github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

[NFC] polish test_2p5d/checks_2p5d/check_operation_2p5d.py code style (#1718)

pull/1743/head
Kai Wang (Victor Kai) 2022-10-17 17:52:28 +08:00 committed by Frank Lee
parent ea961d8fd1
commit b38efe4e8a
1 changed files with 12 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

44
tests/test_layers/test_2p5d/checks_2p5d/check_operation_2p5d.py
View File

@ -39,16 +39,9 @@ def check_AB():
B.requires_grad = True B.requires_grad = True
out_shape = (BATCH_SIZE // TESSERACT_DIM, SEQ_LENGTH, 4 * HIDDEN_SIZE // TESSERACT_DIM) out_shape = (BATCH_SIZE // TESSERACT_DIM, SEQ_LENGTH, 4 * HIDDEN_SIZE // TESSERACT_DIM)
out = Matmul_AB_2p5D.apply( out = Matmul_AB_2p5D.apply(A, B, TESSERACT_DIM, out_shape, i, j, k, ParallelMode.PARALLEL_2P5D_ROW,
A, B, ParallelMode.PARALLEL_2P5D_COL, data_parallel_rank, pipeline_parallel_rank,
TESSERACT_DIM, out_shape, pipeline_parallel_size, tensor_parallel_size)
i, j, k,
ParallelMode.PARALLEL_2P5D_ROW,
ParallelMode.PARALLEL_2P5D_COL,
data_parallel_rank,
pipeline_parallel_rank,
pipeline_parallel_size,
tensor_parallel_size)
C_shape = (BATCH_SIZE, SEQ_LENGTH, 4 * HIDDEN_SIZE) C_shape = (BATCH_SIZE, SEQ_LENGTH, 4 * HIDDEN_SIZE)
A_master = A_master.clone() A_master = A_master.clone()
@ -116,16 +109,10 @@ def check_ABT():
B = B.clone() B = B.clone()
B.requires_grad = True B.requires_grad = True
out = Matmul_ABT_2p5D.apply( out = Matmul_ABT_2p5D.apply(C, B, TESSERACT_DIM,
C, B, (BATCH_SIZE // TESSERACT_DIM, SEQ_LENGTH, HIDDEN_SIZE // TESSERACT_DIM), i, j, k,
TESSERACT_DIM, (BATCH_SIZE // TESSERACT_DIM, SEQ_LENGTH, HIDDEN_SIZE // TESSERACT_DIM), ParallelMode.PARALLEL_2P5D_ROW, ParallelMode.PARALLEL_2P5D_COL, data_parallel_rank,
i, j, k, pipeline_parallel_rank, pipeline_parallel_size, tensor_parallel_size)
ParallelMode.PARALLEL_2P5D_ROW,
ParallelMode.PARALLEL_2P5D_COL,
data_parallel_rank,
pipeline_parallel_rank,
pipeline_parallel_size,
tensor_parallel_size)
A_shape = (BATCH_SIZE, SEQ_LENGTH, HIDDEN_SIZE) A_shape = (BATCH_SIZE, SEQ_LENGTH, HIDDEN_SIZE)
C_master = C_master.clone() C_master = C_master.clone()
@ -191,16 +178,10 @@ def check_ATB():
C = C.clone() C = C.clone()
C.requires_grad = True C.requires_grad = True
out = Matmul_ATB_2p5D.apply( out = Matmul_ATB_2p5D.apply(A, C, TESSERACT_DIM, (HIDDEN_SIZE // TESSERACT_DIM, 4 * HIDDEN_SIZE // TESSERACT_DIM),
A, C, i, j, k, ParallelMode.PARALLEL_2P5D_ROW, ParallelMode.PARALLEL_2P5D_COL,
TESSERACT_DIM, (HIDDEN_SIZE // TESSERACT_DIM, 4 * HIDDEN_SIZE // TESSERACT_DIM), data_parallel_rank, pipeline_parallel_rank, pipeline_parallel_size,
i, j, k, tensor_parallel_size)
ParallelMode.PARALLEL_2P5D_ROW,
ParallelMode.PARALLEL_2P5D_COL,
data_parallel_rank,
pipeline_parallel_rank,
pipeline_parallel_size,
tensor_parallel_size)
B_shape = (HIDDEN_SIZE, 4 * HIDDEN_SIZE) B_shape = (HIDDEN_SIZE, 4 * HIDDEN_SIZE)
A_master = A_master.clone() A_master = A_master.clone()
@ -208,8 +189,7 @@ def check_ATB():
C_master = C_master.clone() C_master = C_master.clone()
C_master.requires_grad = True C_master.requires_grad = True
B_master = torch.matmul( B_master = torch.matmul(
A_master.view(-1, A_master.shape[-1]).transpose(0, 1), A_master.view(-1, A_master.shape[-1]).transpose(0, 1), C_master.view(-1, C_master.shape[-1]))
C_master.view(-1, C_master.shape[-1]))
B = torch.chunk(B_master, TESSERACT_DIM, dim=0)[i] B = torch.chunk(B_master, TESSERACT_DIM, dim=0)[i]
B = torch.chunk(B, TESSERACT_DIM, dim=-1)[j] B = torch.chunk(B, TESSERACT_DIM, dim=-1)[j]
check_equal(out, B) check_equal(out, B)