sans profiling
This commit is contained in:
RyanOnTheInside
+6
-141
@@ -4,134 +4,6 @@ import comfy.samplers
|
||||
import random
|
||||
import time
|
||||
import os
|
||||
import json
|
||||
from datetime import datetime
|
||||
from functools import wraps
|
||||
import cProfile
|
||||
import pstats
|
||||
import io
|
||||
|
||||
# Simple profiling setup - SINGLE FILE
|
||||
PROFILE_FILE = "stream_sampler_profile.json"
|
||||
PROFILE_DATA = []
|
||||
|
||||
def profile_time(func):
|
||||
@wraps(func)
|
||||
def wrapper(*args, **kwargs):
|
||||
# Start timing
|
||||
start_time = time.time()
|
||||
|
||||
# Check memory before
|
||||
if torch.cuda.is_available():
|
||||
torch.cuda.synchronize()
|
||||
mem_before = torch.cuda.memory_allocated() / (1024 * 1024)
|
||||
|
||||
# Execute the function
|
||||
result = func(*args, **kwargs)
|
||||
|
||||
# End timing
|
||||
if torch.cuda.is_available():
|
||||
torch.cuda.synchronize()
|
||||
end_time = time.time()
|
||||
|
||||
# Check memory after
|
||||
if torch.cuda.is_available():
|
||||
mem_after = torch.cuda.memory_allocated() / (1024 * 1024)
|
||||
mem_diff = mem_after - mem_before
|
||||
else:
|
||||
mem_diff = 0
|
||||
|
||||
exec_time = end_time - start_time
|
||||
|
||||
# Get shape info if available
|
||||
shape_info = "unknown"
|
||||
if len(args) > 1 and hasattr(args[1], 'shape'): # If this is sample(), args[1] is noise
|
||||
shape_info = str(args[1].shape)
|
||||
|
||||
# Log results
|
||||
entry = {
|
||||
'function': func.__name__,
|
||||
'execution_time': exec_time,
|
||||
'memory_diff_mb': mem_diff,
|
||||
'timestamp': datetime.now().strftime('%Y-%m-%d %H:%M:%S'),
|
||||
'shape': shape_info
|
||||
}
|
||||
|
||||
PROFILE_DATA.append(entry)
|
||||
|
||||
# Print to console
|
||||
print(f"PROFILE: {func.__name__} - Time: {exec_time:.4f}s, Memory: {mem_diff:.2f}MB")
|
||||
|
||||
# Save to file after each run
|
||||
save_profile_data()
|
||||
|
||||
return result
|
||||
return wrapper
|
||||
|
||||
|
||||
def profile_cprofile(func):
|
||||
@wraps(func)
|
||||
def wrapper(*args, **kwargs):
|
||||
# Create profiler
|
||||
pr = cProfile.Profile()
|
||||
pr.enable()
|
||||
|
||||
# Execute the function
|
||||
result = func(*args, **kwargs)
|
||||
|
||||
# Disable profiler
|
||||
pr.disable()
|
||||
|
||||
# Get stats
|
||||
s = io.StringIO()
|
||||
ps = pstats.Stats(pr, stream=s).sort_stats('cumulative')
|
||||
ps.print_stats(20)
|
||||
|
||||
# Add to the most recent profile entry
|
||||
if PROFILE_DATA:
|
||||
PROFILE_DATA[-1]['cprofile_data'] = s.getvalue()
|
||||
|
||||
return result
|
||||
return wrapper
|
||||
|
||||
|
||||
def save_profile_data():
|
||||
"""Save all profiling data to a single JSON file"""
|
||||
if not PROFILE_DATA:
|
||||
return
|
||||
|
||||
# Convert to serializable format
|
||||
json_data = []
|
||||
for entry in PROFILE_DATA:
|
||||
serializable_entry = {
|
||||
'function': entry['function'],
|
||||
'execution_time': float(entry['execution_time']),
|
||||
'memory_diff_mb': float(entry['memory_diff_mb']),
|
||||
'timestamp': entry['timestamp'],
|
||||
'shape': entry.get('shape', 'unknown')
|
||||
}
|
||||
|
||||
if 'cprofile_data' in entry:
|
||||
# Only store recent cprofile data to keep file size manageable
|
||||
if len(json_data) < 10 or len(json_data) % 10 == 0:
|
||||
serializable_entry['cprofile_data'] = entry['cprofile_data']
|
||||
|
||||
json_data.append(serializable_entry)
|
||||
|
||||
