John Pollock
72a20338ef
Refactor memory management in distorch_2.py to patch load_models_gpu instead of LoadedModel.model_memory_required. Implement correct memory reporting based on model flags (eject_models and is_distorch_model), ensuring proper eviction logic and improved handling of virtual VRAM. This drives behavior purely by either comfy core matching or DisTorch flag, fixing potential issues in multi-GPU setups.
762 lines
35 KiB
Python
762 lines
35 KiB
Python
"""
|
|
DisTorch Safetensor Memory Management Module
|
|
Contains all safetensor related code for distributed memory management
|
|
"""
|
|
|
|
import sys
|
|
import torch
|
|
import logging
|
|
import hashlib
|
|
import re
|
|
import gc
|
|
|
|
logger = logging.getLogger("MultiGPU")
|
|
import copy
|
|
import inspect
|
|
from collections import defaultdict
|
|
import comfy.model_management as mm
|
|
import comfy.model_patcher
|
|
from .device_utils import get_device_list, soft_empty_cache_multigpu
|
|
from .model_management_mgpu import multigpu_memory_log, force_full_system_cleanup
|
|
|
|
|
|
safetensor_allocation_store = {}
|
|
safetensor_settings_store = {}
|
|
|
|
|
|
def create_safetensor_model_hash(model, caller):
|
|
"""Create a unique hash for a safetensor model to track allocations"""
|
|
if hasattr(model, 'model'):
|
|
# For ModelPatcher objects
|
|
actual_model = model.model
|
|
model_type = type(actual_model).__name__
|
|
# Use ComfyUI's model_size if available
|
|
if hasattr(model, 'model_size'):
|
|
model_size = model.model_size()
|
|
else:
|
|
model_size = sum(p.numel() * p.element_size() for p in actual_model.parameters())
|
|
if hasattr(model, 'model_state_dict'):
|
|
first_layers = str(list(model.model_state_dict().keys())[:3])
|
|
else:
|
|
first_layers = str(list(actual_model.state_dict().keys())[:3])
|
|
else:
|
|
# Direct model
|
|
model_type = type(model).__name__
|
|
model_size = sum(p.numel() * p.element_size() for p in model.parameters())
|
|
first_layers = str(list(model.state_dict().keys())[:3])
|
|
|
|
identifier = f"{model_type}_{model_size}_{first_layers}"
|
|
final_hash = hashlib.sha256(identifier.encode()).hexdigest()
|
|
|
|
# DEBUG STATEMENT - ALWAYS LOG THE HASH
|
|
logger.debug(f"[MultiGPU DisTorch V2] Created hash for {caller}: {final_hash[:8]}...")
|
|
return final_hash
|
|
|
|
def register_patched_safetensor_modelpatcher():
|
|
"""Register and patch the ModelPatcher for distributed safetensor loading"""
|
|
from comfy.model_patcher import wipe_lowvram_weight, move_weight_functions
|
|
# Patch ComfyUI's ModelPatcher
|
|
if not hasattr(comfy.model_patcher.ModelPatcher, '_distorch_patched'):
|
|
|
|
|
|
# PATCH load_models_gpu with correct memory calculations per model flags
|
|
original_load_models_gpu = mm.load_models_gpu
|
|
|
|
def patched_load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
|
|
from comfy.model_management import cleanup_models_gc, get_free_memory, free_memory, current_loaded_models
|
|
from comfy.model_management import VRAMState, vram_state, lowvram_available, MIN_WEIGHT_MEMORY_RATIO
|
|
from comfy.model_management import minimum_inference_memory, extra_reserved_memory, is_device_cpu
|
|
|
|
multigpu_memory_log("load_models_gpu_top_level", "start")
|
|
|
|
cleanup_models_gc()
|
|
|
|
inference_memory = minimum_inference_memory()
|
|
extra_reserved_mem = extra_reserved_memory()
|
|
memory_required_total = memory_required + extra_reserved_mem
|
|
extra_mem = max(inference_memory, memory_required_total)
|
|
if minimum_memory_required is None:
|
|
minimum_memory_required = extra_mem
|
|
else:
|
|
minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_mem)
|
|
|
|
models_temp = set()
|
|
for m in models:
|
|
models_temp.add(m)
|
|
for mm_patch in m.model_patches_models():
|
|
models_temp.add(mm_patch)
|
|
|
|
models = models_temp
|
|
|
|
models_to_load = []
|
|
|
|
for x in models:
|
|
loaded_model = mm.LoadedModel(x)
|
|
try:
|
|
loaded_model_index = current_loaded_models.index(loaded_model)
|
|
except:
|
|
loaded_model_index = None
|
|
|
|
if loaded_model_index is not None:
|
|
loaded = current_loaded_models[loaded_model_index]
|
|
loaded.currently_used = True
|
|
models_to_load.append(loaded)
|
|
else:
|
|
if hasattr(x, "model"):
|
|
logging.info(f"Requested to load {x.model.__class__.__name__}")
|
|
models_to_load.append(loaded_model)
|
|
|
|
for loaded_model in models_to_load:
|
|
to_unload = []
|
|
for i in range(len(current_loaded_models)):
|
|
if loaded_model.model.is_clone(current_loaded_models[i].model):
|
|
