fix(setup): pre-install torch so flash-attn can build (v0.1.1)

nano-vllm-voxcpm pulls in flash-attn, which needs torch importable at
build time. pip's default PEP 517 build isolation gives it an empty
sandbox, so the setup from v0.1.0 died at first install with
"ModuleNotFoundError: No module named 'torch'".

Fix:
  * setup.sh pre-installs torch + torchaudio before the main
    requirements pass.
  * requirements.txt drops the torch/torchaudio pins so they come only
    from the pre-install step; avoids double-resolve and version
    conflicts.
  * The main requirements install now uses --no-build-isolation so
    flash-attn picks up torch from the venv.

v0.1.0 never actually installed end-to-end on a clean machine;
v0.1.1 is the first release that does.

Author: fakehec

CHANGELOG.md

@@ -5,6 +5,21 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.1.1] - 2026-04-17
+
+### Fixed
+- `setup.sh` failed during `pip install -r requirements.txt` because
+  `flash-attn` (a transitive dep of `nano-vllm-voxcpm`) requires
+  `torch` at build time, but pip's default PEP 517 build-isolation
+  sandbox does not have it. `setup.sh` now pre-installs torch and
+  torchaudio, then runs `pip install --no-build-isolation -r
+  requirements.txt` so flash-attn picks up the torch in the venv.
+  `requirements.txt` drops the torch pins (they live in the
+  pre-install step).
+
+v0.1.0 never made it past the first install on a clean machine; this
+is the first release that actually installs end-to-end.
+
 ## [0.1.0] - 2026-04-17
 
 First scaffold release. Pre-alpha — active development. API surface may

requirements.txt

@@ -1,10 +1,12 @@
 # nano-vllm-voxcpm pulls in its own torch/vLLM/kernel stack. Pin to a
 # tested commit/version once the project publishes one; for now we
 # depend on the PyPI release.
+#
+# IMPORTANT: torch and torchaudio must already be installed before this
+# file is processed, because nano-vllm-voxcpm -> flash-attn needs torch
+# at build time. See setup.sh for the two-step install that handles it.
 nano-vllm-voxcpm>=2.0.0
 huggingface-hub>=0.20.0
-torch>=2.5.0
-torchaudio>=2.5.0
 
 # Audio IO + format conversion.
 soundfile>=0.12.0

setup.sh

@@ -21,10 +21,18 @@ $PYTHON_BIN -m venv venv
 source venv/bin/activate
 
 echo "[*] Installing build-time dependencies..."
-pip install --upgrade pip setuptools wheel
+pip install --upgrade pip setuptools wheel packaging
+
+# flash-attn is a transitive dependency of nano-vllm-voxcpm. Its build
+# system requires torch to be importable in the build environment, but
+# pip's default PEP 517 build isolation provides an empty env. So we
+# install torch first, then the rest with --no-build-isolation so flash-attn
+# picks up torch from the venv.
+echo "[*] Pre-installing torch (required before flash-attn can build)..."
+pip install "torch>=2.5.0,<2.10.0" "torchaudio>=2.5.0,<2.10.0"
 
 echo "[*] Installing core dependencies from requirements.txt (may take several minutes)..."
-pip install -r requirements.txt
+pip install --no-build-isolation -r requirements.txt
 
 if [ -f "./setup_assets.sh" ]; then
     echo "[*] Python environment ready. Handing over to setup_assets.sh..."