307 lines
13 KiB
Python
307 lines
13 KiB
Python
import torch
|
|
import comfy.model_management
|
|
import comfy.samplers
|
|
import random
|
|
import time
|
|
import os
|
|
|
|
class StreamBatchSampler:
|
|
|
|
|
|
RETURN_TYPES = ("SAMPLER",)
|
|
FUNCTION = "update"
|
|
CATEGORY = "StreamPack/sampling"
|
|
DESCRIPTION = "Implements batched denoising for faster inference by processing multiple frames in parallel at different denoising steps. Also adds temportal consistency to the denoising process."
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"num_steps": ("INT", {
|
|
"default": 4,
|
|
"min": 1,
|
|
"max": 10,
|
|
"step": 1,
|
|
"tooltip": "Number of denoising steps. Should match the frame buffer size."
|
|
}),
|
|
},
|
|
}
|
|
|
|
def __init__(self):
|
|
self.num_steps = None
|
|
self.frame_buffer = []
|
|
self.x_t_latent_buffer = None
|
|
self.stock_noise = None
|
|
self.is_txt2img_mode = False
|
|
|
|
# Initialize all optimization buffers as None
|
|
self.zeros_reference = None
|
|
self.random_noise_buffer = None
|
|
self.sigmas_view_buffer = None
|
|
self.expanded_stock_noise = None
|
|
self.working_buffer = None
|
|
self.output_buffer = None
|
|
|
|
def sample(self, model, noise, sigmas, extra_args=None, callback=None, disable=None):
|
|
"""Sample with staggered batch denoising steps - Optimized version"""
|
|
extra_args = {} if extra_args is None else extra_args
|
|
|
|
# Get number of frames in batch and available sigmas
|
|
batch_size = noise.shape[0]
|
|
num_sigmas = len(sigmas) - 1 # Subtract 1 because last sigma is the target (0.0)
|
|
|
|
# Optimization 1: Reuse zeros buffer for txt2img detection
|
|
if self.zeros_reference is None:
|
|
# We only need a small reference tensor for comparison, not a full tensor
|
|
self.zeros_reference = torch.zeros(1, device=noise.device, dtype=noise.dtype)
|
|
|
|
# Check if noise tensor is all zeros - functionally identical but more efficient
|
|
self.is_txt2img_mode = torch.abs(noise).sum() < 1e-5
|
|
|
|
# Noise handling with memory optimization
|
|
if self.is_txt2img_mode:
|
|
# Optimization 2: If txt2img mode, reuse the noise tensor directly
|
|
# instead of allocating new memory
|
|
if self.random_noise_buffer is None or self.random_noise_buffer.shape != noise.shape:
|
|
self.random_noise_buffer = torch.empty_like(noise)
|
|
|
|
# Generate random noise in-place
|
|
self.random_noise_buffer.normal_()
|
|
x = self.random_noise_buffer # Use pre-allocated buffer
|
|
else:
|
|
# If not txt2img, we'll still need to add noise later
|
|
x = noise # No need to copy, will add noise later
|
|
|
|
# Verify batch size matches number of timesteps
|
|
if batch_size != num_sigmas:
|
|
raise ValueError(f"Batch size ({batch_size}) must match number of timesteps ({num_sigmas})")
|
|
|
|
# Pre-compute alpha and beta terms
|
|
alpha_prod_t = (sigmas[:-1] / sigmas[0]).view(-1, 1, 1, 1) # [B,1,1,1]
|
|
beta_prod_t = (1 - alpha_prod_t)
|
|
|
|
# Optimization 3: Initialize stock noise with reuse
|
|
if self.stock_noise is None or self.stock_noise.shape != noise[0].shape:
|
|
self.stock_noise = torch.empty_like(noise[0])
|
|
self.stock_noise.normal_() # Generate random noise in-place
|
|
|
|
# Optimization 4: Pre-allocate and reuse view buffer for sigmas
|
|
if self.sigmas_view_buffer is None or self.sigmas_view_buffer.shape[0] != len(sigmas)-1:
|
|
self.sigmas_view_buffer = torch.empty((len(sigmas)-1, 1, 1, 1),
|
|
device=sigmas.device,
|
|
dtype=sigmas.dtype)
|
|
# In-place copy of sigmas view
|
|
self.sigmas_view_buffer.copy_(sigmas[:-1].view(-1, 1, 1, 1))
