e2182cc21d
* Integrate ck branch of ck_tile/fa_bwd_opt * Assume dq and q share the same stride * update ck * Integrate more stride of dq_acc * Revert fwd dropout * Fix paremeter order * Integrate ck with more stride * update the limit of hdim of bwd * Check argument * Add test_flash_attn_causal * Support unpad lse * Add test_flash_attn_varlen_causal, test_flash_attn_race_condition, test_flash_attn_bwd_overflow, test_flash_attn_bwd_transpose, test_flash_attn_bwd_varlen_overflow, test_flash_attn_deterministic, test_flash_attn_varlen_deterministic * Fix stride and Kn0 * Fix CK sync issue * Fix typo * Update CK for changing of fmha_fwd_args * Add kvcache tmp * Add kvcache * Fix comment * Sync behavior with ck * Update CK to develop * remove large test case * Add kvcache test * Fix page_block_size in arg * Minor fix * Fix stride error * Update seqlen of kvcache before splitkv * Fix compile error * Fix bug of hdim is not 8x * Fit ck arg * support adaptive num_splits * add more tests * Refine test tolerance * update CK * Move override_num_splits_if_necessary into cpp * update ck * Update ck * Support different flag for different version of hip * remove coerce-illegal, becasue this is not required in FA * Update ck to fix xcratch memory * Add coerce-illegal in some version * Add compile flag for rtn rounding * remove redundant init * Using env var to switch rounding mode * update ck |
||
|---|---|---|
| .. | ||
| composable_kernel@a9b170b541 | ||
| cutlass@756c351b49 | ||
| flash_attn | ||
| flash_attn_ck | ||
| ft_attention | ||
| fused_dense_lib | ||
| fused_softmax | ||
| layer_norm | ||
| rotary | ||
| xentropy | ||