to_unload = [i] + to_unload
|
|
for i in to_unload:
|
|
model_to_unload = current_loaded_models.pop(i)
|
|
model_to_unload.model.detach(unpatch_all=False)
|
|
model_to_unload.model_finalizer.detach()
|
|
|
|
# DisTorch Processing
|
|
total_memory_required = {}
|
|
eject_device = None
|
|
|
|
for loaded_model in models_to_load:
|
|
device = loaded_model.device
|
|
base_memory = loaded_model.model_memory_required(device)
|
|
|
|
# Check DisTorch flags
|
|
is_distorch = hasattr(loaded_model.model.model, '_mgpu_virtual_vram_gb')
|
|
has_eject = hasattr(loaded_model.model.model, '_mgpu_eject_models')
|
|
|
|
if has_eject:
|
|
eject_device = device
|
|
logger.mgpu_mm_log("DisTorch eject_models=True, is_distorch=True - MAX memory eviction")
|
|
|
|
if is_distorch:
|
|
# is_distorch=True: use compute device allocation size
|
|
virtual_vram_gb = loaded_model.model.model._mgpu_virtual_vram_gb
|
|
virtual_vram_bytes = virtual_vram_gb * (1024**3)
|
|
adjusted_memory = max(0, base_memory - virtual_vram_bytes)
|
|
total_memory_required[device] = total_memory_required.get(device, 0) + adjusted_memory
|
|
logger.mgpu_mm_log(f"DisTorch is_distorch=True, model adjusted {(base_memory - virtual_vram_bytes)/(1024**3):.2f}GB for device {device}")
|
|
else:
|
|
# is_distorch=False: use full model size
|
|
total_memory_required[device] = total_memory_required.get(device, 0) + base_memory
|
|
logger.mgpu_mm_log(f"[LOAD_MODELS_GPU] Standard model {(base_memory)/(1024**3):.2f}GB for device {device}")
|
|
|
|
for device in total_memory_required:
|
|
if device != torch.device("cpu"):
|
|
requested_mem = total_memory_required[device] * 1.1 + extra_mem
|
|
logger.mgpu_mm_log(f"[FREE_MEMORY_CALL] Device {device}: requesting {requested_mem/(1024**3):.2f}GB = {total_memory_required[device]/(1024**3):.2f}GB * 1.1 + {extra_mem/(1024**3):.2f}GB inference")
|
|
|
|
|
|
multigpu_memory_log("free_memory", "pre")
|
|
|
|
for device in total_memory_required:
|
|
if device != torch.device("cpu"):
|
|
if device == eject_device:
|
|
total_device_memory = mm.get_total_memory(device)
|
|
logger.mgpu_mm_log(f"[LOAD_MODELS_GPU] eject_models=1, is_distorch=1 → using MAX memory ({total_device_memory/(1024**3):.2f}GB) for eviction")
|
|
free_memory(total_device_memory,device)
|
|
else:
|
|
logger.mgpu_mm_log(f"[LOAD_MODELS_GPU] eject_models=0, using Comfy Core Computed memory ({(total_memory_required[device] * 1.1 + extra_mem)/(1024**3):.2f}GB) for eviction")
|
|
free_memory(total_memory_required[device] * 1.1 + extra_mem, device)
|
|
|
|
multigpu_memory_log("free_memory/minimum_memory_required", "post/pre")
|
|
|
|
for device in total_memory_required:
|
|
if device != torch.device("cpu"):
|
|
free_mem = get_free_memory(device)
|
|
free_mem_gb = free_mem / (1024**3)
|
|
min_required_gb = minimum_memory_required / (1024**3)
|
|
logger.mgpu_mm_log(f"[MIN_MEMORY_CHECK] Device {device}: free={free_mem_gb:.2f}GB, required={min_required_gb:.2f}GB, will_evict={free_mem < minimum_memory_required}")
|
|
|
|
if free_mem < minimum_memory_required:
|
|
models_l = free_memory(minimum_memory_required, device)
|
|
logger.mgpu_mm_log(f"[EVICTION] Device {device}: unloaded {len(models_l)} models due to insufficient memory")
|
|
logging.info("{} models unloaded.".format(len(models_l)))
|
|
|
|
multigpu_memory_log("minimum_memory_required", "post")
|
|
|
|
for loaded_model in models_to_load:
|
|
model = loaded_model.model
|
|
torch_dev = model.load_device
|
|
if is_device_cpu(torch_dev):
|
|
vram_set_state = VRAMState.DISABLED
|
|
else:
|
|
vram_set_state = vram_state
|
|
lowvram_model_memory = 0
|
|
if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM) and not force_full_load:
|
|
loaded_memory = loaded_model.model_loaded_memory()
|
|
current_free_mem = get_free_memory(torch_dev) + loaded_memory
|
|
|
|
lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
|
|
lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory)
|
|
|
|
if vram_set_state == VRAMState.NO_VRAM:
|
|
lowvram_model_memory = 0.1
|
|
|
|
loaded_model.model_load(lowvram_model_memory, force_patch_weights=force_patch_weights)
|
|
current_loaded_models.insert(0, loaded_model)
|
|
|
|
# Replace the module function
|
|
mm.load_models_gpu = patched_load_models_gpu
|
|
|
|
original_partially_load = comfy.model_patcher.ModelPatcher.partially_load
|
|
|
|
def new_partially_load(self, device_to, extra_memory=0, full_load=False, force_patch_weights=False, **kwargs):
|
|
"""Override to use our static device assignments"""
|
|
global safetensor_allocation_store
|
|
|
