Merge branch 'master' into beta

Author: jn-jairo

comfy/clip_model.py

@@ -0,0 +1,126 @@
+import torch
+from comfy.ldm.modules.attention import optimized_attention_for_device
+
+class CLIPAttention(torch.nn.Module):
+    def __init__(self, embed_dim, heads, dtype, device, operations):
+        super().__init__()
+
+        self.heads = heads
+        self.q_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+        self.k_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+        self.v_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+
+        self.out_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x, mask=None, optimized_attention=None):
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        out = optimized_attention(q, k, v, self.heads, mask)
+        return self.out_proj(out)
+
+ACTIVATIONS = {"quick_gelu": lambda a: a * torch.sigmoid(1.702 * a),
+               "gelu": torch.nn.functional.gelu,
+}
+
+class CLIPMLP(torch.nn.Module):
+    def __init__(self, embed_dim, intermediate_size, activation, dtype, device, operations):
+        super().__init__()
+        self.fc1 = operations.Linear(embed_dim, intermediate_size, bias=True, dtype=dtype, device=device)
+        self.activation = ACTIVATIONS[activation]
+        self.fc2 = operations.Linear(intermediate_size, embed_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.activation(x)
+        x = self.fc2(x)
+        return x
+
+class CLIPLayer(torch.nn.Module):
+    def __init__(self, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
+        super().__init__()
+        self.layer_norm1 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+        self.self_attn = CLIPAttention(embed_dim, heads, dtype, device, operations)
+        self.layer_norm2 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+        self.mlp = CLIPMLP(embed_dim, intermediate_size, intermediate_activation, dtype, device, operations)
+
+    def forward(self, x, mask=None, optimized_attention=None):
+        x += self.self_attn(self.layer_norm1(x), mask, optimized_attention)
+        x += self.mlp(self.layer_norm2(x))
+        return x
+
+
+class CLIPEncoder(torch.nn.Module):
+    def __init__(self, num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
+        super().__init__()
+        self.layers = torch.nn.ModuleList([CLIPLayer(embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations) for i in range(num_layers)])
+
+    def forward(self, x, mask=None, intermediate_output=None):
+        optimized_attention = optimized_attention_for_device(x.device, mask=True)
+        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
+        if mask is not None:
+            mask += causal_mask
+        else:
+            mask = causal_mask
+
+        if intermediate_output is not None:
+            if intermediate_output < 0:
+                intermediate_output = len(self.layers) + intermediate_output
+
+        intermediate = None
+        for i, l in enumerate(self.layers):
+            x = l(x, mask, optimized_attention)
+            if i == intermediate_output:
+                intermediate = x.clone()
+        return x, intermediate
+
+class CLIPEmbeddings(torch.nn.Module):
+    def __init__(self, embed_dim, vocab_size=49408, num_positions=77, dtype=None, device=None):
+        super().__init__()
+        self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim, dtype=dtype, device=device)
+        self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
+
+    def forward(self, input_tokens):
+        return self.token_embedding(input_tokens) + self.position_embedding.weight
+
+
+class CLIPTextModel_(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        num_layers = config_dict["num_hidden_layers"]
+        embed_dim = config_dict["hidden_size"]
+        heads = config_dict["num_attention_heads"]
+        intermediate_size = config_dict["intermediate_size"]
+        intermediate_activation = config_dict["hidden_act"]
+
+        super().__init__()
+        self.embeddings = CLIPEmbeddings(embed_dim, dtype=torch.float32, device=device)
+        self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
+        self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+
+    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True):
+        x = self.embeddings(input_tokens)
+        #TODO: attention_mask
+        x, i = self.encoder(x, intermediate_output=intermediate_output)
+        x = self.final_layer_norm(x)
+        if i is not None and final_layer_norm_intermediate:
+            i = self.final_layer_norm(i)
+
+        pooled_output = x[torch.arange(x.shape[0], device=x.device), input_tokens.to(dtype=torch.int, device=x.device).argmax(dim=-1),]
+        return x, i, pooled_output
+
+class CLIPTextModel(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        self.num_layers = config_dict["num_hidden_layers"]
+        self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
+        self.dtype = dtype
+
+    def get_input_embeddings(self):
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, embeddings):
+        self.text_model.embeddings.token_embedding = embeddings
+
+    def forward(self, *args, **kwargs):
+        return self.text_model(*args, **kwargs)

comfy/ldm/modules/attention.py

@@ -112,10 +112,13 @@ def attention_basic(q, k, v, heads, mask=None):
     del q, k
 
     if exists(mask):
-        mask = rearrange(mask, 'b ... -> b (...)')
-        max_neg_value = -torch.finfo(sim.dtype).max
-        mask = repeat(mask, 'b j -> (b h) () j', h=h)
-        sim.masked_fill_(~mask, max_neg_value)
+        if mask.dtype == torch.bool:
+            mask = rearrange(mask, 'b ... -> b (...)') #TODO: check if this bool part matches pytorch attention
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+        else:
+            sim += mask
 
