Upload 13 files
Browse files- blocks.py +41 -0
- custom_embedding.py +10 -0
- fc.py +7 -0
- ffn.py +39 -0
- flash_attn_triton.py +484 -0
- hf_prefixlm_converter.py +180 -0
- norm.py +57 -0
- param_init_fns.py +179 -0
blocks.py
ADDED
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""GPT Blocks used for the GPT Model."""
|
2 |
+
from typing import Any, Dict, Optional, Tuple
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
from .attention import ATTN_CLASS_REGISTRY
|
6 |
+
from .ffn import FFN_CLASS_REGISTRY, build_ffn
|
7 |
+
from .norm import NORM_CLASS_REGISTRY
|
8 |
+
|
9 |
+
class MPTBlock(nn.Module):
|
10 |
+
|
11 |
+
def __init__(self, d_model: int, n_heads: int, expansion_ratio: int, attn_config: Optional[Dict]=None, ffn_config: Optional[Dict]=None, resid_pdrop: float=0.0, norm_type: str='low_precision_layernorm', fc_type: str='torch', device: Optional[str]=None, no_bias: bool=False, **kwargs: Any):
|
12 |
+
if attn_config is None:
|
13 |
+
attn_config = {'attn_type': 'multihead_attention', 'attn_pdrop': 0.0, 'attn_impl': 'triton', 'qk_ln': False, 'clip_qkv': None, 'softmax_scale': None, 'prefix_lm': False, 'attn_uses_sequence_id': False, 'alibi': False, 'alibi_bias_max': 8}
|
14 |
+
if ffn_config is None:
|
15 |
+
ffn_config = {'ffn_type': 'mptmlp'}
|
16 |
+
del kwargs
|
17 |
+
super().__init__()
|
18 |
+
norm_class = NORM_CLASS_REGISTRY[norm_type.lower()]
|
19 |
+
assert isinstance(attn_config['attn_type'], str)
|
20 |
+
attn_class = ATTN_CLASS_REGISTRY[attn_config['attn_type']]
|
21 |
+
args_to_exclude_in_attn_class = {'attn_type', 'prefix_lm', 'alibi', 'attn_uses_sequence_id', 'alibi_bias_max'}
|
22 |
+
attn_config_subset_for_attn_class = {k: v for (k, v) in attn_config.items() if k not in args_to_exclude_in_attn_class}
|
23 |
+
self.norm_1 = norm_class(d_model, device=device)
|
24 |
+
self.attn = attn_class(d_model=d_model, n_heads=n_heads, fc_type=fc_type, device=device, **attn_config_subset_for_attn_class, bias=not no_bias)
|
25 |
+
self.norm_2 = None
|
26 |
+
if not getattr(FFN_CLASS_REGISTRY[ffn_config['ffn_type']], '_has_norm', False):
|
27 |
+
self.norm_2 = norm_class(d_model, device=device)
|
28 |
+
self.ffn = build_ffn(d_model=d_model, expansion_ratio=expansion_ratio, device=device, bias=not no_bias, **ffn_config)
|
29 |
+
self.resid_attn_dropout = nn.Dropout(resid_pdrop)
|
30 |
+
self.resid_ffn_dropout = nn.Dropout(resid_pdrop)
|
31 |
+
|
32 |
+
def forward(self, x: torch.Tensor, past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]]=None, attn_bias: Optional[torch.Tensor]=None, attention_mask: Optional[torch.ByteTensor]=None, is_causal: bool=True, output_attentions: bool=False) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor, torch.Tensor]]]:
|
33 |
+
a = self.norm_1(x)
|
34 |
+
(b, attn_weights, past_key_value) = self.attn(a, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=is_causal, needs_weights=output_attentions)
|
35 |
+
x = x + self.resid_attn_dropout(b)
|
36 |
+
m = x
|
37 |
+
if self.norm_2 is not None:
|
38 |
+
m = self.norm_2(x)
|
39 |
+
n = self.ffn(m)
|
40 |
+
x = x + self.resid_ffn_dropout(n)
|
41 |
+
return (x, attn_weights, past_key_value)
|
custom_embedding.py
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import torch.nn.functional as F
|
3 |
+
from torch import Tensor
|
4 |
+
|
5 |
+
class SharedEmbedding(nn.Embedding):
|
6 |
+
|
7 |
+
def forward(self, input: Tensor, unembed: bool=False) -> Tensor:
|
8 |
+
if unembed:
|
9 |
+
return F.linear(input, self.weight)
|
10 |
+
return super().forward(input)
|
fc.py
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch import nn
|
2 |
+
FC_CLASS_REGISTRY = {'torch': nn.Linear}
|
3 |
+
try:
|
4 |
+
import transformer_engine.pytorch as te
|
5 |
+
FC_CLASS_REGISTRY['te'] = te.Linear
|
6 |
+
except:
|
7 |
+
pass
|
ffn.py
ADDED
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""GPT Blocks used for the GPT Model."""
|
2 |
+
from typing import Any, Optional
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
from .fc import FC_CLASS_REGISTRY
|
6 |
+
try:
|
7 |
+
import transformer_engine.pytorch as te
|
8 |
+
except:
|
9 |
+
te = None
|
10 |
+
|
11 |
+
class MPTMLP(nn.Module):
|
12 |
+
|
13 |
+
def __init__(self, d_model: int, expansion_ratio: int, fc_type: str='torch', device: Optional[str]=None, bias: bool=True):
|
14 |
+
super().__init__()
|
15 |
+
fc_kwargs: dict[str, Any] = {'bias': bias}
|
16 |
+
if fc_type != 'te':
|
17 |
+
fc_kwargs['device'] = device
|
18 |
+
self.up_proj = FC_CLASS_REGISTRY[fc_type](d_model, expansion_ratio * d_model, **fc_kwargs)
|
19 |
+
self.act = nn.GELU(approximate='none')
|
20 |
+
self.down_proj = FC_CLASS_REGISTRY[fc_type](expansion_ratio * d_model, d_model, **fc_kwargs)
|
21 |
+
self.down_proj._is_residual = True
|
22 |
+
|
23 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
24 |
+
return self.down_proj(self.act(self.up_proj(x)))
|
25 |
+
FFN_CLASS_REGISTRY = {'mptmlp': MPTMLP}
|
26 |
+
if te is not None:
|
27 |
+
te.LayerNormMLP._has_norm = True
|
28 |
+
FFN_CLASS_REGISTRY['te_ln_mlp'] = te.LayerNormMLP
|
29 |
+
|
30 |
+
def build_ffn(d_model: int, expansion_ratio: int, fc_type: str='torch', device: Optional[str]=None, bias: bool=True, **kwargs: Any) -> nn.Module:
|
31 |
+
ffn_type = kwargs.pop('ffn_type')
|
32 |
+
if ffn_type == 'mptmlp':
|
33 |
+
if len(kwargs) > 0:
|
34 |
+
raise ValueError(f'MPTMLP got an unexpected keyword argument: {kwargs}')
|
35 |
+
return MPTMLP(d_model=d_model, expansion_ratio=expansion_ratio, fc_type=fc_type, device=device, bias=bias)
|
36 |
+
elif ffn_type == 'te_ln_mlp':
|
37 |
+
assert te is not None
|
38 |
+
return te.LayerNormMLP(hidden_size=d_model, ffn_hidden_size=d_model * expansion_ratio, bias=bias, **kwargs)
|
39 |
+
raise ValueError(f'ffn_type={ffn_type!r} not recognized.')
|
flash_attn_triton.py
ADDED
@@ -0,0 +1,484 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Copied from https://github.com/HazyResearch/flash-attention/blob/eff9fe6b8076df59d64d7a3f464696738a3c7c24/flash_attn/flash_attn_triton.py
|
3 |
+
update imports to use 'triton_pre_mlir'
|
4 |
+
|
5 |
+
*Experimental* implementation of FlashAttention in Triton.
|
6 |
+
Tested with triton==2.0.0.dev20221202.
|
7 |
+
Triton 2.0 has a new backend (MLIR) but seems like it doesn't yet work for head dimensions
|
8 |
+
other than 64:
|
9 |
+
https://github.com/openai/triton/blob/d376020f90002757eea3ea9475d4f7cfc2ec5ead/python/triton/ops/flash_attention.py#L207
|
10 |
+
We'll update this implementation with the new Triton backend once this is fixed.
|
11 |
+
|
12 |
+
We use the FlashAttention implementation from Phil Tillet a starting point.
|
13 |
+
https://github.com/openai/triton/blob/master/python/tutorials/06-fused-attention.py
|
14 |
+
|
15 |
+
Changes:
|
16 |
+
- Implement both causal and non-causal attention.
|
17 |
+
- Implement both self-attention and cross-attention.
|
18 |
+
- Support arbitrary seqlens (not just multiples of 128), for both forward and backward.
|
19 |
+
- Support all head dimensions up to 128 (not just 16, 32, 64, 128), for both forward and backward.
|
20 |
+
- Support attention bias.
|
21 |
+
- Speed up the forward pass a bit, and only store the LSE instead of m and l.
|
22 |
+
- Make the backward for d=128 much faster by reducing register spilling.
|
23 |
+
- Optionally parallelize the backward pass across seqlen_k, to deal with the case of
|
24 |
+
small batch size * nheads.
|
25 |
+
|
26 |
+
Caution:
|
27 |
+
- This is an *experimental* implementation. The forward pass should be quite robust but
|
28 |
+
I'm not 100% sure that the backward pass doesn't have race conditions (due to the Triton compiler).