|
debug_hash = create_safetensor_model_hash(self, "partial_load")
|
|
multigpu_memory_log(f"safetensor:{debug_hash[:8]}", "pre-load")
|
|
allocations = safetensor_allocation_store.get(debug_hash)
|
|
|
|
# Set default precision flag before checking
|
|
if not hasattr(self.model, '_distorch_high_precision_loras'):
|
|
self.model._distorch_high_precision_loras = True
|
|
|
|
if not allocations:
|
|
result = original_partially_load(self, device_to, extra_memory, force_patch_weights)
|
|
multigpu_memory_log(f"safetensor:{debug_hash[:8]}", "post-load")
|
|
if hasattr(self, '_distorch_block_assignments'):
|
|
del self._distorch_block_assignments
|
|
return result
|
|
|
|
if not hasattr(self.model, 'current_weight_patches_uuid'):
|
|
self.model.current_weight_patches_uuid = None
|
|
|
|
unpatch_weights = self.model.current_weight_patches_uuid is not None and (self.model.current_weight_patches_uuid != self.patches_uuid or force_patch_weights)
|
|
|
|
if unpatch_weights:
|
|
logger.debug(f"[MultiGPU DisTorch V2] Patches changed or forced. Unpatching model.")
|
|
self.unpatch_model(self.offload_device, unpatch_weights=True)
|
|
|
|
self.patch_model(load_weights=False)
|
|
|
|
mem_counter = 0
|
|
|
|
is_clip_model = getattr(self, 'is_clip', False)
|
|
device_assignments = analyze_safetensor_loading(self, allocations, is_clip=is_clip_model)
|
|
|
|
model_original_dtype = comfy.utils.weight_dtype(self.model.state_dict())
|
|
high_precision_loras = getattr(self.model, "_distorch_high_precision_loras", True)
|
|
loading = self._load_list()
|
|
loading.sort(reverse=True)
|
|
for module_size, module_name, module_object, params in loading:
|
|
if not unpatch_weights and hasattr(module_object, "comfy_patched_weights") and module_object.comfy_patched_weights == True:
|
|
block_target_device = device_assignments['block_assignments'].get(module_name, device_to)
|
|
current_module_device = None
|
|
try:
|
|
if any(p.numel() > 0 for p in module_object.parameters(recurse=False)):
|
|
current_module_device = next(module_object.parameters(recurse=False)).device
|
|
except StopIteration:
|
|
pass
|
|
|
|
if current_module_device is not None and str(current_module_device) != str(block_target_device):
|
|
logger.info(f"[MultiGPU DisTorch V2] Moving already patched {module_name} to {block_target_device}")
|
|
module_object.to(block_target_device)
|
|
|
|
mem_counter += module_size
|
|
continue
|
|
|
|
# Step 1: Write block/tensor to compute device first
|
|
module_object.to(device_to)
|
|
|
|
# Step 2: Apply LoRa patches while on compute device
|
|
weight_key = "{}.weight".format(module_name)
|
|
bias_key = "{}.bias".format(module_name)
|
|
|
|
if weight_key in self.patches:
|
|
self.patch_weight_to_device(weight_key, device_to=device_to)
|
|
if weight_key in self.weight_wrapper_patches:
|
|
module_object.weight_function.extend(self.weight_wrapper_patches[weight_key])
|
|
|
|
if bias_key in self.patches:
|
|
self.patch_weight_to_device(bias_key, device_to=device_to)
|
|
if bias_key in self.weight_wrapper_patches:
|
|
module_object.bias_function.extend(self.weight_wrapper_patches[bias_key])
|
|
|
|
# Step 3: FP8 casting for CPU storage (if enabled)
|
|
block_target_device = device_assignments['block_assignments'].get(module_name, device_to)
|
|
has_patches = weight_key in self.patches or bias_key in self.patches
|
|
|
|
if not high_precision_loras and block_target_device == "cpu" and has_patches and model_original_dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
|
|
for param_name, param in module_object.named_parameters():
|
|
if param.dtype.is_floating_point:
|
|
cast_data = comfy.float.stochastic_rounding(param.data, torch.float8_e4m3fn)
|
|
new_param = torch.nn.Parameter(cast_data.to(torch.float8_e4m3fn))
|
|
new_param.requires_grad = param.requires_grad
|
|
setattr(module_object, param_name, new_param)
|
|
logger.debug(f"[MultiGPU DisTorch V2] Cast {module_name}.{param_name} to FP8 for CPU storage")
|
|
|
|
# Step 4: Move to ultimate destination based on DisTorch assignment
|
|
if block_target_device != device_to:
|
|
logger.debug(f"[MultiGPU DisTorch V2] Moving {module_name} from {device_to} to {block_target_device}")
|
|
module_object.to(block_target_device)
|
|
module_object.comfy_cast_weights = True
|
|
|
|
# Mark as patched and update memory counter
|
|
module_object.comfy_patched_weights = True
|
|
mem_counter += module_size
|
|
|
|
self.model.current_weight_patches_uuid = self.patches_uuid
|
|
|
|
logger.info("[MultiGPU DisTorch V2] DisTorch loading completed.")
