chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[diffusion] multi-platform: support diffusion on amd and fix encoder loading on MI325 (#13760)

Browse Source 
Co-authored-by: Sabre Shao <sabre.shao@amd.com>
Co-authored-by: Yusheng (Ethan) Su <yushengsu.thu@gmail.com>
Co-authored-by: Hubert Lu <Hubert.Lu@amd.com>
Co-authored-by: xsun <sunxiao04@gmail.com>
This commit is contained in:
Yuzhen Zhou
2025-12-19 15:38:46 +08:00
committed by GitHub
co-authored by Sabre Shao Yusheng Su Hubert Lu xsun
parent f2d64e6782
commit 4bf06635fc
34 changed files with 823 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/multimodal_gen/runtime/platforms/rocm.py
+13
View File
@@ -80,6 +80,19 @@ class RocmPlatform(Platform):
elif selected_backend in (AttentionBackendEnum.FA, None):
pass
elif selected_backend == AttentionBackendEnum.AITER:
if dtype not in (torch.float16, torch.bfloat16):
logger.warning(
"AITer backend only supports fp16/bf16 inputs but got dtype=%s. "
"Falling back to Torch SDPA backend.",
dtype,
)
# TODO: need to compare triton with sdpa as an alternative backend
return "sglang.multimodal_gen.runtime.layers.attention.backends.sdpa.SDPABackend"
logger.info("Using AITer backend on ROCm.")
return "sglang.multimodal_gen.runtime.layers.attention.backends.aiter.AITerBackend"
elif selected_backend in (
AttentionBackendEnum.SLIDING_TILE_ATTN,
AttentionBackendEnum.SAGE_ATTN,