comfyanonymous/ComfyUI

xFormers is heavily optimized for NVIDIA CUDA and frequently lacks stable, pre-built wheels for ROCm, causing runtime or compilation failures. On AMD Instinct GPUs, it is highly recommended to bypass xFormers and rely on PyTorch's native scaled_dot_product_attention (SDPA), which has optimized ROCm backends. This requires wrapping the import in a try-except block or using environment flags to gracefully disable it.

--- a/comfy/ldm/modules/attention.py
+++ b/comfy/ldm/modules/attention.py
@@ -16,7 +16,11 @@
 import torch
 
-import xformers
+try:
+    import xformers
+    XFORMERS_IS_AVAILABLE = True
+except ImportError:
+    XFORMERS_IS_AVAILABLE = False

xFormers is not officially supported on ROCm/AMD Instinct and will fail to import or run correctly. The advisory change guards the import and falls back to PyTorch native scaleddotproduct_attention, which is supported on ROCm.

--- a/comfy/ldm/modules/attention.py
+++ b/comfy/ldm/modules/attention.py
@@ -17,7 +17,11 @@
 import torch
 from torch import nn
 import torch.nn.functional as F
-import xformers.ops
+try:
+    import xformers.ops
+    HAS_XFORMERS = True
+except ImportError:
+    HAS_XFORMERS = False

The flash_attn package import path is compatible with ROCm, but the standard PyPI flash-attn wheel targets NVIDIA GPUs and will not build or run on AMD Instinct. Users must install AMD's ROCm fork of flash-attention (github.com/ROCm/flash-attention) which exposes the same flash_attn_func API, so no source change is strictly required if the correct package is installed.

PyTorch on ROCm maps the HIP backend to the cuda device namespace, so device_type='cuda' is actually compatible and required for torch.autocast on AMD GPUs. Replacing this with 'hip' will cause a runtime error. No code change is strictly necessary for ROCm migration, though dynamically inferring the device type from a tensor is a best practice for cross-platform compatibility.

Hard-coded "cuda" device type checks will fail on AMD Instinct GPUs where the device type is "hip" (or "rocm" in some contexts), causing the conditional to take the wrong branch and potentially skip optimized attention paths. Replace the literal with a portable check that accepts both CUDA and HIP/ROCm backends.

xFormers is primarily optimized for NVIDIA CUDA and may not be available or perform optimally on AMD Instinct GPUs. For ROCm, it is highly recommended to bypass xFormers and rely on PyTorch 2.0+ native F.scaled_dot_product_attention (SDPA), which includes dedicated ROCm optimizations. Guarding the import prevents hard crashes on systems where xFormers is not built for ROCm.

- import xformers
+ try:
+     import xformers
+ except ImportError:
+     xformers = None

xFormers has limited/experimental ROCm support and is not guaranteed to work on AMD Instinct GPUs. The recommended migration path is to replace xformers.ops memoryefficientattention with PyTorch's native torch.nn.functional.scaleddotproduct_attention, which has ROCm-optimized backends (CK/MIOpen). This import will fail or produce wrong results if xFormers is unavailable, so it must be guarded or replaced.

--- a/comfy/ldm/modules/diffusionmodules/model.py
+++ b/comfy/ldm/modules/diffusionmodules/model.py
@@ -14,1 +14,5 @@
-import xformers.ops
+try:
+    import xformers.ops
+    HAS_XFORMERS = True
+except ImportError:
+    HAS_XFORMERS = False

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

xFormers is primarily optimized for NVIDIA CUDA and often lacks stable support or performance benefits on AMD ROCm. Migrating this requires replacing xFormers-specific attention mechanisms with PyTorch's native scaleddotproduct_attention or ROCm-optimized Flash Attention. Failing to handle this will likely cause runtime errors or suboptimal performance on Instinct GPUs.