|
|
logger.info(f"[MultiGPU DisTorch V2] Total memory: {mem_counter / (1024 * 1024):.2f}MB")
|
|
multigpu_memory_log(f"safetensor:{debug_hash[:8]}", "post-load")
|
|
|
|
return 0
|
|
|
|
|
|
comfy.model_patcher.ModelPatcher.partially_load = new_partially_load
|
|
comfy.model_patcher.ModelPatcher._distorch_patched = True
|
|
logger.info("[MultiGPU Core Patching] Successfully patched ModelPatcher.partially_load")
|
|
|
|
def _extract_clip_head_blocks(raw_block_list, compute_device):
|
|
"""Identify and pre-assign CLIP head blocks to compute device returning head_blocks, distributable_blocks, block_assignments, and head_memory."""
|
|
head_keywords = ['embed', 'wte', 'wpe', 'token_embedding', 'position_embedding']
|
|
head_blocks = []
|
|
distributable_blocks = []
|
|
head_memory = 0
|
|
block_assignments = {}
|
|
|
|
for module_size, module_name, module_object, params in raw_block_list:
|
|
if any(kw in module_name.lower() for kw in head_keywords):
|
|
head_blocks.append((module_size, module_name, module_object, params))
|
|
block_assignments[module_name] = compute_device
|
|
head_memory += module_size
|
|
else:
|
|
distributable_blocks.append((module_size, module_name, module_object, params))
|
|
|
|
return head_blocks, distributable_blocks, block_assignments, head_memory
|
|
|
|
def analyze_safetensor_loading(model_patcher, allocations_string, is_clip=False):
|
|
"""
|
|
Analyze and distribute safetensor model blocks across devices.
|
|
Supports CLIP head preservation when is_clip=True.
|
|
"""
|
|
DEVICE_RATIOS_DISTORCH = {}
|
|
device_table = {}
|
|
distorch_alloc = allocations_string
|
|
virtual_vram_gb = 0.0
|
|
|
|
distorch_alloc, virtual_vram_str = allocations_string.split('#')
|
|
|
|
compute_device = virtual_vram_str.split(';')[0]
|
|
logger.debug(f"[MultiGPU DisTorch V2] Compute Device: {compute_device}")
|
|
|
|
if not distorch_alloc:
|
|
mode = "fraction"
|
|
distorch_alloc = calculate_safetensor_vvram_allocation(model_patcher, virtual_vram_str)
|
|
|
|
elif any(c in distorch_alloc.lower() for c in ['g', 'm', 'k', 'b']):
|
|
mode = "byte"
|
|
distorch_alloc = calculate_fraction_from_byte_expert_string(model_patcher, distorch_alloc)
|
|
elif "%" in distorch_alloc:
|
|
mode = "ratio"
|
|
distorch_alloc = calculate_fraction_from_ratio_expert_string(model_patcher, distorch_alloc)
|
|
|
|
all_devices = get_device_list()
|
|
present_devices = {item.split(',')[0] for item in distorch_alloc.split(';') if ',' in item}
|
|
for device in all_devices:
|
|
if device not in present_devices:
|
|
distorch_alloc += f";{device},0.0"
|
|
|
|
eq_line = "=" * 50
|
|
dash_line = "-" * 50
|
|
fmt_assign = "{:<18}{:>7}{:>14}{:>10}"
|
|
|
|
logger.info(eq_line)
|
|
logger.info(f"[MultiGPU DisTorch V2] Final Allocation String:\n{distorch_alloc}")
|
|
|
|
for allocation in distorch_alloc.split(';'):
|
|
if ',' not in allocation:
|
|
continue
|
|
dev_name, fraction = allocation.split(',')
|
|
fraction = float(fraction)
|
|
total_mem_bytes = mm.get_total_memory(torch.device(dev_name))
|
|
alloc_gb = (total_mem_bytes * fraction) / (1024**3)
|
|
DEVICE_RATIOS_DISTORCH[dev_name] = alloc_gb
|
|
device_table[dev_name] = {
|
|
"fraction": fraction,
|
|
"total_gb": total_mem_bytes / (1024**3),
|
|
"alloc_gb": alloc_gb
|
|
}
|
|
|
|
logger.info(eq_line)
|
|
logger.info(" DisTorch2 Model Device Allocations")
|
|
logger.info(eq_line)
|
|
|
|
fmt_rosetta = "{:<8}{:>9}{:>9}{:>11}{:>10}"
|
|
logger.info(fmt_rosetta.format("Device", "VRAM GB", "Dev %", "Model GB", "Dist %"))
