Disable comfy kitchen cuda if pytorch cuda less than 13 (#11681)

comfy/latent_formats.py

@@ -408,7 +408,9 @@ class LTXV(LatentFormat):
         self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512]
 
 class LTXAV(LTXV):
-    pass
+    def __init__(self):
+        self.latent_rgb_factors = None
+        self.latent_rgb_factors_bias = None
 
 class HunyuanVideo(LatentFormat):
     latent_channels = 16

comfy/ldm/flux/math.py

@@ -4,6 +4,7 @@ from torch import Tensor
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
+import logging
 
 
 def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor:
@@ -13,7 +14,6 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme
     x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
     return x
 
-
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
     if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
@@ -28,13 +28,20 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
     return out.to(dtype=torch.float32, device=pos.device)
 
-def apply_rope1(x: Tensor, freqs_cis: Tensor):
-    x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
 
-    x_out = freqs_cis[..., 0] * x_[..., 0]
-    x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
+try:
+    import comfy.quant_ops
+    apply_rope = comfy.quant_ops.ck.apply_rope
+    apply_rope1 = comfy.quant_ops.ck.apply_rope1
+except:
+    logging.warning("No comfy kitchen, using old apply_rope functions.")
+    def apply_rope1(x: Tensor, freqs_cis: Tensor):
+        x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
 
-    return x_out.reshape(*x.shape).type_as(x)
+        x_out = freqs_cis[..., 0] * x_[..., 0]
+        x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
 
-def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-    return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+        return x_out.reshape(*x.shape).type_as(x)
+
+    def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+        return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)

comfy/ldm/lightricks/embeddings_connector.py

@@ -276,7 +276,7 @@ class Embeddings1DConnector(nn.Module):
                 max(1024, hidden_states.shape[1]) / self.num_learnable_registers
             )
             learnable_registers = torch.tile(
-                self.learnable_registers, (num_registers_duplications, 1)
+                self.learnable_registers.to(hidden_states), (num_registers_duplications, 1)
             )
 
             hidden_states = torch.cat((hidden_states, learnable_registers[hidden_states.shape[1]:].unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)), dim=1)

comfy/model_management.py

@@ -1504,6 +1504,16 @@ def supports_fp8_compute(device=None):
 
     return True
 
+def supports_nvfp4_compute(device=None):
+    if not is_nvidia():
+        return False
+
+    props = torch.cuda.get_device_properties(device)
+    if props.major < 10:
+        return False
+
+    return True
+
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):

comfy/ops.py

@@ -493,11 +493,12 @@ from .quant_ops import (
 )
 
 
-def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False):
+def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]):
     class MixedPrecisionOps(manual_cast):
         _quant_config = quant_config
         _compute_dtype = compute_dtype
         _full_precision_mm = full_precision_mm
+        _disabled = disabled
 
         class Linear(torch.nn.Module, CastWeightBiasOp):
             def __init__(
@@ -522,6 +523,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 
                 self.tensor_class = None
                 self._full_precision_mm = MixedPrecisionOps._full_precision_mm
+                self._full_precision_mm_config = False
 
             def reset_parameters(self):
                 return None
@@ -556,8 +558,12 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                     self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False)
                 else:
                     self.quant_format = layer_conf.get("format", None)
+                    self._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False)
                     if not self._full_precision_mm:
-                        self._full_precision_mm = layer_conf.get("full_precision_matrix_mult", False)
+                        self._full_precision_mm = self._full_precision_mm_config
+
+                    if self.quant_format in MixedPrecisionOps._disabled:
+                        self._full_precision_mm = True
 
                     if self.quant_format is None:
                         raise ValueError(f"Unknown quantization format for layer {layer_name}")
@@ -630,7 +636,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                         sd["{}weight_scale".format(prefix)] = self.weight._params.block_scale
 
                     quant_conf = {"format": self.quant_format}
-                    if self._full_precision_mm:
+                    if self._full_precision_mm_config:
                         quant_conf["full_precision_matrix_mult"] = True
                     sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
                 return sd
@@ -711,10 +717,17 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):
     fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
+    nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device)
 
     if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config:
         logging.info("Using mixed precision operations")
-        return mixed_precision_ops(model_config.quant_config, compute_dtype, full_precision_mm=not fp8_compute)
+        disabled = set()
+        if not nvfp4_compute:
+            disabled.add("nvfp4")
+        if not fp8_compute:
+            disabled.add("float8_e4m3fn")
+            disabled.add("float8_e5m2")
+        return mixed_precision_ops(model_config.quant_config, compute_dtype, disabled=disabled)
 
     if (
         fp8_compute and

comfy/quant_ops.py

@@ -13,6 +13,13 @@ try:
         get_layout_class,
     )
     _CK_AVAILABLE = True
+    if torch.version.cuda is None:
+        ck.registry.disable("cuda")
+    else:
+        cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
+        if cuda_version < (13,):
+            ck.registry.disable("cuda")
+
     ck.registry.disable("triton")
     for k, v in ck.list_backends().items():
         logging.info(f"Found comfy_kitchen backend {k}: {v}")

comfy/text_encoders/lt.py

@@ -86,17 +86,19 @@ class LTXAVTEModel(torch.nn.Module):
         )
 
     def set_clip_options(self, options):
+        self.execution_device = options.get("execution_device", self.execution_device)
         self.gemma3_12b.set_clip_options(options)
 
     def reset_clip_options(self):
         self.gemma3_12b.reset_clip_options()
+        self.execution_device = None
 
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs = token_weight_pairs["gemma3_12b"]
 
         out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
         out_device = out.device
-        out = out.movedim(1, -1).to(self.text_embedding_projection.weight.device)
+        out = out.movedim(1, -1).to(self.execution_device)
         out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6)
         out = out.reshape((out.shape[0], out.shape[1], -1))
         out = self.text_embedding_projection(out)

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.35.9
-comfyui-workflow-templates==0.7.66
+comfyui-workflow-templates==0.7.67
 comfyui-embedded-docs==0.3.1
 torch
 torchsde
@@ -21,7 +21,7 @@ psutil
 alembic
 SQLAlchemy
 av>=14.2.0
-comfy-kitchen>=0.2.0
+comfy-kitchen>=0.2.2
 
 #non essential dependencies:
 kornia>=0.7.1