Add environment variable to opt out of #10302 (forced disablement of …

.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat

@@ -0,0 +1,3 @@
+..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
+pause

.ci/windows_nvidia_base_files/run_nvidia_gpu.bat

@@ -1,2 +1,3 @@
 .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
 pause

.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat

@@ -1,2 +1,3 @@
 .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
 pause

.github/workflows/release-stable-all.yml

@@ -18,9 +18,9 @@ jobs:
     uses: ./.github/workflows/stable-release.yml
     with:
       git_tag: ${{ inputs.git_tag }}
-      cache_tag: "cu129"
+      cache_tag: "cu130"
       python_minor: "13"
-      python_patch: "6"
+      python_patch: "9"
       rel_name: "nvidia"
       rel_extra_name: ""
       test_release: true

.github/workflows/windows_release_dependencies.yml

@@ -17,7 +17,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "129"
+        default: "130"
 
       python_minor:
         description: 'python minor version'
@@ -29,7 +29,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "6"
+        default: "9"
 #  push:
 #    branches:
 #      - master

README.md

@@ -176,6 +176,8 @@ Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you
 
 If you have trouble extracting it, right click the file -> properties -> unblock
 
+Update your Nvidia drivers if it doesn't start.
+
 #### Alternative Downloads:
 
 [Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)

comfy/cli_args.py

@@ -105,6 +105,7 @@ class LatentPreviewMethod(enum.Enum):
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
 cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
+cache_group.add_argument("--cache-ram", nargs='?', const=4.0, type=float, default=0, help="Use RAM pressure caching with the specified headroom threshold. If available RAM drops below the threhold the cache remove large items to free RAM. Default 4GB")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
@@ -144,8 +145,9 @@ class PerformanceFeature(enum.Enum):
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
+    PinnedMem = "pinned_memory"
 
-parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
+parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
 
 parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
 parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.")

comfy/controlnet.py

@@ -310,11 +310,13 @@ def __init__(self, in_features: int, out_features: int, bias: bool = True,
             self.bias = None
 
         def forward(self, input):
-            weight, bias = comfy.ops.cast_bias_weight(self, input)
+            weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True)
             if self.up is not None:
-                return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
+                x = torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
             else:
-                return torch.nn.functional.linear(input, weight, bias)
+                x = torch.nn.functional.linear(input, weight, bias)
+            comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
     class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp):
         def __init__(
@@ -350,12 +352,13 @@ def __init__(
 
 
         def forward(self, input):
-            weight, bias = comfy.ops.cast_bias_weight(self, input)
+            weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True)
             if self.up is not None:
-                return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
+                x = torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
             else:
-                return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
-
+                x = torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
+            comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
 class ControlLora(ControlNet):
     def __init__(self, control_weights, global_average_pooling=False, model_options={}): #TODO? model_options

comfy/ldm/flux/layers.py

@@ -195,8 +195,8 @@ def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=N
             txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
-        img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
-        img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
+        img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
+        img += apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
 
         # calculate the txt bloks
         txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt)

comfy/ldm/flux/math.py

@@ -7,15 +7,7 @@
 
 
 def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor:
-    q_shape = q.shape
-    k_shape = k.shape
-
-    if pe is not None:
-        q = q.to(dtype=pe.dtype).reshape(*q.shape[:-1], -1, 1, 2)
-        k = k.to(dtype=pe.dtype).reshape(*k.shape[:-1], -1, 1, 2)
-        q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v)
-        k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
-
+    q, k = apply_rope(q, k, pe)
     heads = q.shape[1]
     x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
     return x

comfy/ldm/lightricks/model.py

@@ -3,12 +3,11 @@
 import comfy.patcher_extension
 import comfy.ldm.modules.attention
 import comfy.ldm.common_dit
-from einops import rearrange
 import math
 from typing import Dict, Optional, Tuple
 
 from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords
-
+from comfy.ldm.flux.math import apply_rope1
 
 def get_timestep_embedding(
     timesteps: torch.Tensor,
@@ -238,20 +237,6 @@ def forward(self, x):
         return self.net(x)
 
