Skip to content

FreeToken brings datacenter-scale model serving to your desktop. Run massive models locally, fast and efficiently.

fix(install): resolve sglang-kernel from PyPI, not the sglang index (#524)

Xiaoze Fana8eb986758a2
diff --git a/install.sh b/install.sh
index a7df8327..60b4439c 100755
--- a/install.sh
+++ b/install.sh
@@ -4,8 +4,8 @@
#
# Installs the `freetoken` runtime (the `ft` CLI) and its prebuilt kernel-cache
# wheel into a managed venv, then wires it up so FreeToken Desktop can find it.
-# Dependencies come from PyPI via uv, except torch and sglang-kernel whose cu130
-# wheels live on dedicated indexes (see CU_INDEX_ARGS below).
+# Dependencies come from PyPI via uv, except torch whose cu130
+# wheels live on a dedicated index (see CU_INDEX_ARGS below).
#
# Typical use (once a release exists):
# curl -fsSL https://<host>/install.sh | bash
@@ -201,12 +201,12 @@ mkdir -p "$FT_HOME"
# re-install can't inherit a stale/mismatched torch (e.g. an old cu128 venv after a cu130 bump).
"$UV" venv "$VENV" --python "$PY_VERSION" --clear
-# PyPI's torch 2.11.0 and sglang-kernel 0.4.5 are the same cu130 builds these indexes
-# serve; the explicit indexes pin provenance to the cu130 channels. `unsafe-best-match`
+# PyPI's torch 2.11.0 is the same cu130 build the pytorch index
+# serves; the explicit index pins provenance to the cu130 channel. `unsafe-best-match`
# is needed because the pytorch index also mirrors stale copies of common deps (e.g.
# packaging<=24.1) that would shadow PyPI under uv's first-index strategy; all indexes
# here are trusted. [tool.uv.sources] does not survive into a built wheel, so the
-# indexes it names must be repeated below.
+# index it names must be repeated below.
# flashinfer JIT-compiles its kernels on first use (e.g. sampling softmax), which needs nvcc --
# breaking driver-only on a box with no CUDA toolkit. flashinfer-cubin + flashinfer-jit-cache ship
# those kernels PREBUILT (multi-arch), so nothing compiles at runtime. They live on flashinfer's
@@ -220,7 +220,6 @@ INSTALL_WHEELS=(
CU_INDEX_ARGS=(
--index-strategy unsafe-best-match
--extra-index-url https://download.pytorch.org/whl/cu130
- --extra-index-url https://docs.sglang.io/whl/cu130
--extra-index-url https://flashinfer.ai/whl
--extra-index-url https://flashinfer.ai/whl/cu130
)
diff --git a/pyproject.toml b/pyproject.toml
index 5511c1f2..6b247a2c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -87,24 +87,18 @@ accel = ["freetoken[fi,sgl]"]
# in a dedicated environment when you specifically need the Marlin path.
# uv-only provenance pins — pip ignores this section and resolves from PyPI, which
-# works: PyPI's torch 2.11.0 and sglang-kernel 0.4.5 are the same cu130 builds these
-# indexes serve. `explicit = true` scopes each index to the one package that needs it
+# works: PyPI's torch 2.11.0 is the same cu130 build this index serves.
+# `explicit = true` scopes the index to the one package that needs it
# (the torch index also mirrors stale copies of common deps, e.g. packaging<=24.1,
# which would otherwise shadow PyPI under uv's first-index strategy).
[tool.uv.sources]
torch = { index = "pytorch-cu130" }
-sglang-kernel = { index = "sglang-cu130" }
[[tool.uv.index]]
name = "pytorch-cu130"
url = "https://download.pytorch.org/whl/cu130"
explicit = true
-[[tool.uv.index]]
-name = "sglang-cu130"
-url = "https://docs.sglang.io/whl/cu130"
-explicit = true
-
[tool.setuptools]
include-package-data = true
package-dir = {"" = "python"}