comfy-env 0.0.9__tar.gz

comfy_env-0.0.9/.github/workflows/publish.yml

@@ -0,0 +1,28 @@
+name: Publish to PyPI
+
+on:
+  release:
+    types: [published]
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write  # Required for trusted publishing
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Install build tools
+        run: pip install build
+
+      - name: Build package
+        run: python -m build
+
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1

comfy_env-0.0.9/.gitignore

@@ -0,0 +1,29 @@
+# Python
+__pycache__/
+*.py[cod]
+*.so
+*.egg
+*.egg-info/
+dist/
+build/
+.eggs/
+
+# Virtual environments
+.venv/
+venv/
+ENV/
+
+# IDE
+.idea/
+.vscode/
+*.swp
+*.swo
+
+# Testing
+.pytest_cache/
+.coverage
+htmlcov/
+
+# OS
+.DS_Store
+Thumbs.db

comfy_env-0.0.9/CLAUDE.md

@@ -0,0 +1,131 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code when working with this repository.
+
+## Project Overview
+
+**comfy-env** is a Python library for ComfyUI custom nodes that provides:
+
+1. **CUDA Wheel Resolution** - Deterministic wheel URL construction for CUDA packages (nvdiffrast, pytorch3d, etc.)
+2. **In-Place Installation** - Install CUDA wheels into current environment without compile
+3. **Process Isolation** - Run nodes in separate venvs with different dependencies
+
+## Architecture
+
+### Type 1 Nodes (Isolated Venv)
+Nodes that need separate venv due to conflicting dependencies:
+```
+ComfyUI Main Process                    Isolated Subprocess
+┌─────────────────────────┐            ┌──────────────────────────┐
+│  @isolated decorator    │            │  runner.py entrypoint    │
+│  intercepts FUNCTION    │   UDS/     │                          │
+│  method calls           │  stdin ──► │  Imports node module     │
+│                         │            │  (decorator is no-op)    │
+│  Tensor IPC via shared  │            │                          │
+│  memory / CUDA IPC      │ ◄──────────│  Returns result          │
+└─────────────────────────┘            └──────────────────────────┘
+```
+
+### Type 2 Nodes (In-Place)
+Nodes that just need CUDA wheels resolved:
+```
+comfy-env.toml
+       │
+       ▼
+┌──────────────────────────────────────────────┐
+│  WheelResolver                               │
+│  - Detects CUDA/PyTorch/Python versions      │
+│  - Constructs exact wheel URLs               │
+│  - pip install --no-deps                     │
+└──────────────────────────────────────────────┘
+```
+
+## Key Components
+
+| File | Purpose |
+|------|---------|
+| `src/comfy_env/install.py` | `install()` function for both modes |
+| `src/comfy_env/resolver.py` | Wheel URL resolution with template expansion |
+| `src/comfy_env/cli.py` | `comfy-env` CLI commands |
+| `src/comfy_env/decorator.py` | `@isolated` decorator for process isolation |
+| `src/comfy_env/workers/` | Worker classes (TorchMPWorker, VenvWorker) |
+| `src/comfy_env/env/manager.py` | venv creation with `uv` |
+| `src/comfy_env/env/config_file.py` | TOML config parsing |
+
+## Development Commands
+
+```bash
+# Install in development mode
+cd /home/shadeform/comfy-env
+pip install -e .
+
+# Run CLI
+comfy-env info
+comfy-env install --dry-run
+comfy-env resolve nvdiffrast==0.4.0
+```
+
+## Usage Patterns
+
+### In-Place Installation (Type 2)
+```python
+from comfy_env import install
+
+# Auto-discover config and install
+install()
+
+# Dry run
+install(dry_run=True)
+```
+
+### Process Isolation (Type 1)
+```python
+from comfy_env import isolated
+
+@isolated(env="myenv")
+class MyGPUNode:
+    FUNCTION = "process"
+    RETURN_TYPES = ("IMAGE",)
+
+    def process(self, image):
+        # Runs in isolated subprocess
+        import heavy_dependency
+        return (result,)
+```
+
+### Direct Worker Usage
+```python
+from comfy_env import TorchMPWorker
+
+worker = TorchMPWorker()
+result = worker.call(my_function, image=tensor)
+```
+
+## Config File Format
+
+```toml
+[env]
+name = "my-node"
+python = "3.10"
+cuda = "auto"
+
+[packages]
+requirements = ["transformers>=4.56"]
+no_deps = ["nvdiffrast==0.4.0"]
+
+[sources]
+wheel_sources = ["https://github.com/.../releases/download/"]
+```
+
+## Key Design Decisions
+
+1. **Deterministic Resolution**: Wheel URLs are constructed, not solved. If URL 404s, fail fast with clear message.
+
+2. **No Compilation on User Machines**: If wheel doesn't exist, fail with actionable error showing what combos are available.
+
+3. **Template Variables**: `{cuda_short}`, `{torch_mm}`, `{py_short}`, `{platform}` for URL construction.
+
+## Related Projects
+
+- **pyisolate** - ComfyUI's official security-focused isolation
+- **comfy-cli** - High-level ComfyUI management