|
|
logger.info(dash_line)
|
|
|
|
sorted_devices = sorted(device_table.keys(), key=lambda d: (d == "cpu", d))
|
|
|
|
total_allocated_model_bytes = sum(d["alloc_gb"] * (1024**3) for d in device_table.values())
|
|
|
|
for dev in sorted_devices:
|
|
total_dev_gb = device_table[dev]["total_gb"]
|
|
alloc_fraction = device_table[dev]["fraction"]
|
|
alloc_gb = device_table[dev]["alloc_gb"]
|
|
|
|
dist_ratio_percent = (alloc_gb * (1024**3) / total_allocated_model_bytes) * 100 if total_allocated_model_bytes > 0 else 0
|
|
|
|
logger.info(fmt_rosetta.format(
|
|
dev,
|
|
f"{total_dev_gb:.2f}",
|
|
f"{alloc_fraction*100:.1f}%",
|
|
f"{alloc_gb:.2f}",
|
|
f"{dist_ratio_percent:.1f}%"
|
|
))
|
|
|
|
logger.info(dash_line)
|
|
|
|
block_summary = {}
|
|
block_list = []
|
|
memory_by_type = defaultdict(int)
|
|
total_memory = 0
|
|
|
|
raw_block_list = model_patcher._load_list()
|
|
total_memory = sum(module_size for module_size, _, _, _ in raw_block_list)
|
|
|
|
MIN_BLOCK_THRESHOLD = total_memory * 0.0001
|
|
logger.debug(f"[MultiGPU DisTorch V2] Total model memory: {total_memory} bytes")
|
|
logger.debug(f"[MultiGPU DisTorch V2] Tiny block threshold (0.01%): {MIN_BLOCK_THRESHOLD} bytes")
|
|
|
|
# CLIP-specific: Extract head blocks and get pre-assignments
|
|
head_memory = 0
|
|
block_assignments = {}
|
|
if is_clip:
|
|
head_blocks, distributable_raw, block_assignments, head_memory = \
|
|
_extract_clip_head_blocks(raw_block_list, compute_device)
|
|
logger.info(f"[MultiGPU DisTorch V2 CLIP] Preserving {len(head_blocks)} head layer(s) ({head_memory/(1024**2):.2f} MB) on compute device: {compute_device}")
|
|
else:
|
|
distributable_raw = raw_block_list
|
|
|
|
# Build all_blocks list for summary (using full raw_block_list)
|
|
all_blocks = []
|
|
for module_size, module_name, module_object, params in raw_block_list:
|
|
block_type = type(module_object).__name__
|
|
# Populate summary dictionaries
|
|
block_summary[block_type] = block_summary.get(block_type, 0) + 1
|
|
memory_by_type[block_type] += module_size
|
|
all_blocks.append((module_name, module_object, block_type, module_size))
|
|
|
|
# Use distributable blocks for actual allocation (for CLIP, this excludes heads)
|
|
distributable_all_blocks = []
|
|
for module_size, module_name, module_object, params in distributable_raw:
|
|
distributable_all_blocks.append((module_name, module_object, type(module_object).__name__, module_size))
|
|
|
|
block_list = [b for b in distributable_all_blocks if b[3] >= MIN_BLOCK_THRESHOLD]
|
|
tiny_block_list = [b for b in distributable_all_blocks if b[3] < MIN_BLOCK_THRESHOLD]
|
|
|
|
logger.debug(f"[MultiGPU DisTorch V2] Total blocks: {len(all_blocks)}")
|
|
logger.debug(f"[MultiGPU DisTorch V2] Distributable blocks: {len(block_list)}")
|
|
logger.debug(f"[MultiGPU DisTorch V2] Tiny blocks (<0.01%): {len(tiny_block_list)}")
|
|
|
|
logger.info(" DisTorch2 Model Layer Distribution")
|
|
logger.info(dash_line)
|
|
fmt_layer = "{:<18}{:>7}{:>14}{:>10}"
|
|
logger.info(fmt_layer.format("Layer Type", "Layers", "Memory (MB)", "% Total"))
|
|
logger.info(dash_line)
|
|
|
|
for layer_type, count in block_summary.items():
|
|
mem_mb = memory_by_type[layer_type] / (1024 * 1024)
|
|
mem_percent = (memory_by_type[layer_type] / total_memory) * 100 if total_memory > 0 else 0
|
|
logger.info(fmt_layer.format(layer_type[:18], str(count), f"{mem_mb:.2f}", f"{mem_percent:.1f}%"))
|
|
|
|
logger.info(dash_line)
