natas_hw
/
openvino
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Move scaling from QK dot product to Q (#24896) 

### Details:
- Move scaling from QK dot product to Q input for better SDPA accuracy
with high-value ranges

### Tickets:
 - [CVS-143256](https://jira.devtools.intel.com/browse/CVS-143256)
This commit is contained in:
Sergey Shlyapnikov 2024-06-07 17:04:12 +04:00 committed by GitHub
parent d996280d5e
commit beb37eb4e1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 1 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/plugins/intel_gpu/src/kernel_selector/cl_kernels/sdpa_opt.cl
View File

@ -202,8 +202,7 @@ KERNEL(sdpa_opt)(
#define QUERY_BLOCK_SIZE 1
INPUT0_TYPE val = BLOCK_READN(INPUT0_TYPE, QUERY_BLOCK_SIZE, query_input, query_offset);
query_local[query_local_offset] = val;
query_local[query_local_offset] = val * scale_val;
query_local_offset += QUERY_STEP_LOCAL;
query_offset += query_pitch;
}
@ -338,7 +337,6 @@ KERNEL(sdpa_opt)(
for (uint seq_len = sgid * SUBGROUP_SIZE + sglid; seq_len < partition_seq_len; seq_len += (HEAD_SIZE)) {
// Read value from SLM and apply scale
qk_val[seq_idx] = qk_local[seq_idx * SEQ_LEN_PARTITION_SIZE + seq_len];
qk_val[seq_idx] *= scale_val;
// Apply attention mask
#if IS_CAUSAL