comfy_env-0.0.9/CRITICISM.md

@@ -0,0 +1,316 @@
+# Critical Analysis: comfyui-isolation Architecture
+
+*Analysis from the perspective of a senior systems engineer with GPU computing expertise*
+
+**Date**: January 2025
+**Compared Against**: pyisolate (ComfyUI's official isolation library)
+
+---
+
+## Executive Summary
+
+The current implementation is a **pragmatic MVP** that solves the immediate problem (dependency isolation) but has fundamental architectural decisions that will become technical debt at scale. It prioritizes simplicity over performance, which is acceptable for prototyping but not for production GPU workloads.
+
+**Overall Grade: B-** — Works, but with significant overhead and several design choices that need iteration.
+
+---
+
+## 1. IPC Mechanism: JSON over stdin/stdout
+
+### The Problem
+
+```python
+# protocol.py - Every tensor transfer does this:
+arr = obj.cpu().numpy()  # GPU → CPU copy
+pickle.dumps(arr)         # Serialize to bytes
+base64.b64encode(...)     # Encode to string (+33% size)
+json.dumps(...)           # Wrap in JSON
+```
+
+### Critique
+
+- **4x memory overhead minimum** for tensor data (original + numpy + pickle + base64)
+- **2 copies minimum** per tensor (GPU→CPU, then pickle serialization)
+- For a 1024×1024 RGBA image: ~16MB becomes ~21MB on wire, with ~64MB peak memory during serialization
+- JSON parsing is CPU-bound and single-threaded
+
+### What pyisolate Does Better
+
+pyisolate uses **CUDA IPC handles** for zero-copy GPU tensor sharing:
+
+```python
+# pyisolate/tensor_serializer.py
+def _serialize_cuda_tensor(t: torch.Tensor) -> dict[str, Any]:
+    func, args = reductions.reduce_tensor(t)
+    # args[7] is the CUDA IPC handle - NO DATA COPY
+    return {
+        "__type__": "TensorRef",
+        "device": "cuda",
+        "handle": base64.b64encode(args[7]).decode('ascii'),
+        # ... metadata only, no tensor data
+    }
+```
+
+**Performance difference for 1GB tensor:**
+- comfyui-isolation: ~500ms, 3.3GB memory
+- pyisolate: ~0ms, ~0 extra memory
+
+### Recommended Fix
+
+```python
+# Priority 1: Adopt CUDA IPC for tensors
+import torch.multiprocessing.reductions as reductions
+
+def serialize_tensor(t: torch.Tensor) -> dict:
+    if t.is_cuda:
+        func, args = reductions.reduce_tensor(t)
+        return {"__type__": "TensorRef", "device": "cuda",
+                "handle": base64.b64encode(args[7]), ...}
+    else:
+        t.share_memory_()
+        # Use file_system strategy for CPU tensors
+        ...
+```
+
+---
+
+## 2. The fd-level Redirection Hack
+
+### Current Code
+
+```python
+# runner.py lines 178-183
+stdout_fd = original_stdout.fileno()
+stderr_fd = sys.stderr.fileno()
+stdout_fd_copy = os.dup(stdout_fd)
+os.dup2(stderr_fd, stdout_fd)
+```
+
+### Critique
+
+This is a **code smell** indicating a deeper architectural problem. We're fighting the IPC design rather than fixing it.
+
+**Why This Exists:**
+- JSON-RPC uses stdout for responses
+- C libraries (pymeshfix) print to fd 1 directly
+- Can't distinguish "library noise" from "JSON response"
+
+### The Real Fix
+
+Don't use stdout for data. Use a **dedicated communication channel**:
+
+```python
+# Option 1: Unix Domain Socket
+sock_path = f"/tmp/comfyui-isolation-{pid}.sock"
+server = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+server.bind(sock_path)
+# Pass sock_path to subprocess, both stdout/stderr go to logging
+
+# Option 2: socketpair at spawn time
+parent_sock, child_sock = socket.socketpair()
