* feature: Set the Ascend NPU to use a single one
* Enable the `--cuda-device` parameter to support both CUDA and Ascend NPUs simultaneously.
* Make the code just set the ASCENT_RT_VISIBLE_DEVICES environment variable without any other edits to master branch
---------
Co-authored-by: Jedrzej Kosinski <kosinkadink1@gmail.com>
* flux: math: Use _addcmul to avoid expensive VRAM intermediate
The rope process can be the VRAM peak and this intermediate
for the addition result before releasing the original can OOM.
addcmul_ it.
* wan: Delete the self attention before cross attention
This saves VRAM when the cross attention and FFN are in play as the
VRAM peak.
- Auto-derive TORCH_CUDA_ARCH_LIST from torch device capabilities (unique, sorted, optional +PTX) to cover all charted GPUs:
Turing 7.5, Ampere 8.0/8.6/8.7, Ada 8.9, Hopper 9.0, and Blackwell 10.0 & 12.0/12.1; add name-based fallbacks for mixed or torch-less scenarios.
- Add user-tunable build parallelism with SAGE_MAX_JOBS (preferred) and MAX_JOBS (alias) that cap PyTorch cpp_extension/ninja -j; fall back to a RAM/CPU heuristic to prevent OOM “Killed” during CUDA/C++ builds.
- Correct Sage flags: SAGE_ATTENTION_AVAILABLE only signals “built/installed,” while FORCE_SAGE_ATTENTION=1 enables Sage at startup; fix logs to reference FORCE_SAGE_ATTENTION.
- Maintain Triton install strategy by GPU generation for compatibility and performance.
- Add first-run dependency installation with COMFY_FORCE_INSTALL override; keep permissions bootstrap and minor logging/URL cleanups.
Adds installed and required workflow templates version information to the
/system_stats endpoint, allowing the frontend to detect and notify users
when their templates package is outdated.
- Add get_installed_templates_version() and get_required_templates_version()
methods to FrontendManager
- Include templates version info in system_stats response
- Add comprehensive unit tests for the new functionality
Introduce a first-run flag to install custom_nodes dependencies only on the
initial container start, with COMFY_FORCE_INSTALL=1 to override on demand;
correct Sage Attention flag semantics so SAGE_ATTENTION_AVAILABLE=1 only
indicates the build is present while FORCE_SAGE_ATTENTION=1 enables it at
startup; fix the misleading log to reference FORCE_SAGE_ATTENTION. Update
TORCH_CUDA_ARCH_LIST mapping to 7.5 (Turing), 8.6 (Ampere), 8.9 (Ada), and
10.0 (Blackwell/RTX 50); retain Triton strategy with a compatibility pin on
Turing and latest for Blackwell, including fallbacks. Clean up git clone URLs,
standardize on python -m pip, and tighten logs; preserve user remapping and
strategy-based rebuild detection via the .built flag.
Replace Markdown-style links in git clone with standard HTTPS URLs so the
repository actually clones under bash.
Derive TORCH_CUDA_ARCH_LIST from PyTorch devices and add +PTX to the
highest architecture for forward-compat extension builds.
Warn explicitly on Blackwell (sm_120) when the active torch/CUDA build
lacks support, prompting an upgrade to torch with CUDA 12.8+.
Keep pip --no-cache-dir, preserve Triton pin for Turing, and retain
idempotent ComfyUI-Manager update logic.
Cap build parallelism via MAX_JOBS (override SAGEATTENTION_MAX_JOBS) and
CMAKE_BUILD_PARALLEL_LEVEL to prevent OOM kills during nvcc/cc1plus when
ninja fanout is high in constrained containers.
Compute TORCH_CUDA_ARCH_LIST from torch.cuda device properties to target
exact GPU SMs across mixed setups; keep human-readable nvidia-smi logs.
Move PATH/PYTHONPATH exports earlier and use `python -m pip` with
`--no-cache-dir` consistently to avoid stale caches and reduce image bloat.
Fix git clone/update commands to standard HTTPS and reset against
origin/HEAD; keep shallow operations for speed and reproducibility.
Clarify Sage Attention flags: set SAGE_ATTENTION_AVAILABLE only when
module import succeeds; require FORCE_SAGE_ATTENTION=1 to enable at boot.
Keep first-run dependency installation with COMFY_AUTO_INSTALL=1 override
to re-run installs on later boots without removing the first-run flag.
Compute TORCH_CUDA_ARCH_LIST from torch.cuda device properties to build
for the exact GPUs present, improving correctness across mixed setups.
Add first-run dependency install gate with a COMFY_AUTO_INSTALL=1
override to re-run installs on later boots without removing the flag.
Use `python -m pip` consistently with `--no-cache-dir` to avoid stale
wheels and reduce container bloat during rebuilds.
Fix git clone commands to standard HTTPS (no Markdown link syntax) and
use shallow fetch/reset against origin/HEAD for speed and reliability.
Clarify Sage Attention flags: set SAGE_ATTENTION_AVAILABLE only when the
module is importable; require FORCE_SAGE_ATTENTION=1 to enable at boot.
Keep readable GPU logs via `nvidia-smi`, while relying on torch for
compile-time arch targeting. Improve logging throughout the flow.
When unloading models in load_models_gpu(), the model finalizer was not
being explicitly detached, leading to a memory leak. This caused
linear memory consumption increase over time as models are repeatedly
loaded and unloaded.
This change prevents orphaned finalizer references from accumulating in
memory during model switching operations.
Updates README to match the Dockerfile and entrypoint: Python 3.12 slim trixie with CUDA 12.9 dev libs and PyTorch via cu129 wheels; SageAttention is built at startup but only enabled when FORCE_SAGE_ATTENTION=1 and the import test passes; Compose example uses Deploy device reservations with driver:nvidia and capabilities:[gpu]; documents PUID/PGID, COMFY_AUTO_INSTALL, and FORCE_SAGE_ATTENTION; clarifies port 8188 mapping and how to change ports.
The entrypoint no longer exports SAGE_ATTENTION_AVAILABLE=1 on successful builds, preventing global attention patching from being forced; instead, it builds/tests SageAttention, sets SAGE_ATTENTION_BUILT=1 for visibility, and only appends --use-sage-attention when SAGE_ATTENTION_AVAILABLE=1 is supplied by the environment, preserving user control across docker run -e/compose env usage while keeping the feature available.