|
29 |
+
- This implementation has only been tested on A100.
|
30 |
+
- If you plan to use headdim other than 64 and 128, you should test for race conditions
|
31 |
+
(due to the Triton compiler), as done in tests/test_flash_attn.py
|
32 |
+
"test_flash_attn_triton_race_condition". I've tested and fixed many race conditions
|
33 |
+
for different head dimensions (40, 48, 64, 128, 80, 88, 96), but I'm still not 100% confident
|
34 |
+
that there are none left for other head dimensions.
|
35 |
+
|
36 |
+
Differences between this Triton version and the CUDA version:
|
37 |
+
- Triton version doesn't support dropout.
|
38 |
+
- Triton forward is generally faster than CUDA forward, while Triton backward is
|
39 |
+
generally slower than CUDA backward. Overall Triton forward + backward is slightly slower
|
40 |
+
than CUDA forward + backward.
|
41 |
+
- Triton version doesn't support different sequence lengths in a batch (i.e., RaggedTensor/NestedTensor).
|
42 |
+
- Triton version supports attention bias, while CUDA version doesn't.
|
43 |
+
"""
|
44 |
+
import math
|
45 |
+
import torch
|
46 |
+
import triton_pre_mlir as triton
|
47 |
+
import triton_pre_mlir.language as tl
|
48 |
+
|
49 |
+
@triton.heuristics({'EVEN_M': lambda args: args['seqlen_q'] % args['BLOCK_M'] == 0, 'EVEN_N': lambda args: args['seqlen_k'] % args['BLOCK_N'] == 0, 'EVEN_HEADDIM': lambda args: args['headdim'] == args['BLOCK_HEADDIM']})
|
50 |
+
@triton.jit
|
51 |
+
def _fwd_kernel(Q, K, V, Bias, Out, Lse, TMP, softmax_scale, stride_qb, stride_qh, stride_qm, stride_kb, stride_kh, stride_kn, stride_vb, stride_vh, stride_vn, stride_bb, stride_bh, stride_bm, stride_ob, stride_oh, stride_om, nheads, seqlen_q, seqlen_k, seqlen_q_rounded, headdim, CACHE_KEY_SEQLEN_Q, CACHE_KEY_SEQLEN_K, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
|
52 |
+
start_m = tl.program_id(0)
|
53 |
+
off_hb = tl.program_id(1)
|
54 |
+
off_b = off_hb // nheads
|
55 |
+
off_h = off_hb % nheads
|
56 |
+
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
|
57 |
+
offs_n = tl.arange(0, BLOCK_N)
|
58 |
+
offs_d = tl.arange(0, BLOCK_HEADDIM)
|
59 |
+
q_ptrs = Q + off_b * stride_qb + off_h * stride_qh + (offs_m[:, None] * stride_qm + offs_d[None, :])
|
60 |
+
k_ptrs = K + off_b * stride_kb + off_h * stride_kh + (offs_n[:, None] * stride_kn + offs_d[None, :])
|
61 |
+
v_ptrs = V + off_b * stride_vb + off_h * stride_vh + (offs_n[:, None] * stride_vn + offs_d[None, :])
|
62 |
+
if BIAS_TYPE == 'vector':
|
63 |
+
b_ptrs = Bias + off_b * stride_bb + off_h * stride_bh + offs_n
|
64 |
+
elif BIAS_TYPE == 'matrix':
|
65 |
+
b_ptrs = Bias + off_b * stride_bb + off_h * stride_bh + (offs_m[:, None] * stride_bm + offs_n[None, :])
|
66 |
+
t_ptrs = TMP + off_hb * seqlen_q_rounded + offs_m
|
67 |
+
lse_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float('inf')
|
68 |
+
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float('inf')
|
69 |
+
acc_o = tl.zeros([BLOCK_M, BLOCK_HEADDIM], dtype=tl.float32)
|
70 |
+
if EVEN_M & EVEN_N:
|
71 |
+
if EVEN_HEADDIM:
|
72 |
+
q = tl.load(q_ptrs)
|
73 |
+
else:
|
74 |
+
q = tl.load(q_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
|
75 |
+
elif EVEN_HEADDIM:
|
76 |
+
q = tl.load(q_ptrs, mask=offs_m[:, None] < seqlen_q, other=0.0)
|
77 |
+
else:
|
78 |
+
q = tl.load(q_ptrs, mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
|
79 |
+
end_n = seqlen_k if not IS_CAUSAL else tl.minimum((start_m + 1) * BLOCK_M, seqlen_k)
|
80 |
+
for start_n in range(0, end_n, BLOCK_N):
|
81 |
+
start_n = tl.multiple_of(start_n, BLOCK_N)
|
82 |
+
if EVEN_N & EVEN_M:
|
83 |
+
if EVEN_HEADDIM:
|
84 |
+
k = tl.load(k_ptrs + start_n * stride_kn)
|
85 |
+
else:
|
86 |
+
k = tl.load(k_ptrs + start_n * stride_kn, mask=offs_d[None, :] < headdim, other=0.0)
|
87 |
+
elif EVEN_HEADDIM:
|
88 |
+
k = tl.load(k_ptrs + start_n * stride_kn, mask=(start_n + offs_n)[:, None] < seqlen_k, other=0.0)
|
89 |
+
else:
|
90 |
+
k = tl.load(k_ptrs + start_n * stride_kn, mask=((start_n + offs_n)[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
|
91 |
+
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
|
92 |
+
qk += tl.dot(q, k, trans_b=True)
|
93 |
+
if not EVEN_N:
|
94 |
+
qk += tl.where((start_n + offs_n)[None, :] < seqlen_k, 0, float('-inf'))
|
95 |
+
if IS_CAUSAL:
|
96 |
+
qk += tl.where(offs_m[:, None] >= (start_n + offs_n)[None, :], 0, float('-inf'))
|
97 |
+
if BIAS_TYPE != 'none':
|
98 |
+
if BIAS_TYPE == 'vector':
|
99 |
+
if EVEN_N:
|
100 |
+
bias = tl.load(b_ptrs + start_n).to(tl.float32)
|
101 |
+
else:
|
102 |
+
bias = tl.load(b_ptrs + start_n, mask=start_n + offs_n < seqlen_k, other=0.0).to(tl.float32)
|
103 |
+
bias = bias[None, :]
|
104 |
+
elif BIAS_TYPE == 'matrix':
|
105 |
+
if EVEN_M & EVEN_N:
|
106 |
+
bias = tl.load(b_ptrs + start_n).to(tl.float32)
|
107 |
+
else:
|
108 |
+
bias = tl.load(b_ptrs + start_n, mask=(offs_m[:, None] < seqlen_q) & ((start_n + offs_n)[None, :] < seqlen_k), other=0.0).to(tl.float32)
|
109 |
+
qk = qk * softmax_scale + bias
|
110 |
+
m_ij = tl.maximum(tl.max(qk, 1), lse_i)
|
111 |
+
p = tl.exp(qk - m_ij[:, None])
|
112 |
+
else:
|
113 |
+
m_ij = tl.maximum(tl.max(qk, 1) * softmax_scale, lse_i)
|
114 |
+
p = tl.exp(qk * softmax_scale - m_ij[:, None])
|
115 |
+
l_ij = tl.sum(p, 1)
|
116 |
+
acc_o_scale = tl.exp(m_i - m_ij)
|
117 |
+
tl.store(t_ptrs, acc_o_scale)
|
118 |
+
acc_o_scale = tl.load(t_ptrs)
|
119 |
+
acc_o = acc_o * acc_o_scale[:, None]
|
120 |
+
if EVEN_N & EVEN_M:
|
121 |
+
if EVEN_HEADDIM:
|
122 |
+
v = tl.load(v_ptrs + start_n * stride_vn)
|
123 |
+
else:
|
124 |
+
v = tl.load(v_ptrs + start_n * stride_vn, mask=offs_d[None, :] < headdim, other=0.0)
|
125 |
+
elif EVEN_HEADDIM:
|
126 |
+
v = tl.load(v_ptrs + start_n * stride_vn, mask=(start_n + offs_n)[:, None] < seqlen_k, other=0.0)
|
127 |
+
else:
|
128 |
+
v = tl.load(v_ptrs + start_n * stride_vn, mask=((start_n + offs_n)[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
|
129 |
+
p = p.to(v.dtype)
|
130 |
+
acc_o += tl.dot(p, v)
|
131 |
+
m_i = m_ij
|
132 |
+
l_i_new = tl.exp(lse_i - m_ij) + l_ij
|
133 |
+
lse_i = m_ij + tl.log(l_i_new)
|
134 |
+
o_scale = tl.exp(m_i - lse_i)
|
135 |
+
tl.store(t_ptrs, o_scale)
|
136 |
+
o_scale = tl.load(t_ptrs)
|
137 |
+
acc_o = acc_o * o_scale[:, None]
|
138 |
+
start_m = tl.program_id(0)
|
139 |
+
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
|
140 |
+
lse_ptrs = Lse + off_hb * seqlen_q_rounded + offs_m
|
141 |
+
tl.store(lse_ptrs, lse_i)
|
142 |
+
offs_d = tl.arange(0, BLOCK_HEADDIM)
|
143 |
+
out_ptrs = Out + off_b * stride_ob + off_h * stride_oh + (offs_m[:, None] * stride_om + offs_d[None, :])
|
144 |
+
if EVEN_M:
|
145 |
+
if EVEN_HEADDIM:
|
146 |
+
tl.store(out_ptrs, acc_o)
|
147 |
+
else:
