File size: 5,647 Bytes
a402b9b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 | import unittest
import torch
from torch.nn.functional import scaled_dot_product_attention
from sglang.test.test_utils import CustomTestCase
torch.manual_seed(1234)
class TestDecodeAttention(CustomTestCase):
def _run_sdpa_forward_decode(
self,
query: torch.Tensor,
output: torch.Tensor,
k_cache: torch.Tensor,
v_cache: torch.Tensor,
req_to_token: torch.Tensor,
req_pool_indices: torch.Tensor,
seq_lens: torch.Tensor,
scaling=None,
enable_gqa=False,
causal=False,
):
# [num_tokens, num_heads, head_size] -> [num_heads, num_tokens, head_size]
query = query.movedim(0, query.dim() - 2)
start_q, start_kv = 0, 0
for seq_idx in range(seq_lens.shape[0]):
seq_len_q = 1
seq_len_kv = seq_lens[seq_idx]
end_q = start_q + seq_len_q
end_kv = start_kv + seq_len_kv
per_req_query = query[:, start_q:end_q, :]
# get key and value from cache. per_req_tokens contains the kv cache
# index for each token in the sequence.
req_pool_idx = req_pool_indices[seq_idx]
per_req_tokens = req_to_token[req_pool_idx, :seq_len_kv]
per_req_key = k_cache[per_req_tokens].movedim(0, query.dim() - 2)
per_req_value = v_cache[per_req_tokens].movedim(0, query.dim() - 2)
per_req_out = (
scaled_dot_product_attention(
per_req_query.unsqueeze(0),
per_req_key.unsqueeze(0),
per_req_value.unsqueeze(0),
enable_gqa=enable_gqa,
scale=scaling,
is_causal=causal,
)
.squeeze(0)
.movedim(query.dim() - 2, 0)
)
output[start_q:end_q, :, :] = per_req_out
start_q, start_kv = end_q, end_kv
return output
def _test_grouped_decode_attention_once(self, B, H_Q, H_KV, D, D_V, dtype, device):
# This represents the number of tokens already in the sequence
seq_len = 1024
total_tokens = B * seq_len
sm_scale = 1.0 / (D**0.5)
logit_cap = 0.0
num_kv_splits = 8
enable_gqa = H_Q != H_KV
# q represents the new token being generated, one per batch
q = torch.randn(B, H_Q, D, dtype=dtype, device=device)
# k_buffer and v_buffer represent all previous tokens
k_buffer = torch.randn(total_tokens, H_KV, D, dtype=dtype, device=device)
v_buffer = torch.randn(total_tokens, H_KV, D_V, dtype=dtype, device=device)
key = torch.randn(B, H_KV, D, dtype=dtype)
value = torch.randn(B, H_KV, D_V, dtype=dtype)
loc = torch.randint(0, 10, (B,)).to(torch.int64)
# set kv cache
k_buffer[loc] = key
v_buffer[loc] = value
# o will have the same shape as q
o = torch.zeros(B, H_Q, D_V, dtype=dtype, device=device)
o_grouped = torch.zeros(B, H_Q, D_V, dtype=dtype, device=device)
req_to_token = (
torch.arange(total_tokens, device=device)
.reshape(B, seq_len)
.to(torch.int32)
)
b_req_idx = torch.arange(B, device=device).to(torch.int64)
b_seq_len = torch.full((B,), seq_len, device=device).to(torch.int64)
attn_logits = torch.empty(
(B, H_Q, num_kv_splits, D_V + 1),
dtype=torch.float32,
device=device,
)
# k_buffer, v_buffer, query, key and value supports non-contiguous tensors
k_buffer = k_buffer.transpose(0, 1).contiguous().transpose(0, 1)
v_buffer = v_buffer.transpose(0, 1).contiguous().transpose(0, 1)
q = q.transpose(0, 1).contiguous().transpose(0, 1)
key = key.transpose(0, 1).contiguous().transpose(0, 1)
value = value.transpose(0, 1).contiguous().transpose(0, 1)
torch.ops.sgl_kernel.decode_attention_cpu(
q,
k_buffer,
v_buffer,
o,
key,
value,
loc,
attn_logits,
req_to_token,
b_req_idx,
b_seq_len,
sm_scale,
logit_cap,
)
self._run_sdpa_forward_decode(
q,
o_grouped,
k_buffer,
v_buffer,
req_to_token,
b_req_idx,
b_seq_len,
scaling=sm_scale,
enable_gqa=enable_gqa,
)
cos_sim = torch.nn.functional.cosine_similarity(
o.flatten(), o_grouped.flatten(), dim=0
)
self.assertGreater(cos_sim.item(), 0.99)
torch.testing.assert_close(o, o_grouped, atol=3e-2, rtol=1e-6)
def _test_grouped_decode_attention(self, device="cuda"):
configs = [
(2, 16, 16, 64, 64),
(2, 16, 1, 16, 16),
(2, 32, 8, 33, 55),
(2, 16, 1, 64, 64),
(2, 64, 1, 13, 13),
(2, 128, 1, 80, 80),
(2, 128, 2, 512, 512),
(1, 16, 1, 576, 512),
(1, 16, 16, 576, 512),
(1, 22, 1, 576, 512),
(1, 40, 8, 128, 128),
]
for B, H_Q, H_KV, D, D_V in configs:
for dtype in [torch.bfloat16, torch.float16]:
self._test_grouped_decode_attention_once(
B, H_Q, H_KV, D, D_V, dtype=dtype, device=device
)
def test_grouped_decode_attention(self):
self._test_grouped_decode_attention("cpu")
if __name__ == "__main__":
unittest.main()
|