Merge branch 'master' into feat/api-nodes/TopazVideo-Astra2

Author: Kosinkadink

comfy/ldm/lightricks/av_model.py

@@ -16,6 +16,7 @@
 from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
 from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
 import comfy.ldm.common_dit
+import comfy.model_prefetch
 
 class CompressedTimestep:
     """Store video timestep embeddings in compressed form using per-frame indexing."""
@@ -907,9 +908,11 @@ def _process_transformer_blocks(
         """Process transformer blocks for LTXAV."""
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        prefetch_queue = comfy.model_prefetch.make_prefetch_queue(list(self.transformer_blocks), vx.device, transformer_options)
 
         # Process transformer blocks
         for i, block in enumerate(self.transformer_blocks):
+            comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, block)
             if ("double_block", i) in blocks_replace:
 
                 def block_wrap(args):
@@ -982,6 +985,8 @@ def block_wrap(args):
                     a_prompt_timestep=a_prompt_timestep,
                 )
 
+        comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, None)
+
         return [vx, ax]
 
     def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):

comfy/lora.py

@@ -17,6 +17,7 @@
 """
 
 from __future__ import annotations
+import comfy.memory_management
 import comfy.utils
 import comfy.model_management
 import comfy.model_base
@@ -473,3 +474,17 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
             weight = old_weight
 
     return weight
+
+def prefetch_prepared_value(value, allocate_buffer, stream):
+    if isinstance(value, torch.Tensor):
+        dest = allocate_buffer(comfy.memory_management.vram_aligned_size(value))
+        comfy.model_management.cast_to_gathered([value], dest, non_blocking=True, stream=stream)
+        return comfy.memory_management.interpret_gathered_like([value], dest)[0]
+    elif isinstance(value, weight_adapter.WeightAdapterBase):
+        return type(value)(value.loaded_keys, prefetch_prepared_value(value.weights, allocate_buffer, stream))
+    elif isinstance(value, tuple):
+        return tuple(prefetch_prepared_value(item, allocate_buffer, stream) for item in value)
+    elif isinstance(value, list):
+        return [prefetch_prepared_value(item, allocate_buffer, stream) for item in value]
+
+    return value

comfy/model_base.py

@@ -214,6 +214,11 @@ def _apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, tran
         if "latent_shapes" in extra_conds:
             xc = utils.unpack_latents(xc, extra_conds.pop("latent_shapes"))
 
+        transformer_options = transformer_options.copy()
+        transformer_options["prefetch_dynamic_vbars"] = (
+            self.current_patcher is not None and self.current_patcher.is_dynamic()
+        )
+
         model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds)
         if len(model_output) > 1 and not torch.is_tensor(model_output):
             model_output, _ = utils.pack_latents(model_output)

comfy/model_management.py

@@ -31,6 +31,7 @@
 import comfy.memory_management
 import comfy.utils
 import comfy.quant_ops
+import comfy_aimdo.vram_buffer
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -1175,6 +1176,10 @@ def current_stream(device):
 
 STREAM_CAST_BUFFERS = {}
 LARGEST_CASTED_WEIGHT = (None, 0)
+STREAM_AIMDO_CAST_BUFFERS = {}
+LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+
+DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE = 16 * 1024 ** 3
 
 def get_cast_buffer(offload_stream, device, size, ref):
     global LARGEST_CASTED_WEIGHT
@@ -1208,13 +1213,26 @@ def get_cast_buffer(offload_stream, device, size, ref):
 
     return cast_buffer
 
+def get_aimdo_cast_buffer(offload_stream, device):
+    cast_buffer = STREAM_AIMDO_CAST_BUFFERS.get(offload_stream, None)
+    if cast_buffer is None:
+        cast_buffer = comfy_aimdo.vram_buffer.VRAMBuffer(DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE, device.index)
+        STREAM_AIMDO_CAST_BUFFERS[offload_stream] = cast_buffer
+
+    return cast_buffer
 def reset_cast_buffers():
     global LARGEST_CASTED_WEIGHT
+    global LARGEST_AIMDO_CASTED_WEIGHT
+
     LARGEST_CASTED_WEIGHT = (None, 0)
-    for offload_stream in STREAM_CAST_BUFFERS:
-        offload_stream.synchronize()
+    LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+    for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS):
+        if offload_stream is not None:
+            offload_stream.synchronize()
     synchronize()
+
     STREAM_CAST_BUFFERS.clear()
+    STREAM_AIMDO_CAST_BUFFERS.clear()
     soft_empty_cache()
 
 def get_offload_stream(device):

comfy/model_patcher.py

@@ -121,9 +121,20 @@ def __init__(self, key, patches, convert_func=None, set_func=None):
         self.patches = patches
         self.convert_func = convert_func # TODO: remove
         self.set_func = set_func
+        self.prepared_patches = None
+
+    def prepare(self, allocate_buffer, stream):
+        self.prepared_patches = [
+            (patch[0], comfy.lora.prefetch_prepared_value(patch[1], allocate_buffer, stream), patch[2], patch[3], patch[4])
+            for patch in self.patches[self.key]
+        ]
+
+    def clear_prepared(self):
+        self.prepared_patches = None
 
     def __call__(self, weight):
-        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+        patches = self.prepared_patches if self.prepared_patches is not None else self.patches[self.key]
+        return comfy.lora.calculate_weight(patches, weight, self.key, intermediate_dtype=weight.dtype)
 
 LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2
 

comfy/model_prefetch.py

@@ -0,0 +1,65 @@
+import comfy_aimdo.model_vbar
+import comfy.model_management
+import comfy.ops
+
+PREFETCH_QUEUES = []
+
+def cleanup_prefetched_modules(comfy_modules):
+    for s in comfy_modules:
+        prefetch = getattr(s, "_prefetch", None)
+        if prefetch is None:
+            continue
+        for param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                lowvram_fn.clear_prepared()
+        if prefetch["signature"] is not None:
+            comfy_aimdo.model_vbar.vbar_unpin(s._v)
+        delattr(s, "_prefetch")
+
+def cleanup_prefetch_queues():
+    global PREFETCH_QUEUES
+
+    for queue in PREFETCH_QUEUES:
+        for entry in queue:
+            if entry is None or not isinstance(entry, tuple):
+                continue
+            _, prefetch_state = entry
+            comfy_modules = prefetch_state[1]
+            if comfy_modules is not None:
+                cleanup_prefetched_modules(comfy_modules)
+    PREFETCH_QUEUES = []
+
+def prefetch_queue_pop(queue, device, module):
+    if queue is None:
+        return
+
+    consumed = queue.pop(0)
+    if consumed is not None:
+        offload_stream, prefetch_state = consumed
+        offload_stream.wait_stream(comfy.model_management.current_stream(device))
+        _, comfy_modules = prefetch_state
+        if comfy_modules is not None:
+            cleanup_prefetched_modules(comfy_modules)
+
+    prefetch = queue[0]
+    if prefetch is not None:
+        comfy_modules = []
+        for s in prefetch.modules():
+            if hasattr(s, "_v"):
+                comfy_modules.append(s)
+
+        offload_stream = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True)
+        comfy.model_management.sync_stream(device, offload_stream)
+        queue[0] = (offload_stream, (prefetch, comfy_modules))
+
+def make_prefetch_queue(queue, device, transformer_options):
+    if (not transformer_options.get("prefetch_dynamic_vbars", False)
+        or comfy.model_management.NUM_STREAMS == 0
+        or comfy.model_management.is_device_cpu(device)
+        or not comfy.model_management.device_supports_non_blocking(device)):
+        return None
+
+    queue = [None] + queue + [None]
+    PREFETCH_QUEUES.append(queue)
+    return queue