     # attention, what we cannot get enough of
     sim = sim.softmax(dim=-1)
@@ -347,6 +350,18 @@ def attention_pytorch(q, k, v, heads, mask=None):
 if model_management.pytorch_attention_enabled():
     optimized_attention_masked = attention_pytorch
 
+def optimized_attention_for_device(device, mask=False):
+    if device == torch.device("cpu"): #TODO
+        if model_management.pytorch_attention_enabled():
+            return attention_pytorch
+        else:
+            return attention_basic
+    if mask:
+        return optimized_attention_masked
+
+    return optimized_attention
+
+
 class CrossAttention(nn.Module):
     def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., dtype=None, device=None, operations=comfy.ops):
         super().__init__()
@@ -391,7 +406,7 @@ def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=
         self.is_res = inner_dim == dim
 
         if self.ff_in:
-            self.norm_in = nn.LayerNorm(dim, dtype=dtype, device=device)
+            self.norm_in = operations.LayerNorm(dim, dtype=dtype, device=device)
             self.ff_in = FeedForward(dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)
 
         self.disable_self_attn = disable_self_attn

comfy/model_management.py

@@ -564,12 +564,12 @@ def cast_to_device(tensor, device, dtype, copy=False):
     if device_supports_cast:
         if copy:
             if tensor.device == device:
-                return tensor.to(dtype, copy=copy)
-            return tensor.to(device, copy=copy).to(dtype)
+                return tensor.to(dtype, copy=copy, non_blocking=True)
+            return tensor.to(device, copy=copy, non_blocking=True).to(dtype, non_blocking=True)
         else:
-            return tensor.to(device).to(dtype)
+            return tensor.to(device, non_blocking=True).to(dtype, non_blocking=True)
     else:
-        return tensor.to(dtype).to(device, copy=copy)
+        return tensor.to(device, dtype, copy=copy, non_blocking=True)
 
 def xformers_enabled():
     global directml_enabled

comfy/sd1_clip.py

@@ -1,12 +1,14 @@
 import os
 
-from transformers import CLIPTokenizer, CLIPTextModel, CLIPTextConfig, modeling_utils
+from transformers import CLIPTokenizer
 import comfy.ops
 import torch
 import traceback
 import zipfile
 from . import model_management
 import contextlib
+import comfy.clip_model
+import json
 
 def gen_empty_tokens(special_tokens, length):
     start_token = special_tokens.get("start", None)
@@ -65,35 +67,19 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         "hidden"
     ]
     def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77,
-                 freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, textmodel_path=None, dtype=None,
-                 special_tokens={"start": 49406, "end": 49407, "pad": 49407},layer_norm_hidden_state=True, config_class=CLIPTextConfig,
-                 model_class=CLIPTextModel, inner_name="text_model"):  # clip-vit-base-patch32
+                 freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=comfy.clip_model.CLIPTextModel,
+                 special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True):  # clip-vit-base-patch32
         super().__init__()
         assert layer in self.LAYERS
-        self.num_layers = 12
-        if textmodel_path is not None:
-            self.transformer = model_class.from_pretrained(textmodel_path)
-        else:
-            if textmodel_json_config is None:
-                textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json")
-            config = config_class.from_json_file(textmodel_json_config)
-            self.num_layers = config.num_hidden_layers
-            with comfy.ops.use_comfy_ops(device, dtype):
-                with modeling_utils.no_init_weights():
-                    self.transformer = model_class(config)
-
-        self.inner_name = inner_name
-        if dtype is not None:
-            inner_model = getattr(self.transformer, self.inner_name)
-            if hasattr(inner_model, "embeddings"):
-                embeddings_bak = inner_model.embeddings.to(torch.float32)
-                inner_model.embeddings = None
-                self.transformer.to(dtype)
-                inner_model.embeddings = embeddings_bak
-            else:
-                previous_inputs = self.transformer.get_input_embeddings().to(torch.float32, copy=True)
-                self.transformer.to(dtype)
-                self.transformer.set_input_embeddings(previous_inputs)
+
+        if textmodel_json_config is None:
+            textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json")
+
+        with open(textmodel_json_config) as f:
+            config = json.load(f)
+
+        self.transformer = model_class(config, dtype, device, comfy.ops)
+        self.num_layers = self.transformer.num_layers
 
         self.max_length = max_length
         if freeze:
@@ -108,7 +94,7 @@ def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_le
         self.layer_norm_hidden_state = layer_norm_hidden_state
         if layer == "hidden":
             assert layer_idx is not None
-            assert abs(layer_idx) <= self.num_layers
+            assert abs(layer_idx) < self.num_layers
             self.clip_layer(layer_idx)
         self.layer_default = (self.layer, self.layer_idx)
 