 
-def apply_rotary_emb(input_tensor, freqs_cis): #TODO: remove duplicate funcs and pick the best/fastest one
-    cos_freqs = freqs_cis[0]
-    sin_freqs = freqs_cis[1]
-
-    t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2)
-    t1, t2 = t_dup.unbind(dim=-1)
-    t_dup = torch.stack((-t2, t1), dim=-1)
-    input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)")
-
-    out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs
-
-    return out
-
-
 class CrossAttention(nn.Module):
     def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None):
         super().__init__()
@@ -281,8 +266,8 @@ def forward(self, x, context=None, mask=None, pe=None, transformer_options={}):
         k = self.k_norm(k)
 
         if pe is not None:
-            q = apply_rotary_emb(q, pe)
-            k = apply_rotary_emb(k, pe)
+            q = apply_rope1(q.unsqueeze(1), pe).squeeze(1)
+            k = apply_rope1(k.unsqueeze(1), pe).squeeze(1)
 
         if mask is None:
             out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options)
@@ -306,12 +291,17 @@ def __init__(self, dim, n_heads, d_head, context_dim=None, attn_precision=None,
     def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None, transformer_options={}):
         shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + timestep.reshape(x.shape[0], timestep.shape[1], self.scale_shift_table.shape[0], -1)).unbind(dim=2)
 
-        x += self.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe, transformer_options=transformer_options) * gate_msa
+        attn1_input = comfy.ldm.common_dit.rms_norm(x)
+        attn1_input = torch.addcmul(attn1_input, attn1_input, scale_msa).add_(shift_msa)
+        attn1_input = self.attn1(attn1_input, pe=pe, transformer_options=transformer_options)
+        x.addcmul_(attn1_input, gate_msa)
+        del attn1_input
 
         x += self.attn2(x, context=context, mask=attention_mask, transformer_options=transformer_options)
 
-        y = comfy.ldm.common_dit.rms_norm(x) * (1 + scale_mlp) + shift_mlp
-        x += self.ff(y) * gate_mlp
+        y = comfy.ldm.common_dit.rms_norm(x)
+        y = torch.addcmul(y, y, scale_mlp).add_(shift_mlp)
+        x.addcmul_(self.ff(y), gate_mlp)
 
         return x
 
@@ -327,41 +317,35 @@ def get_fractional_positions(indices_grid, max_pos):
 
 
 def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]):
-    dtype = torch.float32 #self.dtype
+    dtype = torch.float32
+    device = indices_grid.device
 
+    # Get fractional positions and compute frequency indices
     fractional_positions = get_fractional_positions(indices_grid, max_pos)
+    indices = theta ** torch.linspace(0, 1, dim // 6, device=device, dtype=dtype) * math.pi / 2
 
-    start = 1
-    end = theta
-    device = fractional_positions.device
-
-    indices = theta ** (
-        torch.linspace(
-            math.log(start, theta),
-            math.log(end, theta),
-            dim // 6,
-            device=device,
-            dtype=dtype,
-        )
-    )
-    indices = indices.to(dtype=dtype)
+    # Compute frequencies and apply cos/sin
+    freqs = (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)).transpose(-1, -2).flatten(2)
+    cos_vals = freqs.cos().repeat_interleave(2, dim=-1)
+    sin_vals = freqs.sin().repeat_interleave(2, dim=-1)
+
+    # Pad if dim is not divisible by 6
+    if dim % 6 != 0:
+        padding_size = dim % 6
+        cos_vals = torch.cat([torch.ones_like(cos_vals[:, :, :padding_size]), cos_vals], dim=-1)
+        sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1)
 
-    indices = indices * math.pi / 2
+    # Reshape and extract one value per pair (since repeat_interleave duplicates each value)
+    cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0].to(out_dtype)  # [B, N, dim//2]
+    sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0].to(out_dtype)  # [B, N, dim//2]
 
-    freqs = (
-        (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
-        .transpose(-1, -2)
-        .flatten(2)
-    )
+    # Build rotation matrix [[cos, -sin], [sin, cos]] and add heads dimension
+    freqs_cis = torch.stack([
+        torch.stack([cos_vals, -sin_vals], dim=-1),
+        torch.stack([sin_vals, cos_vals], dim=-1)
+    ], dim=-2).unsqueeze(1)  # [B, 1, N, dim//2, 2, 2]
 
-    cos_freq = freqs.cos().repeat_interleave(2, dim=-1)
-    sin_freq = freqs.sin().repeat_interleave(2, dim=-1)
-    if dim % 6 != 0:
-        cos_padding = torch.ones_like(cos_freq[:, :, : dim % 6])
-        sin_padding = torch.zeros_like(cos_freq[:, :, : dim % 6])
-        cos_freq = torch.cat([cos_padding, cos_freq], dim=-1)
-        sin_freq = torch.cat([sin_padding, sin_freq], dim=-1)
-    return cos_freq.to(out_dtype), sin_freq.to(out_dtype)
+    return freqs_cis
 
 
 class LTXVModel(torch.nn.Module):
@@ -501,7 +485,7 @@ def block_wrap(args):
         shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]
         x = self.norm_out(x)
         # Modulation
-        x = x * (1 + scale) + shift
+        x = torch.addcmul(x, x, scale).add_(shift)
         x = self.proj_out(x)
 
         x = self.patchifier.unpatchify(

comfy/ldm/lumina/model.py

@@ -522,7 +522,7 @@ def patchify_and_embed(
         max_cap_len = max(l_effective_cap_len)
         max_img_len = max(l_effective_img_len)
 
-        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.int32, device=device)
+        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device)
 
         for i in range(bsz):
             cap_len = l_effective_cap_len[i]
@@ -531,10 +531,22 @@ def patchify_and_embed(
             H_tokens, W_tokens = H // pH, W // pW
             assert H_tokens * W_tokens == img_len
 
-            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.int32, device=device)
+            rope_options = transformer_options.get("rope_options", None)
+            h_scale = 1.0
+            w_scale = 1.0
+            h_start = 0
+            w_start = 0
+            if rope_options is not None:
+                h_scale = rope_options.get("scale_y", 1.0)
+                w_scale = rope_options.get("scale_x", 1.0)
+
+                h_start = rope_options.get("shift_y", 0.0)
+                w_start = rope_options.get("shift_x", 0.0)
+
+            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.float32, device=device)
             position_ids[i, cap_len:cap_len+img_len, 0] = cap_len
-            row_ids = torch.arange(H_tokens, dtype=torch.int32, device=device).view(-1, 1).repeat(1, W_tokens).flatten()
-            col_ids = torch.arange(W_tokens, dtype=torch.int32, device=device).view(1, -1).repeat(H_tokens, 1).flatten()
+            row_ids = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
+            col_ids = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
             position_ids[i, cap_len:cap_len+img_len, 1] = row_ids
             position_ids[i, cap_len:cap_len+img_len, 2] = col_ids
 

comfy/ldm/qwen_image/model.py

@@ -10,6 +10,7 @@
 from comfy.ldm.flux.layers import EmbedND
 import comfy.ldm.common_dit
 import comfy.patcher_extension
+from comfy.ldm.flux.math import apply_rope1
 
 class GELU(nn.Module):
     def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None):
@@ -134,33 +135,34 @@ def forward(
         image_rotary_emb: Optional[torch.Tensor] = None,
         transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        batch_size = hidden_states.shape[0]
+        seq_img = hidden_states.shape[1]
         seq_txt = encoder_hidden_states.shape[1]
 
-        img_query = self.to_q(hidden_states).unflatten(-1, (self.heads, -1))
-        img_key = self.to_k(hidden_states).unflatten(-1, (self.heads, -1))
-        img_value = self.to_v(hidden_states).unflatten(-1, (self.heads, -1))
+        # Project and reshape to BHND format (batch, heads, seq, dim)
+        img_query = self.to_q(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
+        img_key = self.to_k(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
+        img_value = self.to_v(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2)
 
-        txt_query = self.add_q_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
-        txt_key = self.add_k_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
-        txt_value = self.add_v_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+        txt_query = self.add_q_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
+        txt_key = self.add_k_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
+        txt_value = self.add_v_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2)
 
         img_query = self.norm_q(img_query)
         img_key = self.norm_k(img_key)
         txt_query = self.norm_added_q(txt_query)
         txt_key = self.norm_added_k(txt_key)
 
-        joint_query = torch.cat([txt_query, img_query], dim=1)
-        joint_key = torch.cat([txt_key, img_key], dim=1)
-        joint_value = torch.cat([txt_value, img_value], dim=1)
+        joint_query = torch.cat([txt_query, img_query], dim=2)
+        joint_key = torch.cat([txt_key, img_key], dim=2)
+        joint_value = torch.cat([txt_value, img_value], dim=2)
 
-        joint_query = apply_rotary_emb(joint_query, image_rotary_emb)
-        joint_key = apply_rotary_emb(joint_key, image_rotary_emb)
+        joint_query = apply_rope1(joint_query, image_rotary_emb)
+        joint_key = apply_rope1(joint_key, image_rotary_emb)
 
-        joint_query = joint_query.flatten(start_dim=2)
-        joint_key = joint_key.flatten(start_dim=2)
-        joint_value = joint_value.flatten(start_dim=2)
-
-        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask, transformer_options=transformer_options)
+        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads,
+                                                         attention_mask, transformer_options=transformer_options,
+                                                         skip_reshape=True)
 
         txt_attn_output = joint_hidden_states[:, :seq_txt, :]
         img_attn_output = joint_hidden_states[:, seq_txt:, :]
@@ -413,7 +415,7 @@ def _forward(
         txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
         txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
         ids = torch.cat((txt_ids, img_ids), dim=1)
-        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous()
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states)