|
148 |
+
tl.store(out_ptrs, acc_o, mask=offs_d[None, :] < headdim)
|
149 |
+
elif EVEN_HEADDIM:
|
150 |
+
tl.store(out_ptrs, acc_o, mask=offs_m[:, None] < seqlen_q)
|
151 |
+
else:
|
152 |
+
tl.store(out_ptrs, acc_o, mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim))
|
153 |
+
|
154 |
+
@triton.jit
|
155 |
+
def _bwd_preprocess_do_o_dot(Out, DO, Delta, stride_ob, stride_oh, stride_om, stride_dob, stride_doh, stride_dom, nheads, seqlen_q, seqlen_q_rounded, headdim, BLOCK_M: tl.constexpr, BLOCK_HEADDIM: tl.constexpr):
|
156 |
+
start_m = tl.program_id(0)
|
157 |
+
off_hb = tl.program_id(1)
|
158 |
+
off_b = off_hb // nheads
|
159 |
+
off_h = off_hb % nheads
|
160 |
+
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
|
161 |
+
offs_d = tl.arange(0, BLOCK_HEADDIM)
|
162 |
+
o = tl.load(Out + off_b * stride_ob + off_h * stride_oh + offs_m[:, None] * stride_om + offs_d[None, :], mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0).to(tl.float32)
|
163 |
+
do = tl.load(DO + off_b * stride_dob + off_h * stride_doh + offs_m[:, None] * stride_dom + offs_d[None, :], mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0).to(tl.float32)
|
164 |
+
delta = tl.sum(o * do, axis=1)
|
165 |
+
tl.store(Delta + off_hb * seqlen_q_rounded + offs_m, delta)
|
166 |
+
|
167 |
+
@triton.jit
|
168 |
+
def _bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr):
|
169 |
+
if EVEN_N & EVEN_M:
|
170 |
+
if EVEN_HEADDIM:
|
171 |
+
tl.store(dv_ptrs, dv)
|
172 |
+
tl.store(dk_ptrs, dk)
|
173 |
+
else:
|
174 |
+
tl.store(dv_ptrs, dv, mask=offs_d[None, :] < headdim)
|
175 |
+
tl.store(dk_ptrs, dk, mask=offs_d[None, :] < headdim)
|
176 |
+
elif EVEN_HEADDIM:
|
177 |
+
tl.store(dv_ptrs, dv, mask=offs_n[:, None] < seqlen_k)
|
178 |
+
tl.store(dk_ptrs, dk, mask=offs_n[:, None] < seqlen_k)
|
179 |
+
else:
|
180 |
+
tl.store(dv_ptrs, dv, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim))
|
181 |
+
tl.store(dk_ptrs, dk, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim))
|
182 |
+
|
183 |
+
@triton.jit
|
184 |
+
def _bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD: tl.constexpr, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
|
185 |
+
begin_m = 0 if not IS_CAUSAL else start_n * BLOCK_N // BLOCK_M * BLOCK_M
|
186 |
+
offs_qm = begin_m + tl.arange(0, BLOCK_M)
|
187 |
+
offs_n = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
|
188 |
+
offs_m = tl.arange(0, BLOCK_M)
|
189 |
+
offs_d = tl.arange(0, BLOCK_HEADDIM)
|
190 |
+
q_ptrs = Q + (offs_qm[:, None] * stride_qm + offs_d[None, :])
|
191 |
+
k_ptrs = K + (offs_n[:, None] * stride_kn + offs_d[None, :])
|
192 |
+
v_ptrs = V + (offs_n[:, None] * stride_vn + offs_d[None, :])
|
193 |
+
do_ptrs = DO + (offs_qm[:, None] * stride_dom + offs_d[None, :])
|
194 |
+
dq_ptrs = DQ + (offs_qm[:, None] * stride_dqm + offs_d[None, :])
|
195 |
+
if BIAS_TYPE == 'vector':
|
196 |
+
b_ptrs = Bias + offs_n
|
197 |
+
elif BIAS_TYPE == 'matrix':
|
198 |
+
b_ptrs = Bias + (offs_qm[:, None] * stride_bm + offs_n[None, :])
|
199 |
+
dv = tl.zeros([BLOCK_N, BLOCK_HEADDIM], dtype=tl.float32)
|
200 |
+
dk = tl.zeros([BLOCK_N, BLOCK_HEADDIM], dtype=tl.float32)
|
201 |
+
if begin_m >= seqlen_q:
|
202 |
+
dv_ptrs = DV + (offs_n[:, None] * stride_dvn + offs_d[None, :])
|
203 |
+
dk_ptrs = DK + (offs_n[:, None] * stride_dkn + offs_d[None, :])
|
204 |
+
_bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM)
|
205 |
+
return
|
206 |
+
if EVEN_N & EVEN_M:
|
207 |
+
if EVEN_HEADDIM:
|
208 |
+
k = tl.load(k_ptrs)
|
209 |
+
v = tl.load(v_ptrs)
|
210 |
+
else:
|
211 |
+
k = tl.load(k_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
|
212 |
+
v = tl.load(v_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
|
213 |
+
elif EVEN_HEADDIM:
|
214 |
+
k = tl.load(k_ptrs, mask=offs_n[:, None] < seqlen_k, other=0.0)
|
215 |
+
v = tl.load(v_ptrs, mask=offs_n[:, None] < seqlen_k, other=0.0)
|
216 |
+
else:
|
217 |
+
k = tl.load(k_ptrs, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
|
218 |
+
v = tl.load(v_ptrs, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
|
219 |
+
num_block_m = tl.cdiv(seqlen_q, BLOCK_M)
|
220 |
+
for start_m in range(begin_m, num_block_m * BLOCK_M, BLOCK_M):
|
221 |
+
start_m = tl.multiple_of(start_m, BLOCK_M)
|
222 |
+
offs_m_curr = start_m + offs_m
|
223 |
+
if EVEN_M & EVEN_HEADDIM:
|
224 |
+
q = tl.load(q_ptrs)
|
225 |
+
elif EVEN_HEADDIM:
|
226 |
+
q = tl.load(q_ptrs, mask=offs_m_curr[:, None] < seqlen_q, other=0.0)
|
227 |
+
else:
|
228 |
+
q = tl.load(q_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
|
229 |
+
qk = tl.dot(q, k, trans_b=True)
|
230 |
+
if not EVEN_N:
|
231 |
+
qk = tl.where(offs_n[None, :] < seqlen_k, qk, float('-inf'))
|
232 |
+
if IS_CAUSAL:
|
233 |
+
qk = tl.where(offs_m_curr[:, None] >= offs_n[None, :], qk, float('-inf'))
|
234 |
+
if BIAS_TYPE != 'none':
|
235 |
+
tl.debug_barrier()
|
236 |
+
if BIAS_TYPE == 'vector':
|
237 |
+
if EVEN_N:
|
238 |
+
bias = tl.load(b_ptrs).to(tl.float32)
|
239 |
+
else:
|
240 |
+
bias = tl.load(b_ptrs, mask=offs_n < seqlen_k, other=0.0).to(tl.float32)
|
241 |
+
bias = bias[None, :]
|
242 |
+
elif BIAS_TYPE == 'matrix':
|
243 |
+
if EVEN_M & EVEN_N:
|
244 |
+
bias = tl.load(b_ptrs).to(tl.float32)
|
245 |
+
else:
|
246 |
+
bias = tl.load(b_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_n[None, :] < seqlen_k), other=0.0).to(tl.float32)
|
247 |
+
qk = qk * softmax_scale + bias
|
248 |
+
if not EVEN_M & EVEN_HEADDIM:
|
249 |
+
tl.debug_barrier()
|
250 |
+
lse_i = tl.load(LSE + offs_m_curr)
|
251 |
+
if BIAS_TYPE == 'none':
|
252 |
+
p = tl.exp(qk * softmax_scale - lse_i[:, None])
|
253 |
+
else:
|
254 |
+
p = tl.exp(qk - lse_i[:, None])
|
255 |
+
if EVEN_M & EVEN_HEADDIM:
|
256 |
+
do = tl.load(do_ptrs)
|
257 |
+
else:
|
258 |
+
do = tl.load(do_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
|
259 |
+
dv += tl.dot(p.to(do.dtype), do, trans_a=True)
|
260 |
+
if not EVEN_M & EVEN_HEADDIM:
|
261 |
+
tl.debug_barrier()
|
262 |
+
dp = tl.dot(do, v, trans_b=True)
|
263 |
+
if not EVEN_HEADDIM:
|
264 |
+
tl.debug_barrier()
|
265 |
+
Di = tl.load(D + offs_m_curr)
|
266 |
+
ds = (p * (dp - Di[:, None]) * softmax_scale).to(q.dtype)
|
267 |
+
dk += tl.dot(ds, q, trans_a=True)
|
268 |
+
if not EVEN_M & EVEN_HEADDIM:
|
269 |
+
tl.debug_barrier()
|
270 |
+
if not ATOMIC_ADD:
|
271 |
+
if EVEN_M & EVEN_HEADDIM:
|
272 |
+
dq = tl.load(dq_ptrs, eviction_policy='evict_last')
|
273 |
+
dq += tl.dot(ds, k)
|
274 |
+
tl.store(dq_ptrs, dq, eviction_policy='evict_last')
|
275 |
+
elif EVEN_HEADDIM:
|
276 |
+
dq = tl.load(dq_ptrs, mask=offs_m_curr[:, None] < seqlen_q, other=0.0, eviction_policy='evict_last')
|
277 |
+
dq += tl.dot(ds, k)
|
278 |
+
tl.store(dq_ptrs, dq, mask=offs_m_curr[:, None] < seqlen_q, eviction_policy='evict_last')
|
279 |
+
else:
|
280 |
+
dq = tl.load(dq_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0, eviction_policy='evict_last')
|
281 |
+
dq += tl.dot(ds, k)
|
282 |
+
tl.store(dq_ptrs, dq, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), eviction_policy='evict_last')
|
283 |
+
else:
|
284 |
+
dq = tl.dot(ds, k)
|
285 |
+
if EVEN_M & EVEN_HEADDIM:
|
286 |
+
tl.atomic_add(dq_ptrs, dq)
|
287 |
+
elif EVEN_HEADDIM:
|
288 |
+
tl.atomic_add(dq_ptrs, dq, mask=offs_m_curr[:, None] < seqlen_q)
|
289 |
+
else:
|
290 |
+
tl.atomic_add(dq_ptrs, dq, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim))
|
291 |
+
dq_ptrs += BLOCK_M * stride_dqm
|
292 |
+
q_ptrs += BLOCK_M * stride_qm
|
293 |
+
do_ptrs += BLOCK_M * stride_dom
|
294 |
+
if BIAS_TYPE == 'matrix':
|
295 |
+
b_ptrs += BLOCK_M * stride_bm
|
296 |
+
dv_ptrs = DV + (offs_n[:, None] * stride_dvn + offs_d[None, :])
|
297 |
+
dk_ptrs = DK + (offs_n[:, None] * stride_dkn + offs_d[None, :])
|
298 |
+
_bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM)
|
299 |
+
|
300 |
+
def init_to_zero(name):
|
301 |
+
return lambda nargs: nargs[name].zero_()
|
302 |
+
|
303 |
+
@triton.autotune(configs=[triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'SEQUENCE_PARALLEL': False}, num_warps=8, num_stages=1, pre_hook=init_to_zero('DQ')), triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'SEQUENCE_PARALLEL': True}, num_warps=8, num_stages=1, pre_hook=init_to_zero('DQ'))], key=['CACHE_KEY_SEQLEN_Q', 'CACHE_KEY_SEQLEN_K', 'BIAS_TYPE', 'IS_CAUSAL', 'BLOCK_HEADDIM'])
|
304 |
+
@triton.heuristics({'EVEN_M': lambda args: args['seqlen_q'] % args['BLOCK_M'] == 0, 'EVEN_N': lambda args: args['seqlen_k'] % args['BLOCK_N'] == 0, 'EVEN_HEADDIM': lambda args: args['headdim'] == args['BLOCK_HEADDIM']})
|
305 |
+
@triton.jit
|
306 |
+
def _bwd_kernel(Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qb, stride_qh, stride_qm, stride_kb, stride_kh, stride_kn, stride_vb, stride_vh, stride_vn, stride_bb, stride_bh, stride_bm, stride_dob, stride_doh, stride_dom, stride_dqb, stride_dqh, stride_dqm, stride_dkb, stride_dkh, stride_dkn, stride_dvb, stride_dvh, stride_dvn, nheads, seqlen_q, seqlen_k, seqlen_q_rounded, headdim, CACHE_KEY_SEQLEN_Q, CACHE_KEY_SEQLEN_K, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, SEQUENCE_PARALLEL: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
|
307 |
+
off_hb = tl.program_id(1)
|
308 |
+
off_b = off_hb // nheads
|
309 |
+
off_h = off_hb % nheads
|
310 |
+
Q += off_b * stride_qb + off_h * stride_qh
|
311 |
+
K += off_b * stride_kb + off_h * stride_kh
|
312 |
+
V += off_b * stride_vb + off_h * stride_vh
|
313 |
+
DO += off_b * stride_dob + off_h * stride_doh
|
314 |
+
DQ += off_b * stride_dqb + off_h * stride_dqh
|
315 |
+
DK += off_b * stride_dkb + off_h * stride_dkh
|
316 |
+
DV += off_b * stride_dvb + off_h * stride_dvh
|
317 |
+
if BIAS_TYPE != 'none':
|
318 |
+
Bias += off_b * stride_bb + off_h * stride_bh
|
319 |
+
D += off_hb * seqlen_q_rounded
|
320 |
+
LSE += off_hb * seqlen_q_rounded
|
321 |
+
if not SEQUENCE_PARALLEL:
|
322 |
+
num_block_n = tl.cdiv(seqlen_k, BLOCK_N)
|
323 |
+
for start_n in range(0, num_block_n):
|
324 |
+
_bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD=False, BIAS_TYPE=BIAS_TYPE, IS_CAUSAL=IS_CAUSAL, BLOCK_HEADDIM=BLOCK_HEADDIM, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM, BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N)
|
325 |
+
else:
|
326 |
+
start_n = tl.program_id(0)
|
327 |
+
_bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD=True, BIAS_TYPE=BIAS_TYPE, IS_CAUSAL=IS_CAUSAL, BLOCK_HEADDIM=BLOCK_HEADDIM, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM, BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N)
|
328 |
+
|
329 |
+
def _flash_attn_forward(q, k, v, bias=None, causal=False, softmax_scale=None):
|
330 |
+
(batch, seqlen_q, nheads, d) = q.shape
|
331 |
+
(_, seqlen_k, _, _) = k.shape
|
332 |
+
assert k.shape == (batch, seqlen_k, nheads, d)
|
333 |
+
assert v.shape == (batch, seqlen_k, nheads, d)
|
334 |
+
assert d <= 128, 'FlashAttention only support head dimensions up to 128'
|
335 |
+
assert q.dtype == k.dtype == v.dtype, 'All tensors must have the same type'
|
336 |
+
assert q.dtype in [torch.float16, torch.bfloat16], 'Only support fp16 and bf16'
|
337 |
+
assert q.is_cuda and k.is_cuda and v.is_cuda
|
338 |
+
softmax_scale = softmax_scale or 1.0 / math.sqrt(d)
|
339 |
+
has_bias = bias is not None
|
340 |
+
bias_type = 'none'
|
341 |
+
if has_bias:
|
342 |
+
assert bias.dtype in [q.dtype, torch.float]
|
343 |
+
assert bias.is_cuda
|
344 |
+
assert bias.dim() == 4
|
345 |
+
if bias.stride(-1) != 1:
|
346 |
+
bias = bias.contiguous()
|
347 |
+
if bias.shape[2:] == (1, seqlen_k):
|
348 |
+
bias_type = 'vector'
|
349 |
+
elif bias.shape[2:] == (seqlen_q, seqlen_k):
|
350 |
+
bias_type = 'matrix'
|
351 |
+
else:
|
352 |
+
raise RuntimeError('Last 2 dimensions of bias must be (1, seqlen_k) or (seqlen_q, seqlen_k)')
|
353 |
+
bias = bias.expand(batch, nheads, seqlen_q, seqlen_k)
|
354 |
+
bias_strides = (bias.stride(0), bias.stride(1), bias.stride(2)) if has_bias else (0, 0, 0)
|
355 |
+
seqlen_q_rounded = math.ceil(seqlen_q / 128) * 128
|
356 |
+
lse = torch.empty((batch, nheads, seqlen_q_rounded), device=q.device, dtype=torch.float32)
|
357 |
+
tmp = torch.empty((batch, nheads, seqlen_q_rounded), device=q.device, dtype=torch.float32)
|
358 |
+
o = torch.empty_like(q)
|
359 |
+
BLOCK_HEADDIM = max(triton.next_power_of_2(d), 16)
|
360 |
+
BLOCK = 128
|
361 |
+
num_warps = 4 if d <= 64 else 8
|
362 |
+
grid = lambda META: (triton.cdiv(seqlen_q, META['BLOCK_M']), batch * nheads)
|
363 |
+
_fwd_kernel[grid](q, k, v, bias, o, lse, tmp, softmax_scale, q.stride(0), q.stride(2), q.stride(1), k.stride(0), k.stride(2), k.stride(1), v.stride(0), v.stride(2), v.stride(1), *bias_strides, o.stride(0), o.stride(2), o.stride(1), nheads, seqlen_q, seqlen_k, seqlen_q_rounded, d, seqlen_q // 32, seqlen_k // 32, bias_type, causal, BLOCK_HEADDIM, BLOCK_M=BLOCK, BLOCK_N=BLOCK, num_warps=num_warps, num_stages=1)
|
364 |
+
return (o, lse, softmax_scale)
|
365 |
+
|
366 |
+
def _flash_attn_backward(do, q, k, v, o, lse, dq, dk, dv, bias=None, causal=False, softmax_scale=None):
|
367 |
+
if do.stride(-1) != 1:
|
368 |
+
do = do.contiguous()
|
369 |
+
(batch, seqlen_q, nheads, d) = q.shape
|
370 |
+
(_, seqlen_k, _, _) = k.shape
|
371 |
+
assert d <= 128
|
372 |
+
seqlen_q_rounded = math.ceil(seqlen_q / 128) * 128
|
373 |
+
assert lse.shape == (batch, nheads, seqlen_q_rounded)
|
374 |
+
assert q.stride(-1) == k.stride(-1) == v.stride(-1) == o.stride(-1) == 1
|
375 |
+
assert dq.stride(-1) == dk.stride(-1) == dv.stride(-1) == 1
|
376 |