@@ -119,7 +105,7 @@ def freeze(self):
             param.requires_grad = False
 
     def clip_layer(self, layer_idx):
-        if abs(layer_idx) >= self.num_layers:
+        if abs(layer_idx) > self.num_layers:
             self.layer = "last"
         else:
             self.layer = "hidden"
@@ -174,7 +160,7 @@ def forward(self, tokens):
         tokens = self.set_up_textual_embeddings(tokens, backup_embeds)
         tokens = torch.LongTensor(tokens).to(device)
 
-        if getattr(self.transformer, self.inner_name).final_layer_norm.weight.dtype != torch.float32:
+        if self.transformer.dtype != torch.float32:
             precision_scope = torch.autocast
         else:
             precision_scope = lambda a, dtype: contextlib.nullcontext(a)
@@ -190,20 +176,16 @@ def forward(self, tokens):
                         if tokens[x, y] == max_token:
                             break
 
-            outputs = self.transformer(input_ids=tokens, attention_mask=attention_mask, output_hidden_states=self.layer=="hidden")
+            outputs = self.transformer(tokens, attention_mask, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state)
             self.transformer.set_input_embeddings(backup_embeds)
 
             if self.layer == "last":
-                z = outputs.last_hidden_state
-            elif self.layer == "pooled":
-                z = outputs.pooler_output[:, None, :]
+                z = outputs[0]
             else:
-                z = outputs.hidden_states[self.layer_idx]
-                if self.layer_norm_hidden_state:
-                    z = getattr(self.transformer, self.inner_name).final_layer_norm(z)
+                z = outputs[1]
 
-            if hasattr(outputs, "pooler_output"):
-                pooled_output = outputs.pooler_output.float()
+            if outputs[2] is not None:
+                pooled_output = outputs[2].float()
             else:
                 pooled_output = None
 

comfy/sd2_clip.py

@@ -3,13 +3,13 @@
 import os
 
 class SD2ClipHModel(sd1_clip.SDClipModel):
-    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, textmodel_path=None, dtype=None):
+    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
         if layer == "penultimate":
             layer="hidden"
-            layer_idx=23
+            layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd2_clip_config.json")
-        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, textmodel_path=textmodel_path, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
 
 class SD2ClipHTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, tokenizer_path=None, embedding_directory=None):

comfy/sdxl_clip.py

@@ -3,13 +3,13 @@
 import os
 
 class SDXLClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, textmodel_path=None, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
-        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, textmodel_path=textmodel_path, dtype=dtype,
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
                          special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False)
 
     def load_sd(self, sd):
@@ -37,7 +37,7 @@ def untokenize(self, token_weight_pair):
 class SDXLClipModel(torch.nn.Module):
     def __init__(self, device="cpu", dtype=None):
         super().__init__()
-        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=11, device=device, dtype=dtype, layer_norm_hidden_state=False)
+        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False)
         self.clip_g = SDXLClipG(device=device, dtype=dtype)
 
     def clip_layer(self, layer_idx):

comfy_extras/nodes_hypertile.py

@@ -2,18 +2,18 @@
 
 import math
 from einops import rearrange
-import random
+# Use torch rng for consistency across generations
+from torch import randint
 
-def random_divisor(value: int, min_value: int, /, max_options: int = 1, counter = 0) -> int:
+def random_divisor(value: int, min_value: int, /, max_options: int = 1) -> int:
     min_value = min(min_value, value)
 
     # All big divisors of value (inclusive)
     divisors = [i for i in range(min_value, value + 1) if value % i == 0]
 
     ns = [value // i for i in divisors[:max_options]]  # has at least 1 element
 
-    random.seed(counter)
-    idx = random.randint(0, len(ns) - 1)
+    idx = randint(low=0, high=len(ns) - 1, size=(1,)).item()
 
     return ns[idx]
 
@@ -42,7 +42,6 @@ def patch(self, model, tile_size, swap_size, max_depth, scale_depth):
 
         latent_tile_size = max(32, tile_size) // 8
         self.temp = None
-        self.counter = 1
 
         def hypertile_in(q, k, v, extra_options):
             if q.shape[-1] in apply_to:
@@ -53,10 +52,8 @@ def hypertile_in(q, k, v, extra_options):
                 h, w = round(math.sqrt(hw * aspect_ratio)), round(math.sqrt(hw / aspect_ratio))
 
                 factor = 2**((q.shape[-1] // model_channels) - 1) if scale_depth else 1
-                nh = random_divisor(h, latent_tile_size * factor, swap_size, self.counter)
-                self.counter += 1
-                nw = random_divisor(w, latent_tile_size * factor, swap_size, self.counter)
-                self.counter += 1
+                nh = random_divisor(h, latent_tile_size * factor, swap_size)
+                nw = random_divisor(w, latent_tile_size * factor, swap_size)
 
                 if nh * nw > 1:
                     q = rearrange(q, "b (nh h nw w) c -> (b nh nw) (h w) c", h=h // nh, w=w // nw, nh=nh, nw=nw)