Revert "Fix verify kernel when num_heads_kv != num_heads_q"

min-jean-cho · min-jean-cho · commit 770669af878d · 2025-04-22T15:52:48.000+08:00
This reverts commit 2c74e22.
diff --git a/examples/sycl/06_pvc_flash_attention/pvc_flash_attn_runner.hpp b/examples/sycl/06_pvc_flash_attention/pvc_flash_attn_runner.hpp
@@ -371,10 +371,10 @@ template <class GemmKernel, bool isVarLen> struct ExampleRunner {
         if(kv_group_update % q_group_size==0) {
           offset_k += seq_len_kv * head_size_qk;
           offset_v += seq_len_kv * head_size_vo;
-          offset_k_cache += seq_len_kv_cache * head_size_qk;
-          offset_v_cache += seq_len_kv_cache * head_size_vo;
         }
         kv_group_update++;
+        offset_k_cache += seq_len_kv_cache * head_size_qk;
+        offset_v_cache += seq_len_kv_cache * head_size_vo;
         offset_o += seq_len_qo * head_size_vo;
       }
     }

Original file line number	Diff line number	Diff line change
`@@ -371,10 +371,10 @@ template <class GemmKernel, bool isVarLen> struct ExampleRunner {`
`371`	`371`	`if(kv_group_update % q_group_size==0) {`
`372`	`372`	`offset_k += seq_len_kv * head_size_qk;`
`373`	`373`	`offset_v += seq_len_kv * head_size_vo;`
`374`		`- offset_k_cache += seq_len_kv_cache * head_size_qk;`
`375`		`- offset_v_cache += seq_len_kv_cache * head_size_vo;`
`376`	`374`	`}`
`377`	`375`	`kv_group_update++;`
	`376`	`+ offset_k_cache += seq_len_kv_cache * head_size_qk;`
	`377`	`+ offset_v_cache += seq_len_kv_cache * head_size_vo;`
`378`	`378`	`offset_o += seq_len_qo * head_size_vo;`
`379`	`379`	`}`
`380`	`380`	`}`