Files
sglang/test/registered/cpu/test_qwen3.py
T

179 lines
6.1 KiB
Python

import sys
import pytest
import sgl_kernel # noqa: F401
import torch
from sglang.srt.utils import is_host_cpu_arm64
from sglang.test.ci.ci_register import register_cpu_ci
from sglang.test.cpu_test_utils import precision
register_cpu_ci(est_time=5, suite="stage-a-test-cpu-intel")
register_cpu_ci(est_time=10, suite="base-b-test-cpu-arm64")
torch.manual_seed(1234)
def fix_query_key_value_ordering_reshape_cat(
mixed_qkvz, mixed_ba, num_k_heads, num_v_heads, attn_tp_size, head_k_dim, head_v_dim
):
new_tensor_shape_qkvz = mixed_qkvz.size()[:-1] + (
num_k_heads // attn_tp_size,
(
head_k_dim
+ head_k_dim
+ (head_v_dim + head_v_dim) * num_v_heads // num_k_heads
),
)
new_tensor_shape_ba = mixed_ba.size()[:-1] + (
num_k_heads // attn_tp_size,
2 * num_v_heads // num_k_heads,
)
mixed_qkvz = mixed_qkvz.view(*new_tensor_shape_qkvz)
mixed_ba = mixed_ba.view(*new_tensor_shape_ba)
split_arg_list_qkvz = [
head_k_dim,
head_k_dim,
(num_v_heads // num_k_heads * head_v_dim),
(num_v_heads // num_k_heads * head_v_dim),
]
split_arg_list_ba = [
num_v_heads // num_k_heads,
num_v_heads // num_k_heads,
]
# [b, sq, ng, (hn + hn + np/ng * hn + np/ng + np/ng)]
# --> [b, sq, ng, hn], [b, sq, ng, hn], [b, sq, ng, np/ng * hn], [b, sq, ng, np/ng * hn], [b, sq, ng, np/ng], [b, sq, ng, np/ng]
query, key, value, z = torch.split(mixed_qkvz, split_arg_list_qkvz, dim=2)
b, a = torch.split(mixed_ba, split_arg_list_ba, dim=2)
# [b, sq, ng, np/ng * hn] -> [b, sq, np, hn]
value = value.reshape(value.size(0), -1, head_v_dim)
z = z.reshape(z.size(0), -1, head_v_dim)
b = b.reshape(b.size(0), num_v_heads // attn_tp_size)
a = a.reshape(a.size(0), num_v_heads // attn_tp_size)
query, key, value = map(lambda x: x.reshape(x.shape[0], -1), (query, key, value))
mixed_qkv = torch.cat((query, key, value), dim=-1)
return mixed_qkv, z, b, a
def fix_query_key_value_ordering_reshape_cat_contiguous(
mixed_qkvz: torch.Tensor,
mixed_ba: torch.Tensor,
key_dim: int,
value_dim: int,
num_v_heads: int,
head_v_dim: int,
attn_tp_size: int,
):
"""
Derives `query`, `key` and `value` tensors from `mixed_qkvzba`.
