Merge branch 'master' into dr-support-pip-cm

comfy/cli_args.py

@@ -148,6 +148,8 @@ class PerformanceFeature(enum.Enum):
 
 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: fp16_accumulation fp8_matrix_mult cublas_ops")
 
+parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
+
 parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
 parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")
 parser.add_argument("--windows-standalone-build", action="store_true", help="Windows standalone build: Enable convenient things that most people using the standalone windows build will probably enjoy (like auto opening the page on startup).")

comfy/k_diffusion/sampling.py

@@ -1277,6 +1277,7 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
     phi1_fn = lambda t: torch.expm1(t) / t
     phi2_fn = lambda t: (phi1_fn(t) - 1.0) / t
 
+    old_sigma_down = None
     old_denoised = None
     uncond_denoised = None
     def post_cfg_function(args):
@@ -1304,9 +1305,9 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
                 x = x + d * dt
         else:
             # Second order multistep method in https://arxiv.org/pdf/2308.02157
-            t, t_next, t_prev = t_fn(sigmas[i]), t_fn(sigma_down), t_fn(sigmas[i - 1])
+            t, t_old, t_next, t_prev = t_fn(sigmas[i]), t_fn(old_sigma_down), t_fn(sigma_down), t_fn(sigmas[i - 1])
             h = t_next - t
-            c2 = (t_prev - t) / h
+            c2 = (t_prev - t_old) / h
 
             phi1_val, phi2_val = phi1_fn(-h), phi2_fn(-h)
             b1 = torch.nan_to_num(phi1_val - phi2_val / c2, nan=0.0)
@@ -1326,6 +1327,7 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
             old_denoised = uncond_denoised
         else:
             old_denoised = denoised
+        old_sigma_down = sigma_down
     return x
 
 @torch.no_grad()

comfy/ldm/ace/attention.py

@@ -19,6 +19,7 @@ import torch.nn.functional as F
 from torch import nn
 
 import comfy.model_management
+from comfy.ldm.modules.attention import optimized_attention
 
 class Attention(nn.Module):
     def __init__(
@@ -326,10 +327,6 @@ class CustomerAttnProcessor2_0:
     Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
     """
 
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
-
     def apply_rotary_emb(
         self,
         x: torch.Tensor,
@@ -435,13 +432,9 @@ class CustomerAttnProcessor2_0:
             attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
 
         # the output of sdp = (batch, num_heads, seq_len, head_dim)
-        # TODO: add support for attn.scale when we move to Torch 2.1
-        hidden_states = F.scaled_dot_product_attention(
-            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
-        )
-
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-        hidden_states = hidden_states.to(query.dtype)
+        hidden_states = optimized_attention(
+            query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True,
+        ).to(query.dtype)
 
         # linear proj
         hidden_states = attn.to_out[0](hidden_states)

comfy/utils.py

@@ -28,6 +28,9 @@ import logging
 import itertools
 from torch.nn.functional import interpolate
 from einops import rearrange
+from comfy.cli_args import args
+
+MMAP_TORCH_FILES = args.mmap_torch_files
 
 ALWAYS_SAFE_LOAD = False
 if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in pytorch 2.4, the unsafe path should be removed once earlier versions are deprecated
@@ -67,8 +70,12 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
                     raise ValueError("{}\n\nFile path: {}\n\nThe safetensors file is corrupt/incomplete. Check the file size and make sure you have copied/downloaded it correctly.".format(message, ckpt))
             raise e
     else:
+        torch_args = {}
+        if MMAP_TORCH_FILES:
+            torch_args["mmap"] = True
+
         if safe_load or ALWAYS_SAFE_LOAD:
-            pl_sd = torch.load(ckpt, map_location=device, weights_only=True)
+            pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args)
         else:
             pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle)
         if "global_step" in pl_sd: