comfy aimdo 0.2.11 + Improved RAM Pressure release strategies - Windows speedups (Comfy-Org#12925)

* Implement seek and read for pins

Source pins from an mmap is pad because its its a CPU->CPU copy that
attempts to fully buffer the same data twice. Instead, use seek and
read which avoids the mmap buffering while usually being a faster
read in the first place (avoiding mmap faulting etc).

* pinned_memory: Use Aimdo pinner

The aimdo pinner bypasses pytorches CPU allocator which can leak
windows commit charge.

* ops: bypass init() of weight for embedding layer

This similarly consumes large commit charge especially for TEs. It can
cause a permanement leaked commit charge which can destabilize on
systems close to the commit ceiling and generally confuses the RAM
stats.

* model_patcher: implement pinned memory counter

Implement a pinned memory counter for better accounting of what volume
of memory pins have.

* implement touch accounting

Implement accounting of touching mmapped tensors.

* mm+mp: add residency mmap getter

* utils: use the aimdo mmap to load sft files

* model_management: Implement tigher RAM pressure semantics

Implement a pressure release on entire MMAPs as windows does perform
faster when mmaps are unloaded and model loads free ramp into fully
unallocated RAM.

Make the concept of freeing for pins a completely separate concept.
Now that pins are loadable directly from original file and don' touch
the mmap, tighten the freeing budget to just the current loaded model
- what you have left over. This still over-frees pins, but its a lot
better than before.

So after the pins are freed with that algorithm, bounce entire MMAPs
to free RAM based on what the model needs, deducting off any known
resident-in-mmap tensors to the free quota to keep it as tight as
possible.

* comfy-aimdo 0.2.11

Comfy aimdo 0.2.11

* mm: Implement file_slice path for QT

* ruff

* ops: put meta-tensors in place to allow custom nodes to check geo

Author: rattus128

comfy/memory_management.py

@@ -1,9 +1,68 @@
 import math
+import ctypes
+import threading
+import dataclasses
 import torch
 from typing import NamedTuple
 
 from comfy.quant_ops import QuantizedTensor
 
+
+class TensorFileSlice(NamedTuple):
+    file_ref: object
+    thread_id: int
+    offset: int
+    size: int
+
+
+def read_tensor_file_slice_into(tensor, destination):
+
+    if isinstance(tensor, QuantizedTensor):
+        if not isinstance(destination, QuantizedTensor):
+            return False
+        if tensor._layout_cls != destination._layout_cls:
+            return False
+
+        if not read_tensor_file_slice_into(tensor._qdata, destination._qdata):
+            return False
+
+        dst_orig_dtype = destination._params.orig_dtype
+        destination._params.copy_from(tensor._params, non_blocking=False)
+        destination._params = dataclasses.replace(destination._params, orig_dtype=dst_orig_dtype)
+        return True
+
+    info = getattr(tensor.untyped_storage(), "_comfy_tensor_file_slice", None)
+    if info is None:
+        return False
+
+    file_obj = info.file_ref
+    if (destination.device.type != "cpu"
+            or file_obj is None
+            or threading.get_ident() != info.thread_id
+            or destination.numel() * destination.element_size() < info.size):
+        return False
+
+    if info.size == 0:
+        return True
+
+    buf_type = ctypes.c_ubyte * info.size
+    view = memoryview(buf_type.from_address(destination.data_ptr()))
+
+    try:
+        file_obj.seek(info.offset)
+        done = 0
+        while done < info.size:
+            try:
+                n = file_obj.readinto(view[done:])
+            except OSError:
+                return False
+            if n <= 0:
+                return False
+            done += n
+        return True
+    finally:
+        view.release()
+
 class TensorGeometry(NamedTuple):
     shape: any
     dtype: torch.dtype

comfy/model_management.py

@@ -505,6 +505,28 @@ def module_size(module):
         module_mem += t.nbytes
     return module_mem
 
+def module_mmap_residency(module, free=False):
+    mmap_touched_mem = 0
+    module_mem = 0
+    bounced_mmaps = set()
+    sd = module.state_dict()
+    for k in sd:
+        t = sd[k]
+        module_mem += t.nbytes
+        storage = t._qdata.untyped_storage() if isinstance(t, comfy.quant_ops.QuantizedTensor) else t.untyped_storage()
+        if not getattr(storage, "_comfy_tensor_mmap_touched", False):
+            continue
+        mmap_touched_mem += t.nbytes
+        if not free:
+            continue
+        storage._comfy_tensor_mmap_touched = False
+        mmap_obj = storage._comfy_tensor_mmap_refs[0]
+        if mmap_obj in bounced_mmaps:
+            continue
+        mmap_obj.bounce()
+        bounced_mmaps.add(mmap_obj)
+    return mmap_touched_mem, module_mem
+
 class LoadedModel:
     def __init__(self, model):
         self._set_model(model)
@@ -532,6 +554,9 @@ def model(self):
     def model_memory(self):
         return self.model.model_size()
 
+    def model_mmap_residency(self, free=False):
+        return self.model.model_mmap_residency(free=free)
+
     def model_loaded_memory(self):
         return self.model.loaded_size()
 
@@ -633,7 +658,7 @@ def extra_reserved_memory():
 def minimum_inference_memory():
     return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory()
 
-def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_required=0):
+def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins_required=0, ram_required=0):
     cleanup_models_gc()
     unloaded_model = []
     can_unload = []
@@ -646,13 +671,14 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
                 can_unload.append((-shift_model.model_offloaded_memory(), sys.getrefcount(shift_model.model), shift_model.model_memory(), i))
                 shift_model.currently_used = False
 
-    for x in sorted(can_unload):
+    can_unload_sorted = sorted(can_unload)
+    for x in can_unload_sorted:
         i = x[-1]
         memory_to_free = 1e32
-        ram_to_free = 1e32
+        pins_to_free = 1e32
         if not DISABLE_SMART_MEMORY:
             memory_to_free = memory_required - get_free_memory(device)
-            ram_to_free = ram_required - get_free_ram()
+            pins_to_free = pins_required - get_free_ram()
             if current_loaded_models[i].model.is_dynamic() and for_dynamic:
                 #don't actually unload dynamic models for the sake of other dynamic models
                 #as that works on-demand.
@@ -661,9 +687,18 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
         if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
             logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
             unloaded_model.append(i)
-        if ram_to_free > 0:
+        if pins_to_free > 0:
+            logging.debug(f"PIN Unloading {current_loaded_models[i].model.model.__class__.__name__}")
+            current_loaded_models[i].model.partially_unload_ram(pins_to_free)
+
+    for x in can_unload_sorted:
+        i = x[-1]
+        ram_to_free = ram_required - psutil.virtual_memory().available
+        if ram_to_free <= 0 and i not in unloaded_model:
+            continue
+        resident_memory, _ = current_loaded_models[i].model_mmap_residency(free=True)
+        if resident_memory > 0:
             logging.debug(f"RAM Unloading {current_loaded_models[i].model.model.__class__.__name__}")
-            current_loaded_models[i].model.partially_unload_ram(ram_to_free)
 
     for i in sorted(unloaded_model, reverse=True):
         unloaded_models.append(current_loaded_models.pop(i))
@@ -729,17 +764,27 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
 
 
     total_memory_required = {}
+    total_pins_required = {}
     total_ram_required = {}
     for loaded_model in models_to_load:
-        total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
-        #x2, one to make sure the OS can fit the model for loading in disk cache, and for us to do any pinning we
-        #want to do.
-        #FIXME: This should subtract off the to_load current pin consumption.
-        total_ram_required[loaded_model.device] = total_ram_required.get(loaded_model.device, 0) + loaded_model.model_memory() * 2
+        device = loaded_model.device
+        total_memory_required[device] = total_memory_required.get(device, 0) + loaded_model.model_memory_required(device)
+        resident_memory, model_memory = loaded_model.model_mmap_residency()
+        pinned_memory = loaded_model.model.pinned_memory_size()
+        #FIXME: This can over-free the pins as it budgets to pin the entire model. We should
+        #make this JIT to keep as much pinned as possible.
+        pins_required = model_memory - pinned_memory
+        ram_required = model_memory - resident_memory
+        total_pins_required[device] = total_pins_required.get(device, 0) + pins_required
+        total_ram_required[device] = total_ram_required.get(device, 0) + ram_required
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
-            free_memory(total_memory_required[device] * 1.1 + extra_mem, device, for_dynamic=free_for_dynamic, ram_required=total_ram_required[device])
+            free_memory(total_memory_required[device] * 1.1 + extra_mem,
+                        device,
+                        for_dynamic=free_for_dynamic,
+                        pins_required=total_pins_required[device],
+                        ram_required=total_ram_required[device])
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
@@ -1225,6 +1270,11 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None):
             dest_view = dest_views.pop(0)
             if tensor is None:
                 continue
+            if comfy.memory_management.read_tensor_file_slice_into(tensor, dest_view):
+                continue
+            storage = tensor._qdata.untyped_storage() if isinstance(tensor, comfy.quant_ops.QuantizedTensor) else tensor.untyped_storage()
+            if hasattr(storage, "_comfy_tensor_mmap_touched"):
+                storage._comfy_tensor_mmap_touched = True
             dest_view.copy_(tensor, non_blocking=non_blocking)
 
 

comfy/model_patcher.py

@@ -297,6 +297,9 @@ def model_size(self):
         self.size = comfy.model_management.module_size(self.model)
         return self.size
 
+    def model_mmap_residency(self, free=False):
+        return comfy.model_management.module_mmap_residency(self.model, free=free)
+
     def get_ram_usage(self):
         return self.model_size()
 
@@ -1063,6 +1066,10 @@ def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
 
             return self.model.model_loaded_weight_memory - current_used
 
+    def pinned_memory_size(self):
+        # Pinned memory pressure tracking is only implemented for DynamicVram loading
+        return 0
+
     def partially_unload_ram(self, ram_to_unload):
         pass
 
@@ -1653,6 +1660,16 @@ def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=Fals
 
         return freed
 
+    def pinned_memory_size(self):
+        total = 0
+        loading = self._load_list(for_dynamic=True)
+        for x in loading:
+            _, _, _, _, m, _ = x
+            pin = comfy.pinned_memory.get_pin(m)
+            if pin is not None:
+                total += pin.numel() * pin.element_size()
+        return total
+
     def partially_unload_ram(self, ram_to_unload):
         loading = self._load_list(for_dynamic=True, default_device=self.offload_device)
         for x in loading:

comfy/ops.py

@@ -306,6 +306,33 @@ class CastWeightBiasOp:
     bias_function = []
 
 class disable_weight_init:
+    @staticmethod
+    def _lazy_load_from_state_dict(module, state_dict, prefix, local_metadata,
+                                   missing_keys, unexpected_keys, weight_shape,
+                                   bias_shape=None):
+        assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
+        prefix_len = len(prefix)
+        for k, v in state_dict.items():
+            key = k[prefix_len:]
+            if key == "weight":
+                if not assign_to_params_buffers:
+                    v = v.clone()
+                module.weight = torch.nn.Parameter(v, requires_grad=False)
+            elif bias_shape is not None and key == "bias" and v is not None:
+                if not assign_to_params_buffers:
+                    v = v.clone()
+                module.bias = torch.nn.Parameter(v, requires_grad=False)
+            else:
+                unexpected_keys.append(k)
+
+        if module.weight is None:
+            module.weight = torch.nn.Parameter(torch.zeros(weight_shape), requires_grad=False)
+            missing_keys.append(prefix + "weight")
+
+        if bias_shape is not None and module.bias is None and getattr(module, "comfy_need_lazy_init_bias", False):
+            module.bias = torch.nn.Parameter(torch.zeros(bias_shape), requires_grad=False)
+            missing_keys.append(prefix + "bias")
+
     class Linear(torch.nn.Linear, CastWeightBiasOp):
 