+
softmax_scale = softmax_scale or 1.0 / math.sqrt(d)
|
377 |
+
dq_accum = torch.empty_like(q, dtype=torch.float32)
|
378 |
+
delta = torch.empty_like(lse)
|
379 |
+
BLOCK_HEADDIM = max(triton.next_power_of_2(d), 16)
|
380 |
+
grid = lambda META: (triton.cdiv(seqlen_q, META['BLOCK_M']), batch * nheads)
|
381 |
+
_bwd_preprocess_do_o_dot[grid](o, do, delta, o.stride(0), o.stride(2), o.stride(1), do.stride(0), do.stride(2), do.stride(1), nheads, seqlen_q, seqlen_q_rounded, d, BLOCK_M=128, BLOCK_HEADDIM=BLOCK_HEADDIM)
|
382 |
+
has_bias = bias is not None
|
383 |
+
bias_type = 'none'
|
384 |
+
if has_bias:
|
385 |
+
assert bias.dtype in [q.dtype, torch.float]
|
386 |
+
assert bias.is_cuda
|
387 |
+
assert bias.dim() == 4
|
388 |
+
assert bias.stride(-1) == 1
|
389 |
+
if bias.shape[2:] == (1, seqlen_k):
|
390 |
+
bias_type = 'vector'
|
391 |
+
elif bias.shape[2:] == (seqlen_q, seqlen_k):
|
392 |
+
bias_type = 'matrix'
|
393 |
+
else:
|
394 |
+
raise RuntimeError('Last 2 dimensions of bias must be (1, seqlen_k) or (seqlen_q, seqlen_k)')
|
395 |
+
bias = bias.expand(batch, nheads, seqlen_q, seqlen_k)
|
396 |
+
bias_strides = (bias.stride(0), bias.stride(1), bias.stride(2)) if has_bias else (0, 0, 0)
|
397 |
+
grid = lambda META: (triton.cdiv(seqlen_k, META['BLOCK_N']) if META['SEQUENCE_PARALLEL'] else 1, batch * nheads)
|
398 |
+
_bwd_kernel[grid](q, k, v, bias, do, dq_accum, dk, dv, lse, delta, softmax_scale, q.stride(0), q.stride(2), q.stride(1), k.stride(0), k.stride(2), k.stride(1), v.stride(0), v.stride(2), v.stride(1), *bias_strides, do.stride(0), do.stride(2), do.stride(1), dq_accum.stride(0), dq_accum.stride(2), dq_accum.stride(1), dk.stride(0), dk.stride(2), dk.stride(1), dv.stride(0), dv.stride(2), dv.stride(1), nheads, seqlen_q, seqlen_k, seqlen_q_rounded, d, seqlen_q // 32, seqlen_k // 32, bias_type, causal, BLOCK_HEADDIM)
|
399 |
+
dq.copy_(dq_accum)
|
400 |
+
|
401 |
+
class FlashAttnQKVPackedFunc(torch.autograd.Function):
|
402 |
+
|
403 |
+
@staticmethod
|
404 |
+
def forward(ctx, qkv, bias=None, causal=False, softmax_scale=None):
|
405 |
+
"""
|
406 |
+
qkv: (batch, seqlen, 3, nheads, headdim)
|
407 |
+
bias: optional, shape broadcastible to (batch, nheads, seqlen, seqlen).
|
408 |
+
For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen).
|
409 |
+
ALiBi mask for non-causal would have shape (1, nheads, seqlen, seqlen)
|
410 |
+
"""
|
411 |
+
if qkv.stride(-1) != 1:
|
412 |
+
qkv = qkv.contiguous()
|
413 |
+
(o, lse, ctx.softmax_scale) = _flash_attn_forward(qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2], bias=bias, causal=causal, softmax_scale=softmax_scale)
|
414 |
+
ctx.save_for_backward(qkv, o, lse, bias)
|
415 |
+
ctx.causal = causal
|
416 |
+
return o
|
417 |
+
|
418 |
+
@staticmethod
|
419 |
+
def backward(ctx, do):
|
420 |
+
(qkv, o, lse, bias) = ctx.saved_tensors
|
421 |
+
assert not ctx.needs_input_grad[1], 'FlashAttention does not support bias gradient yet'
|
422 |
+
with torch.inference_mode():
|
423 |
+
dqkv = torch.empty_like(qkv)
|
424 |
+
_flash_attn_backward(do, qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2], o, lse, dqkv[:, :, 0], dqkv[:, :, 1], dqkv[:, :, 2], bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
|
425 |
+
return (dqkv, None, None, None)
|
426 |
+
flash_attn_qkvpacked_func = FlashAttnQKVPackedFunc.apply
|
427 |
+
|
428 |
+
class FlashAttnKVPackedFunc(torch.autograd.Function):
|
429 |
+
|
430 |
+
@staticmethod
|
431 |
+
def forward(ctx, q, kv, bias=None, causal=False, softmax_scale=None):
|
432 |
+
"""
|
433 |
+
q: (batch, seqlen_q, nheads, headdim)
|
434 |
+
kv: (batch, seqlen_k, 2, nheads, headdim)
|
435 |
+
bias: optional, shape broadcastible to (batch, nheads, seqlen_q, seqlen_k).
|
436 |
+
For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen_k).
|
437 |
+
ALiBi mask for non-causal would have shape (1, nheads, seqlen_q, seqlen_k)
|
438 |
+
"""
|
439 |
+
(q, kv) = [x if x.stride(-1) == 1 else x.contiguous() for x in [q, kv]]
|
440 |
+
(o, lse, ctx.softmax_scale) = _flash_attn_forward(q, kv[:, :, 0], kv[:, :, 1], bias=bias, causal=causal, softmax_scale=softmax_scale)
|
441 |
+
ctx.save_for_backward(q, kv, o, lse, bias)
|
442 |
+
ctx.causal = causal
|
443 |
+
return o
|
444 |
+
|
445 |
+
@staticmethod
|
446 |
+
def backward(ctx, do):
|
447 |
+
(q, kv, o, lse, bias) = ctx.saved_tensors
|
448 |
+
if len(ctx.needs_input_grad) >= 3:
|
449 |
+
assert not ctx.needs_input_grad[2], 'FlashAttention does not support bias gradient yet'
|
450 |
+
with torch.inference_mode():
|
451 |
+
dq = torch.empty_like(q)
|
452 |
+
dkv = torch.empty_like(kv)
|
453 |
+
_flash_attn_backward(do, q, kv[:, :, 0], kv[:, :, 1], o, lse, dq, dkv[:, :, 0], dkv[:, :, 1], bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
|
454 |
+
return (dq, dkv, None, None, None)
|
455 |
+
flash_attn_kvpacked_func = FlashAttnKVPackedFunc.apply
|
456 |
+
|
457 |
+
class FlashAttnFunc(torch.autograd.Function):
|
458 |
+
|
459 |
+
@staticmethod
|
460 |
+
def forward(ctx, q, k, v, bias=None, causal=False, softmax_scale=None):
|
461 |
+
"""
|
462 |
+
q: (batch_size, seqlen_q, nheads, headdim)
|
463 |
+
k, v: (batch_size, seqlen_k, nheads, headdim)
|
464 |
+
bias: optional, shape broadcastible to (batch, nheads, seqlen_q, seqlen_k).
|
465 |
+
For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen_k).
|
466 |
+
ALiBi mask for non-causal would have shape (1, nheads, seqlen_q, seqlen_k)
|
467 |
+
"""
|
468 |
+
(q, k, v) = [x if x.stride(-1) == 1 else x.contiguous() for x in [q, k, v]]
|
469 |
+
(o, lse, ctx.softmax_scale) = _flash_attn_forward(q, k, v, bias=bias, causal=causal, softmax_scale=softmax_scale)
|
470 |
+
ctx.save_for_backward(q, k, v, o, lse, bias)
|
471 |
+
ctx.causal = causal
|
472 |
+
return o
|
473 |
+
|
474 |
+
@staticmethod
|
475 |
+
def backward(ctx, do):
|
476 |
+
(q, k, v, o, lse, bias) = ctx.saved_tensors
|
477 |
+
assert not ctx.needs_input_grad[3], 'FlashAttention does not support bias gradient yet'
|
478 |
+
with torch.inference_mode():
|
479 |
+
dq = torch.empty_like(q)
|
480 |
+
dk = torch.empty_like(k)
|
481 |
+
dv = torch.empty_like(v)
|
482 |
+
_flash_attn_backward(do, q, k, v, o, lse, dq, dk, dv, bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
|
483 |
+
return (dq, dk, dv, None, None, None)
|
484 |
+
flash_attn_func = FlashAttnFunc.apply
|
hf_prefixlm_converter.py
ADDED
@@ -0,0 +1,180 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Converts Huggingface Causal LM to Prefix LM.
|
2 |
+
|
3 |
+
Conversion does lightweight surgery on a HuggingFace
|
4 |
+
Causal LM to convert it to a Prefix LM.
|
5 |
+
|
6 |
+
Prefix LMs accepts a `bidirectional_mask` input in `forward`
|
7 |
+
and treat the input prompt as the prefix in `generate`.
