Skip to content

Write TiledAttention outputs in token-major row order. #1061

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
copybara-service merged 1 commit into from
Oct 15, 2026
Merged

Write TiledAttention outputs in token-major row order. #1061

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
3 changes: 2 additions & 1 deletion gemma/tiled_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1254,8 +1254,9 @@ void LocalAttentionForAllHeadsTokensAndBatch(
size_t token_idx = div_heads_per_kv_head.Divide(q_idx);
size_t head_in_group_idx = div_heads_per_kv_head.Remainder(q_idx);

// Rows are interleaved token-major, as for the Q read and KV write.
const size_t batch_index =
current_qbatch_idx * num_query_tokens + token_idx;
token_idx * qbatch.Size() + current_qbatch_idx;
const size_t q_head_idx =
kv_head_idx * heads_per_kv_head + head_in_group_idx;
const size_t activations_att_out_start_idx = q_head_idx * qkv_dim;
Expand Down
66 changes: 34 additions & 32 deletions gemma/tiled_attention_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -433,6 +433,8 @@ void TestLocalAttentionForAllHeadsTokensAndBatch() {
}
}

// Expected `att_out` of the tests below for 2 tokens and qbatch 2. Rows are
// token-major, i.e. row `token_idx * qbatch_size + qi`, as in `TiledAttention`.
const std::vector<float> AttentionMultipleTokensAttentionGoldens = {
34.7414, 34.7717, 34.8022, 34.8327, 34.8631, 34.8936, 34.9241, 34.9545,
34.985, 35.0156, 35.046, 35.0765, 35.1068, 35.1373, 35.1678, 35.1982,
Expand Down Expand Up @@ -466,38 +468,6 @@ const std::vector<float> AttentionMultipleTokensAttentionGoldens = {
38.1612, 38.1903, 38.2197, 38.249, 38.2781, 38.3075, 38.3368, 38.366,
38.3955, 38.4248, 38.4539, 38.4834, 38.5127, 38.5419, 38.5714, 38.6008,
38.63, 38.6595, 38.6889, 38.7181, 38.7477, 38.777, 38.8063, 38.8358,
39.0984, 39.1479, 39.1976, 39.2475, 39.297, 39.3468, 39.3967, 39.4463,
39.4961, 39.546, 39.5957, 39.6455, 39.695, 39.7447, 39.7946, 39.8441,
39.8939, 39.9438, 39.9934, 40.0431, 40.0931, 40.1427, 40.1925, 40.2425,
40.2921, 40.342, 40.3915, 40.4412, 40.4911, 40.5407, 40.5904, 40.6403,
40.6899, 40.7397, 40.7897, 40.8393, 40.8892, 40.9387, 40.9884, 41.0382,
41.0878, 41.1375, 41.1874, 41.237, 41.2868, 41.3367, 41.3863, 41.4361,
41.4861, 41.5358, 41.5856, 41.6351, 41.6849, 41.7347, 41.7843, 41.834,
41.884, 41.9336, 41.9834, 42.0333, 42.083, 42.1328, 42.1823, 42.232,
38.9699, 39.0188, 39.068, 39.1173, 39.1663, 39.2155, 39.2648, 39.3138,
39.3631, 39.4124, 39.4615, 39.5108, 39.5597, 39.6089, 39.6581, 39.7071,
39.7563, 39.8056, 39.8546, 39.9039, 39.9532, 40.0023, 40.0515, 40.1009,
40.15, 40.1993, 40.2483, 40.2974, 40.3467, 40.3957, 40.4449, 40.4942,
40.5433, 40.5925, 40.6419, 40.691, 40.7402, 40.7892, 40.8383, 40.8876,
40.9366, 40.9857, 41.035, 41.0841, 41.1333, 41.1826, 41.2317, 41.2809,
41.3303, 41.3794, 41.4287, 41.4777, 41.5268, 41.5761, 41.6251, 41.6743,
41.7237, 41.7727, 41.8219, 41.8713, 41.9204, 41.9697, 42.0186, 42.0677,
43.4945, 43.5425, 43.5902, 43.6376, 43.6856, 43.7334, 43.7808, 43.8289,
43.8766, 43.9241, 43.9722, 44.02, 44.0675, 44.1157, 44.1635, 44.2111,
44.2583, 44.3062, 44.3538, 44.4011, 44.449, 44.4966, 44.544, 44.5919,
44.6396, 44.6869, 44.735, 44.7826, 44.8301, 44.8781, 44.9258, 44.9733,
45.0213, 45.0691, 45.1166, 45.1647, 45.2125, 45.26, 45.3081, 45.356,
45.4035, 45.4508, 45.4987, 45.5462, 45.5936, 45.6415, 45.6891, 45.7364,
45.7844, 45.832, 45.8794, 45.9274, 45.9751, 46.0225, 46.0705, 46.1183,
46.1657, 46.2138, 46.2615, 46.309, 46.3571, 46.4049, 46.4525, 46.5006,
