nodes/Runware-ComfyUI-Runware
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

feat: Replace TeaCache/DeepCache with unified Accelerator Options

Browse Source 
- Remove individual TeaCache and DeepCache modules
- Create new unified 'Runware Accelerator Options' node
- Support all accelerator parameters: fbcache, cacheDistance, teaCache, teaCacheDistance, deepCache, deepCacheInterval, deepCacheBranchId
- Add cache step control: cacheStartStep, cacheStartStepPercentage, cacheEndStep, cacheEndStepPercentage, cacheMaxConsecutiveSteps
- Update imageInference.py to support new accelerator system
This commit is contained in:
Sirshendu Ganguly
2025-10-01 12:16:58 -04:00
parent 02c84044b2
commit e589062ea4
5 changed files with 147 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/acceleratorOptions.py
+143
View File
@@ -0,0 +1,143 @@
from .utils import runwareUtils as rwUtils
class acceleratorOptions:
@classmethod
def INPUT_TYPES(cls):
return {
"required": {},
"optional": {
# FBCache parameters
"fbcache": ("BOOLEAN", {
"tooltip": "Enable Frame Buffer Cache for faster generation",
"default": None,
}),
"cacheDistance": ("FLOAT", {
"tooltip": "Cache distance parameter for FBCache",
"default": None,
"min": 0.0,
"max": 1.0,
"step": 0.01,
}),
# TeaCache parameters
"teaCache": ("BOOLEAN", {
"tooltip": "Enable TeaCache for transformer-based models (Flux, SD3). Does not work with UNet models like SDXL or SD1.5. Particularly effective for iterative editing workflows.",
"default": None,
}),
"teaCacheDistance": ("FLOAT", {
"tooltip": "Controls TeaCache aggressiveness: 0.0 (conservative/quality) to 1.0 (aggressive/speed). Lower values maintain quality, higher values prioritize speed.",
"default": None,
"min": 0.0,
"max": 1.0,
"step": 0.01,
}),
# DeepCache parameters
"deepCache": ("BOOLEAN", {
"tooltip": "Enable DeepCache for UNet-based models (SDXL, SD1.5). Not applicable to transformer models. Provides performance improvements for high-throughput scenarios.",
"default": None,
}),
"deepCacheInterval": ("INT", {
"tooltip": "Frequency of feature caching - steps between cache operations. Larger values = faster but may impact quality. Smaller values = better quality but slower.",
"default": None,
"min": 1,
"max": 20,
}),
"deepCacheBranchId": ("INT", {
"tooltip": "Network branch for caching (0=shallowest, deeper=more conservative). Lower IDs = aggressive caching/faster, higher IDs = conservative caching/better quality.",
"default": None,
"min": 0,
"max": 10,
}),
# Cache step control parameters
"cacheStartStep": ("INT", {
"tooltip": "Inference step number where caching begins (0 to total steps). Alternative to cacheStartStepPercentage.",
"default": None,
"min": 0,
"max": 200,
}),
"cacheStartStepPercentage": ("INT", {
"tooltip": "Percentage of total steps where caching begins (0-99). Alternative to cacheStartStep.",
"default": None,
"min": 0,
"max": 99,
}),
"cacheEndStep": ("INT", {
"tooltip": "Inference step number where caching stops (must be > cacheStartStep). Alternative to cacheEndStepPercentage.",
"default": None,
"min": 1,
"max": 200,
}),
"cacheEndStepPercentage": ("INT", {
"tooltip": "Percentage of total steps where caching stops (must be > cacheStartStepPercentage). Alternative to cacheEndStep.",
"default": None,
"min": 1,
"max": 100,
}),
"cacheMaxConsecutiveSteps": ("INT", {
"tooltip": "Max consecutive steps using cached computations before forcing fresh computation. Prevents quality degradation from extended cache reuse.",
"default": None,
"min": 1,
"max": 5,
}),
},
}
DESCRIPTION = "Configure Runware accelerator options for faster image generation. Supports TeaCache (transformer models like Flux/SD3) and DeepCache (UNet models like SDXL/SD1.5) with fine-grained control over caching parameters and step ranges. Note: These caching methods work best with deterministic schedulers (Euler, DDIM) rather than stochastic schedulers."
FUNCTION = "acceleratorOptions"
RETURN_TYPES = ("RUNWAREACCELERATOR",)
RETURN_NAMES = ("Runware Accelerator",)
CATEGORY = "Runware"
def acceleratorOptions(self, **kwargs):
# Get all optional parameters
fbcache = kwargs.get("fbcache")
cacheDistance = kwargs.get("cacheDistance")
teaCache = kwargs.get("teaCache")
teaCacheDistance = kwargs.get("teaCacheDistance")
deepCache = kwargs.get("deepCache")
deepCacheInterval = kwargs.get("deepCacheInterval")
deepCacheBranchId = kwargs.get("deepCacheBranchId")
cacheStartStep = kwargs.get("cacheStartStep")
cacheStartStepPercentage = kwargs.get("cacheStartStepPercentage")
cacheEndStep = kwargs.get("cacheEndStep")
cacheEndStepPercentage = kwargs.get("cacheEndStepPercentage")
cacheMaxConsecutiveSteps = kwargs.get("cacheMaxConsecutiveSteps")
# Build accelerator options dictionary
acceleratorOptions = {}
# FBCache options
if fbcache is not None:
acceleratorOptions["fbcache"] = fbcache
if cacheDistance is not None:
acceleratorOptions["cacheDistance"] = cacheDistance
# TeaCache options
if teaCache is not None:
acceleratorOptions["teaCache"] = teaCache
if teaCacheDistance is not None:
acceleratorOptions["teaCacheDistance"] = teaCacheDistance
# DeepCache options
if deepCache is not None:
acceleratorOptions["deepCache"] = deepCache
if deepCacheInterval is not None:
acceleratorOptions["deepCacheInterval"] = deepCacheInterval
if deepCacheBranchId is not None:
acceleratorOptions["deepCacheBranchId"] = deepCacheBranchId
# Cache step control options
if cacheStartStep is not None:
acceleratorOptions["cacheStartStep"] = cacheStartStep
if cacheStartStepPercentage is not None:
acceleratorOptions["cacheStartStepPercentage"] = cacheStartStepPercentage
if cacheEndStep is not None:
acceleratorOptions["cacheEndStep"] = cacheEndStep
if cacheEndStepPercentage is not None:
acceleratorOptions["cacheEndStepPercentage"] = cacheEndStepPercentage
if cacheMaxConsecutiveSteps is not None:
acceleratorOptions["cacheMaxConsecutiveSteps"] = cacheMaxConsecutiveSteps
return (acceleratorOptions,)