|
8 |
+
"""
|
9 |
+
from types import MethodType
|
10 |
+
from typing import Any, List, MutableMapping, Optional, Tuple, Union
|
11 |
+
import torch
|
12 |
+
from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel
|
13 |
+
from transformers.models.gpt_neo.modeling_gpt_neo import GPTNeoForCausalLM
|
14 |
+
from transformers.models.gpt_neox.modeling_gpt_neox import GPTNeoXForCausalLM
|
15 |
+
from transformers.models.gptj.modeling_gptj import GPTJForCausalLM
|
16 |
+
_SUPPORTED_GPT_MODELS = (GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM)
|
17 |
+
CAUSAL_GPT_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM]
|
18 |
+
|
19 |
+
def _convert_gpt_causal_lm_to_prefix_lm(model: CAUSAL_GPT_TYPES) -> CAUSAL_GPT_TYPES:
|
20 |
+
"""Converts a GPT-style Causal LM to a Prefix LM.
|
21 |
+
|
22 |
+
Supported HuggingFace model classes:
|
23 |
+
- `GPT2LMHeadModel`
|
24 |
+
- `GPTNeoForCausalLM`
|
25 |
+
- `GPTNeoXForCausalLM`
|
26 |
+
- `GPTJForCausalLM`
|
27 |
+
|
28 |
+
See `convert_hf_causal_lm_to_prefix_lm` for more details.
|
29 |
+
"""
|
30 |
+
if hasattr(model, '_prefix_lm_converted'):
|
31 |
+
return model
|
32 |
+
assert isinstance(model, _SUPPORTED_GPT_MODELS)
|
33 |
+
assert model.config.add_cross_attention == False, 'Only supports GPT-style decoder-only models'
|
34 |
+
|
35 |
+
def _get_attn_modules(model: CAUSAL_GPT_TYPES) -> List[torch.nn.Module]:
|
36 |
+
"""Helper that gets a list of the model's attention modules.
|
37 |
+
|
38 |
+
Each module has a `bias` buffer used for causal masking. The Prefix LM
|
39 |
+
conversion adds logic to dynamically manipulate these biases to support
|
40 |
+
Prefix LM attention masking.
|
41 |
+
"""
|
42 |
+
attn_modules = []
|
43 |
+
if isinstance(model, GPTNeoXForCausalLM):
|
44 |
+
blocks = model.gpt_neox.layers
|
45 |
+
else:
|
46 |
+
blocks = model.transformer.h
|
47 |
+
for block in blocks:
|
48 |
+
if isinstance(model, GPTNeoForCausalLM):
|
49 |
+
if block.attn.attention_type != 'global':
|
50 |
+
continue
|
51 |
+
attn_module = block.attn.attention
|
52 |
+
elif isinstance(model, GPTNeoXForCausalLM):
|
53 |
+
attn_module = block.attention
|
54 |
+
else:
|
55 |
+
attn_module = block.attn
|
56 |
+
attn_modules.append(attn_module)
|
57 |
+
return attn_modules
|
58 |
+
setattr(model, '_original_forward', getattr(model, 'forward'))
|
59 |
+
setattr(model, '_original_generate', getattr(model, 'generate'))
|
60 |
+
|
61 |
+
def forward(self: CAUSAL_GPT_TYPES, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]]=None, attention_mask: Optional[torch.FloatTensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, token_type_ids: Optional[torch.LongTensor]=None, position_ids: Optional[torch.LongTensor]=None, head_mask: Optional[torch.FloatTensor]=None, inputs_embeds: Optional[torch.FloatTensor]=None, labels: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None):
|
62 |
+
"""Wraps original forward to enable PrefixLM attention."""
|
63 |
+
|
64 |
+
def call_og_forward():
|
65 |
+
if isinstance(self, GPTNeoXForCausalLM):
|
66 |
+
return self._original_forward(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, head_mask=head_mask, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
|
67 |
+
else:
|
68 |
+
return self._original_forward(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
|
69 |
+
if bidirectional_mask is None:
|
70 |
+
return call_og_forward()
|
71 |
+
assert isinstance(bidirectional_mask, torch.Tensor)
|
72 |
+
attn_modules = _get_attn_modules(model)
|
73 |
+
(b, s) = bidirectional_mask.shape
|
74 |
+
max_length = attn_modules[0].bias.shape[-1]
|
75 |
+
if s > max_length:
|
76 |
+
raise ValueError(f'bidirectional_mask sequence length (={s}) exceeds the ' + f'max length allowed by the model ({max_length}).')
|
77 |
+
assert s <= max_length
|
78 |
+
if s < max_length:
|
79 |
+
pad = torch.zeros((int(b), int(max_length - s)), dtype=bidirectional_mask.dtype, device=bidirectional_mask.device)
|
80 |
+
bidirectional_mask = torch.cat([bidirectional_mask, pad], dim=1)
|
81 |
+
bidirectional = bidirectional_mask.unsqueeze(1).unsqueeze(1)
|
82 |
+
for attn_module in attn_modules:
|
83 |
+
assert isinstance(attn_module.bias, torch.Tensor)
|
84 |
+
attn_module.bias.data = torch.logical_or(attn_module.bias.data, bidirectional)
|
85 |
+
output = call_og_forward()
|
86 |
+
for attn_module in attn_modules:
|
87 |
+
attn_module.bias.data = torch.tril(attn_module.bias.data[0, 0])[None, None]
|
88 |
+
return output
|
89 |
+
|
90 |
+
def generate(self: CAUSAL_GPT_TYPES, *args: Any, **kwargs: Any):
|
91 |
+
"""Wraps original generate to enable PrefixLM attention."""
|
92 |
+
attn_modules = _get_attn_modules(model)
|
93 |
+
for attn_module in attn_modules:
|
94 |
+
attn_module.bias.data[:] = 1
|
95 |
+
output = self._original_generate(*args, **kwargs)
|
96 |
+
for attn_module in attn_modules:
|
97 |
+
attn_module.bias.data = torch.tril(attn_module.bias.data[0, 0])[None, None]
|
98 |
+
return output
|
99 |
+
setattr(model, 'forward', MethodType(forward, model))
|
100 |
+
setattr(model, 'generate', MethodType(generate, model))
|
101 |
+
setattr(model, '_prefix_lm_converted', True)
|
102 |
+
return model
|
103 |
+
_SUPPORTED_HF_MODELS = _SUPPORTED_GPT_MODELS
|
104 |
+
CAUSAL_LM_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM]
|
105 |
+
|
106 |
+
def convert_hf_causal_lm_to_prefix_lm(model: CAUSAL_LM_TYPES) -> CAUSAL_LM_TYPES:
|
107 |
+
"""Converts a HuggingFace Causal LM to a Prefix LM.
|
108 |
+
|
109 |
+
Supported HuggingFace model classes:
|
110 |
+
- `GPT2LMHeadModel`
|
111 |
+
- `GPTNeoForCausalLM`
|
112 |
+
- `GPTNeoXForCausalLM`
|
113 |
+
- `GPTJForCausalLM`
|
114 |
+
|
115 |
+
Conversion to a Prefix LM is done by modifying the `forward` method, and possibly also the
|
116 |
+
`generate` method and/or select underlying methods depending on the model class.
|
117 |
+
|
118 |
+
These changes preserve the model API, but add a new input to `forward`: "bidirectional_mask".
|
119 |
+
|
120 |
+
Notes on training:
|
121 |
+
To actually train the converted model as a Prefix LM, training batches will need to indicate
|
122 |
+
the prefix/target structure by including `bidirectional_mask` as part of the batch inputs.
|
123 |
+
|
124 |
+
**This is not a standard input and requires custom layers either within or after your dataloader.**
|
125 |
+
|
126 |
+
In addition to adding `bidirectional_mask` to the batch, this custom code should modify `labels`
|
127 |
+
such that `batch['labels'][batch['bidirectional_mask'] == 1] == -100`.
|
128 |
+
That is, the prefix portion of the sequence should not generate any loss. Loss should only be
|
129 |
+
generated by the target portion of the sequence.
|
130 |
+
|
131 |
+
Notes on `GPTNeoForCausalLM`:
|
132 |
+
To simplify the implementation, "global" and "local" attention layers are handled differently.
|
133 |
+
For "global" layers, we handle conversion as described above. For "local" layers, which use a
|
134 |
+
causal attention mask within a restricted local window, we do not alter the masking.
|
135 |
+
|
136 |
+
Notes on `forward` method conversion:
|
137 |
+
After conversion, the `forward` method will handle a new input, `bidirectional_mask`,
|
138 |
+
which should be a [batch_size, seq_length] byte tensor, where 1 indicates token positions
|
139 |
+
belonging to the prefix (prefix tokens can attend to one another bidirectionally), and
|
140 |
+
0 indicates token positions belonging to the target.
|
141 |
+
|
142 |
+
The new `forward` method will incorporate `bidirectional_mask` (if supplied) into the existing
|
143 |
+
causal mask, call the original `forward` method, and (if the causal mask is a buffer) reset
|
144 |
+
the causal masks before returning the result.
|
145 |
+
|
146 |
+
Notes on `generate` method conversion:
|
147 |
+
After conversion, the `generate` method will have the same signature but will internally
|
148 |
+
convert all causal masks to be purely bidirectional, call the original `generate` method, and
|
149 |
+
(where appropriate) reset the causal masks before returning the result.
