sequence_id to True.'
|
| 362 |
+
)
|
| 363 |
+
|
| 364 |
+
S = input_ids.size(1)
|
| 365 |
+
|
| 366 |
+
assert (
|
| 367 |
+
S <= self.config.max_seq_len
|
| 368 |
+
), f'Cannot forward input with seq_len={S}, this model only supports seq_len<={self.config.max_seq_len}'
|
| 369 |
+
|
| 370 |
+
tok_emb = self.wte(input_ids) # type: ignore
|
| 371 |
+
if self.alibi:
|
| 372 |
+
x = tok_emb
|
| 373 |
+
else:
|
| 374 |
+
past_position = 0
|
| 375 |
+
if past_key_values is not None:
|
| 376 |
+
if len(past_key_values) != self.config.n_layers:
|
| 377 |
+
raise ValueError(
|
| 378 |
+
f'past_key_values must provide a past_key_value for each attention '
|
| 379 |
+
+
|
| 380 |
+
f'layer in the network ({len(past_key_values)=}; {self.config.n_layers=}).'
|
| 381 |
+
)
|
| 382 |
+
# For attn_impl: triton and flash the past key tensor spec is (batch, seq, dim).
|
| 383 |
+
# For attn_impl: torch the past key tensor spec is (batch, heads, head_dim, seq).
|
| 384 |
+
# Here we shift position embedding using the `seq` dim of the past key
|
| 385 |
+
past_position = past_key_values[0][0].size(1)
|
| 386 |
+
if self.attn_impl == 'torch':
|
| 387 |
+
past_position = past_key_values[0][0].size(3)
|
| 388 |
+
|
| 389 |
+
if S + past_position > self.config.max_seq_len:
|
| 390 |
+
raise ValueError(
|
| 391 |
+
f'Cannot forward input with past sequence length {past_position} and current sequence length '
|
| 392 |
+
f'{S + 1}, this model only supports total sequence length <= {self.config.max_seq_len}.'
|
| 393 |
+
)
|
| 394 |
+
pos = torch.arange(
|
| 395 |
+
past_position,
|
| 396 |
+
S + past_position,
|
| 397 |
+
dtype=torch.long,
|
| 398 |
+
device=input_ids.device,
|
| 399 |
+
).unsqueeze(0)
|
| 400 |
+
if attention_mask is not None:
|
| 401 |
+
# adjust the position indices to account for padding tokens
|
| 402 |
+
pos = torch.clamp(
|
| 403 |
+
pos - torch.cumsum((~attention_mask).to(torch.int32),
|
| 404 |
+
dim=1)[:, past_position:],
|
| 405 |
+
min=0,
|
| 406 |
+
)
|
| 407 |
+
|
| 408 |
+
pos_emb = self.wpe(pos) # type: ignore
|
| 409 |
+
x = tok_emb + pos_emb
|
| 410 |
+
|
| 411 |
+
if self.embedding_fraction == 1:
|
| 412 |
+
x = self.emb_drop(x) # type: ignore
|
| 413 |
+
else:
|
| 414 |
+
# this implementation is proposed on page 7 of the GLM-130B paper https://arxiv.org/abs/2210.02414
|
| 415 |
+
x_shrunk = (x * self.embedding_fraction) + (
|
| 416 |
+
x.detach() * (1 - self.embedding_fraction))
|
| 417 |
+
assert isinstance(self.emb_drop, nn.Module) # pyright
|
| 418 |
+
x = self.emb_drop(x_shrunk)
|
| 419 |
+
|
| 420 |
+
# self._attn_bias_initialized = False #right now this needs to run each step
|
| 421 |
+
|
| 422 |
+
seq_len = S
|
| 423 |
+
if past_key_values is not None:
|
| 424 |
+
past_position = past_key_values[0][0].size(-1)