         def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
@@ -333,29 +360,16 @@ def _load_from_state_dict(self, state_dict, prefix, local_metadata,
             if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
                 return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
                                                      missing_keys, unexpected_keys, error_msgs)
-            assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
-            prefix_len = len(prefix)
-            for k,v in state_dict.items():
-                if k[prefix_len:] == "weight":
-                    if not assign_to_params_buffers:
-                        v = v.clone()
-                    self.weight = torch.nn.Parameter(v, requires_grad=False)
-                elif k[prefix_len:] == "bias" and v is not None:
-                    if not assign_to_params_buffers:
-                        v = v.clone()
-                    self.bias = torch.nn.Parameter(v, requires_grad=False)
-                else:
-                    unexpected_keys.append(k)
-
-            #Reconcile default construction of the weight if its missing.
-            if self.weight is None:
-                v = torch.zeros(self.in_features, self.out_features)
-                self.weight = torch.nn.Parameter(v, requires_grad=False)
-                missing_keys.append(prefix+"weight")
-            if self.bias is None and self.comfy_need_lazy_init_bias:
-                v = torch.zeros(self.out_features,)
-                self.bias = torch.nn.Parameter(v, requires_grad=False)
-                missing_keys.append(prefix+"bias")
+            disable_weight_init._lazy_load_from_state_dict(
+                self,
+                state_dict,
+                prefix,
+                local_metadata,
+                missing_keys,
+                unexpected_keys,
+                weight_shape=(self.in_features, self.out_features),
+                bias_shape=(self.out_features,),
+            )
 
 
         def reset_parameters(self):
@@ -547,6 +561,48 @@ def forward(self, *args, **kwargs):
                 return super().forward(*args, **kwargs)
 
     class Embedding(torch.nn.Embedding, CastWeightBiasOp):
+        def __init__(self, num_embeddings, embedding_dim, padding_idx=None, max_norm=None,
+                     norm_type=2.0, scale_grad_by_freq=False, sparse=False, _weight=None,
+                     _freeze=False, device=None, dtype=None):
+            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+                super().__init__(num_embeddings, embedding_dim, padding_idx, max_norm,
+                                 norm_type, scale_grad_by_freq, sparse, _weight,
+                                 _freeze, device, dtype)
+                return
+
+            torch.nn.Module.__init__(self)
+            self.num_embeddings = num_embeddings
+            self.embedding_dim = embedding_dim
+            self.padding_idx = padding_idx
+            self.max_norm = max_norm
+            self.norm_type = norm_type
+            self.scale_grad_by_freq = scale_grad_by_freq
+            self.sparse = sparse
+            # Keep shape/dtype visible for module introspection without reserving storage.
+            embedding_dtype = dtype if dtype is not None else torch.get_default_dtype()
+            self.weight = torch.nn.Parameter(
+                torch.empty((num_embeddings, embedding_dim), device="meta", dtype=embedding_dtype),
+                requires_grad=False,
+            )
+            self.bias = None
+            self.weight_comfy_model_dtype = dtype
+
+        def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                strict, missing_keys, unexpected_keys, error_msgs):
+
+            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+                return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
+                                                     missing_keys, unexpected_keys, error_msgs)
+            disable_weight_init._lazy_load_from_state_dict(
+                self,
+                state_dict,
+                prefix,
+                local_metadata,
+                missing_keys,
+                unexpected_keys,
+                weight_shape=(self.num_embeddings, self.embedding_dim),
+            )
+
         def reset_parameters(self):
             self.bias = None
             return None