[diffusion] fix: keep MiniMax-H3 VAE fused norms off ROCm (#43350)

This commit is contained in:
WenhaoZhang
2026-10-13 12:26:16 +08:00
committed by GitHub
parent b4822bae03
commit cda267ab0f
3 changed files with 41 additions and 3 deletions
@@ -45,6 +45,9 @@ def _vit_norm_input(module, hidden_states):
def _try_fused_qk_rmsnorm_rope(norm_q, norm_k, query, key, rotary_pos_emb):
if not _env_flag("MINIMAX_H3_VAE_DECODER_FUSED_NORM", "1"):
return None
# NVIDIA only: see _fused_blocks_available in vae_vit.py.
if not current_platform.is_cuda():
return None
if (
rotary_pos_emb is None
or torch.is_grad_enabled()
@@ -9,6 +9,8 @@ from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
from diffusers.utils import logging
from sglang.multimodal_gen.runtime.platforms import current_platform
from .base_module import RotaryEmbeddingND, TransformerBlock, _scaled_residual_add
from .vit_utils import _env_flag, create_token_ids, prepare_rotary_pos_emb
@@ -31,6 +33,10 @@ def _cuda_autocast_disabled(tensor: torch.Tensor):
def _fused_blocks_available(blocks, hidden_states) -> bool:
if not _env_flag("MINIMAX_H3_VAE_DECODER_FUSED_NORM", "1"):
return False
# The fused Triton norms decode to all-NaN on ROCm (MI300), where
# tensor.is_cuda is also true; keep ROCm on the unfused fp32-norm path.
if not current_platform.is_cuda():
return False
if (
not hidden_states.is_cuda
or hidden_states.dim() != 3
@@ -44,10 +44,15 @@ from sglang.multimodal_gen.runtime.models.vaes.minimax_h3_video_vae.vit_utils im
create_token_ids,
prepare_rotary_pos_emb,
)
from sglang.multimodal_gen.runtime.platforms import AttentionBackendEnum
from sglang.multimodal_gen.runtime.platforms import (
AttentionBackendEnum,
current_platform,
)
# torch.cuda.is_available() is also true on ROCm, where these JIT/Triton
# kernels do not compile; gate on the platform instead.
requires_cuda = pytest.mark.skipif(
not torch.cuda.is_available(), reason="fused qk-norm+RoPE kernel needs CUDA"
not current_platform.is_cuda(), reason="fused VAE kernels need CUDA"
)
@@ -306,7 +311,7 @@ def _h3_shaped_decoder():
return decoder
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
@requires_cuda
def test_fused_decoder_matches_unfused_forward():
from sglang.multimodal_gen.runtime.managers.forward_context import (
set_forward_context,
@@ -346,3 +351,27 @@ def test_fused_decoder_matches_unfused_forward():
os.environ["MINIMAX_H3_VAE_DECODER_FUSED_NORM"] = previous
torch.testing.assert_close(fused, eager, rtol=1e-3, atol=1e-3)
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires a GPU")
def test_fused_decoder_is_disabled_off_nvidia():
# On ROCm tensor.is_cuda is true, but the fused Triton norms decoded to NaN
# on MI300; the decoder must keep the unfused path there.
decoder = _h3_shaped_decoder()
latent = torch.randn(2, 4, 1, 2, 2, device="cuda", dtype=torch.float16)
from sglang.kernels.ops import diffusion as diffusion_ops
vae_dir = "sglang.multimodal_gen.runtime.models.vaes.minimax_h3_video_vae"
with (
torch.no_grad(),
torch.autocast("cuda", dtype=torch.float16),
set_forward_context(current_timestep=0, attn_metadata=None),
mock.patch.dict(os.environ, {"MINIMAX_H3_VAE_DECODER_FUSED_NORM": "1"}),
mock.patch(f"{vae_dir}.vae_vit.current_platform.is_cuda", return_value=False),
mock.patch(f"{vae_dir}.attention.current_platform.is_cuda", return_value=False),
mock.patch.object(diffusion_ops, "h3_vae_scale_add_rmsnorm") as scale_add,
mock.patch.object(diffusion_ops, "h3_vae_qk_rmsnorm_rope") as qk_rope,
):
decoder(latent)
scale_add.assert_not_called()
qk_rope.assert_not_called()