feat: add essential subgraph blueprints

Add 24 non-cloud essential blueprints from comfyui-wiki/Subgraph-Blueprints.
These cover common workflows: text/image/video generation, editing,
inpainting, outpainting, upscaling, depth maps, pose, captioning, and more.

Cloud-only blueprints (5) are excluded and will be added once
client-side distribution filtering lands.

Amp-Thread-ID: https://ampcode.com/threads/T-019c6f43-6212-7308-bea6-bfc35a486cbf

app/subgraph_manager.py

@@ -53,7 +53,7 @@ class SubgraphManager:
         return entry_id, entry
 
     async def load_entry_data(self, entry: SubgraphEntry):
-        with open(entry['path'], 'r', encoding='utf-8') as f:
+        with open(entry['path'], 'r') as f:
             entry['data'] = f.read()
         return entry
 

comfy/memory_management.py

@@ -78,4 +78,4 @@ def interpret_gathered_like(tensors, gathered):
 
     return dest_views
 
-aimdo_enabled = False
+aimdo_allocator = None

comfy/model_management.py

@@ -836,7 +836,7 @@ def unet_inital_load_device(parameters, dtype):
 
     mem_dev = get_free_memory(torch_dev)
     mem_cpu = get_free_memory(cpu_dev)
-    if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_enabled:
+    if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_allocator is None:
         return torch_dev
     else:
         return cpu_dev
@@ -1121,6 +1121,7 @@ def get_cast_buffer(offload_stream, device, size, ref):
             synchronize()
             del STREAM_CAST_BUFFERS[offload_stream]
             del cast_buffer
+            #FIXME: This doesn't work in Aimdo because mempool cant clear cache
             soft_empty_cache()
         with wf_context:
             cast_buffer = torch.empty((size), dtype=torch.int8, device=device)

comfy/utils.py

@@ -1154,7 +1154,7 @@ def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_am
     return tiled_scale_multidim(samples, function, (tile_y, tile_x), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=output_device, pbar=pbar)
 
 def model_trange(*args, **kwargs):
-    if not comfy.memory_management.aimdo_enabled:
+    if comfy.memory_management.aimdo_allocator is None:
         return trange(*args, **kwargs)
 
     pbar = trange(*args, **kwargs, smoothing=1.0)

comfy_extras/nodes_glsl.py

@@ -716,12 +716,12 @@ def _render_shader_batch(
         gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, 0)
         gl.glUseProgram(0)
 
-        for tex in input_textures:
-            gl.glDeleteTextures(tex)
-        for tex in output_textures:
-            gl.glDeleteTextures(tex)
-        for tex in ping_pong_textures:
-            gl.glDeleteTextures(tex)
+        if input_textures:
+            gl.glDeleteTextures(len(input_textures), input_textures)
+        if output_textures:
+            gl.glDeleteTextures(len(output_textures), output_textures)
+        if ping_pong_textures:
+            gl.glDeleteTextures(len(ping_pong_textures), ping_pong_textures)
         if fbo is not None:
             gl.glDeleteFramebuffers(1, [fbo])
         for pp_fbo in ping_pong_fbos:

comfy_extras/nodes_primitive.py

@@ -29,7 +29,6 @@ class StringMultiline(io.ComfyNode):
             node_id="PrimitiveStringMultiline",
             display_name="String (Multiline)",
             category="utils/primitive",
-            essentials_category="Basics",
             inputs=[
                 io.String.Input("value", multiline=True),
             ],

cuda_malloc.py

@@ -1,8 +1,10 @@
 import os
 import importlib.util
-from comfy.cli_args import args, PerformanceFeature
+from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
 import subprocess
 
+import comfy_aimdo.control
+
 #Can't use pytorch to get the GPU names because the cuda malloc has to be set before the first import.
 def get_gpu_names():
     if os.name == 'nt':
@@ -85,6 +87,10 @@ if not args.cuda_malloc:
     except:
         pass
 
+if enables_dynamic_vram() and comfy_aimdo.control.init():
+    args.cuda_malloc = False
+    os.environ['PYTORCH_CUDA_ALLOC_CONF'] = ""
+
 if args.disable_cuda_malloc:
     args.cuda_malloc = False
 

execution.py

@@ -9,6 +9,7 @@ import traceback
 from enum import Enum
 from typing import List, Literal, NamedTuple, Optional, Union
 import asyncio
+from contextlib import nullcontext
 
 import torch
 
@@ -520,14 +521,19 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                 # TODO - How to handle this with async functions without contextvars (which requires Python 3.12)?
                 GraphBuilder.set_default_prefix(unique_id, call_index, 0)
 
-            try:
-                output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
-            finally:
-                if comfy.memory_management.aimdo_enabled:
-                    if args.verbose == "DEBUG":
-                        comfy_aimdo.control.analyze()
-                    comfy.model_management.reset_cast_buffers()
-                    comfy_aimdo.model_vbar.vbars_reset_watermark_limits()
+            #Do comfy_aimdo mempool chunking here on the per-node level. Multi-model workflows
+            #will cause all sorts of incompatible memory shapes to fragment the pytorch alloc
+            #that we just want to cull out each model run.
+            allocator = comfy.memory_management.aimdo_allocator
+            with nullcontext() if allocator is None else torch.cuda.use_mem_pool(torch.cuda.MemPool(allocator.allocator())):
+                try:
+                    output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
+                finally:
+                    if allocator is not None:
+                        if args.verbose == "DEBUG":
+                            comfy_aimdo.model_vbar.vbars_analyze()
+                        comfy.model_management.reset_cast_buffers()
+                        comfy_aimdo.model_vbar.vbars_reset_watermark_limits()
 
             if has_pending_tasks:
                 pending_async_nodes[unique_id] = output_data

main.py

@@ -173,10 +173,6 @@ import gc
 if 'torch' in sys.modules:
     logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.")
 
-import comfy_aimdo.control
-
-if enables_dynamic_vram():
-    comfy_aimdo.control.init()
 
 import comfy.utils
 
@@ -192,9 +188,13 @@ import hook_breaker_ac10a0
 import comfy.memory_management
 import comfy.model_patcher
 
+import comfy_aimdo.control
+import comfy_aimdo.torch
+
 if enables_dynamic_vram():
     if comfy.model_management.torch_version_numeric < (2, 8):
         logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
+        comfy.memory_management.aimdo_allocator = None
     elif comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index):
         if args.verbose == 'DEBUG':
             comfy_aimdo.control.set_log_debug()
@@ -208,10 +208,11 @@ if enables_dynamic_vram():
             comfy_aimdo.control.set_log_info()
 
         comfy.model_patcher.CoreModelPatcher = comfy.model_patcher.ModelPatcherDynamic
-        comfy.memory_management.aimdo_enabled = True
+        comfy.memory_management.aimdo_allocator = comfy_aimdo.torch.get_torch_allocator()
         logging.info("DynamicVRAM support detected and enabled")
     else:
         logging.warning("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
+        comfy.memory_management.aimdo_allocator = None
 
 
 def cuda_malloc_warning():

nodes.py

@@ -70,6 +70,7 @@ class CLIPTextEncode(ComfyNodeABC):
     FUNCTION = "encode"
 
     CATEGORY = "conditioning"
+    ESSENTIALS_CATEGORY = "Basics"
     DESCRIPTION = "Encodes a text prompt using a CLIP model into an embedding that can be used to guide the diffusion model towards generating specific images."
     SEARCH_ALIASES = ["text", "prompt", "text prompt", "positive prompt", "negative prompt", "encode text", "text encoder", "encode prompt"]
 

requirements.txt

@@ -22,7 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
-comfy-aimdo>=0.2.0
+comfy-aimdo>=0.1.8
 requests
 
 #non essential dependencies: