Files
ComfyUI/comfy/model_prefetch.py

150 lines
6.1 KiB
Python
Raw Permalink Normal View History

import torch
import weakref
import comfy_aimdo.model_vbar
from comfy.cli_args import args
Threaded Loader performance fixes / improvements (+ Aimdo 0.4.6) (#14116) * memory_management: Add direct to read GPU mode Make destination optional (or make it optionally GPU) and use aimdo to file_read direct to GPU. * ops: Remove stream pin buffers and use aimdo reads This consumed too much RAM and its better to just take the hit on the CPU syncing back the stream on a short ring buffer. Aimdo implements this so just rip the stream pin buffer from comfy. * model_management: all active pin registration movement Its better to just let the active model load past the pin limit as pins and let the pins move around. The saves the HDD and SATA people disk traffic while only costing a few GPU syncs. * utils: use aimdo file handle This opens on windows with more favourable flags * mp: only count the model proper for loaded_ram and vram Exclude live loras from the numbers to avoid the case where the reported loaded memory exceeds the size of the model. This causes me confusion in the Kijai visualizer when it looked fully loaded but was hitting disk due to this accounding disrepency. * utils: add bit reverse utility useful for max scattering something ordered. * pinned_memory: Implement offload balancing Use a max scatter alogorithm to prioritize pins of the same size such that when doing a little bit of offloading it gets scattered, allowing the prefetcher to more evenly swollow the offload. * comfy-aimdo 0.4.7 Aimdo 0.4.7 implement VRAM buffer exhaustion predection to avoid early speculative load of weights that definately wont fix once the inference gets further in. * model-prefetch: consolidate pin ensures on the sync point This could happen mid prefetch block, cause a sync of the entire block and lose overlap. Get ahead of the problem with a free down at the natural compute stream sync point. * mm: Put a 2GB min on the pin ceiling This is reasonably bad if it starts causing swap pressure, moreso than during normal ram-cache proceedings. Clamp it. * add --fast-disk
2026-05-31 05:20:04 +10:00
import comfy.memory_management
import comfy.model_management
import comfy.ops
PREFETCH_QUEUES = []
GRAPH_MODULES = weakref.WeakSet()
GRAPH_WARMED_MODULES = weakref.WeakSet()
GRAPH_CAPTURE_STREAMS = {}
def cleanup_prefetched_modules(module, 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")
if getattr(module, "_v_block_faulted", False):
comfy_aimdo.model_vbar.vbar_unpin(module._v_block)
del module._v_block_faulted
def cleanup_prefetch_queues():
global PREFETCH_QUEUES, GRAPH_CAPTURE_STREAMS
for queue in PREFETCH_QUEUES:
for entry in queue:
if entry is None or not isinstance(entry, tuple):
continue
_, prefetch_state = entry
prefetched_module, comfy_modules = prefetch_state
if comfy_modules is not None:
cleanup_prefetched_modules(prefetched_module, comfy_modules)
PREFETCH_QUEUES = []
for module in GRAPH_MODULES:
del module._comfy_graph
GRAPH_MODULES.clear()
GRAPH_WARMED_MODULES.clear()
GRAPH_CAPTURE_STREAMS = {}
def prefetch_queue_pop(queue, device, module, dtype=None, core=None, enable_graph=False, generator=None):
enable_graph = enable_graph and not args.disable_cuda_graphs and comfy.model_management.is_device_cuda(device) and getattr(module, "_v_block", None) is not None
if queue is None:
if core is not None:
core()
return
capture_stream = None
if enable_graph:
capture_stream = GRAPH_CAPTURE_STREAMS.get(device)
if capture_stream is None:
capture_stream = torch.cuda.Stream(device=device)
GRAPH_CAPTURE_STREAMS[device] = capture_stream
signature = None
graph_hit = False
graph = getattr(module, "_comfy_graph", None) if enable_graph else None
if graph is not None:
signature = comfy_aimdo.model_vbar.vbar_fault(module._v_block)
if signature is not None:
module._v_block_faulted = True
graph_hit = comfy_aimdo.model_vbar.vbar_signature_compare(signature, graph["signature"])
consumed = queue.pop(0)
if consumed is not None:
offload_stream, prefetch_state = consumed
if offload_stream is not None:
offload_stream.wait_stream(comfy.model_management.current_stream(device))
prefetched_module, comfy_modules = prefetch_state
if comfy_modules is not None:
cleanup_prefetched_modules(prefetched_module, comfy_modules)
if graph_hit:
queue[0] = (None, (module, []))
graph["graph"].replay()
return
fully_faulted = False
prefetch = queue[0]
if prefetch is not None:
comfy_modules = []
prefetch_modules = prefetch if isinstance(prefetch, (list, tuple)) else (prefetch,)
for root in prefetch_modules:
for s in root.modules():
if hasattr(s, "_v"):
comfy_modules.append(s)