43.4125, 43.4603, 43.5077, 43.5549, 43.6027, 43.6502, 43.6974, 43.7453,
43.7928, 43.84, 43.8879, 43.9355, 43.9828, 44.0307, 44.0783, 44.1256,
44.1726, 44.2203, 44.2676, 44.3147, 44.3624, 44.4098, 44.4569, 44.5046,
44.552, 44.5992, 44.6469, 44.6944, 44.7416, 44.7894, 44.8369, 44.8841,
44.9319, 44.9795, 45.0267, 45.0746, 45.1222, 45.1694, 45.2173, 45.265,
45.3123, 45.3593, 45.407, 45.4543, 45.5014, 45.5491, 45.5965, 45.6436,
45.6913, 45.7387, 45.7859, 45.8336, 45.8811, 45.9283, 45.9761, 46.0236,
46.0708, 46.1186, 46.1661, 46.2134, 46.2613, 46.3088, 46.3561, 46.404,
34.7729, 34.8035, 34.8341, 34.8648, 34.8953, 34.9259, 34.9567, 34.9872,
35.0179, 35.0486, 35.0792, 35.1098, 35.1404, 35.171, 35.2016, 35.2322,
35.2628, 35.2935, 35.324, 35.3547, 35.3854, 35.416, 35.4466, 35.4774,
Expand Down Expand Up @@ -530,6 +500,38 @@ const std::vector<float> AttentionMultipleTokensAttentionGoldens = {
38.2159, 38.2451, 38.2747, 38.3042, 38.3334, 38.363, 38.3925, 38.4218,
38.4514, 38.4809, 38.5102, 38.5398, 38.5693, 38.5986, 38.6283, 38.6578,
38.6872, 38.7168, 38.7464, 38.7757, 38.8054, 38.835, 38.8644, 38.8941,
39.0984, 39.1479, 39.1976, 39.2475, 39.297, 39.3468, 39.3967, 39.4463,
39.4961, 39.546, 39.5957, 39.6455, 39.695, 39.7447, 39.7946, 39.8441,
39.8939, 39.9438, 39.9934, 40.0431, 40.0931, 40.1427, 40.1925, 40.2425,
40.2921, 40.342, 40.3915, 40.4412, 40.4911, 40.5407, 40.5904, 40.6403,
40.6899, 40.7397, 40.7897, 40.8393, 40.8892, 40.9387, 40.9884, 41.0382,
41.0878, 41.1375, 41.1874, 41.237, 41.2868, 41.3367, 41.3863, 41.4361,
41.4861, 41.5358, 41.5856, 41.6351, 41.6849, 41.7347, 41.7843, 41.834,
41.884, 41.9336, 41.9834, 42.0333, 42.083, 42.1328, 42.1823, 42.232,
38.9699, 39.0188, 39.068, 39.1173, 39.1663, 39.2155, 39.2648, 39.3138,
39.3631, 39.4124, 39.4615, 39.5108, 39.5597, 39.6089, 39.6581, 39.7071,
39.7563, 39.8056, 39.8546, 39.9039, 39.9532, 40.0023, 40.0515, 40.1009,
40.15, 40.1993, 40.2483, 40.2974, 40.3467, 40.3957, 40.4449, 40.4942,
40.5433, 40.5925, 40.6419, 40.691, 40.7402, 40.7892, 40.8383, 40.8876,
40.9366, 40.9857, 41.035, 41.0841, 41.1333, 41.1826, 41.2317, 41.2809,
41.3303, 41.3794, 41.4287, 41.4777, 41.5268, 41.5761, 41.6251, 41.6743,
41.7237, 41.7727, 41.8219, 41.8713, 41.9204, 41.9697, 42.0186, 42.0677,
43.4945, 43.5425, 43.5902, 43.6376, 43.6856, 43.7334, 43.7808, 43.8289,
43.8766, 43.9241, 43.9722, 44.02, 44.0675, 44.1157, 44.1635, 44.2111,
44.2583, 44.3062, 44.3538, 44.4011, 44.449, 44.4966, 44.544, 44.5919,
44.6396, 44.6869, 44.735, 44.7826, 44.8301, 44.8781, 44.9258, 44.9733,
45.0213, 45.0691, 45.1166, 45.1647, 45.2125, 45.26, 45.3081, 45.356,
45.4035, 45.4508, 45.4987, 45.5462, 45.5936, 45.6415, 45.6891, 45.7364,
45.7844, 45.832, 45.8794, 45.9274, 45.9751, 46.0225, 46.0705, 46.1183,
46.1657, 46.2138, 46.2615, 46.309, 46.3571, 46.4049, 46.4525, 46.5006,
43.4125, 43.4603, 43.5077, 43.5549, 43.6027, 43.6502, 43.6974, 43.7453,
43.7928, 43.84, 43.8879, 43.9355, 43.9828, 44.0307, 44.0783, 44.1256,
44.1726, 44.2203, 44.2676, 44.3147, 44.3624, 44.4098, 44.4569, 44.5046,
44.552, 44.5992, 44.6469, 44.6944, 44.7416, 44.7894, 44.8369, 44.8841,
44.9319, 44.9795, 45.0267, 45.0746, 45.1222, 45.1694, 45.2173, 45.265,
45.3123, 45.3593, 45.407, 45.4543, 45.5014, 45.5491, 45.5965, 45.6436,
45.6913, 45.7387, 45.7859, 45.8336, 45.8811, 45.9283, 45.9761, 46.0236,
46.0708, 46.1186, 46.1661, 46.2134, 46.2613, 46.3088, 46.3561, 46.404,
39.1594, 39.2093, 39.2593, 39.3095, 39.3594, 39.4094, 39.4597, 39.5096,
39.5597, 39.61, 39.6599, 39.7101, 39.7599, 39.8099, 39.8601, 39.91,
39.96, 40.0102, 40.0601, 40.1102, 40.1605, 40.2104, 40.2605, 40.3108,
Expand Down
Loading