|
| 425 |
+
seq_len += past_position
|
| 426 |
+
|
| 427 |
+
attn_bias, attn_bias_ae, attention_mask = self._attn_bias(
|
| 428 |
+
device=x.device,
|
| 429 |
+
dtype=torch.float32,
|
| 430 |
+
attention_mask=attention_mask,
|
| 431 |
+
prefix_mask=prefix_mask,
|
| 432 |
+
sequence_id=sequence_id,
|
| 433 |
+
seq_len = seq_len,
|
| 434 |
+
use_active_externalism=use_active_externalism,
|
| 435 |
+
topk=topk
|
| 436 |
+
)
|
| 437 |
+
|
| 438 |
+
# initialize the past key values cache if it should be used
|
| 439 |
+
if use_cache and past_key_values is None:
|
| 440 |
+
past_key_values = [() for _ in range(self.config.n_layers)
|
| 441 |
+
] # type: ignore
|
| 442 |
+
|
| 443 |
+
all_hidden_states = () if output_hidden_states else None
|
| 444 |
+
all_self_attns = () if output_attentions else None
|
| 445 |
+
all_idx = () if output_attentions else None
|
| 446 |
+
for b_idx, block in enumerate(self.blocks): # type: ignore
|
| 447 |
+
if output_hidden_states:
|
| 448 |
+
assert all_hidden_states is not None # pyright
|
| 449 |
+
all_hidden_states = all_hidden_states + (x,)
|
| 450 |
+
past_key_value = (past_key_values[b_idx]
|
| 451 |
+
if past_key_values is not None else None)
|
| 452 |
+
long_range_past_key_value = (long_range_past_key_values[b_idx]
|
| 453 |
+
if (long_range_past_key_values is not None and self.use_active_externalism_by_layer[b_idx] and use_active_externalism is True) else None)
|
| 454 |
+
|
| 455 |
+
if long_range_past_key_value is not None and faiss_indexes is not None:
|
| 456 |
+
raise NotImplementedError(
|
| 457 |
+
'Using faiss and passing key value pairs manually are mutually exclusive right now.')
|
| 458 |
+
|
| 459 |
+
x, attn_weights, past_key_value, reshaped_idx = block(
|
| 460 |
+
x,
|
| 461 |
+
past_key_value=past_key_value,
|
| 462 |
+
long_range_past_key_value=long_range_past_key_value,
|
| 463 |
+
attn_bias=attn_bias,
|
| 464 |
+
attention_mask=attention_mask,
|
| 465 |
+
attn_bias_ae=attn_bias_ae,
|
| 466 |
+
is_causal=self.is_causal,
|
| 467 |
+
topk=topk,
|
| 468 |
+
needs_weights=output_attentions,
|
| 469 |
+
faiss_indexes=faiss_indexes,
|
| 470 |
+
n_layers=self.config.n_layers,
|
| 471 |
+
current_layer=b_idx,
|
| 472 |
+
mask_by_sim=self.mask_by_sim,
|
| 473 |
+
sim_threshold=self.sim_threshold,
|
| 474 |
+
)
|
| 475 |
+
if past_key_values is not None:
|
| 476 |
+
past_key_values[b_idx] = past_key_value
|
| 477 |
+
|
| 478 |
+
if output_attentions:
|
| 479 |
+
assert all_self_attns is not None # pyright
|
| 480 |
+
all_self_attns = all_self_attns + (attn_weights,)
|
| 481 |
+
|
| 482 |
+
assert all_idx is not None
|
| 483 |
+
all_idx = all_idx + (reshaped_idx,)
|
| 484 |
+
|
| 485 |
+
x = self.norm_f(x) # type: ignore
|
| 486 |
+
|
| 487 |
+
# add hidden states from the last decoder layer
|