|
|
|
|
# Optimization 5: Eliminate unsqueeze allocation by pre-expanding stock noise
|
|
if self.expanded_stock_noise is None or self.expanded_stock_noise.shape[0] != batch_size:
|
|
self.expanded_stock_noise = self.stock_noise.expand(batch_size, *self.stock_noise.shape)
|
|
|
|
# Apply noise with pre-allocated buffers - no new memory allocation
|
|
if not self.is_txt2img_mode: # Already handled txt2img case above
|
|
# If we need a working buffer separate from noise input:
|
|
if id(x) == id(noise): # They're the same object, need a separate buffer
|
|
if self.working_buffer is None or self.working_buffer.shape != noise.shape:
|
|
self.working_buffer = torch.empty_like(noise)
|
|
x = self.working_buffer
|
|
# Add noise to input
|
|
torch.add(noise, self.expanded_stock_noise * self.sigmas_view_buffer, out=x)
|
|
|
|
# Initialize and manage latent buffer with memory optimization
|
|
if (self.x_t_latent_buffer is None or self.is_txt2img_mode) and num_sigmas > 1:
|
|
# Optimization 6: Pre-allocate or resize as needed
|
|
if self.x_t_latent_buffer is None or self.x_t_latent_buffer.shape != x[0].shape:
|
|
self.x_t_latent_buffer = torch.empty_like(x[0])
|
|
# In-place copy instead of clone
|
|
self.x_t_latent_buffer.copy_(x[0])
|
|
|
|
# Use buffer for first frame to maintain temporal consistency
|
|
if num_sigmas > 1:
|
|
# In-place update - no new allocation
|
|
x[0].copy_(self.x_t_latent_buffer)
|
|
|
|
# Run model on entire batch at once
|
|
with torch.no_grad():
|
|
# Process all frames in parallel
|
|
sigma_batch = sigmas[:-1]
|
|
|
|
denoised_batch = model(x, sigma_batch, **extra_args)
|
|
|
|
# Update buffer with intermediate results
|
|
if num_sigmas > 1:
|
|
# Store result from first frame as buffer for next iteration
|
|
self.x_t_latent_buffer.copy_(denoised_batch[0]) # In-place update
|
|
|
|
# Optimization 7: Pre-allocate output buffer
|
|
if self.output_buffer is None or self.output_buffer.shape != (1, *denoised_batch[-1].shape):
|
|
self.output_buffer = torch.empty(1, *denoised_batch[-1].shape,
|
|
device=denoised_batch.device,
|
|
dtype=denoised_batch.dtype)
|
|
# Copy the result directly to pre-allocated buffer
|
|
self.output_buffer[0].copy_(denoised_batch[-1])
|
|
x_0_pred_out = self.output_buffer
|
|
else:
|
|
x_0_pred_out = denoised_batch
|
|
self.x_t_latent_buffer = None
|
|
|
|
# Call callback if provided
|
|
if callback is not None:
|
|
callback({'x': x_0_pred_out, 'i': 0, 'sigma': sigmas[0], 'sigma_hat': sigmas[0], 'denoised': denoised_batch[-1:]})
|
|
|
|
return x_0_pred_out
|
|
|
|
def update(self, num_steps=4):
|
|
"""Create sampler with specified settings"""
|
|
self.num_steps = num_steps
|
|
sampler = comfy.samplers.KSAMPLER(self.sample)
|
|
return (sampler,)
|
|
|
|
|
|
class StreamScheduler:
|
|
|
|
RETURN_TYPES = ("SIGMAS",)
|
|
FUNCTION = "update"
|
|
CATEGORY = "StreamPack/sampling"
|
|
DESCRIPTION = "Implements StreamDiffusion's efficient timestep selection. Use in conjunction with StreamBatchSampler."
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"model": ("MODEL",),
|
|
"t_index_list": ("STRING", {
|
|
"default": "32,45",
|
|
"tooltip": "Comma-separated list of timesteps to actually use for denoising. Examples: '32,45' for img2img or '0,16,32,45' for txt2img"
|
|
}),
|
|
"num_inference_steps": ("INT", {
|
|
"default": 50,
|
|
"min": 1,
|
|
"max": 1000,
|
|
"step": 1,
|
|
"tooltip": "Total number of timesteps in schedule. StreamDiffusion uses 50 by default. Only timesteps specified in t_index_list are actually used."
