[ROCm] Bump torch 2.13, triton 3.8, torchaudio, torchvision (#50605)

Signed-off-by: Rohan Potdar <rohan.potdar@amd.com>
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Signed-off-by: jpvillam <juan.villamizar@amd.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: jpvillam <juan.villamizar@amd.com>
This commit is contained in:
Rohan Potdar
2026-09-25 14:55:52 -05:00
committed by GitHub
co-authored by Claude Fable 5 jpvillam
parent 31cc226401
commit e440b75bb6
5 changed files with 52 additions and 14 deletions
@@ -120,12 +120,10 @@ def _mxfp8_dot_scaled_linear(
return out
# Triton 3.8 enables TRITON_HIP_USE_ASYNC_COPY by default on gfx950; its extra LDS
# buffer leaves no room for num_stages=3 at BLOCK_K=256.
# TODO(rasmith)(Rohan138): Remove the 3.8 check once
# https://github.com/vllm-project/vllm/pull/50605 merges.
# gfx950's async-copy LDS (default in the triton 3.8 this branch pins) leaves no
# room for num_stages=3 at BLOCK_K=256; use 2 there.
_BK256_STAGES = 3
if triton.__version__.startswith("3.8") and current_platform.is_rocm():
if current_platform.is_rocm():
from vllm.platforms.rocm import on_gfx950
if on_gfx950():