| 488 |
+
if output_hidden_states:
|
| 489 |
+
assert all_hidden_states is not None # pyright
|
| 490 |
+
all_hidden_states = all_hidden_states + (x,)
|
| 491 |
+
|
| 492 |
+
return BaseModelOutputWithPast(
|
| 493 |
+
last_hidden_state=x,
|
| 494 |
+
past_key_values=past_key_values,
|
| 495 |
+
hidden_states=all_hidden_states,
|
| 496 |
+
attentions=(all_self_attns, all_idx),
|
| 497 |
+
)
|
| 498 |
+
|
| 499 |
+
# Param Initialization, needed for device='meta' fast initialization
|
| 500 |
+
def param_init_fn(self, module):
|
| 501 |
+
init_fn_name = self.config.init_config['name']
|
| 502 |
+
MODEL_INIT_REGISTRY[init_fn_name](
|
| 503 |
+
module=module,
|
| 504 |
+
n_layers=self.config.n_layers,
|
| 505 |
+
d_model=self.config.d_model,
|
| 506 |
+
**self.config.init_config,
|
| 507 |
+
)
|
| 508 |
+
|
| 509 |
+
# FSDP Wrap function
|
| 510 |
+
def fsdp_wrap_fn(self, module):
|
| 511 |
+
return isinstance(module, MPTBlock)
|
| 512 |
+
|
| 513 |
+
# Activation Checkpointing
|
| 514 |
+
def activation_checkpointing_fn(self, module):
|
| 515 |
+
return isinstance(module, MPTBlock)
|
| 516 |
+
|
| 517 |
+
class ExtendedMPTForCausalLM(MPTPreTrainedModel):
|
| 518 |
+
|
| 519 |
+
def __init__(self, config:ExtendedMPTConfig, external_memories=None):
|
| 520 |
+
if isinstance(config, DictConfig):
|
| 521 |
+
config = instantiate_from_config(config)
|
| 522 |
+
|
| 523 |
+
super().__init__(config)
|
| 524 |
+
if not config.tie_word_embeddings:
|
| 525 |
+
raise ValueError(
|
| 526 |
+
'MPTForCausalLM only supports tied word embeddings')
|
| 527 |
+
|
| 528 |
+
print(f'Instantiating an MPTForCausalLM model from {__file__}')
|
| 529 |
+
|
| 530 |
+
self.transformer: ExtendedMPTModel = ExtendedMPTModel(config)
|
| 531 |
+
|
| 532 |
+
self.use_active_externalism = config.attn_config['use_active_externalism']
|
| 533 |
+
self.memory_type = config.attn_config['memory_type']
|
| 534 |
+
self._memories = None
|
| 535 |
+
self.memory_device = config.memory_device
|
| 536 |
+
|
| 537 |
+
for child in self.transformer.children():
|
| 538 |
+
if isinstance(child, torch.nn.ModuleList):
|
| 539 |
+
continue
|
| 540 |
+
if isinstance(child, torch.nn.Module):
|
| 541 |
+
child._fsdp_wrap = True
|
| 542 |
+
|
| 543 |
+
# enables scaling output logits; similar to a softmax "temperature"
|
| 544 |
+
# PaLM paper uses scale 1/sqrt(config.d_model)
|
| 545 |
+
self.logit_scale = None
|
| 546 |
+
if config.logit_scale is not None:
|
| 547 |
+
logit_scale = config.logit_scale
|
| 548 |
+
if isinstance(logit_scale, str):
|
| 549 |
+
if logit_scale == 'inv_sqrt_d_model':
|
| 550 |
+
logit_scale = 1 / math.sqrt(config.d_model)
|
| 551 |
+
else:
|
| 552 |
+
raise ValueError(
|
| 553 |
+
f"{logit_scale=} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'."
