fix(deps): make torch/ML stack opt-in instead of a mandatory install

requirements-amd.txt/requirements-nvidia.txt were pulling a multi-GB
torch wheel by default even though nothing in synapse/ imports torch,
transformers, accelerate, bitsandbytes, or PySide6 - dead weight that
made the pip batch fragile (one failed download could take unrelated
base deps down with it on a slow connection). Split the unused ML/GUI
stack out of requirements-base.txt into a new opt-in requirements-ml.txt,
and dropped the torch lines from the AMD/NVIDIA overlays and generator.

Also: recreate the venv if it exists but pip is missing, instead of
silently reusing a half-built one (ensurepip can fail during venv
creation and leave an interpreter with no pip).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

bin/gen-nvidia-reqs.py

@@ -53,10 +53,9 @@ def main():
 
 -r requirements-base.txt
 
-# GPU Compute Stack
-torch
-torchaudio
-torchvision
+# Torch itself is opt-in — see requirements-ml.txt. The --index-url above is
+# what makes `pip install -r requirements-nvidia.txt -r requirements-ml.txt`
+# resolve torch as the matching CUDA build instead of CPU-only.
 """, newline="\n")
 
     print(f"\nWritten: {NVIDIA_REQS}")

bin/sync.py

@@ -61,6 +61,16 @@ def venv_python() -> Path:
     """Path to the Promethean interpreter, creating the venv if it's missing."""
     venv = ROOT / "Promethean"
     py = venv / ("Scripts/python.exe" if os.name == "nt" else "bin/python")
+    if py.exists() and subprocess.run(
+            [str(py), "-m", "pip", "--version"], capture_output=True).returncode:
+        # venv creation can partially succeed: the interpreter gets built but
+        # ensurepip's bootstrap fails (e.g. the matching pythonX.Y-venv package
+        # wasn't installed yet), leaving pip missing. Existence of `py` alone
+        # can't tell a venv like that apart from a good one, so a prior failed
+        # run would otherwise be reused forever instead of getting rebuilt now
+        # that whatever broke ensurepip is fixed.
+        print("Existing venv has no pip - recreating it...")
+        shutil.rmtree(venv)
     if not py.exists():
         result = subprocess.run([sys.executable, "-m", "venv", str(venv)])
         if result.returncode:

requirements-amd.txt

@@ -4,7 +4,6 @@
 
 -r requirements-base.txt
 
-# GPU Compute Stack
-torch
-torchaudio
-torchvision
+# Torch itself is opt-in — see requirements-ml.txt. The --index-url above is
+# what makes `pip install -r requirements-amd.txt -r requirements-ml.txt`
+# resolve torch as the ROCm build instead of CPU-only.

requirements-base.txt

@@ -1,11 +1,6 @@
 # --- Shared Base (GPU-agnostic) ---
 
-# AI Stack
-transformers
-accelerate
-bitsandbytes
-safetensors
-sentencepiece
+# AI Stack (tokenizers/huggingface-hub: faster-whisper deps, not transformers)
 tokenizers
 huggingface-hub
 
@@ -28,10 +23,6 @@ numpy
 scipy
 pandas
 psutil
-PySide6
-PySide6_Addons
-PySide6_Essentials
-shiboken6
 tqdm
 rich
 python-dotenv

requirements-ml.txt

@@ -0,0 +1,20 @@
+# --- Optional: local ML inference stack ---
+# Nothing in synapse/ imports any of this — Ollama handles all inference over
+# HTTP. Only install this if you're doing local model work outside Ollama
+# (fine-tuning, direct transformers inference, etc). Skipping it is what keeps
+# the default install light and out of a multi-GB torch download.
+#
+# This file has no --index-url of its own, so torch resolves as CPU-only
+# unless you combine it with the GPU overlay that sets one:
+#   pip install -r requirements-amd.txt    -r requirements-ml.txt   # ROCm
+#   pip install -r requirements-nvidia.txt -r requirements-ml.txt   # CUDA
+#   pip install -r requirements-ml.txt                              # CPU only
+
+transformers
+accelerate
+bitsandbytes
+safetensors
+sentencepiece
+torch
+torchaudio
+torchvision

requirements-nvidia.txt

@@ -4,7 +4,6 @@
 
 -r requirements-base.txt
 
-# GPU Compute Stack
-torch
-torchaudio
-torchvision
+# Torch itself is opt-in — see requirements-ml.txt. The --index-url above is
+# what makes `pip install -r requirements-nvidia.txt -r requirements-ml.txt`
+# resolve torch as the matching CUDA build instead of CPU-only.