+subprocess.Popen(..., pass_fds=[child_sock.fileno()])
+```
+
+### Why This Matters
+
+- Current hack is **fragile** — any library could still break it with buffered writes
+- **Not portable** — `select()` on stdin doesn't work the same on Windows
+- **Race conditions** — fd manipulation during execution is not thread-safe
+
+---
+
+## 3. Process Lifecycle Management
+
+### Current Design
+
+- One subprocess per `(env_name, node_package_dir)` tuple
+- Process kept alive, reused for multiple calls
+- Killed on timeout or error
+
+### Issues
+
+**A. No GPU Memory Management:**
+```python
+# After node execution, GPU memory is NOT freed
+# The subprocess stays alive, holding VRAM
+# Next node in workflow inherits fragmented GPU state
+```
+
+**B. No Graceful Degradation:**
+```python
+# If subprocess dies, you lose ALL state
+# No checkpointing, no recovery
+```
+
+**C. Single-Process Bottleneck:**
+```python
+# One subprocess per env = sequential execution
+# Can't parallelize across nodes even if they're independent
+```
+
+### Recommendations
+
+- **Process pool** per environment with configurable size
+- **Explicit VRAM management** — option to kill subprocess after each call
+- **Health checks** — periodic GPU memory queries, automatic restart if fragmented
+
+---
+
+## 4. Serialization Protocol Design
+
+### Current Issues
+
+**A. Type Detection by Shape Heuristics:**
+```python
+# protocol.py lines 105-110
+if len(shape) == 4 and shape[-1] in (3, 4):
+    obj_type = "comfyui_image"
+elif len(shape) in (2, 3) and arr.dtype in ('float32', 'float64'):
+    if arr.min() >= 0 and arr.max() <= 1:
+        obj_type = "comfyui_mask"
+```
+
+This is **brittle**. A 4D tensor that happens to have shape `(1, 100, 100, 3)` will be misidentified as an image even if it's not.
+
+**B. Pickle Security:**
+Using pickle for arbitrary objects is a **security risk**. Malicious pickle payloads can execute arbitrary code.
+
+**C. SimpleNamespace Fallback:**
+```python
+# Objects become SimpleNamespace after round-trip
+ns = SimpleNamespace(**data)
+ns._class_name = obj.get("_class", "unknown")
+```
+
+This **loses type identity**. Method calls on reconstructed objects will fail.
+
+### What pyisolate Does Better
+
+- Explicit type registry with custom serializers
+- Attempts to reconstruct original classes
+- Separate serializers for known ComfyUI types
+
+---
+
+## 5. Comparison: comfyui-isolation vs pyisolate
+
+| Aspect | comfyui-isolation | pyisolate |
+|--------|-------------------|-----------|
+| **Process Model** | `subprocess.Popen` | `multiprocessing.Process` (spawn) |
+| **IPC Channel** | stdin/stdout (JSON) | `multiprocessing.Queue` OR Unix Domain Sockets |
+| **Tensor Transfer** | CPU copy → pickle → base64 | `share_memory_()` + CUDA IPC handles |
+| **Serialization** | Custom JSON protocol | Pickle (Queue) OR JSON (Sandbox mode) |
+| **Security** | JSON-only (no pickle RCE) | Pickle + bwrap sandbox option |
+| **API** | Simple decorator | Async + inheritance |
+
+### What We Do Better
+
+1. **Simpler User API** - Just add `@isolated` decorator
+2. **Complete Process Isolation** - subprocess.Popen means truly separate processes
+3. **JSON-only Security** - No pickle deserialization RCE risk by default
+4. **Config File Discovery** - Auto-discovers isolation config files
+
+### What pyisolate Does Better
+
+1. **Zero-Copy Tensor Sharing** - CUDA IPC handles, no data copy
+2. **Transport Abstraction** - Pluggable transports (Queue, UDS, JSON Socket)
+3. **TensorKeeper Pattern** - Prevents GC race conditions on shared tensors
