Skip to content

perf: Tier 1 hot-path allocation elimination & text encoder stutter fix #7

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
forkni wants to merge 6 commits into
base: pr2/fp8-tensorrt-build
Choose a base branch
from
Open

perf: Tier 1 hot-path allocation elimination & text encoder stutter fix #7

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
ce0d51c
perf: pre-allocate KV clone buffers in CachedSTAttnProcessor2_0
INTER-NYC Apr 4, 2026
90c48a2
perf: add CUDA/PyTorch env var tuning and cudnn.benchmark
INTER-NYC Apr 4, 2026
6fe3fd6
perf: Tier 1 hot-path allocation elimination (Phase A-C)
INTER-NYC Apr 5, 2026
350b089
perf: skip text encoder reload on identical prompt; seed FPS EMA from…
INTER-NYC Apr 5, 2026
20572f8
fix: remove empty_cache() from text encoder offload to prevent prompt…
INTER-NYC Apr 5, 2026
b19dba5
perf: keep text encoders on GPU during inference; add force_offload f…
INTER-NYC Apr 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 23 additions & 0 deletions Start_StreamDiffusion.bat
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,23 @@
@echo off
cd /d %~dp0

:: ─── CUDA / PyTorch Performance Tuning ───
:: Prevents memory fragmentation from per-frame torch.cat allocations
set PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128,expandable_segments:True
:: Defers CUDA module loading until first use (~1-5s faster startup)
set CUDA_MODULE_LOADING=LAZY
:: Enables cuDNN v8 graph API for better conv kernel selection (VAE, preprocessors)
set TORCH_CUDNN_V8_API_ENABLED=1
:: Ensures async kernel launches (default=0, but explicit protects against debug leftovers)
set CUDA_LAUNCH_BLOCKING=0
:: Caches compiled Triton kernels to disk (eliminates 30-60s JIT warmup on restart)
set TORCHINDUCTOR_FX_GRAPH_CACHE=1

if exist venv (
call venv\Scripts\activate.bat
venv\Scripts\python.exe streamdiffusionTD\td_main.py
) else (
call .venv\Scripts\activate.bat
.venv\Scripts\python.exe streamdiffusionTD\td_main.py
)
pause
Loading