|
|
|
|
# Distribute blocks sequentially from the tail of the model
|
|
|
|
device_assignments = {device: [] for device in DEVICE_RATIOS_DISTORCH.keys()}
|
|
block_assignments = {}
|
|
|
|
# Create a memory quota for each donor device based on its calculated allocation.
|
|
donor_devices = [d for d in sorted_devices]
|
|
donor_quotas = {
|
|
dev: device_table[dev]["alloc_gb"] * (1024**3)
|
|
for dev in donor_devices
|
|
}
|
|
|
|
# CLIP-specific: Adjust compute_device quota to account for locked head blocks
|
|
if is_clip and compute_device in donor_quotas and head_memory > 0:
|
|
donor_quotas[compute_device] = max(0, donor_quotas[compute_device] - head_memory)
|
|
logger.debug(f"[MultiGPU DisTorch V2 CLIP] Adjusted {compute_device} quota by -{head_memory/(1024**2):.2f} MB for head preservation")
|
|
|
|
# Iterate from the TAIL of the model, assigning blocks to donors until their quotas are filled.
|
|
for block_name, module, block_type, block_memory in reversed(block_list):
|
|
assigned_to_donor = False
|
|
for donor in donor_devices:
|
|
if donor_quotas[donor] >= block_memory:
|
|
block_assignments[block_name] = donor
|
|
donor_quotas[donor] -= block_memory
|
|
assigned_to_donor = True
|
|
break # Move to the next block
|
|
|
|
if not assigned_to_donor: #Note - small rounding errors and tensor-fitting on devices make a block occasionally an orphan. We treat orphans the same as tiny_block_list as they are generally small rounding errors
|
|
block_assignments[block_name] = compute_device
|
|
|
|
if tiny_block_list:
|
|
for block_name, module, block_type, block_memory in tiny_block_list:
|
|
block_assignments[block_name] = compute_device
|
|
|
|
# Populate device_assignments from the final block_assignments
|
|
for block_name, device in block_assignments.items():
|
|
# Find the block in the original list to get all its info
|
|
for b_name, b_module, b_type, b_mem in all_blocks:
|
|
if b_name == block_name:
|
|
device_assignments[device].append((b_name, b_module, b_type, b_mem))
|
|
break
|
|
|
|
logger.info("DisTorch2 Model Final Device/Layer Assignments")
|
|
logger.info(dash_line)
|
|
logger.info(fmt_assign.format("Device", "Layers", "Memory (MB)", "% Total"))
|
|
logger.info(dash_line)
|
|
|
|
if tiny_block_list:
|
|
tiny_block_memory = sum(b[3] for b in tiny_block_list)
|
|
tiny_mem_mb = tiny_block_memory / (1024 * 1024)
|
|
tiny_mem_percent = (tiny_block_memory / total_memory) * 100 if total_memory > 0 else 0
|
|
device_label = f"{compute_device} (<0.01%)"
|
|
logger.info(fmt_assign.format(device_label, str(len(tiny_block_list)), f"{tiny_mem_mb:.2f}", f"{tiny_mem_percent:.1f}%"))
|
|
logger.debug(f"[MultiGPU DisTorch V2] Tiny block memory breakdown: {tiny_block_memory} bytes ({tiny_mem_mb:.2f} MB), which is {tiny_mem_percent:.4f}% of total model memory.")
|
|
|
|
total_assigned_memory = 0
|
|
device_memories = {}
|
|
|
|
for device, blocks in device_assignments.items():
|
|
dist_blocks = [b for b in blocks if b[3] >= MIN_BLOCK_THRESHOLD]
|
|
if not dist_blocks:
|
|
continue
|
|
|
|
device_memory = sum(b[3] for b in dist_blocks)
|
|
device_memories[device] = device_memory
|
|
total_assigned_memory += device_memory
|
|
|
|
sorted_assignments = sorted(device_memories.keys(), key=lambda d: (d == "cpu", d))
|
|
|
|
for dev in sorted_assignments:
|
|
# Get only the distributed blocks for the count
|
|
dist_blocks = [b for b in device_assignments[dev] if b[3] >= MIN_BLOCK_THRESHOLD]
|
|
if not dist_blocks:
|
|
continue
|
|
|
|
mem_mb = device_memories[dev] / (1024 * 1024)
|
|
mem_percent = (device_memories[dev] / total_memory) * 100 if total_memory > 0 else 0
|
|
logger.info(fmt_assign.format(dev, str(len(dist_blocks)), f"{mem_mb:.2f}", f"{mem_percent:.1f}%"))
|
|
|
|
logger.info(dash_line)
|
|
|
|
return {
|
|
"device_assignments": device_assignments,
|
|
"block_assignments": block_assignments
|
|
}
|
|
|
|
def parse_memory_string(mem_str):
|
|
"""Parses a memory string (e.g., '4.0g', '512M') and returns bytes."""
|
|
mem_str = mem_str.strip().lower()
|
|
match = re.match(r'(\d+\.?\d*)\s*([gmkb]?)', mem_str)
|
|
if not match:
|
|
raise ValueError(f"Invalid memory string format: {mem_str}")
|
|
|
|
val, unit = match.groups()
|
|
val = float(val)
|
|
|
|
if unit == 'g':
|
|
return val * (1024**3)
|
|
elif unit == 'm':
|
|
return val * (1024**2)
|
|
elif unit == 'k':
|
|
return val * 1024
|
|
else: # b or no unit
|
|
return val
|
|
|
|
def calculate_fraction_from_byte_expert_string(model_patcher, byte_str):
|
|
"""Convert byte allocation string (e.g. 'cuda:1,4gb;cpu,*') to fractional VRAM allocation string respecting device order and byte quotas."""
