Merge 43e9509856 into 3cd7b32f1b

.github/workflows/stable-release.yml

@@ -117,7 +117,7 @@ jobs:
           ./python.exe get-pip.py
           ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
 
-          grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
+          grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
           ./python.exe -s -m pip install -r requirements_comfyui.txt
           rm requirements_comfyui.txt
 

comfy/text_encoders/lt.py

@@ -36,10 +36,10 @@ class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
 
 class Gemma3_12BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
-        llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None)
-        if llama_scaled_fp8 is not None:
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
             model_options = model_options.copy()
-            model_options["scaled_fp8"] = llama_scaled_fp8
+            model_options["quantization_metadata"] = llama_quantization_metadata
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
@@ -119,12 +119,12 @@ class LTXAVTEModel(torch.nn.Module):
             return self.load_state_dict(sdo, strict=False)
 
 
-def ltxav_te(dtype_llama=None, llama_scaled_fp8=None):
+def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
     class LTXAVTEModel_(LTXAVTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
+            if llama_quantization_metadata is not None:
                 model_options = model_options.copy()
-                model_options["llama_scaled_fp8"] = llama_scaled_fp8
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
             if dtype_llama is not None:
                 dtype = dtype_llama
             super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)

comfy_extras/nodes_sage3.py

@@ -0,0 +1,78 @@
+from typing import Callable
+
+import torch
+from typing_extensions import override
+
+from comfy.ldm.modules.attention import get_attention_function
+from comfy.model_patcher import ModelPatcher
+from comfy_api.latest import ComfyExtension, io
+from server import PromptServer
+
+
+class Sage3PatchModel(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="Sage3PatchModel",
+            display_name="Patch SageAttention 3",
+            description="Patch the model to use `attention3_sage` during the middle blocks and steps, keeping the default attention function for the first/last blocks and steps",
+            category="_for_testing",
+            inputs=[
+                io.Model.Input("model"),
+            ],
+            outputs=[io.Model.Output()],
+            is_experimental=True,
+        )
+
+    @classmethod
+    def execute(cls, model: ModelPatcher) -> io.NodeOutput:
+        sage3: Callable | None = get_attention_function("sage3", default=None)
+
+        if sage3 is None:
+            PromptServer.instance.send_progress_text(
+                "`sageattn3` is not installed / available...",
+                cls.hidden.unique_id,
+            )
+            return io.NodeOutput(model)
+
+        def attention_override(func: Callable, *args, **kwargs):
+            transformer_options: dict = kwargs.get("transformer_options", {})
+
+            block_index: int = transformer_options.get("block_index", 0)
+            total_blocks: int = transformer_options.get("total_blocks", 1)
+
+            if block_index == 0 or block_index >= (total_blocks - 1):
+                return func(*args, **kwargs)
+
+            sample_sigmas: torch.Tensor = transformer_options["sample_sigmas"]
+            sigmas: torch.Tensor = transformer_options["sigmas"]
+
+            total_steps: int = sample_sigmas.size(0)
+            step: int = 0
+
+            for i in range(total_steps):
+                if torch.allclose(sample_sigmas[i], sigmas):
+                    step = i
+                    break
+
+            if step == 0 or step >= (total_steps - 1):
+                return func(*args, **kwargs)
+
+            return sage3(*args, **kwargs)
+
+        model = model.clone()
+        model.model_options["transformer_options"][
+            "optimized_attention_override"
+        ] = attention_override
+
+        return io.NodeOutput(model)
+
+
+class Sage3Extension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [Sage3PatchModel]
+
+
+async def comfy_entrypoint():
+    return Sage3Extension()

nodes.py

@@ -2369,6 +2369,7 @@ async def init_builtin_extra_nodes():
         "nodes_nop.py",
         "nodes_kandinsky5.py",
         "nodes_wanmove.py",
+        "nodes_sage3.py",
     ]
 
     import_failed = []