[AMD][diffusion] VAE attention to the math SDPA backend on gfx1250 (#42858)

This commit is contained in:
YC Yen-Ching Tseng
2026-10-07 02:32:02 -07:00
committed by GitHub
parent 11972e5207
commit 49d7d60dfa
@@ -14,6 +14,7 @@ from typing import Any
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.attention import SDPBackend, sdpa_kernel
import sglang.multimodal_gen.envs as envs
from sglang.multimodal_gen.runtime.platforms.interface import (
@@ -23,6 +24,7 @@ from sglang.multimodal_gen.runtime.platforms.interface import (
PlatformEnum,
)
from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
from sglang.srt.utils import is_gfx1250_supported
logger = init_logger(__name__)
@@ -244,6 +246,15 @@ class RocmPlatform(Platform):
exc_info=True,
)
if is_gfx1250_supported():
count = cls._force_math_sdpa_in_attention(vae)
if count > 0:
logger.info(
"Pinned %d VAE attention modules to the math SDPA backend "
"(AOTriton fails above head_dim 256 on gfx1250)",
count,
)
use_bf16 = envs.SGLANG_USE_ROCM_VAE_CONV2D_BF16
use_conv2d = envs.SGLANG_USE_ROCM_VAE_CONV2D or use_bf16
if use_conv2d:
@@ -259,6 +270,35 @@ class RocmPlatform(Platform):
return vae
@staticmethod
def _force_math_sdpa_in_attention(module: torch.nn.Module) -> int:
"""Wrap VAE attention forwards so SDPA resolves to the math backend.
On gfx1250, AOTriton's flash and mem_efficient SDPA backends fail when
head_dim > 256. On ROCm 10.0 (AOTriton 0.13.50) the VAE returns wrong
values with no error. On ROCm 10.1 (0.14.50) it raises
hipErrorProfilerNotInitialized after the full denoise.
"""
count = 0
for child in module.modules():
if getattr(child, "_sgl_math_sdpa", False):
continue
has_qkv = hasattr(child, "to_qkv") or all(
getattr(child, a, None) is not None for a in ("to_q", "to_k", "to_v")
)
if not has_qkv:
continue
original = child.forward
def wrapped(*args, _orig=original, **kwargs):
with sdpa_kernel(SDPBackend.MATH):
return _orig(*args, **kwargs)
child.forward = wrapped
child._sgl_math_sdpa = True
count += 1
return count
@staticmethod
def _replace_groupnorm(module: torch.nn.Module, aiter_gn_cls: type) -> int:
count = 0