NVIDIA V100+3090 Not working

[FearL0rd]

I'm running using TORCH, but it complains about flash
If I change flash to V1, the node doesn't load.
Can we use xformers or sdpa? I see this working on way lower card like intels

System

• OS Ubuntu
• PyTorch and its CUDA Version. 2.8/cu128
• GPU and System RAM 3090+V100/128Gb

Terminal Output
XFuserKSamplerAdvanced
�[36mray::RayWorker.common_ksampler()�[39m (pid=13974, ip=192.168.200.15, actor_id=34d2e499e943e8dba3f975a601000000, repr=<raylight.distributed_worker.ray_worker.RayWorker object at 0x701d4feff0e0>)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/tmp/latentsync_f0ab1d40/ray/session_2025-12-27_11-45-51_453917_12859/runtime_resources/py_modules_files/_ray_pkg_79f9abf44dbbf81b/raylight/distributed_worker/utils.py", line 56, in wrapper
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/tmp/latentsync_f0ab1d40/ray/session_2025-12-27_11-45-51_453917_12859/runtime_resources/py_modules_files/_ray_pkg_79f9abf44dbbf81b/raylight/distributed_worker/ray_worker.py", line 441, in common_ksampler
samples = comfy.sample.sample(
^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/sample.py", line 60, in sample
samples = sampler.sample(noise, positive, negative, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 1178, in sample
return sample(self.model, noise, positive, negative, cfg, self.device, sampler, sigmas, self.model_options, latent_image=latent_image, denoise_mask=denoise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 1068, in sample
return cfg_guider.sample(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 1050, in sample
output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/patcher_extension.py", line 112, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 994, in outer_sample
output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 980, in inner_sample
samples = executor.execute(self, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/patcher_extension.py", line 112, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 752, in sample
samples = self.sampler_function(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **self.extra_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/k_diffusion/sampling.py", line 199, in sample_euler
denoised = model(x, sigma_hat * s_in, **extra_args)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 401, in call
out = self.inner_model(x, sigma, model_options=model_options, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 953, in call
return self.outer_predict_noise(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 960, in outer_predict_noise
).execute(x, timestep, model_options, seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/patcher_extension.py", line 112, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 963, in predict_noise
return sampling_function(self.inner_model, x, timestep, self.conds.get("negative", None), self.conds.get("positive", None), self.cfg, model_options=model_options, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 381, in sampling_function
out = calc_cond_batch(model, conds, x, timestep, model_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 206, in calc_cond_batch
return _calc_cond_batch_outer(model, conds, x_in, timestep, model_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 214, in _calc_cond_batch_outer
return executor.execute(model, conds, x_in, timestep, model_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/patcher_extension.py", line 112, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/samplers.py", line 326, in calc_cond_batch
output = model.apply_model(input_x, timestep, **c).chunk(batch_chunks)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/model_base.py", line 162, in apply_model
return comfy.patcher_extension.WrapperExecutor.new_class_executor(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/patcher_extension.py", line 112, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/model_base.py", line 204, in _apply_model
model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1879, in _call_impl
return inner()
^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1827, in inner
result = forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/ldm/lumina/model.py", line 599, in forward
return comfy.patcher_extension.WrapperExecutor.new_class_executor(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/patcher_extension.py", line 112, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/tmp/latentsync_f0ab1d40/ray/session_2025-12-27_11-45-51_453917_12859/runtime_resources/py_modules_files/_ray_pkg_79f9abf44dbbf81b/raylight/diffusion_models/lumina/xdit_context_parallel.py", line 71, in usp_dit_forward
x = layer(
^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1879, in _call_impl
return inner()
^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1827, in inner
result = forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy/ComfyUI/comfy/ldm/lumina/model.py", line 280, in forward
clamp_fp16(self.attention(
^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/tmp/latentsync_f0ab1d40/ray/session_2025-12-27_11-45-51_453917_12859/runtime_resources/py_modules_files/_ray_pkg_79f9abf44dbbf81b/raylight/diffusion_models/lumina/xdit_context_parallel.py", line 117, in usp_joint_attention_forward
output = xfuser_optimized_attention(xq.movedim(1, 2),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/tmp/latentsync_f0ab1d40/ray/session_2025-12-27_11-45-51_453917_12859/runtime_resources/py_modules_files/_ray_pkg_79f9abf44dbbf81b/raylight/distributed_modules/attention.py", line 103, in _attention_xfuser_unmask
out = xfuser_attn(
^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py", line 929, in _fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/xfuser/core/long_ctx_attention/hybrid/attn_layer.py", line 186, in forward
out = self.ring_attn_fn(
^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/xfuser/core/long_ctx_attention/ring/ring_flash_attn.py", line 252, in xdit_ring_flash_attn_func
return xFuserRingFlashAttnFunc.apply(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/autograd/function.py", line 576, in apply
return super().apply(*args, **kwargs) # type: ignore[misc]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/xfuser/core/long_ctx_attention/ring/ring_flash_attn.py", line 173, in forward
out, softmax_lse = xdit_ring_flash_attn_forward(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/xfuser/core/long_ctx_attention/ring/ring_flash_attn.py", line 116, in xdit_ring_flash_attn_forward
block_out, block_lse = fn(
^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/yunchang/kernels/attention.py", line 54, in pytorch_attn_forward
out, lse = aten._scaled_dot_product_flash_attention(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/cesar/comfy-env/lib/python3.12/site-packages/torch/_ops.py", line 1243, in call
return self._op(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: FlashAttention only supports Ampere GPUs or newer.

[komikndr]

There is a new version of the enum type in yunchang (the core of xfuser) for sequence parallelism that uses torch_efficient attention. This is the thing that guy on Bilibili are using to run Raylight on V100 GPUs.

At the moment, introducing this change breaks compatibility with the normal Torch attention enum used in Ray init actors. As a temporary bandaid fix, you can edit the following file:

/home/cesar/comfy-env/lib/python3.12/site-packages/yunchang/kernels/attention.py

def pytorch_attn_forward(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    dropout_p=0.0,
    softmax_scale=None,
    causal=True,
    window_size=(-1, -1),
    softcap=None,
    alibi_slopes=None,
    return_softmax=False,
    op_type="flash",
):

change the op_type from flash into efficient

I'll find a way to fix the fallback between this new version and stable version

[FearL0rd]

Thanks. Let me know if you need any special testing on my hardware.

[komikndr]

is it working? if so, I might add additional attention backends for our guys that use Turing

[FearL0rd]

this mod in the attention.py made it work