apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

【Inference Optimize】DeepSeek-V3-model MLA Optimize (#3886)

Browse Source 
* support MLA chunk_size auto search & cuda_graph
This commit is contained in:
AIbin
2025-09-11 10:46:09 +08:00
committed by GitHub
parent 637d96c6ae
commit a7392a0ff9
23 changed files with 375 additions and 310 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/operators/test_tree_mask.py
View File

@@ -192,7 +192,7 @@ class TestTreeMask(unittest.TestCase):
decoder_block_shape_q = 16
group_size = self.num_q_head // self.num_kv_head
decode_max_tile_size = (
self.bsz * (decoder_step_token_num * group_size + decoder_block_shape_q - 1) / decoder_block_shape_q
1024 * self.bsz * (decoder_step_token_num * group_size + decoder_block_shape_q - 1) / decoder_block_shape_q
)
encode_max_tile_size = (
self.bsz * (self.max_seq_len * group_size + encoder_block_shape_q - 1) / encoder_block_shape_q
@@ -202,6 +202,8 @@ class TestTreeMask(unittest.TestCase):
decoder_batch_ids = paddle.full([int(decode_max_tile_size)], 0, dtype="int32")
decoder_tile_ids_per_batch = paddle.full([int(decode_max_tile_size)], 0, dtype="int32")
decoder_num_blocks = paddle.full([1], 0, dtype="int32").pin_memory()
decoder_num_blocks_device = paddle.full([1], 0, dtype="int32")
decoder_chunk_size_device = paddle.full([1], 64, dtype="int32")
max_len_tensor_cpu = paddle.full([8], 0, dtype="int32").cpu()
encoder_batch_ids = paddle.full([int(encode_max_tile_size)], 0, dtype="int32")
encoder_tile_ids_per_batch = paddle.full([int(encode_max_tile_size)], 0, dtype="int32")
@@ -222,6 +224,8 @@ class TestTreeMask(unittest.TestCase):
decoder_batch_ids,
decoder_tile_ids_per_batch,
decoder_num_blocks,
decoder_num_blocks_device,
decoder_chunk_size_device,
max_len_tensor_cpu,
encoder_batch_ids,
encoder_tile_ids_per_batch,