Import latest aicc hipcc fp8 pa snapshot.

Source: feature/aicc-hipcc-unified-attn-fp8-pa @ fc89765

Import latest aicc hipcc fp8 pa snapshot.
Source: feature/aicc-hipcc-unified-attn-fp8-pa @ fc89765
5248d7d2 · hly · c2a1b310 · 5248d7d2 · 5248d7d2 · 5248d7d2
Commit 5248d7d2 authored Jun 05, 2026 by hly
20 changed files
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim224_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim224_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim224_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim224_fp16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim256_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim256_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim256_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim256_fp16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim32_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim32_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim32_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim32_fp16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim64_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim64_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim64_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim64_fp16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim96_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim96_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdim96_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdim96_fp16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdimqk192_hdimv128_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdimqk192_hdimv128_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_bwd_hdimqk192_hdimv128_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_bwd_hdimqk192_hdimv128_fp16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_fp8_fwd_hdim128_prefix_prefill_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_fp8_fwd_hdim128_prefix_prefill_bf16.cpp
+// Copyright (c) 2025, Wenjian Zhang.
+// Splitting the different head dimensions to different files to speed up compilation.
+// This file is auto-generated. See "generate_kernels.py"
+
+#include "../flash_fwd_launch_template.h"
+
+template<>
+void run_fp8_mha_fwd_prefix_prefill_<BFloat16, 128, 128>(Flash_fwd_params &params, hipStream_t stream) {
+#ifdef BUILD_FA_FWD
+    run_fp8_flash_fwd_prefix_prefill<BFloat16, 128, 128>(params, stream);
+#endif
+}
\ No newline at end of file
--- a/csrc/flash_attn_hg/src/target/flash_fp8_fwd_hdim128_prefix_prefill_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_fp8_fwd_hdim128_prefix_prefill_fp16.cpp
+// Copyright (c) 2025, Wenjian Zhang.
+// Splitting the different head dimensions to different files to speed up compilation.
+// This file is auto-generated. See "generate_kernels.py"
+
+#include "../flash_fwd_launch_template.h"
+
+template<>
+void run_fp8_mha_fwd_prefix_prefill_<Float16, 128, 128>(Flash_fwd_params &params, hipStream_t stream) {
+#ifdef BUILD_FA_FWD
+    run_fp8_flash_fwd_prefix_prefill<Float16, 128, 128>(params, stream);
+#endif
+}
\ No newline at end of file
--- a/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_attn_mask_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_attn_mask_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_attn_mask_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_attn_mask_fp16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_fp16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_padding_mask_bf16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_padding_mask_bf16.cpp
--- a/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_padding_mask_fp16.cpp
+++ b/csrc/flash_attn_hg/src/target/flash_fwd_hdim128_padding_mask_fp16.cpp