- import xformers
+ import torch
+ # xFormers is not reliably supported on ROCm; fallback to native PyTorch SDPA
+ if torch.version.hip is None:
+     try:
+         import xformers
+     except ImportError:
+         xformers = None
+ else:
+     xformers = None

xFormers is not built or supported for ROCm on AMD Instinct GPUs, so this import will fail at runtime. The import should be guarded with a try/except and call sites should fall back to PyTorch native scaleddotproduct_attention (SDPA) or AMD's ROCm flash-attention port.

--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -403,1 +403,4 @@
-import xformers.ops
+try:
+    import xformers.ops
+except ImportError:
+    xformers = None

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

Triton supports AMD Instinct GPUs via ROCm, so this import is generally compatible provided a ROCm-enabled version of Triton is installed in the environment. However, any custom Triton kernels used in this file may require tuning for AMD architecture differences, such as warp size (64 on AMD vs 32 on NVIDIA).

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

The hardcoded default device string "cuda" will still function under ROCm since PyTorch maps the "cuda" device namespace to HIP, but it is not portable across non-CUDA backends. For maintainability and future-proofing, consider defaulting to a device-agnostic resolution (e.g., via torch or a framework helper) rather than a literal CUDA string.

The default argument device="cuda" hardcodes a CUDA device string, which should be changed to "hip" for ROCm or made configurable. On ROCm, PyTorch may still accept "cuda" via its compatibility layer, but using "hip" is the correct and explicit convention for AMD Instinct GPUs.

torch.cuda.amp.autocast is deprecated since PyTorch 1.13 and should be replaced with the device-agnostic torch.amp.autocast, which accepts a devicetype argument. On ROCm, devicetype='cuda' works because PyTorch's HIP layer maps the 'cuda' device string, so this change is portable across NVIDIA and AMD. Advisory only — verify ctx.gpuautocastkwargs contents (e.g., enabled, dtype) remain compatible with the new signature.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

The hardcoded device_type='cuda' in torch.autocast will not enable ROCm's native autocast path; on AMD Instinct it should be 'rocm' or dynamically derived. A portable fix is to read the device type from a tensor in scope so the same code works on both CUDA and ROCm.

Using .cuda() hardcodes the tensor allocation to NVIDIA GPUs and relies on HIP's compatibility layer on ROCm, which can break device-agnostic execution. Replacing it with device=labels.device ensures the tensor is allocated directly on the correct active device, improving portability across CUDA, ROCm, and CPU.

The .cuda() call hardcodes the CUDA device backend and will fail or be non-portable on ROCm. Replace it with a device-agnostic pattern that respects the module's or tensor's existing device, or use torch.cuda.current_device() only if explicitly targeting the HIP device.

The hardcoded .cuda() forces CUDA device placement; on ROCm, PyTorch/HIP usually maps .cuda() to the current AMD device, but this is not device-agnostic and can break in CPU-fallback or multi-device contexts. Deriving the device from an existing tensor (e.g., global_caption.device) is the recommended advisory change.

PyTorch's ROCm implementation aliases the torch.cuda namespace to HIP, meaning torch.cuda.Stream will generally work as-is on AMD Instinct GPUs. No code modification is strictly required for this specific API call, provided the PyTorch build includes ROCm support.

torch.cuda.currentstream is part of PyTorch's CUDA stream API, which is supported on ROCm through PyTorch's HIP compatibility layer — no source change is required for functionality. However, if you want to make the code explicitly ROCm-aware or device-agnostic, you could use torch.cuda.currentstream with a device argument (already present here) or wrap it in a helper that abstracts the backend.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

The literal device string "cuda" resolves to the active AMD GPU on ROCm PyTorch. .to("cuda") / device="cuda" need no edits.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.

Calls under torch.cuda.* are aliased by ROCm builds of PyTorch. The same code runs on AMD Instinct GPUs unchanged — torch.cuda.is_available(), streams, events and AMP all map onto HIP.