|
|
}),
|
|
},
|
|
}
|
|
|
|
def update(self, model, t_index_list="32,45", num_inference_steps=50):
|
|
# Get model's sampling parameters
|
|
model_sampling = model.get_model_object("model_sampling")
|
|
|
|
# Parse timestep list
|
|
try:
|
|
t_index_list = [int(t.strip()) for t in t_index_list.split(",")]
|
|
except ValueError as e:
|
|
|
|
t_index_list = [32, 45]
|
|
|
|
# Create full schedule using normal scheduler
|
|
full_sigmas = comfy.samplers.normal_scheduler(model_sampling, num_inference_steps)
|
|
|
|
# Select only the sigmas at our desired indices, but in reverse order
|
|
# This ensures we go from high noise to low noise
|
|
selected_sigmas = []
|
|
for t in sorted(t_index_list, reverse=True): # Sort in reverse to go from high noise to low
|
|
if t < 0 or t >= num_inference_steps:
|
|
|
|
continue
|
|
selected_sigmas.append(float(full_sigmas[t]))
|
|
|
|
# Add final sigma
|
|
selected_sigmas.append(0.0)
|
|
|
|
# Convert to tensor and move to appropriate device
|
|
selected_sigmas = torch.FloatTensor(selected_sigmas).to(comfy.model_management.get_torch_device())
|
|
return (selected_sigmas,)
|
|
|
|
|
|
class StreamFrameBuffer:
|
|
|
|
|
|
RETURN_TYPES = ("LATENT",)
|
|
FUNCTION = "update"
|
|
CATEGORY = "StreamPack/sampling"
|
|
DESCRIPTION = "Accumulates frames to enable staggered batch denoising like StreamDiffusion. Use in conjunction with StreamBatchSampler"
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"latent": ("LATENT",),
|
|
"buffer_size": ("INT", {
|
|
"default": 4,
|
|
"min": 1,
|
|
"max": 10,
|
|
"step": 1,
|
|
"tooltip": "Number of frames to buffer before starting batch processing. Should match number of denoising steps."
|
|
}),
|
|
},
|
|
}
|
|
|
|
def __init__(self):
|
|
self.frame_buffer = None # Tensor of shape [buffer_size, C, H, W]
|
|
self.buffer_size = None
|
|
self.buffer_pos = 0 # Current position in ring buffer
|
|
self.is_initialized = False # Track buffer initialization
|
|
self.is_txt2img_mode = False
|
|
|
|
def update(self, latent, buffer_size=4):
|
|
"""Add new frame to buffer and return batch when ready"""
|
|
self.buffer_size = buffer_size
|
|
|
|
# Extract latent tensor from input and remove batch dimension if present
|
|
x = latent["samples"]
|
|
|
|
# Check if this is an empty latent (for txt2img)
|
|
is_empty_latent = x.numel() == 0 or (x.dim() > 0 and x.shape[0] == 0)
|
|
|
|
if is_empty_latent:
|
|
self.is_txt2img_mode = True
|
|
|
|
# Create empty latents with correct dimensions for txt2img
|
|
# Get dimensions from latent dict
|
|
height = latent.get("height", 512)
|
|
width = latent.get("width", 512)
|
|
|
|
# Calculate latent dimensions (typically 1/8 of image dimensions for SD)
|
|
latent_height = height // 8
|
|
latent_width = width // 8
|
|
|
|
# Create zero tensor with correct shape
|
|
x = torch.zeros((4, latent_height, latent_width),
|
|
device=comfy.model_management.get_torch_device())
|
|
|
|
elif x.dim() == 4: # [B,C,H,W]
|
|
self.is_txt2img_mode = False
|
|
x = x.squeeze(0) # Remove batch dimension -> [C,H,W]
|
|
|
|
# Optimization: Initialize or resize frame_buffer as a tensor
|
|
if not self.is_initialized or self.frame_buffer.shape[0] != self.buffer_size or \
|
|
self.frame_buffer.shape[1:] != x.shape:
|
|
# Pre-allocate buffer with correct shape
|
|
self.frame_buffer = torch.zeros(
|
|
(self.buffer_size, *x.shape),
|
|
device=x.device,
|
|
dtype=x.dtype
|
|
)
|
|
if self.is_txt2img_mode or not self.is_initialized:
|
|
# Optimization: Use broadcasting to fill buffer with copies
|
|
self.frame_buffer[:] = x.unsqueeze(0) # Broadcast x to [buffer_size, C, H, W]
|
|
|
|
self.is_initialized = True
|
|
self.buffer_pos = 0
|
|
else:
|
|
# Add new frame to buffer using ring buffer logic
|
|
self.frame_buffer[self.buffer_pos] = x # In-place update
|
|
self.buffer_pos = (self.buffer_pos + 1) % self.buffer_size # Circular increment
|
|
|
|
# Optimization: frame_buffer is already a tensor batch, no need to stack
|
|
batch = self.frame_buffer
|
|
|
|
|
|
# Return as latent dict with preserved dimensions
|
|
result = {"samples": batch}
|
|
|
|
# Preserve height and width if present in input
|
|
if "height" in latent:
|
|
result["height"] = latent["height"]
|
|
if "width" in latent:
|
|
result["width"] = latent["width"]
|
|
|
|
return (result,) |