+4. **Sandboxing** - bwrap integration for untrusted code
+
+---
+
+## 6. Recommended Improvements (Prioritized)
+
+| Priority | Change | Effort | Impact |
+|----------|--------|--------|--------|
+| **P0** | Adopt CUDA IPC for tensor serialization | High | 100x+ faster for large tensors |
+| **P0** | Replace stdout with Unix Domain Socket | Medium | Eliminates fd hack, cleaner design |
+| **P1** | Add transport abstraction layer | Medium | Flexibility for future transports |
+| **P1** | Explicit type registry for serialization | Low | Eliminates shape-guessing bugs |
+| **P2** | Add `fresh_process=True` option | Low | Guaranteed VRAM cleanup when needed |
+| **P2** | Process pool with GPU affinity | High | Multi-GPU support, parallelism |
+| **P3** | Optional bwrap sandboxing | High | Security for untrusted extensions |
+
+---
+
+## 7. Implementation Notes
+
+### Adopting pyisolate's Tensor Serializer
+
+The key file is `pyisolate/_internal/tensor_serializer.py`. Key patterns:
+
+```python
+# TensorKeeper - prevents GC race condition
+class TensorKeeper:
+    def keep(self, t: torch.Tensor) -> None:
+        # Hold reference for 30s to ensure receiver can open shared memory
+        self._keeper.append((time.time(), t))
+
+# CPU tensor via file_system shared memory
+def _serialize_cpu_tensor(t: torch.Tensor) -> dict:
+    _tensor_keeper.keep(t)
+    if not t.is_shared():
+        t.share_memory_()
+    storage = t.untyped_storage()
+    sfunc, sargs = reductions.reduce_storage(storage)
+    return {
+        "__type__": "TensorRef",
+        "strategy": "file_system",
+        "manager_path": sargs[1].decode('utf-8'),
+        "storage_key": sargs[2].decode('utf-8'),
+        ...
+    }
+
+# CUDA tensor via IPC handle
+def _serialize_cuda_tensor(t: torch.Tensor) -> dict:
+    _tensor_keeper.keep(t)
+    func, args = reductions.reduce_tensor(t)
+    return {
+        "__type__": "TensorRef",
+        "device": "cuda",
+        "handle": base64.b64encode(args[7]).decode('ascii'),
+        ...
+    }
+```
+
+### Unix Domain Socket Transport
+
+```python
+class UDSTransport:
+    def __init__(self, sock: socket.socket):
+        self._sock = sock
+        self._lock = threading.Lock()
+
+    def send(self, obj: Any) -> None:
+        data = json.dumps(obj).encode('utf-8')
+        msg = struct.pack('>I', len(data)) + data  # Length-prefixed
+        with self._lock:
+            self._sock.sendall(msg)
+
+    def recv(self) -> Any:
+        raw_len = self._recvall(4)
+        msg_len = struct.unpack('>I', raw_len)[0]
+        data = self._recvall(msg_len)
+        return json.loads(data.decode('utf-8'))
+```
+
+---
+
+## 8. References
+
+- **pyisolate source**: `/home/shadeform/pyisolate/`
+- **pyisolate PR #3**: Production-ready update with RPC, sandboxing, tensor serialization
+- **torch.multiprocessing.reductions**: PyTorch's IPC serialization primitives
+- **CUDA IPC**: `cudaIpcGetMemHandle` / `cudaIpcOpenMemHandle`
+
+---
+
+## Conclusion
+
+The core insight: **pyisolate solves the hard problems (tensor IPC, sandboxing) but has UX issues. comfyui-isolation has good UX but needs to adopt their tensor handling.**
+
+Don't rewrite from scratch. Instead:
+1. Steal the tensor serializer from pyisolate
+2. Add Unix Domain Socket transport
+3. Keep the decorator API
+4. Consider optional sandbox mode for security

comfy_env-0.0.9/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Andrea Pozzetti
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.