drbh
commited on
Commit
Β·
0518250
1
Parent(s):
876ac68
fix: bump readme
Browse files
README.md
CHANGED
@@ -6,18 +6,21 @@ tags:
|
|
6 |
|
7 |
<!--  -->
|
8 |
|
9 |
-
> [!WARNING]
|
10 |
-
> The latest build b58ed97 may contain an accuracy issue, which is currently being addressed. Please use with caution, and be aware that corrected outputs will be available soon.
|
11 |
-
|
12 |
# Flash Attention
|
13 |
|
14 |
Flash Attention is a fast and memory-efficient implementation of the attention mechanism, designed to work with large models and long sequences. This is a Hugging Face compliant kernel build of Flash Attention.
|
15 |
|
16 |
Original code here [https://github.com/Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention).
|
17 |
|
|
|
|
|
18 |
```python
|
19 |
# /// script
|
20 |
-
# dependencies = [
|
|
|
|
|
|
|
|
|
21 |
# ///
|
22 |
import torch
|
23 |
from kernels import get_kernel
|
@@ -27,23 +30,87 @@ torch.manual_seed(42)
|
|
27 |
flash_attn = get_kernel("kernels-community/flash-attn")
|
28 |
device = torch.device("cuda")
|
29 |
|
30 |
-
# Show available functions
|
31 |
print("Flash Attention functions:", [i for i in dir(flash_attn) if i.startswith("mha")])
|
32 |
|
33 |
-
#
|
34 |
-
print("\n1. Standard attention:")
|
35 |
B, S, H, D = 2, 5, 4, 8 # batch, seq_len, heads, head_dim
|
36 |
q = k = v = torch.randn(B, S, H, D, device=device, dtype=torch.float16)
|
37 |
-
out = flash_attn.mha_fwd(q=q, k=k, v=v, is_causal=False)[0]
|
38 |
-
print(f"Output: {out.shape}")
|
39 |
|
40 |
-
#
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
q_var = torch.randn(10, H, D, device=device, dtype=torch.float16) # total_q=10
|
43 |
k_var = v_var = torch.randn(12, H, D, device=device, dtype=torch.float16) # total_k=12
|
44 |
-
|
45 |
-
cu_q = torch.tensor([0, 3, 7, 10], device=device, dtype=torch.int32)
|
46 |
cu_k = torch.tensor([0, 4, 9, 12], device=device, dtype=torch.int32)
|
|
|
|
|
|
|
47 |
out_var = flash_attn.mha_varlen_fwd(
|
48 |
q=q_var,
|
49 |
k=k_var,
|
@@ -52,42 +119,37 @@ out_var = flash_attn.mha_varlen_fwd(
|
|
52 |
cu_seqlens_k=cu_k,
|
53 |
max_seqlen_q=4,
|
54 |
max_seqlen_k=5,
|
|
|
55 |
)[0]
|
56 |
-
print(f"
|
57 |
-
|
58 |
-
|
59 |
-
print("\n3. KV-cache:")
|
60 |
-
cache_len, new_len = 10, 2
|
61 |
-
kcache = vcache = torch.randn(B, cache_len, H, D, device=device, dtype=torch.float16)
|
62 |
-
q_new = k_new = v_new = torch.randn(
|
63 |
-
B, new_len, H, D, device=device, dtype=torch.float16
|
64 |
-
)
|
65 |
-
seqlens = torch.full((B,), cache_len + new_len, device=device, dtype=torch.int32)
|
66 |
-
out_kv = flash_attn.mha_fwd_kvcache(
|
67 |
-
q=q_new,
|
68 |
-
kcache=kcache,
|
69 |
-
vcache=vcache,
|
70 |
-
k=k_new,
|
71 |
-
v=v_new,
|
72 |
-
seqlens_k=seqlens,
|
73 |
-
is_causal=True,
|
74 |
-
)[0]
|
75 |
-
print(f"Output: {out_kv.shape}")
|
76 |
```
|
77 |
|
78 |
-
|
|
|
|
|
|
|
|
|
79 |
|
80 |
```txt
|
81 |
-
|
|
|
82 |
Flash Attention functions: ['mha_bwd', 'mha_fwd', 'mha_fwd_kvcache', 'mha_varlen_bwd', 'mha_varlen_fwd']
|
83 |
|
84 |
1. Standard attention:
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
92 |
```
|
93 |
|
|
|
6 |
|
7 |
<!--  -->
|
8 |
|
|
|
|
|
|
|
9 |
# Flash Attention
|
10 |
|
11 |
Flash Attention is a fast and memory-efficient implementation of the attention mechanism, designed to work with large models and long sequences. This is a Hugging Face compliant kernel build of Flash Attention.