# Write to a single file
|
||||
with open(PROFILE_FILE, 'w') as f:
|
||||
json.dump(json_data, f, indent=2)
|
||||
|
||||
# Calculate stats
|
||||
sample_times = [x['execution_time'] for x in PROFILE_DATA if x['function'] == 'sample']
|
||||
if sample_times:
|
||||
avg_time = sum(sample_times) / len(sample_times)
|
||||
min_time = min(sample_times)
|
||||
max_time = max(sample_times)
|
||||
print(f"PROFILE SUMMARY: {len(sample_times)} runs, Avg: {avg_time:.4f}s, Min: {min_time:.4f}s, Max: {max_time:.4f}s")
|
||||
|
||||
print(f"Profiling data saved to {PROFILE_FILE}")
|
||||
|
||||
|
||||
class StreamBatchSampler:
|
||||
|
||||
@@ -169,8 +41,6 @@ class StreamBatchSampler:
|
||||
self.working_buffer = None
|
||||
self.output_buffer = None
|
||||
|
||||
@profile_time
|
||||
@profile_cprofile
|
||||
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
|
||||
@@ -279,16 +149,12 @@ class StreamBatchSampler:
|
||||
|
||||
return x_0_pred_out
|
||||
|
||||
@profile_time
|
||||
def update(self, num_steps=4):
|
||||
"""Create sampler with specified settings"""
|
||||
self.num_steps = num_steps
|
||||
sampler = comfy.samplers.KSAMPLER(self.sample)
|
||||
return (sampler,)
|
||||
|
||||
# Print setup info when module is imported
|
||||
print(f"StreamBatchSampler profiling enabled. Results will be saved to {PROFILE_FILE}")
|
||||
|
||||
|
||||
class StreamScheduler:
|
||||
|
||||
@@ -323,7 +189,7 @@ class StreamScheduler:
|
||||
try:
|
||||
t_index_list = [int(t.strip()) for t in t_index_list.split(",")]
|
||||
except ValueError as e:
|
||||
print(f"Error parsing timesteps: {e}. Using default [32,45]")
|
||||
|
||||
t_index_list = [32, 45]
|
||||
|
||||
# Create full schedule using normal scheduler
|
||||
@@ -334,7 +200,7 @@ class StreamScheduler:
|
||||
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:
|
||||
print(f"Warning: timestep {t} out of range [0,{num_inference_steps}), skipping")
|
||||
|
||||
continue
|
||||
selected_sigmas.append(float(full_sigmas[t]))
|
||||
|
||||
@@ -387,7 +253,7 @@ class StreamFrameBuffer:
|
||||
|
||||
if is_empty_latent:
|
||||
self.is_txt2img_mode = True
|
||||
print(f"[StreamFrameBuffer] Detected empty latent for text-to-image mode")
|
||||
|
||||
# Create empty latents with correct dimensions for txt2img
|
||||
# Get dimensions from latent dict
|
||||
height = latent.get("height", 512)
|
||||
@@ -400,7 +266,7 @@ class StreamFrameBuffer:
|
||||
# Create zero tensor with correct shape
|
||||
x = torch.zeros((4, latent_height, latent_width),
|
||||
device=comfy.model_management.get_torch_device())
|
||||
print(f"[StreamFrameBuffer] Created empty latent with shape: {x.shape}")
|
||||
|
||||
elif x.dim() == 4: # [B,C,H,W]
|
||||
self.is_txt2img_mode = False
|
||||
x = x.squeeze(0) # Remove batch dimension -> [C,H,W]
|
||||
@@ -417,18 +283,17 @@ class StreamFrameBuffer:
|
||||
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]
|
||||
print(f"[StreamFrameBuffer] Initialized buffer with {self.buffer_size} copies of frame")
|
||||
|
||||
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
|
||||
print(f"[StreamFrameBuffer] Added new frame to buffer at position {self.buffer_pos}")
|
||||
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
|
||||
print(f"[StreamFrameBuffer] Created batch with shape: {batch.shape}")
|
||||
|
||||
|
||||
# Return as latent dict with preserved dimensions
|
||||
result = {"samples": batch}
|
||||
|
||||
Reference in New Issue
Block a user