|
| 554 |
+
)
|
| 555 |
+
self.logit_scale = logit_scale
|
| 556 |
+
|
| 557 |
+
if external_memories is not None:
|
| 558 |
+
self._memories = external_memories
|
| 559 |
+
self.memories = None
|
| 560 |
+
|
| 561 |
+
def set_memories(self, memories):
|
| 562 |
+
self.memories = memories
|
| 563 |
+
|
| 564 |
+
def empty_memories(self):
|
| 565 |
+
self.memories = None
|
| 566 |
+
|
| 567 |
+
def get_input_embeddings(self):
|
| 568 |
+
return self.transformer.wte
|
| 569 |
+
|
| 570 |
+
def set_input_embeddings(self, value):
|
| 571 |
+
self.transformer.wte = value
|
| 572 |
+
|
| 573 |
+
def get_output_embeddings(self):
|
| 574 |
+
return self.transformer.wte
|
| 575 |
+
|
| 576 |
+
def set_output_embeddings(self, new_embeddings):
|
| 577 |
+
self.transformer.wte = new_embeddings
|
| 578 |
+
|
| 579 |
+
def set_decoder(self, decoder):
|
| 580 |
+
self.transformer = decoder
|
| 581 |
+
|
| 582 |
+
def get_decoder(self):
|
| 583 |
+
return self.transformer
|
| 584 |
+
|
| 585 |
+
def forward(
|
| 586 |
+
self,
|
| 587 |
+
input_ids: torch.LongTensor,
|
| 588 |
+
past_key_values: Optional[List[Tuple[torch.FloatTensor]]] = None,
|
| 589 |
+
attention_mask: Optional[torch.ByteTensor] = None,
|
| 590 |
+
prefix_mask: Optional[torch.ByteTensor] = None,
|
| 591 |
+
sequence_id: Optional[torch.LongTensor] = None,
|
| 592 |
+
labels: Optional[torch.LongTensor] = None,
|
| 593 |
+
return_dict: Optional[bool] = None,
|
| 594 |
+
output_attentions: Optional[bool] = None,
|
| 595 |
+
output_hidden_states: Optional[bool] = None,
|
| 596 |
+
use_cache: Optional[bool] = None,
|
| 597 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 598 |
+
use_active_externalism: Optional[bool]=None,
|
| 599 |
+
topk:int=None
|
| 600 |
+
):
|
| 601 |
+
if self._memories is not None and self.memories is None:
|
| 602 |
+
self.memories = self.generate_cache(self._memories, cache_type=self.memory_type)
|
| 603 |
+
|
| 604 |
+
return_dict = (return_dict
|
| 605 |
+
if return_dict is not None else self.config.return_dict)
|
| 606 |
+
use_cache = (use_cache
|
| 607 |
+
if use_cache is not None else self.config.use_cache)
|
| 608 |
+
use_active_externalism = (use_active_externalism
|
| 609 |
+
if use_active_externalism is not None else self.use_active_externalism)
|
| 610 |
+
|
| 611 |
+
topk = topk if topk is not None else None
|
| 612 |
+
|
| 613 |
+
# if input_embeds is not none, raise a not implemented error
|
| 614 |
+
if inputs_embeds is not None:
|
| 615 |
+
raise NotImplementedError(
|
| 616 |
+
'inputs_embeds has to be None (for hf/peft support).')
|
| 617 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
| 618 |
+
|
| 619 |
+
if hasattr(self, "memories") and type(self.memories)==list:
|
| 620 |
+
long_range_past_key_values = self.memories
|
| 621 |
+
faiss_indexes = None
|
| 622 |
+
elif hasattr(self, "memories"):
|
| 623 |
+
long_range_past_key_values = None
|
| 624 |
+
faiss_indexes = self.memories
|
| 625 |
+
else:
|
| 626 |
+
long_range_past_key_values = None
|
| 627 |
+
faiss_indexes = None
|
| 628 |
+
|
| 629 |
+
outputs = self.transformer(
|
| 630 |
+
input_ids=input_ids,
|
| 631 |
+
past_key_values=past_key_values,
|
| 632 |
+
long_range_past_key_values=long_range_past_key_values,
|
| 633 |
+
faiss_indexes=faiss_indexes,
|
| 634 |
+
attention_mask=attention_mask,
|
| 635 |
+
prefix_mask=prefix_mask,
|
| 636 |
+
sequence_id=sequence_id,
|
| 637 |
+
return_dict=return_dict,
|
| 638 |
+
output_attentions=output_attentions,
|
| 639 |
+
output_hidden_states=output_hidden_states,
|
| 640 |
+
use_cache=use_cache,
|
| 641 |
+
use_active_externalism=use_active_externalism,
|
| 642 |
+
topk=topk
|
| 643 |
+
)
|
| 644 |
+
|
| 645 |
+
# move outputs to same device as weights for token embedding
|
| 646 |
+
# needed to support HF `device_map`
|
| 647 |
+
logits = self.transformer.wte(
|
| 648 |
+
outputs.last_hidden_state.to(self.transformer.wte.weight.device),
|
| 649 |
+
True,
|
| 650 |
+
)
|
| 651 |
+
|
| 652 |
+
if self.logit_scale is not None:
|
| 653 |
+
if self.logit_scale == 0:
|
| 654 |
+
warnings.warn(
|
| 655 |
+
f'Multiplying logits by {self.logit_scale=}. This will produce uniform (uninformative) outputs.'
