Skip to content

[CUDA] Correct after_gather_dim for nibbled uint4 index #26484

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Nov 6, 2025
Merged

[CUDA] Correct after_gather_dim for nibbled uint4 index #26484

Changes from 1 commit
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
76d8632
Matched after_gather_dim in gather_block_quantized to correct CUDA ke…
jixiongdeng Oct 30, 2025
4092a09
Added general u8/i4x2 tests from @xiaomsft
jixiongdeng Nov 6, 2025
e37e5b9
Added UINT4X2 cases in Test_GatherAxis_WithZeroPoints_NoPading
jixiongdeng Nov 6, 2025
d9d8a29
Added test for uint8_t with two packed u4.
jixiongdeng Nov 6, 2025
12a8bb2
Added testcases that mimic usecase for shared emb_tokens/lm_head & Up…
jixiongdeng Nov 6, 2025
7aa39a2
Limited testcases to CUDA backend
jixiongdeng Nov 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Added testcases that mimic usecase for shared emb_tokens/lm_head & Up… 
…dated comments
  • Loading branch information
@jixiongdeng
jixiongdeng committed Nov 6, 2025
commit 12a8bb2eb528c194dca357493d3f9f29b9af16bc
78 changes: 72 additions & 6 deletions onnxruntime/test/contrib_ops/gather_block_quantized_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -573,6 +573,78 @@ TEST(GatherBlockQuantizedOpTest, GatherAxis0NoZeroPoints_8Bits) {
}
#endif

template <typename T1, typename T2, typename Tind>
void Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_4Bits() {
// This test case specific to shared 4bit token_embedding/lm_head use case on CUDA
std::vector<int> data = {-8, -7, -6, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7,
0, 1, 2, 3, 4, 5, 6, 7, -8, -7, -6, -5, -4, -3, -2, -1};
std::vector<int64_t> data_shape = {2, 16};
std::vector<int> indices = {1};
std::vector<int64_t> indices_shape = {1};
std::vector<float> scales = {2.0f, 1.0f};
std::vector<int64_t> scales_shape = {2, 1};
// Explicit zero points for each row
std::vector<int> zero_points = {-2, 1};

// With explicit zero points:
// Unpacked data (row 1): [0, 1, 2, 3, 4, 5, 6, 7, -8, -7, -6, -5, -4, -3, -2, -1] ---add offset 8--->
// Packed (add offset 8): [8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7]
// Gathered scales (row 1): scale = 1.0f, zero_point (row 1): packed: [1] ---add offset 8---> unpacked: [9]
// Expected (CUDA doesn't subtract zero point): [8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7]
std::vector<float> output = {8.f, 9.f, 10.f, 11.f, 12.f, 13.f, 14.f, 15.f, 0.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f};
std::vector<int64_t> output_shape = {1, 16};

constexpr int64_t gather_axis = 0;
constexpr int64_t quantize_axis = 1; // Last axis (required for CUDA)
constexpr int64_t block_size = 16;
constexpr int64_t bits = 4;
RunUnpackedData<T1, T2, Tind>(data, data_shape, indices, indices_shape, scales, scales_shape, zero_points,
gather_axis, quantize_axis, block_size, bits, output, output_shape, true);
}

template <typename T1, typename T2, typename Tind>
void Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_8Bits() {
// This test case specific to shared 8bit token_embedding/lm_head use case on CUDA
std::vector<int> data = {-128, -127, -126, -125, -124, -123, -122, -121, -120, -119, -118, -117, -116, -115, -114, -113,
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
std::vector<int64_t> data_shape = {2, 16};
std::vector<int> indices = {1};
std::vector<int64_t> indices_shape = {1};
std::vector<float> scales = {1.0f, 2.0f};
std::vector<int64_t> scales_shape = {2, 1};
// Explicit zero points
std::vector<int> zero_points = {10, -5};

// With explicit zero points:
// Unpacked data (row 1): [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] ---add offset 128--->
// Packed (row1): [128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143]
// Zero point unpacked: [-5] ---add offset 128---> packed: [123]
// Dequantization: [(128-123)*2, (129-123)*2, ..., (143-123)*2] = [10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32, 34, 36, 38, 40]
std::vector<float> output = {10.f, 12.f, 14.f, 16.f, 18.f, 20.f, 22.f, 24.f, 26.f, 28.f, 30.f, 32.f, 34.f, 36.f, 38.f, 40.f};
std::vector<int64_t> output_shape = {1, 16};

constexpr int64_t gather_axis = 0;
constexpr int64_t quantize_axis = 1;
constexpr int64_t block_size = 16;
constexpr int64_t bits = 8;
RunUnpackedData<T1, T2, Tind>(data, data_shape, indices, indices_shape, scales, scales_shape, zero_points,
gather_axis, quantize_axis, block_size, bits, output, output_shape, true);
}

TEST(GatherBlockQuantizedOpTest, GatherAxis0_QuantizedAxis1_Uint8_4Bits_WithZeroPoints) {
Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_4Bits<uint8_t, float, int32_t>();
Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_4Bits<uint8_t, MLFloat16, int32_t>();
Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_4Bits<uint8_t, float, int64_t>();
Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_4Bits<uint8_t, MLFloat16, int64_t>();
}

TEST(GatherBlockQuantizedOpTest, GatherAxis0_QuantizedAxis1_Uint8_8Bits_WithZeroPoints) {
Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_8Bits<uint8_t, float, int32_t>();
Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_8Bits<uint8_t, MLFloat16, int32_t>();
Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_8Bits<uint8_t, float, int64_t>();
Test_GatherAxis0_QuantizedAxis1_WithZeroPoints_8Bits<uint8_t, MLFloat16, int64_t>();
}

template <typename T1, typename T2, typename Tind>
void Test_GatherAxis1_WithZeroPoints() {
std::vector<int> data = {-8, -7, -6, -5,
Expand Down Expand Up @@ -709,12 +781,6 @@ TEST(GatherBlockQuantizedOpTest, GatherAxisWithZeroPointsNoPading) {

template <typename T1, typename T2, typename Tind>
void Test_GatherAxis_NoPading_4bit() {
// BUG: CUDA kernel does NOT subtract default zero point (8 for 4-bit) when no explicit zero_points provided
// Row 0 data: [-8,-7,-6,-5,...] repeated
// Packed (add offset 8): [0,1,2,3,...] repeated
// Row 0 scales: [1.0, 2.0, 1.0] (indices 0,1,2 from [1,2,1,2,1,2])
// CUDA output: [0*1, 1*1, 2*1, 3*1, ...] [0*2, 1*2, 2*2, 3*2, ...] [0*1, 1*1, 2*1, 3*1, ...]
// = [0, 1, 2, 3, ...] [0, 2, 4, 6, ...] [0, 1, 2, 3, ...]
std::vector<int> data = {
-8, -7, -6, -5, -8, -7, -6, -5, -8, -7, -6, -5, -8, -7, -6, -5,
-8, -7, -6, -5, -8, -7, -6, -5, -8, -7, -6, -5, -8, -7, -6, -5,
Expand Down
Loading