From fb2f80d9c52fed6417038965713d27eff3fad01b Mon Sep 17 00:00:00 2001
From: "Gabriel A. Devenyi" <gdevenyi@gmail.com>
Date: Sat, 22 Aug 2026 08:19:26 -0400
Subject: [PATCH] build(deps): make the torch CUDA build a uv group choice

`[tool.uv.sources]` pinned torch to the cu130 index unconditionally, so `uv sync` in
a checkout could only ever resolve the CUDA 13 build. CUDA 13 dropped Maxwell, Pascal
and Volta, and its wheels carry no cubin below sm_75, so on a pre-Turing card that
resolution produces a torch that cannot address the GPU at all.

Split the pin into two conflicting dependency groups. cu130 is the default group, so
a plain `uv sync` resolves byte-for-byte as it did before; a pre-Turing GPU asks for
the other one:

    uv sync --no-default-groups --group cu126

torch 2.11.0+cu126 exists on the pytorch index and satisfies the existing
`torch>=2.11,<2.12` range, so no version constraint moves. Its arch list is
sm_50/60/70/75/80/86/90, and the sm_60 cubin runs on sm_61 under CUDA's
minor-version binary compatibility guarantee.

Declaring the groups conflicting means the two can never resolve into one
environment; asking for both is a clean error rather than a silent winner.

`sglang-kernel` keeps its single cu130 pin deliberately: it publishes cu130 wheels
only and its AOT kernels are sm_75+ regardless, so a pre-Turing install leaves the
`sgl` extra off (same for flashinfer's `fi`) and falls back to the pure-triton
kernels.

Verified against the real dependency set (130 packages resolved):

    uv export                                    -> torch==2.11.0+cu130
    uv export --no-default-groups --group cu126   -> torch==2.11.0+cu126
    uv export --group cu126 --group cu130         -> error: groups are incompatible

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KQy3DziJN9peJ7nA8L59ns
---
 pyproject.toml | 31 ++++++++++++++++++++++++++++++-
 1 file changed, 30 insertions(+), 1 deletion(-)

diff --git a/pyproject.toml b/pyproject.toml
index 8bd653f87..c46edd23c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -89,10 +89,39 @@ accel = ["freetoken[fi,sgl]"]
 # indexes serve. `explicit = true` scopes each 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).
+#
+# Which CUDA build of torch to pin is a *group* choice. cu130 is the default, so a plain
+# `uv sync` resolves exactly as it did before this was split out. CUDA 13 dropped Maxwell,
+# Pascal and Volta, so its wheels carry no cubin below sm_75 and simply cannot run on those
+# cards; the cu126 build still ships sm_50/60/70/75/80/86/90, and its sm_60 cubin runs on
+# sm_61 under CUDA's minor-version binary compatibility guarantee. For a pre-Turing GPU:
+#
+#     uv sync --no-default-groups --group cu126
+#
+# The groups are declared conflicting, so the two can never resolve into one environment.
+[dependency-groups]
+cu130 = ["torch>=2.11,<2.12"]
+cu126 = ["torch>=2.11,<2.12"]
+
+[tool.uv]
+default-groups = ["cu130"]
+conflicts = [[{ group = "cu126" }, { group = "cu130" }]]
+
 [tool.uv.sources]
-torch = { index = "pytorch-cu130" }
+torch = [
+    { index = "pytorch-cu126", group = "cu126" },
+    { index = "pytorch-cu130", group = "cu130" },
+]
+# No cu126 counterpart on purpose: sglang-kernel publishes cu130 wheels only, and its AOT
+# kernels are sm_75+ regardless, so a pre-Turing install leaves the `sgl` extra off and
+# falls back to the pure-triton kernels. Same for flashinfer's `fi` extra (cu13 cubins).
 sglang-kernel = { index = "sglang-cu130" }
 
+[[tool.uv.index]]
+name = "pytorch-cu126"
+url = "https://download.pytorch.org/whl/cu126"
+explicit = true
+
 [[tool.uv.index]]
 name = "pytorch-cu130"
 url = "https://download.pytorch.org/whl/cu130"