|
| 656 |
+
)
|
| 657 |
+
logits *= self.logit_scale
|
| 658 |
+
|
| 659 |
+
loss = None
|
| 660 |
+
if labels is not None:
|
| 661 |
+
_labels = torch.roll(labels, shifts=-1)
|
| 662 |
+
_labels[:, -1] = -100
|
| 663 |
+
loss = F.cross_entropy(
|
| 664 |
+
logits.view(-1, logits.size(-1)),
|
| 665 |
+
_labels.to(logits.device).view(-1),
|
| 666 |
+
)
|
| 667 |
+
|
| 668 |
+
return CausalLMOutputWithPast(
|
| 669 |
+
loss=loss,
|
| 670 |
+
logits=logits,
|
| 671 |
+
past_key_values=outputs.past_key_values,
|
| 672 |
+
hidden_states=outputs.hidden_states,
|
| 673 |
+
attentions=outputs.attentions,
|
| 674 |
+
)
|
| 675 |
+
|
| 676 |
+
# Param Initialization, needed for device='meta' fast initialization
|
| 677 |
+
def param_init_fn(self, module):
|
| 678 |
+
init_fn_name = self.config.init_config['name']
|
| 679 |
+
MODEL_INIT_REGISTRY[init_fn_name](
|
| 680 |
+
module=module,
|
| 681 |
+
n_layers=self.config.n_layers,
|
| 682 |
+
d_model=self.config.d_model,
|
| 683 |
+
**self.config.init_config,
|
| 684 |
+
)
|
| 685 |
+
|
| 686 |
+
# FSDP Wrap function
|
| 687 |
+
def fsdp_wrap_fn(self, module):
|
| 688 |
+
return isinstance(module, MPTBlock)
|
| 689 |
+
|
| 690 |
+
# Activation Checkpointing
|
| 691 |
+
def activation_checkpointing_fn(self, module):
|
| 692 |
+
return isinstance(module, MPTBlock)
|
| 693 |
+
|
| 694 |
+
def generate_cache(self,
|
| 695 |
+
input_ids:torch.LongTensor,
|
| 696 |
+
stride:int=512,
|
| 697 |
+
max_len:int=2048,
|
| 698 |
+
cache_type:str='manual'):
|
| 699 |
+
if cache_type not in ['manual', 'faiss']:
|
| 700 |
+
raise NotImplementedError(f"Cache type {cache_type} not implemented.")
|
| 701 |
+
|
| 702 |
+
prev_end_loc=0
|
| 703 |
+
long_range_past_key_values = None
|
| 704 |
+
faiss_indexes= None
|
| 705 |
+
for b_idx in range(0, input_ids.size(-1), stride):
|
| 706 |
+
end_loc = min(b_idx + max_len, input_ids.size(-1))
|
| 707 |
+
|
| 708 |
+
trg_len = end_loc - prev_end_loc
|
| 709 |
+
subseq = input_ids[:, b_idx:end_loc].to(self.device)
|
| 710 |
+
with torch.no_grad():
|
| 711 |
+
outputs = self.transformer(subseq, use_cache=True, use_active_externalism=False)
|
| 712 |
+
to_cache = [(
|
| 713 |
+
kv[0][:,:,:,-trg_len:],
|
| 714 |
+
kv[1][:,:,-trg_len:])
|
| 715 |
+
for kv in outputs.past_key_values
|
| 716 |
+
]
|
| 717 |
+
long_range_past_key_values, faiss_indexes = self.cache(to_cache, cache_type, long_range_past_key_values=long_range_past_key_values, faiss_indexes=faiss_indexes)
|
| 718 |
+
|
| 719 |
+
prev_end_loc = end_loc
|
| 720 |
+
if end_loc == input_ids.size(-1):
|
| 721 |
+
break
|
| 722 |
+
if long_range_past_key_values is not None:
|
| 723 |
+
return long_range_past_key_values
|
| 724 |
+
else:
|
| 725 |
+
return faiss_indexes
|
| 726 |
+
|
| 727 |
+
def cache(self,
|
| 728 |
+
to_cache:List,
|
| 729 |
+
cache_type:str='manual',
|
| 730 |
+
long_range_past_key_values:List=None,
|
| 731 |
+
faiss_indexes:faiss.IndexFlatIP=None,
|
| 732 |
+
max_length_cache=100000,
|
| 733 |
+
verbose=False):
|
| 734 |
+
if long_range_past_key_values is not None and faiss_indexes is not None:
|
| 735 |
+
raise NotImplementedError("Using faiss and passing key value pairs manually are mutually exclusive right now.")
