Keep Half and BFloat16 sums in float in portable matmul, var, cumsum and pooling kernels, and fix portable cat by shoumikhin · Pull Request #24401 · pytorch/executorch

Sum Half and BFloat16 matmul products in float in portable kernels

shoumikhin

committed

a91745e

Sum Half and BFloat16 in float in portable var, var_mean, cumsum and avg_pool2d

shoumikhin

committed

6fb304b

Add the row bias of portable addmm in the same pass

shoumikhin

committed

2e0352f

Check empty inputs to the portable cat kernel the way PyTorch does

shoumikhin

committed

c8e8e02

Convert mixed input dtypes in the portable cat kernel

shoumikhin

committed

ecb4cd1

Keep copying after an empty input in the complex cat kernel

shoumikhin

committed

6c63c1b

Use one accumulator type in the portable sum and matmul kernels

shoumikhin

committed

bff5f20

Round every portable addmm bias, and alpha and beta, once

shoumikhin

committed

319b236

Use a wider block in the Half and BFloat16 cumsum loop

shoumikhin

committed

40d26f4

Keep vec_ops.h free of promote_types

shoumikhin

committed

814145b

Take vec_addmm's alpha and beta in its accumulator type

shoumikhin

committed

80fdb95

Test a portable addmm column bias with different values per row

shoumikhin

committed

96018f9

Run the mixed_linear dtype test for Half input and Half output

shoumikhin

committed

86d3b03

Test a BFloat16 avg_pool2d divisor that BFloat16 cannot hold

shoumikhin

committed

7565dfe

State the stack cost of the Half and BFloat16 cumsum block

shoumikhin

committed

e390afd

Ignore the addmm bias when beta is 0

shoumikhin

committed

badca00

Navigation Menu

Repository navigation

Keep Half and BFloat16 sums in float in portable matmul, var, cumsum and pooling kernels, and fix portable cat - #24401

Keep Half and BFloat16 sums in float in portable matmul, var, cumsum and pooling kernels, and fix portable cat#24401
shoumikhin merged 23 commits into
pytorch:mainfrom
shoumikhin:portable-bf16-accumulate

Commits on Nov 1, 2026

Sum Half and BFloat16 matmul products in float in portable kernels

Sum Half and BFloat16 in float in portable var, var_mean, cumsum and avg_pool2d

Add the row bias of portable addmm in the same pass

Check empty inputs to the portable cat kernel the way PyTorch does

Convert mixed input dtypes in the portable cat kernel

Keep copying after an empty input in the complex cat kernel

Use one accumulator type in the portable sum and matmul kernels

Round every portable addmm bias, and alpha and beta, once

Use a wider block in the Half and BFloat16 cumsum loop

Keep vec_ops.h free of promote_types

Take vec_addmm's alpha and beta in its accumulator type

Test a portable addmm column bias with different values per row

Run the mixed_linear dtype test for Half input and Half output

Test a BFloat16 avg_pool2d divisor that BFloat16 cannot hold

State the stack cost of the Half and BFloat16 cumsum block

Ignore the addmm bias when beta is 0

Commits on Nov 2, 2026

Keep the cumsum block out of the frame of other dtypes

Merge the two portable addmm BFloat16 tests

Test the dim check in get_cat_out_target_size

Multiply a Float mixed_linear input in float for a Double output

Check the float divisor in the BFloat16 var and var_mean tests

Keep the cumsum block only for dims with several trailing positions

Convert mixed-dtype var inputs to the output dtype before summing