minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Revert "[AMD] Fix RotaryEmbedding crash on AMD/ROCm (regression from #17934)" (#18922)

Browse Source
This commit is contained in:
HAI
2026-02-17 01:07:50 -08:00
committed by GitHub
parent 14c95d255c
commit b158f5d4a2
1 changed files with 0 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/rotary_embedding.py
-13
View File
@@ -389,19 +389,6 @@ class RotaryEmbedding(MultiPlatformOp):
) )
return query, key return query, key
def forward_hip(self, *args, **kwargs):
"""HIP/ROCm implementation.
The JIT kernels (sglang.jit_kernel.pos_enc) used in forward_cuda's
fallback path depend on tvm_ffi which invokes nvidia-smi to detect
CUDA compute capability. This fails on AMD GPUs, so we use the
pure-PyTorch native implementation instead.
Uses *args/**kwargs because subclasses (MRotaryEmbedding, etc.)
have different forward_native() signatures.
"""
return self.forward_native(*args, **kwargs)
def extra_repr(self) -> str: def extra_repr(self) -> str:
s = f"head_size={self.head_size}, rotary_dim={self.rotary_dim}" s = f"head_size={self.head_size}, rotary_dim={self.rotary_dim}"
s += f", max_position_embeddings={self.max_position_embeddings}" s += f", max_position_embeddings={self.max_position_embeddings}"