"""
k_tp = key_dim // attn_tp_size
v_tp = value_dim // attn_tp_size
nv_tp = num_v_heads // attn_tp_size
# Directly split, no head group reshape
query, key, value, z = mixed_qkvz.split([k_tp, k_tp, v_tp, v_tp], dim=-1)
b, a = mixed_ba.split([nv_tp, nv_tp], dim=-1)
# value / z reshape to (seq, num_v_heads/tp, head_v_dim)
value = value.reshape(value.size(0), -1, head_v_dim)
z = z.reshape(z.size(0), -1, head_v_dim)
query, key, value = map(lambda x: x.reshape(x.shape[0], -1), (query, key, value))
mixed_qkv = torch.cat((query, key, value), dim=-1)
return mixed_qkv, z, b, a
@pytest.mark.skipif(
is_host_cpu_arm64(), reason="fused_input_proj_cpu is an x86 AMX kernel"
)
def test_fused_input_proj():
batch = 7
hidden_size = 256
qkvz_dim = 128
ba_dim = 64
hidden_states = torch.randn(batch, hidden_size, dtype=torch.bfloat16)
qkvz_weight = torch.randn(qkvz_dim, hidden_size, dtype=torch.bfloat16)
ba_weight = torch.randn(ba_dim, hidden_size, dtype=torch.bfloat16)
qkvz_ref = torch.nn.functional.linear(hidden_states, qkvz_weight)
ba_ref = torch.nn.functional.linear(hidden_states, ba_weight)
qkvz, ba = torch.ops.sgl_kernel.fused_input_proj_cpu(
hidden_states, qkvz_weight, ba_weight, False
)
atol = rtol = precision[qkvz.dtype]
torch.testing.assert_close(qkvz, qkvz_ref, atol=atol, rtol=rtol)
torch.testing.assert_close(ba, ba_ref, atol=atol, rtol=rtol)
def test_fused_qkvzba_split_reshape_cat():
mixed_qkvz = torch.rand(1024, 12288, dtype=torch.bfloat16)
mixed_ba = torch.rand(1024, 64, dtype=torch.bfloat16)
head_k_dim = 128
head_v_dim = 128
num_v_heads = 32
num_k_heads = 16
attn_tp_size = 1
mixed_qkv_ref, z_ref, b_ref, a_ref = fix_query_key_value_ordering_reshape_cat(
mixed_qkvz,
mixed_ba,
num_k_heads,
num_v_heads,
attn_tp_size,
head_k_dim,
head_v_dim,
)
num_heads_qk = num_k_heads // attn_tp_size
num_heads_v = num_v_heads // attn_tp_size
mixed_qkv, z, b, a = torch.ops.sgl_kernel.fused_qkvzba_split_reshape_cat_cpu(
mixed_qkvz, mixed_ba, num_heads_qk, num_heads_v, head_k_dim, head_v_dim
)
atol = rtol = precision[mixed_qkv.dtype]
torch.testing.assert_close(mixed_qkv, mixed_qkv_ref, atol=atol, rtol=rtol)
torch.testing.assert_close(z, z_ref, atol=atol, rtol=rtol)
torch.testing.assert_close(b, b_ref, atol=atol, rtol=rtol)
torch.testing.assert_close(a, a_ref, atol=atol, rtol=rtol)
def test_fused_qkvzba_split_reshape_cat_contiguous():
mixed_qkvz = torch.rand(1, 12288, dtype=torch.bfloat16)
mixed_ba = torch.rand(1, 64, dtype=torch.bfloat16)
head_k_dim = 128
head_v_dim = 128
num_v_heads = 32
num_k_heads = 16
attn_tp_size = 1
key_dim = head_k_dim * num_k_heads
value_dim = head_v_dim * num_v_heads
mixed_qkv_ref, z_ref, b_ref, a_ref = (
fix_query_key_value_ordering_reshape_cat_contiguous(
mixed_qkvz,
mixed_ba,
key_dim,
value_dim,
num_v_heads,
head_v_dim,
attn_tp_size,
)
)
num_heads_qk = num_k_heads // attn_tp_size
num_heads_v = num_v_heads // attn_tp_size
mixed_qkv, z, b, a = (
torch.ops.sgl_kernel.fused_qkvzba_split_reshape_cat_contiguous_cpu(
mixed_qkvz, mixed_ba, num_heads_qk, num_heads_v, head_k_dim, head_v_dim
)
)
atol = rtol = precision[mixed_qkv.dtype]
torch.testing.assert_close(mixed_qkv, mixed_qkv_ref, atol=atol, rtol=rtol)
torch.testing.assert_close(z, z_ref, atol=atol, rtol=rtol)
torch.testing.assert_close(b, b_ref, atol=atol, rtol=rtol)
torch.testing.assert_close(a, a_ref, atol=atol, rtol=rtol)
if __name__ == "__main__":
sys.exit(pytest.main([__file__]))