|
150 |
+
|
151 |
+
This works thanks to the logic of the HuggingFace `generate` API, which first encodes the token
|
152 |
+
"prompt" passed to `generate` (which is treated as the prefix) and then sequentially generates
|
153 |
+
each new token. Encodings are cached as generation happens, so all prefix tokens can attend to one
|
154 |
+
another (as expected in a Prefix LM) and generated tokens can only attend to prefix tokens and
|
155 |
+
previously-generated tokens (also as expected in a Prefix LM).
|
156 |
+
|
157 |
+
To preserve the API, the original methods are renamed to `_original_forward` and
|
158 |
+
`_original_generate`, and replaced with new `forward` and `generate` methods that wrap
|
159 |
+
them, respectively. Although implementation details vary by model class.
|
160 |
+
"""
|
161 |
+
if isinstance(model, _SUPPORTED_GPT_MODELS):
|
162 |
+
return _convert_gpt_causal_lm_to_prefix_lm(model)
|
163 |
+
else:
|
164 |
+
raise TypeError(f'Cannot convert model to Prefix LM. ' + f'Model does not belong to set of supported HF models:' + f'\n{_SUPPORTED_HF_MODELS}')
|
165 |
+
|
166 |
+
def add_bidirectional_mask_if_missing(batch: MutableMapping):
|
167 |
+
"""Attempts to add bidirectional_mask to batch if missing.
|
168 |
+
|
169 |
+
Raises:
|
170 |
+
KeyError if bidirectional_mask is missing and can't be inferred
|
171 |
+
"""
|
172 |
+
if 'bidirectional_mask' not in batch:
|
173 |
+
if batch.get('mode', None) == 'icl_task':
|
174 |
+
batch['bidirectional_mask'] = batch['attention_mask'].clone()
|
175 |
+
for (i, continuation_indices) in enumerate(batch['continuation_indices']):
|
176 |
+
batch['bidirectional_mask'][i, continuation_indices] = 0
|
177 |
+
elif 'labels' in batch and 'attention_mask' in batch:
|
178 |
+
batch['bidirectional_mask'] = torch.logical_and(torch.eq(batch['attention_mask'], 1), torch.eq(batch['labels'], -100)).type_as(batch['attention_mask'])
|
179 |
+
else:
|
180 |
+
raise KeyError('No bidirectional_mask in batch and not sure how to construct one.')
|
norm.py
ADDED
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, List, Optional, Type, Union
|
2 |
+
import torch
|
3 |
+
|
4 |
+
def _cast_if_autocast_enabled(tensor: torch.Tensor) -> torch.Tensor:
|
5 |
+
if torch.is_autocast_enabled():
|
6 |
+
if tensor.device.type == 'cuda':
|
7 |
+
dtype = torch.get_autocast_gpu_dtype()
|
8 |
+
elif tensor.device.type == 'cpu':
|
9 |
+
dtype = torch.get_autocast_cpu_dtype()
|
10 |
+
else:
|
11 |
+
raise NotImplementedError()
|
12 |
+
return tensor.to(dtype=dtype)
|
13 |
+
return tensor
|
14 |
+
|
15 |
+
class LPLayerNorm(torch.nn.LayerNorm):
|
16 |
+
|
17 |
+
def __init__(self, normalized_shape: Union[int, List[int], torch.Size], eps: float=1e-05, elementwise_affine: bool=True, device: Optional[torch.device]=None, dtype: Optional[torch.dtype]=None):
|
18 |
+
super().__init__(normalized_shape=normalized_shape, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
|
19 |
+
|
20 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
21 |
+
module_device = x.device
|
22 |
+
downcast_x = _cast_if_autocast_enabled(x)
|
23 |
+
downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
|
24 |
+
downcast_bias = _cast_if_autocast_enabled(self.bias) if self.bias is not None else self.bias
|
25 |
+
with torch.autocast(enabled=False, device_type=module_device.type):
|
26 |
+
return torch.nn.functional.layer_norm(downcast_x, self.normalized_shape, downcast_weight, downcast_bias, self.eps)
|
27 |
+
|
28 |
+
def rms_norm(x: torch.Tensor, weight: Optional[torch.Tensor]=None, eps: float=1e-05) -> torch.Tensor:
|
29 |
+
output = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + eps)
|
30 |
+
if weight is not None:
|
31 |
+
return output * weight
|
32 |
+
return output
|
33 |
+
|
34 |
+
class RMSNorm(torch.nn.Module):
|
35 |
+
|
36 |
+
def __init__(self, normalized_shape: Union[int, List[int], torch.Size], eps: float=1e-05, weight: bool=True, dtype: Optional[torch.dtype]=None, device: Optional[torch.device]=None):
|
37 |
+
super().__init__()
|
38 |
+
self.eps = eps
|
39 |
+
if weight:
|
40 |
+
self.weight = torch.nn.Parameter(torch.ones(normalized_shape, dtype=dtype, device=device))
|
41 |
+
else:
|
42 |
+
self.register_parameter('weight', None)
|
43 |
+
|
44 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
45 |
+
return rms_norm(x.float(), self.weight, self.eps).to(dtype=x.dtype)
|
46 |
+
|
47 |
+
class LPRMSNorm(RMSNorm):
|
48 |
+
|
49 |
+
def __init__(self, normalized_shape: Union[int, List[int], torch.Size], eps: float=1e-05, weight: bool=True, dtype: Optional[torch.dtype]=None, device: Optional[torch.device]=None):
|
50 |
+
super().__init__(normalized_shape=normalized_shape, eps=eps, weight=weight, dtype=dtype, device=device)
|
51 |
+
|
52 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
53 |
+
downcast_x = _cast_if_autocast_enabled(x)
|
54 |
+
downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
|
55 |
+
with torch.autocast(enabled=False, device_type=x.device.type):
|
56 |
+
return rms_norm(downcast_x, downcast_weight, self.eps).to(dtype=x.dtype)
|
57 |
+
NORM_CLASS_REGISTRY: Dict[str, Type[torch.nn.Module]] = {'layernorm': torch.nn.LayerNorm, 'low_precision_layernorm': LPLayerNorm, 'rmsnorm': RMSNorm, 'low_precision_rmsnorm': LPRMSNorm}
|
param_init_fns.py
ADDED
@@ -0,0 +1,179 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import warnings
|
3 |
+
from collections.abc import Sequence
|
4 |
+
from functools import partial
|
5 |
+
from typing import Any, Callable, Optional, Tuple, Union
|
6 |
+
import torch
|
7 |
+
from torch import nn
|
8 |
+
from .fc import FC_CLASS_REGISTRY
|
9 |
+
from .norm import NORM_CLASS_REGISTRY
|
10 |
+
try:
|
11 |
+
import transformer_engine.pytorch as te
|
12 |
+
except:
|
13 |
+
te = None
|
14 |
+
|
15 |
+
def torch_default_param_init_fn_(module: nn.Module, **kwargs: Any) -> None:
|
16 |
+
del kwargs
|
17 |
+
if hasattr(module, 'reset_parameters') and isinstance(module.reset_parameters, Callable):
|
18 |
+
module.reset_parameters()
|
19 |
+
|
20 |
+
def fused_init_helper_(module: nn.Module, init_fn_: Callable) -> None:
|
21 |
+
_fused = getattr(module, '_fused', None)
|
22 |
+
if _fused is None:
|
23 |
+
raise RuntimeError(f'Internal logic error')
|
24 |
+
assert isinstance(module.weight, torch.Tensor)
|
25 |
+
(dim, splits) = _fused
|
26 |
+
splits = (0, *splits, module.weight.size(dim))
|
27 |
+
for (s, e) in zip(splits[:-1], splits[1:]):
|
28 |
+
slice_indices = [slice(None)] * module.weight.ndim
|
29 |
+
slice_indices[dim] = slice(s, e)
|
30 |
+
init_fn_(module.weight[slice_indices])
|
31 |
+
|
32 |
+
def generic_param_init_fn_(module: nn.Module, init_fn_: Callable, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, **kwargs: Any) -> None:
|
33 |
+
del kwargs
|
34 |
+
init_div_is_residual = init_div_is_residual
|
35 |
+
if init_div_is_residual is False:
|
36 |
+
div_is_residual = 1.0
|
37 |
+
elif init_div_is_residual is True:
|
38 |
+
div_is_residual = math.sqrt(2 * n_layers)
|
39 |
+
elif isinstance(init_div_is_residual, float) or isinstance(init_div_is_residual, int):
|
40 |
+
div_is_residual = init_div_is_residual
|
41 |
+
elif init_div_is_residual.isnumeric():
|
42 |
+
div_is_residual = float(init_div_is_residual)
|
43 |
+
else:
|
44 |
+
div_is_residual = 1.0
|
45 |
+
raise ValueError(f'Expected init_div_is_residual to be boolean or numeric, got {init_div_is_residual}')
|
46 |
+
if isinstance(module, tuple(set(FC_CLASS_REGISTRY.values()))):
|
47 |
+
if hasattr(module, '_fused'):
|
48 |
+
fused_init_helper_(module, init_fn_)
|
49 |
+
else:
|
50 |
+
init_fn_(module.weight)
|
51 |
+
if module.bias is not None:
|
52 |
+
assert isinstance(module.bias, torch.Tensor)
|
53 |
+
torch.nn.init.zeros_(module.bias)
|
54 |
+
if init_div_is_residual is not False and getattr(module, '_is_residual', False):
|
55 |
+
with torch.no_grad():
|
56 |
+
module.weight.div_(div_is_residual)
|
57 |
+
elif isinstance(module, nn.Embedding):
|
58 |
+
if emb_init_std is not None:
|
59 |
+
std = emb_init_std
|
60 |
+
if std == 0:
|
61 |
+
warnings.warn(f'Embedding layer initialized to 0.')