|
12 |
|
13 |
Original code here [https://github.com/Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention).
|
14 |
|
15 |
+
|
16 |
+
[`scripts/readme_example.py`](scripts/readme_example.py) provides a simple example of how to use the Flash Attention kernel in PyTorch. It demonstrates standard attention, causal attention, and variable-length sequences.
|
17 |
```python
|
18 |
# /// script
|
19 |
+
# dependencies = [
|
20 |
+
# "numpy",
|
21 |
+
# "torch",
|
22 |
+
# "kernels"
|
23 |
+
# ]
|
24 |
# ///
|
25 |
import torch
|
26 |
from kernels import get_kernel
|
|
|
30 |
flash_attn = get_kernel("kernels-community/flash-attn")
|
31 |
device = torch.device("cuda")
|
32 |
|
|
|
33 |
print("Flash Attention functions:", [i for i in dir(flash_attn) if i.startswith("mha")])
|
34 |
|
35 |
+
# Create test tensors
|
|
|
36 |
B, S, H, D = 2, 5, 4, 8 # batch, seq_len, heads, head_dim
|
37 |
q = k = v = torch.randn(B, S, H, D, device=device, dtype=torch.float16)
|
|
|
|
|
38 |
|
39 |
+
# Reference implementation using PyTorch SDPA
|
40 |
+
def reference_attention(query, key, value, causal=False):
|
41 |
+
query, key, value = (x.transpose(1, 2).contiguous() for x in (query, key, value))
|
42 |
+
with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
|
43 |
+
out = torch.nn.functional.scaled_dot_product_attention(query, key, value, is_causal=causal)
|
44 |
+
return out.transpose(1, 2).contiguous()
|
45 |
+
|
46 |
+
# 1. Standard attention
|
47 |
+
print("\n1. Standard attention:")
|
48 |
+
out_ref = reference_attention(q, k, v)
|
49 |
+
out_flash = flash_attn.mha_fwd(
|
50 |
+
q=q,
|
51 |
+
k=k,
|
52 |
+
v=v,
|
53 |
+
is_causal=False,
|
54 |
+
softmax_scale=1.0 / (D ** 0.5), # scale factor
|
55 |
+
)[0]
|
56 |
+
print(f"Reference output: {out_ref.shape}")
|
57 |
+
print(f"Flash output: {out_flash.shape}")
|
58 |
+
print(f"Outputs close: {torch.allclose(out_flash, out_ref, atol=1e-2, rtol=1e-3)}")
|
59 |
+
|
60 |
+
# 2. Causal attention (for autoregressive models)
|
61 |
+
print("\n2. Causal attention:")
|
62 |
+
|
63 |
+
out_ref_causal = reference_attention(q, k, v, causal=True)
|
64 |
+
out_causal = flash_attn.mha_fwd(
|
65 |
+
q=q,
|
66 |
+
k=k,
|
67 |
+
v=v,
|
68 |
+
is_causal=True,
|
69 |
+
softmax_scale=1.0 / (D ** 0.5), # scale factor
|
70 |
+
)[0]
|
71 |
+
print(f"Reference causal output: {out_ref_causal.shape}")
|
72 |
+
print(f"Flash causal output: {out_causal.shape}")
|
73 |
+
print(f"Outputs close: {torch.allclose(out_causal, out_ref_causal, atol=1e-2, rtol=1e-3)}")
|
74 |
+
|
75 |
+
def var_reference_attention(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, causal=False):
|
76 |
+
batch_size = cu_seqlens_q.shape[0] - 1
|
77 |
+
# Return output in packed format
|
78 |
+
total_tokens_q = q.shape[0]
|
79 |
+
out = torch.zeros((total_tokens_q, q.shape[1], q.shape[2]), device=q.device, dtype=q.dtype)
|
80 |
+
|
81 |
+
for b in range(batch_size):
|
82 |
+
start_q, end_q = cu_seqlens_q[b], cu_seqlens_q[b + 1]
|
83 |
+
start_k, end_k = cu_seqlens_k[b], cu_seqlens_k[b + 1]
|
84 |
+
|
85 |
+
# Extract slices for this batch
|
86 |
+