|
|
raw_block_list = model_patcher._load_list()
|
|
total_model_memory = sum(module_size for module_size, _, _, _ in raw_block_list)
|
|
remaining_model_bytes = total_model_memory
|
|
|
|
# Use a list of tuples to preserve the user-defined order
|
|
parsed_allocations = []
|
|
wildcard_device = "cpu" # Default wildcard device
|
|
|
|
for allocation in byte_str.split(';'):
|
|
if ',' not in allocation:
|
|
continue
|
|
dev_name, val_str = allocation.split(',', 1)
|
|
is_wildcard = '*' in val_str
|
|
|
|
if is_wildcard:
|
|
wildcard_device = dev_name
|
|
# Don't add wildcard to the priority list yet
|
|
else:
|
|
byte_val = parse_memory_string(val_str)
|
|
parsed_allocations.append({'device': dev_name, 'bytes': byte_val})
|
|
|
|
final_byte_allocations = defaultdict(int)
|
|
|
|
# Process devices with specific byte allocations first, in order
|
|
for alloc in parsed_allocations:
|
|
dev = alloc['device']
|
|
requested_bytes = alloc['bytes']
|
|
|
|
# Determine the actual bytes to allocate to this device
|
|
bytes_to_assign = min(requested_bytes, remaining_model_bytes)
|
|
|
|
if bytes_to_assign > 0:
|
|
final_byte_allocations[dev] = bytes_to_assign
|
|
remaining_model_bytes -= bytes_to_assign
|
|
logger.info(f"[MultiGPU DisTorch V2] Assigning {bytes_to_assign / (1024**2):.2f}MB of model to {dev} (requested {requested_bytes / (1024**2):.2f}MB).")
|
|
|
|
if remaining_model_bytes <= 0:
|
|
logger.info("[MultiGPU DisTorch V2] All model blocks have been allocated. Subsequent devices in the string will receive no assignment.")
|
|
break
|
|
|
|
# Assign any leftover model bytes to the wildcard device
|
|
if remaining_model_bytes > 0:
|
|
final_byte_allocations[wildcard_device] += remaining_model_bytes
|
|
logger.info(f"[MultiGPU DisTorch V2] Assigning remaining {remaining_model_bytes / (1024**2):.2f}MB of model to wildcard device '{wildcard_device}'.")
|
|
|
|
# Convert the final byte allocations to VRAM fractions
|
|
allocation_parts = []
|
|
for dev, bytes_alloc in final_byte_allocations.items():
|
|
total_device_vram = mm.get_total_memory(torch.device(dev))
|
|
if total_device_vram > 0:
|
|
fraction = bytes_alloc / total_device_vram
|
|
allocation_parts.append(f"{dev},{fraction:.4f}")
|
|
|
|
allocations_string = ";".join(allocation_parts)
|
|
|
|
return allocations_string
|
|
|
|
def calculate_fraction_from_ratio_expert_string(model_patcher, ratio_str):
|
|
"""Convert ratio allocation string (e.g. 'cuda:0,25%;cpu,75%') describing model split to fractional VRAM allocation string."""
|
|
raw_block_list = model_patcher._load_list()
|
|
total_model_memory = sum(module_size for module_size, _, _, _ in raw_block_list)
|
|
|
|
raw_ratios = {}
|
|
for allocation in ratio_str.split(';'):
|
|
if ',' not in allocation: continue
|
|
dev_name, val_str = allocation.split(',', 1)