|
62 |
+
emb_init_fn_ = partial(torch.nn.init.normal_, mean=0.0, std=std)
|
63 |
+
elif emb_init_uniform_lim is not None:
|
64 |
+
lim = emb_init_uniform_lim
|
65 |
+
if isinstance(lim, Sequence):
|
66 |
+
if len(lim) > 2:
|
67 |
+
raise ValueError(f'Uniform init requires a min and a max limit. User input: {lim}.')
|
68 |
+
if lim[0] == lim[1]:
|
69 |
+
warnings.warn(f'Embedding layer initialized to {lim[0]}.')
|
70 |
+
else:
|
71 |
+
if lim == 0:
|
72 |
+
warnings.warn(f'Embedding layer initialized to 0.')
|
73 |
+
lim = [-lim, lim]
|
74 |
+
(a, b) = lim
|
75 |
+
emb_init_fn_ = partial(torch.nn.init.uniform_, a=a, b=b)
|
76 |
+
else:
|
77 |
+
emb_init_fn_ = init_fn_
|
78 |
+
emb_init_fn_(module.weight)
|
79 |
+
elif isinstance(module, tuple(set(NORM_CLASS_REGISTRY.values()))):
|
80 |
+
if hasattr(module, 'weight') and isinstance(module.weight, torch.Tensor):
|
81 |
+
torch.nn.init.ones_(module.weight)
|
82 |
+
if hasattr(module, 'bias') and isinstance(module.bias, torch.Tensor):
|
83 |
+
torch.nn.init.zeros_(module.bias)
|
84 |
+
elif isinstance(module, nn.MultiheadAttention):
|
85 |
+
if module._qkv_same_embed_dim:
|
86 |
+
assert module.in_proj_weight is not None
|
87 |
+
assert module.q_proj_weight is None and module.k_proj_weight is None and (module.v_proj_weight is None)
|
88 |
+
assert d_model is not None
|
89 |
+
_d = d_model
|
90 |
+
splits = (0, _d, 2 * _d, 3 * _d)
|
91 |
+
for (s, e) in zip(splits[:-1], splits[1:]):
|
92 |
+
init_fn_(module.in_proj_weight[s:e])
|
93 |
+
else:
|
94 |
+
assert module.q_proj_weight is not None and module.k_proj_weight is not None and (module.v_proj_weight is not None)
|
95 |
+
assert module.in_proj_weight is None
|
96 |
+
init_fn_(module.q_proj_weight)
|
97 |
+
init_fn_(module.k_proj_weight)
|
98 |
+
init_fn_(module.v_proj_weight)
|
99 |
+
if module.in_proj_bias is not None:
|
100 |
+
torch.nn.init.zeros_(module.in_proj_bias)
|
101 |
+
if module.bias_k is not None:
|
102 |
+
torch.nn.init.zeros_(module.bias_k)
|
103 |
+
if module.bias_v is not None:
|
104 |
+
torch.nn.init.zeros_(module.bias_v)
|
105 |
+
init_fn_(module.out_proj.weight)
|
106 |
+
if init_div_is_residual is not False and getattr(module.out_proj, '_is_residual', False):
|
107 |
+
with torch.no_grad():
|
108 |
+
module.out_proj.weight.div_(div_is_residual)
|
109 |
+
if module.out_proj.bias is not None:
|
110 |
+
torch.nn.init.zeros_(module.out_proj.bias)
|
111 |
+
elif te is not None and isinstance(module, te.LayerNormMLP):
|
112 |
+
if isinstance(module.layer_norm_weight, torch.Tensor):
|
113 |
+
torch.nn.init.ones_(module.layer_norm_weight)
|
114 |
+
if isinstance(module.layer_norm_bias, torch.Tensor):
|
115 |
+
torch.nn.init.zeros_(module.layer_norm_bias)
|
116 |
+
init_fn_(module.fc1_weight)
|
117 |
+
if module.fc1_bias is not None:
|
118 |
+
assert isinstance(module.fc1_bias, torch.Tensor)
|
119 |
+
torch.nn.init.zeros_(module.fc1_bias)
|
120 |
+
init_fn_(module.fc2_weight)
|
121 |
+
if module.fc2_bias is not None:
|
122 |
+
assert isinstance(module.fc2_bias, torch.Tensor)
|
123 |
+
torch.nn.init.zeros_(module.fc2_bias)
|
124 |
+
with torch.no_grad():
|
125 |
+
module.fc2_weight.div_(div_is_residual)
|
126 |
+
else:
|
127 |
+
for _ in module.parameters(recurse=False):
|
128 |
+
raise NotImplementedError(f'{module.__class__.__name__} parameters are not initialized by param_init_fn.')
|
129 |
+
|
130 |
+
def _normal_init_(std: float, mean: float=0.0) -> Callable:
|
131 |
+
return partial(torch.nn.init.normal_, mean=mean, std=std)
|
132 |
+
|
133 |
+
def _normal_param_init_fn_(module: nn.Module, std: float, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, **kwargs: Any) -> None:
|
134 |
+
del kwargs
|
135 |
+
init_fn_ = _normal_init_(std=std)
|
136 |
+
generic_param_init_fn_(module=module, init_fn_=init_fn_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim)
|
137 |
+
|
138 |
+
def baseline_param_init_fn_(module: nn.Module, init_std: Optional[float], n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, **kwargs: Any) -> None:
|
139 |
+
del kwargs
|
140 |
+
if init_std is None:
|
141 |
+
raise ValueError("You must set model.init_config['init_std'] to a float value to use the default initialization scheme.")
|
142 |
+
_normal_param_init_fn_(module=module, std=init_std, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim)
|
143 |
+
|
144 |
+
def small_param_init_fn_(module: nn.Module, n_layers: int, d_model: int, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, **kwargs: Any) -> None:
|
145 |
+
del kwargs
|
146 |
+
std = math.sqrt(2 / (5 * d_model))
|
147 |
+
_normal_param_init_fn_(module=module, std=std, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim)
|
148 |
+
|
149 |
+
def neox_param_init_fn_(module: nn.Module, n_layers: int, d_model: int, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, **kwargs: Any) -> None:
|
150 |
+
"""From section 2.3.1 of GPT-NeoX-20B:
|
151 |
+
|
152 |
+
An Open-Source AutoregressiveLanguage Model — Black et. al. (2022)
|
153 |
+
see https://github.com/EleutherAI/gpt-neox/blob/9610391ab319403cef079b438edd016a2443af54/megatron/model/init_functions.py#L151
|
154 |
+
and https://github.com/EleutherAI/gpt-neox/blob/main/megatron/model/transformer.py
|
155 |
+
"""
|
156 |
+
del kwargs
|
157 |
+
residual_div = n_layers / math.sqrt(10)
|
158 |
+
small_param_init_fn_(module=module, d_model=d_model, n_layers=n_layers, init_div_is_residual=residual_div, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim)
|
159 |
+
|
160 |
+
def kaiming_uniform_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, fan_mode: str='fan_in', init_nonlinearity: str='leaky_relu', **kwargs: Any) -> None:
|
161 |
+
del kwargs
|
162 |
+
kaiming_uniform_ = partial(nn.init.kaiming_uniform_, a=init_gain, mode=fan_mode, nonlinearity=init_nonlinearity)
|
163 |
+
generic_param_init_fn_(module=module, init_fn_=kaiming_uniform_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim)
|
164 |
+
|
165 |
+
def kaiming_normal_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, fan_mode: str='fan_in', init_nonlinearity: str='leaky_relu', **kwargs: Any) -> None:
|
166 |
+
del kwargs
|
167 |
+
kaiming_normal_ = partial(torch.nn.init.kaiming_normal_, a=init_gain, mode=fan_mode, nonlinearity=init_nonlinearity)
|
168 |
+
generic_param_init_fn_(module=module, init_fn_=kaiming_normal_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim)
|
169 |
+
|
170 |
+
def xavier_uniform_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, **kwargs: Any) -> None:
|
171 |
+
del kwargs
|
172 |
+
xavier_uniform_ = partial(torch.nn.init.xavier_uniform_, gain=init_gain)
|
173 |
+
generic_param_init_fn_(module=module, init_fn_=xavier_uniform_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim)
|
174 |
+
|
175 |
+
def xavier_normal_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, **kwargs: Any) -> None:
|
176 |
+
del kwargs
|
177 |
+
xavier_normal_ = partial(torch.nn.init.xavier_normal_, gain=init_gain)
|
178 |
+
generic_param_init_fn_(module=module, init_fn_=xavier_normal_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim)
|
179 |
+
MODEL_INIT_REGISTRY = {'default_': torch_default_param_init_fn_, 'baseline_': baseline_param_init_fn_, 'kaiming_uniform_': kaiming_uniform_param_init_fn_, 'kaiming_normal_': kaiming_normal_param_init_fn_, 'neox_init_': neox_param_init_fn_, 'small_init_': small_param_init_fn_, 'xavier_uniform_': xavier_uniform_param_init_fn_, 'xavier_normal_': xavier_normal_param_init_fn_}
|