|
| 736 |
+
|
| 737 |
+
if cache_type=='faiss':
|
| 738 |
+
one_hot_encodings = F.one_hot(torch.arange(0, self.config.n_heads*self.config.n_layers))*10
|
| 739 |
+
if faiss_indexes is None:
|
| 740 |
+
faiss_indexes = (faiss.IndexFlatIP(to_cache[0][0].size(-2)+one_hot_encodings.size(-1)), faiss.IndexFlatIP(to_cache[0][1].size(-1)*2))
|
| 741 |
+
kn_index, kv_index = faiss_indexes
|
| 742 |
+
for b_idx, (k, v) in enumerate(to_cache):
|
| 743 |
+
k_n = (k/vector_norm(k, ord=2, dim=-2, keepdim=True)).to('cpu')
|
| 744 |
+
k_n = torch.concat([rearrange(k_n, 'b h d s -> b (h s) d', h=self.config.n_heads), one_hot_encodings[self.config.n_heads*b_idx:self.config.n_heads*(b_idx+1)].unsqueeze(0).repeat_interleave(repeats=k.size(-1), dim=-2)], dim=-1)
|
| 745 |
+
kn_index.add(k_n.squeeze().numpy())
|
| 746 |
+
|
| 747 |
+
k= rearrange(k, 'b h d s -> b (h s) d', h=self.config.n_heads)
|
| 748 |
+
v= rearrange(v, 'b h s d -> b (h s) d', h=self.config.n_heads)
|
| 749 |
+
kv_index.add(torch.concat([v.squeeze(), k.squeeze()], dim=1).to('cpu').numpy())
|
| 750 |
+
|
| 751 |
+
else:
|
| 752 |
+
if long_range_past_key_values is None:
|
| 753 |
+
long_range_past_key_values = [(k.to(self.memory_device),v.to(self.memory_device)) for k,v in to_cache]
|
| 754 |
+
else:
|
| 755 |
+
long_range_past_key_values = [
|
| 756 |
+
(
|
| 757 |
+
torch.concat([kv[0], to_cache[ind][0].to(self.memory_device)], dim=3),
|
| 758 |
+
torch.concat([kv[1], to_cache[ind][1].to(self.memory_device)], dim=2)
|
| 759 |
+
)
|
| 760 |
+
for ind, kv in enumerate(long_range_past_key_values)
|
| 761 |
+
]
|
| 762 |
+
if long_range_past_key_values is not None:
|
| 763 |
+
if long_range_past_key_values[0][0].size(-1) > max_length_cache: #set a limit on manual memory length
|
| 764 |
+
long_range_past_key_values = [
|
| 765 |
+
(
|
| 766 |
+
kv[0][:, :, :, -max_length_cache:],
|
| 767 |
+
kv[1][:, :, -max_length_cache:]
|
| 768 |
+
)
|
| 769 |
+
for kv in long_range_past_key_values]
|
| 770 |
+
if verbose:
|
| 771 |
+
if cache_type == 'faiss':
|
| 772 |
+
print(f"{kn_index.ntotal} keys in faiss index")
|
| 773 |
+
else:
|
| 774 |
+
print(f"{long_range_past_key_values[0][0].size(-1)} cached kvs")
|
| 775 |
+
|
| 776 |
+
return long_range_past_key_values, (kn_index, kv_index) if cache_type == 'faiss' else None
|
| 777 |
+
|
| 778 |
+
def prepare_inputs_for_generation(
|
| 779 |
+
self,
|
| 780 |
+
input_ids,
|
| 781 |
+
past_key_values=None,
|
| 782 |
+
inputs_embeds=None,
|
| 783 |
+
**kwargs,
|
| 784 |
+
):
|
| 785 |
+
if inputs_embeds is not None:
|
| 786 |
+
raise NotImplementedError(
|
| 787 |
+
'inputs_embeds is not implemented for MPT yet')
|
| 788 |
+
|
| 789 |
+
attention_mask = kwargs['attention_mask'].bool()
|
| 790 |
+
if attention_mask[:, -1].sum() != attention_mask.shape[0]:
|
| 791 |
+
raise NotImplementedError(
|
| 792 |
+
'MPT does not support generation with right padding.')