|
|
# Assumes the value is a unitless ratio number, ignores '%' for simplicity.
|
|
value = float(val_str.replace('%','').strip())
|
|
raw_ratios[dev_name] = value
|
|
|
|
total_ratio_parts = sum(raw_ratios.values())
|
|
allocation_parts = []
|
|
|
|
for dev, ratio_val in raw_ratios.items():
|
|
bytes_of_model_for_device = (ratio_val / total_ratio_parts) * total_model_memory
|
|
|
|
total_vram_of_device = mm.get_total_memory(torch.device(dev))
|
|
|
|
if total_vram_of_device > 0:
|
|
required_fraction = bytes_of_model_for_device / total_vram_of_device
|
|
allocation_parts.append(f"{dev},{required_fraction:.4f}")
|
|
|
|
ratio_values = [str(v) for v in raw_ratios.values()]
|
|
ratio_string = ":".join(ratio_values)
|
|
|
|
normalized_pcts = [(v / total_ratio_parts) * 100 for v in raw_ratios.values()]
|
|
|
|
put_parts = []
|
|
for i, dev_name in enumerate(raw_ratios.keys()):
|
|
put_parts.append(f"{int(normalized_pcts[i])}% on {dev_name}")
|
|
|
|
if len(put_parts) == 1:
|
|
put_part = put_parts[0]
|
|
elif len(put_parts) == 2:
|
|
put_part = f"{put_parts[0]} and {put_parts[1]}"
|
|
else:
|
|
put_part = ", ".join(put_parts[:-1]) + f", and {put_parts[-1]}"
|
|
|
|
logger.info(f"[MultiGPU DisTorch V2] Ratio(%) Mode - {ratio_str} -> {ratio_string} ratio, put {put_part}")
|
|
|
|
allocations_string = ";".join(allocation_parts)
|
|
|
|
return allocations_string
|
|
|
|
def calculate_safetensor_vvram_allocation(model_patcher, virtual_vram_str):
|
|
"""Calculate virtual VRAM allocation string for distributed safetensor loading"""
|
|
recipient_device, vram_amount, donors = virtual_vram_str.split(';')
|
|
virtual_vram_gb = float(vram_amount)
|
|
|
|
eq_line = "=" * 47
|
|
dash_line = "-" * 47
|
|
fmt_assign = "{:<8} {:<6} {:>11} {:>9} {:>9}"
|
|
|
|
logger.info(eq_line)
|
|
logger.info(" DisTorch2 Model Virtual VRAM Analysis")
|
|
logger.info(eq_line)
|
|
logger.info(fmt_assign.format("Object", "Role", "Original(GB)", "Total(GB)", "Virt(GB)"))
|
|
logger.info(dash_line)
|
|
|
|
# Calculate recipient VRAM
|
|
recipient_vram = mm.get_total_memory(torch.device(recipient_device)) / (1024**3)
|
|
recipient_virtual = recipient_vram + virtual_vram_gb
|
|
|
|
logger.info(fmt_assign.format(recipient_device, 'recip', f"{recipient_vram:.2f}GB",f"{recipient_virtual:.2f}GB", f"+{virtual_vram_gb:.2f}GB"))
|
|
|
|
# Handle donor devices
|
|
ram_donors = [d for d in donors.split(',')]
|
|
remaining_vram_needed = virtual_vram_gb
|
|
|
|
donor_device_info = {}
|
|
donor_allocations = {}
|
|
|
|
for donor in ram_donors:
|
|
donor_vram = mm.get_total_memory(torch.device(donor)) / (1024**3)
|
|
max_donor_capacity = donor_vram
|
|
|
|
donation = min(remaining_vram_needed, max_donor_capacity)
|
|
donor_virtual = donor_vram - donation
|
|
remaining_vram_needed -= donation
|
|
donor_allocations[donor] = donation
|
|
|
|
donor_device_info[donor] = (donor_vram, donor_virtual)
|
|
logger.info(fmt_assign.format(donor, 'donor', f"{donor_vram:.2f}GB", f"{donor_virtual:.2f}GB", f"-{donation:.2f}GB"))
|
|
|
|
|
|
logger.info(dash_line)
|
|
|
|
# Calculate model size
|
|
model = model_patcher.model if hasattr(model_patcher, 'model') else model_patcher
|
|
total_memory = 0
|
|
|
|
for name, module in model.named_modules():
|
|
if hasattr(module, "weight"):
|
|
if module.weight is not None:
|
|
total_memory += module.weight.numel() * module.weight.element_size()
|
|
if hasattr(module, "bias") and module.bias is not None:
|
|
total_memory += module.bias.numel() * module.bias.element_size()
|
|
|
|
model_size_gb = total_memory / (1024**3)
|
|
new_model_size_gb = max(0, model_size_gb - virtual_vram_gb)
|
|
|
|
logger.info(fmt_assign.format('model', 'model', f"{model_size_gb:.2f}GB",f"{new_model_size_gb:.2f}GB", f"-{virtual_vram_gb:.2f}GB"))
|
|
|
|
# Warning if model too large
|
|
if model_size_gb > (recipient_vram * 0.9):
|
|
required_offload_gb = model_size_gb - (recipient_vram * 0.9)
|
|
logger.warning(f"\n\n[MultiGPU DisTorch V2] Model size ({model_size_gb:.2f}GB) is larger than 90% of available VRAM on: {recipient_device} ({recipient_vram * 0.9:.2f}GB).")
|
|
logger.warning(f"[MultiGPU DisTorch V2] To prevent an OOM error, set 'virtual_vram_gb' to at least {required_offload_gb:.2f}.\n\n")
|
|
|
|
new_on_recipient = max(0, model_size_gb - virtual_vram_gb)
|
|
|
|
# Build allocation string
|
|
allocation_parts = []
|
|
recipient_percent = new_on_recipient / recipient_vram
|
|
allocation_parts.append(f"{recipient_device},{recipient_percent:.4f}")
|
|
|
|
for donor in ram_donors:
|
|
donor_vram = donor_device_info[donor][0]
|
|
donor_percent = donor_allocations[donor] / donor_vram
|
|
allocation_parts.append(f"{donor},{donor_percent:.4f}")
|
|
|
|
allocations_string = ";".join(allocation_parts)
|
|
return allocations_string
|