BlehGlobalSageAttention

The README calls this node deprecated in its own description, in favor of BlehSageAttentionSampler. That's worth reading as a strong hint before building anything new around it, and it's worth explaining exactly why, because the reason is a genuinely sharp edge rather than just "the newer one is nicer."

This node globally replaces ComfyUI's default attention computation with SageAttention - a memory-efficient, quantized attention kernel from the thu-ml team - across every model, for the rest of the session, until you turn it off the same way. BlehSageAttentionSampler does the identical underlying job but scoped to a single sampler call and an optional percentage window, which is a much easier thing to reason about.

Here's the sharp edge specifically. This is not a normal model patch that lives on the MODEL object and travels naturally with your graph. The README is explicit: settings only apply when the node actually runs. So if you toggle it on, execute your workflow once, and then later bypass or mute the node - the completely normal way you'd expect to "turn something off" in ComfyUI - SageAttention stays active globally, because nothing re-executed to flip it back off. That's the "less sharp edges" comment the README makes about the sampler-wrapper alternative: this node's on/off state doesn't behave the way a toggle normally would in ComfyUI.

Inputs: model (required - passed through unaffected by the attention swap itself, since this isn't really patching the model object), enabled (default true - the actual toggle, with the caveat above: flipping it to false only takes effect the next time this specific node executes, not retroactively). yaml_parameters is optional and works the same way as on the sampler version - raw keys passed mostly straight through to the SageAttention function, no validation.

Output is a MODEL.

Install requires the same manual step as BlehSageAttentionSampler: SageAttention does not ship with this pack and has to be installed separately into your ComfyUI Python environment as a compiled CUDA extension. It should work with SageAttention 1.0 and 2.0.x. On Windows, that typically means also getting Triton working, which is a known source of friction across the ComfyUI ecosystem generally - budget real setup time the first time through, and expect an easier path if you're on Linux.

The single most common trap with this node, and it's expected behavior per the README rather than a bug: "I set enabled to false and nothing changed." The fix is to actually re-run the node with enabled: false, not to bypass or mute it and assume that's equivalent. If that behavior sounds like more mental overhead than you want to carry around, that's exactly the case the deprecation notice is making - for new workflows, use BlehSageAttentionSampler instead and skip this node entirely.