|
| 793 |
+
|
| 794 |
+
if self.transformer.attn_uses_sequence_id and self.training:
|
| 795 |
+
sequence_id = torch.zeros_like(input_ids[:1])
|
| 796 |
+
else:
|
| 797 |
+
sequence_id = None
|
| 798 |
+
|
| 799 |
+
if past_key_values is not None:
|
| 800 |
+
input_ids = input_ids[:, -1].unsqueeze(-1)
|
| 801 |
+
|
| 802 |
+
if self.transformer.prefix_lm:
|
| 803 |
+
# Leverage a convenience of sequential generation!
|
| 804 |
+
prefix_mask = torch.ones_like(attention_mask)
|
| 805 |
+
# This requires that we're using the cache
|
| 806 |
+
if kwargs.get('use_cache') == False:
|
| 807 |
+
raise NotImplementedError(
|
| 808 |
+
'MPT with prefix_lm=True does not support use_cache=False.')
|
| 809 |
+
else:
|
| 810 |
+
prefix_mask = None
|
| 811 |
+
|
| 812 |
+
return {
|
| 813 |
+
'input_ids': input_ids,
|
| 814 |
+
'attention_mask': attention_mask,
|
| 815 |
+
'prefix_mask': prefix_mask,
|
| 816 |
+
'sequence_id': sequence_id,
|
| 817 |
+
'past_key_values': past_key_values,
|
| 818 |
+
'use_cache': kwargs.get('use_cache', True),
|
| 819 |
+
'use_active_externalism': kwargs.get('use_active_externalism'),
|
| 820 |
+
'topk': kwargs.get('topk', None),
|
| 821 |
+
}
|
| 822 |
+
|
| 823 |
+
@staticmethod
|
| 824 |
+
def _reorder_cache(past_key_values, beam_idx):
|
| 825 |
+
"""Used by HuggingFace generate when using beam search with kv-caching.
|
| 826 |
+
|
| 827 |
+
See https://github.com/huggingface/transformers/blob/3ec7a47664ebe40c40f4b722f6bb1cd30c3821ec/src/transformers/models/gpt2/modeling_gpt2.py#L1122-L1133
|
| 828 |
+
for an example in transformers.
|
| 829 |
+
"""
|
| 830 |
+
reordered_past = []
|
| 831 |
+
for layer_past in past_key_values:
|
| 832 |
+
reordered_past += [
|
| 833 |
+
tuple(
|
| 834 |
+
past_state.index_select(0, beam_idx)
|
| 835 |
+
for past_state in layer_past)
|
| 836 |
+
]
|
| 837 |
+
return reordered_past
|
utils.py
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from .utils import *
|
| 2 |
+
|
| 3 |
+
import importlib
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
def instantiate_from_config(config):
|
| 7 |
+
if not "target" in config:
|
| 8 |
+
raise KeyError("Expected key `target` to instantiate.")
|
| 9 |
+
return get_obj_from_str(config["target"])(**config.get("params", dict()))
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
def get_obj_from_str(string, reload=False):
|
| 13 |
+
module, cls = string.rsplit(".", 1)
|
| 14 |
+
if reload:
|
| 15 |
+
module_imp = importlib.import_module(module)
|
| 16 |
+
importlib.reload(module_imp)
|
| 17 |
+
return getattr(importlib.import_module(module, package=None), cls)
|