Skip to content

Commit b9c38a2

Browse files
committed
Update on "[ET-VK][ez] Fix linear weight int4 test due to change in ATen API"
## Context Recently the ATen API for 4-bit quantized linear has changed, so our test must adapt to the change in API. Concretely, the changes in API were: * The `_for_cpu` suffix was added to the operator name * The `_convert_weight_to_int4pack_mm` operator now expects unpacked 4-bit weights instead of a packed scheme where 2 4-bit values are packed into a single 8-bit value. Differential Revision: [D68333687](https://our.internmc.facebook.com/intern/diff/D68333687/) [ghstack-poisoned]
1 parent c917568 commit b9c38a2

File tree

1 file changed

+0
-1
lines changed

1 file changed

+0
-1
lines changed

backends/vulkan/test/op_tests/linear_weight_int4_test.cpp

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -50,7 +50,6 @@ at::Tensor unpack_weights_4x2(const at::Tensor& weights_4x2) {
5050

5151
for (int n = 0; n < N; n++) {
5252
for (int k = 0; k < K; k += 2) {
53-
// const int scale_idx = k_groups * n + group_idx;
5453
const uint8_t packed_val = weights_4x2[n][k / 2].item().to<uint8_t>();
5554
const uint8_t second_val = packed_val & 0x0F;
5655
const uint8_t first_val = (packed_val & 0xF0) >> 4;

0 commit comments

Comments
 (0)