apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-09-27 04:46:16 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

fix dynamic Cfp8 computing error (#4119)

Browse Source 
Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com>
This commit is contained in:
Yuan Xiaolan
2025-09-16 20:21:49 +08:00
committed by GitHub
parent 4f8901489c
commit de8638b1e9
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_impl.cuh
View File

@@ -1049,7 +1049,7 @@ __global__ void append_decode_cache_int8_rope_qk_norm_kernel(
local_max = __hmax(local_max, __habs(out_vec2[i]));
}
#pragma unroll
for (int m_offset = 16; m_offset > 1; m_offset /= 2) {
for (int m_offset = 16; m_offset > 0; m_offset /= 2) {
local_max = __hmax(local_max, __shfl_xor_sync(0xffffffff, local_max, m_offset));
}

4
tests/layers/test_append_attention.py
View File

@@ -379,7 +379,7 @@ class TestAppendGroupQueryAttnWithRope(unittest.TestCase):
)
self.max_enc_len_this_time = paddle.to_tensor([self.max_enc_len_this_time], "int32", place=paddle.CPUPlace())
self.max_dec_len_this_time = paddle.to_tensor([self.max_dec_len_this_time], "int32", place=paddle.CPUPlace())
self.seq_lens_this_time = self.seq_lens_encoder
self.seq_lens_this_time = copy.deepcopy(self.seq_lens_encoder)
decode_max_tile_size = 1024 * self.batch_size * np.ceil((2 * 10) / 12)
self.decoder_batch_ids = paddle.full([int(decode_max_tile_size)], 0, dtype="int32")
self.decoder_tile_ids_per_batch = paddle.full([int(decode_max_tile_size)], 0, dtype="int32")
@@ -652,7 +652,7 @@ class TestAppendGroupQueryAttnWithRope(unittest.TestCase):
)
# encoder
# self.seq_lens_encoder,self.seq_lens_decoder,self.max_enc_len_this_time,self.max_dec_len_this_time=get_encoder_decoder_len(self.batch_size,self.seq_len)
self.seq_lens_this_time = self.seq_lens_encoder
self.seq_lens_this_time = copy.deepcopy(self.seq_lens_encoder)
if self.use_mask_offset:
print("encoder mask_offset: ", self.mask_offset)
self.cmp_append_attention(attn_mask=self.attention_mask)