q_slice = q[start_q:end_q] # Shape: (seq_len_q, H, D)
|
87 |
+
k_slice = k[start_k:end_k] # Shape: (seq_len_k, H, D)
|
88 |
+
v_slice = v[start_k:end_k] # Shape: (seq_len_k, H, D)
|
89 |
+
|
90 |
+
# Add batch dimension for reference_attention
|
91 |
+
q_slice = q_slice.unsqueeze(0) # Shape: (1, seq_len_q, H, D)
|
92 |
+
k_slice = k_slice.unsqueeze(0) # Shape: (1, seq_len_k, H, D)
|
93 |
+
v_slice = v_slice.unsqueeze(0) # Shape: (1, seq_len_k, H, D)
|
94 |
+
|
95 |
+
# Compute attention and remove batch dimension
|
96 |
+
attn_out = reference_attention(q_slice, k_slice, v_slice, causal=causal)
|
97 |
+
attn_out = attn_out.squeeze(0) # Shape: (seq_len_q, H, D)
|
98 |
+
|
99 |
+
# Place result in output tensor (packed format)
|
100 |
+
out[start_q:end_q] = attn_out
|
101 |
+
|
102 |
+
return out
|
103 |
+
|
104 |
+
# 3. Variable length sequences (packed format)
|
105 |
+
print("\n3. Variable length sequences:")
|
106 |
+
# Pack sequences of lengths [3,4,3] for q and [4,5,3] for k into single tensors
|
107 |
q_var = torch.randn(10, H, D, device=device, dtype=torch.float16) # total_q=10
|
108 |
k_var = v_var = torch.randn(12, H, D, device=device, dtype=torch.float16) # total_k=12
|
109 |
+
cu_q = torch.tensor([0, 3, 7, 10], device=device, dtype=torch.int32) # cumulative sequence lengths
|
|
|
110 |
cu_k = torch.tensor([0, 4, 9, 12], device=device, dtype=torch.int32)
|
111 |
+
|
112 |
+
out_var_ref = var_reference_attention(q_var, k_var, v_var, cu_q, cu_k, max_seqlen_q=4, max_seqlen_k=5, causal=False)
|
113 |
+
# Custom function to handle variable
|
114 |
out_var = flash_attn.mha_varlen_fwd(
|
115 |
q=q_var,
|
116 |
k=k_var,
|
|
|
119 |
cu_seqlens_k=cu_k,
|
120 |
max_seqlen_q=4,
|
121 |
max_seqlen_k=5,
|
122 |
+
softmax_scale=1.0 / (D ** 0.5), # scale factor
|
123 |
)[0]
|
124 |
+
print(f"Variable length output: {out_var.shape}")
|
125 |
+
print(f"Reference variable length output: {out_var_ref.shape}")
|
126 |
+
print(f"Outputs close: {torch.allclose(out_var, out_var_ref, atol=1e-2, rtol=1e-3)}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
127 |
```
|
128 |
|
129 |
+
run it using the following command:
|
130 |
+
|
131 |
+
```bash
|
132 |
+
uv run scripts/readme_example.py
|
133 |
+
```
|
134 |
|
135 |
```txt
|
136 |
+
Reading inline script metadata from `flash-attn/scripts/readme_example.py`
|
137 |
+
Fetching 4 files: 100%|βββββββββββββββββββββββββββββββββββββββββ| 4/4 [00:00<00:00, 33354.31it/s]
|
138 |
Flash Attention functions: ['mha_bwd', 'mha_fwd', 'mha_fwd_kvcache', 'mha_varlen_bwd', 'mha_varlen_fwd']
|
139 |
|
140 |
1. Standard attention:
|
141 |
+
Reference output: torch.Size([2, 5, 4, 8])
|
142 |
+
Flash output: torch.Size([2, 5, 4, 8])
|
143 |
+
Outputs close: True
|
144 |
+
|
145 |
+
1. Causal attention:
|
146 |
+
Reference causal output: torch.Size([2, 5, 4, 8])
|
147 |
+
Flash causal output: torch.Size([2, 5, 4, 8])
|
148 |
+
Outputs close: True
|
149 |
+
|
150 |
+
1. Variable length sequences:
|
151 |
+
Variable length output: torch.Size([10, 4, 8])
|
152 |
+
Reference variable length output: torch.Size([10, 4, 8])
|
153 |
+
Outputs close: True
|
154 |
```
|
155 |
|