Threaded Loader performance fixes / improvements (+ Aimdo 0.4.6) (#14116) * memory_management: Add direct to read GPU mode Make destination optional (or make it optionally GPU) and use aimdo to file_read direct to GPU. * ops: Remove stream pin buffers and use aimdo reads This consumed too much RAM and its better to just take the hit on the CPU syncing back the stream on a short ring buffer. Aimdo implements this so just rip the stream pin buffer from comfy. * model_management: all active pin registration movement Its better to just let the active model load past the pin limit as pins and let the pins move around. The saves the HDD and SATA people disk traffic while only costing a few GPU syncs. * utils: use aimdo file handle This opens on windows with more favourable flags * mp: only count the model proper for loaded_ram and vram Exclude live loras from the numbers to avoid the case where the reported loaded memory exceeds the size of the model. This causes me confusion in the Kijai visualizer when it looked fully loaded but was hitting disk due to this accounding disrepency. * utils: add bit reverse utility useful for max scattering something ordered. * pinned_memory: Implement offload balancing Use a max scatter alogorithm to prioritize pins of the same size such that when doing a little bit of offloading it gets scattered, allowing the prefetcher to more evenly swollow the offload. * comfy-aimdo 0.4.7 Aimdo 0.4.7 implement VRAM buffer exhaustion predection to avoid early speculative load of weights that definately wont fix once the inference gets further in. * model-prefetch: consolidate pin ensures on the sync point This could happen mid prefetch block, cause a sync of the entire block and lose overlap. Get ahead of the problem with a free down at the natural compute stream sync point. * mm: Put a 2GB min on the pin ceiling This is reasonably bad if it starts causing swap pressure, moreso than during normal ram-cache proceedings. Clamp it. * add --fast-disk
2026-05-31 05:20:04 +10:00
registerable_size = 0
for s in comfy_modules:
registerable_size += comfy.memory_management.vram_aligned_size([s.weight, s.bias])
for param_key in ("weight", "bias"):
lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
if lowvram_fn is not None:
registerable_size += lowvram_fn.memory_required()
offload_stream, fully_faulted = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True, return_faulted=True)
Threaded Loader performance fixes / improvements (+ Aimdo 0.4.6) (#14116) * memory_management: Add direct to read GPU mode Make destination optional (or make it optionally GPU) and use aimdo to file_read direct to GPU. * ops: Remove stream pin buffers and use aimdo reads This consumed too much RAM and its better to just take the hit on the CPU syncing back the stream on a short ring buffer. Aimdo implements this so just rip the stream pin buffer from comfy. * model_management: all active pin registration movement Its better to just let the active model load past the pin limit as pins and let the pins move around. The saves the HDD and SATA people disk traffic while only costing a few GPU syncs. * utils: use aimdo file handle This opens on windows with more favourable flags * mp: only count the model proper for loaded_ram and vram Exclude live loras from the numbers to avoid the case where the reported loaded memory exceeds the size of the model. This causes me confusion in the Kijai visualizer when it looked fully loaded but was hitting disk due to this accounding disrepency. * utils: add bit reverse utility useful for max scattering something ordered. * pinned_memory: Implement offload balancing Use a max scatter alogorithm to prioritize pins of the same size such that when doing a little bit of offloading it gets scattered, allowing the prefetcher to more evenly swollow the offload. * comfy-aimdo 0.4.7 Aimdo 0.4.7 implement VRAM buffer exhaustion predection to avoid early speculative load of weights that definately wont fix once the inference gets further in. * model-prefetch: consolidate pin ensures on the sync point This could happen mid prefetch block, cause a sync of the entire block and lose overlap. Get ahead of the problem with a free down at the natural compute stream sync point. * mm: Put a 2GB min on the pin ceiling This is reasonably bad if it starts causing swap pressure, moreso than during normal ram-cache proceedings. Clamp it. * add --fast-disk
2026-05-31 05:20:04 +10:00
if not comfy.model_management.args.fast_disk:
comfy.model_management.ensure_pin_registerable(registerable_size)
comfy.model_management.sync_stream(device, offload_stream)
if fully_faulted and dtype is not None:
for comfy_module in comfy_modules:
comfy.ops.resolve_cast_module_with_vbar(comfy_module, dtype, device, dtype, None, False, return_weights=False)
queue[0] = (offload_stream, (module, comfy_modules))
if core is not None:
if enable_graph and fully_faulted and module in GRAPH_WARMED_MODULES:
if signature is None:
signature = comfy_aimdo.model_vbar.vbar_fault(module._v_block)
if signature is not None:
module._v_block_faulted = True
if signature is not None:
graph = torch.cuda.CUDAGraph()
if generator is not None:
graph.register_generator_state(generator)
capture_stream.wait_stream(comfy.model_management.current_stream(device))
with torch.cuda.graph(graph, stream=capture_stream, capture_error_mode="thread_local"):
core()
comfy.model_management.current_stream(device).wait_stream(capture_stream)
graph.replay()
module._comfy_graph = {"graph": graph, "signature": signature}
GRAPH_MODULES.add(module)
return
if capture_stream is None:
core()
else:
capture_stream.wait_stream(comfy.model_management.current_stream(device))
with torch.cuda.stream(capture_stream):
core()
comfy.model_management.current_stream(device).wait_stream(capture_stream)
GRAPH_WARMED_MODULES.add(module)
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