[Bugfix] Fix platform-specific routing in CustomOp implementations (#24444)

Signed-off-by: Konrad Zawora <kzawora@habana.ai>
This commit is contained in:
Konrad Zawora
2025-09-11 17:15:01 +00:00
committed by GitHub
parent 1fdd5c42d7
commit 4aa23892d6
8 changed files with 53 additions and 30 deletions
@@ -8,7 +8,6 @@ import numpy as np
import torch
from transformers import PretrainedConfig
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from .base import RotaryEmbedding
@@ -202,28 +201,6 @@ class MRotaryEmbedding(RotaryEmbedding):
if self.mrope_section:
assert sum(self.mrope_section) == rotary_dim // 2
self.use_triton = current_platform.is_cuda_alike()
def forward(
self,
positions: torch.Tensor,
query: torch.Tensor,
key: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
"""MRope forward.
Args:
positions:
[num_tokens,] (text only) or
[3, num_tokens] (T/H/W positions with multimodal inputs)
query: [num_tokens, num_heads * head_size]
key: [num_tokens, num_kv_heads * head_size]
"""
if self.use_triton:
return self.forward_cuda(positions, query, key)
else:
return self.forward_native(positions, query, key)
def forward_native(
self,
positions: torch.Tensor,