apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[Metax] refactor cutlass moe and optimize flash attention (#5361)

Browse Source 
* [Metax] refactor moe and flash attention backend
---------

Co-authored-by: zhangchenyi_dl <16219492+zhangchenyidl@user.noreply.gitee.com>
This commit is contained in:
Neil Zhu
2025-12-10 17:15:17 +08:00
committed by GitHub
parent fbc9bce1e9
commit 4403a21d4b
19 changed files with 3087 additions and 1727 deletions
Download Patch File Download Diff File Expand all files Collapse all files
custom_ops/setup_ops.py
+12 -1
View File
@@ -627,11 +627,17 @@ elif paddle.device.is_compiled_with_custom_device("metax_gpu"):
"gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu",
"gpu_ops/moe/tritonmoe_preprocess.cu",
"gpu_ops/moe/moe_topk_select.cu",
"gpu_ops/get_img_boundaries.cc",
"gpu_ops/remote_cache_kv_ipc.cc",
"gpu_ops/sample_kernels/rejection_top_p_sampling.cu",
"gpu_ops/sample_kernels/top_k_renorm_probs.cu",
"gpu_ops/sample_kernels/min_p_sampling_from_probs.cu",
"metax_ops/moe_dispatch.cu",
"metax_ops/moe_ffn.cu",
"metax_ops/moe_reduce.cu",
"metax_ops/fused_moe.cu",
"metax_ops/apply_rope.cu",
"metax_ops/apply_rope_qkv.cu",
"metax_ops/cache_kv_with_rope.cu",
]
sources += find_end_files("gpu_ops/speculate_decoding", ".cu")
@@ -657,6 +663,11 @@ elif paddle.device.is_compiled_with_custom_device("metax_gpu"):
os.path.join(maca_path, "include"),
os.path.join(maca_path, "include/mcr"),
os.path.join(maca_path, "include/common"),
os.path.join(maca_path, "include/mcfft"),
os.path.join(maca_path, "include/mcrand"),
os.path.join(maca_path, "include/mcsparse"),
os.path.join(maca_path, "include/mcblas"),
os.path.join(maca_path, "include/mcsolver"),
],
),
)