From 3bb19cd8a32f9cd25f04883028ab3dc719f3edab Mon Sep 17 00:00:00 2001
From: Jung Jiyu <142137932+jungjiyu@users.noreply.github.com>
Date: Fri, 28 Aug 2026 18:00:48 +0900
Subject: [PATCH] [Bugfix][CI/Build] Fail closed when selected precompiled CUDA
 variant is unavailable (#52545)

Signed-off-by: Jung Jiyu <jiyu.jung@navercorp.com>
---
 docs/contributing/ci/nightly_builds.md |  4 ++--
 setup.py                               | 30 ++++++++++----------------
 2 files changed, 13 insertions(+), 21 deletions(-)

diff --git a/docs/contributing/ci/nightly_builds.md b/docs/contributing/ci/nightly_builds.md
index c6d83fd9fc..795255ec69 100644
--- a/docs/contributing/ci/nightly_builds.md
+++ b/docs/contributing/ci/nightly_builds.md
@@ -134,9 +134,9 @@ When installing vLLM with `VLLM_USE_PRECOMPILED=1`, the `setup.py` script:
 
 1. **Determines wheel location** via `precompiled_wheel_utils.determine_wheel_url()`:
     - Env var `VLLM_PRECOMPILED_WHEEL_LOCATION` (user-specified URL/path) always takes precedence and skips all other steps.
-    - Determines the variant from `VLLM_MAIN_CUDA_VERSION` (can be overridden with env var `VLLM_PRECOMPILED_WHEEL_VARIANT`); the default variant will also be tried as a fallback.
+    - For CUDA, uses `VLLM_PRECOMPILED_WHEEL_VARIANT` when specified. Otherwise, an explicitly set `VLLM_MAIN_CUDA_VERSION` selects the variant; if unset, the CUDA version is detected from PyTorch or nvidia-smi, with the default `VLLM_MAIN_CUDA_VERSION` used if detection fails.
     - Determines the _base commit_ (explained later) of this branch (can be overridden with env var `VLLM_PRECOMPILED_WHEEL_COMMIT`).
-2. **Fetches metadata** from `https://wheels.vllm.ai/{commit}/vllm/metadata.json` (for the default variant) or `https://wheels.vllm.ai/{commit}/{variant}/vllm/metadata.json` (for a specific variant).
+2. **For CUDA, fetches metadata** for the selected variant from `https://wheels.vllm.ai/{commit}/{variant}/vllm/metadata.json`.
 3. **Selects compatible wheel** based on:
     - Package name (`vllm`)
     - Platform tag (architecture match)
diff --git a/setup.py b/setup.py
index 5463dcfcaf..4b3488b5fb 100644
--- a/setup.py
+++ b/setup.py
@@ -892,8 +892,7 @@ class precompiled_wheel_utils:
         1. user-specified wheel location (can be either local or remote, via
            VLLM_PRECOMPILED_WHEEL_LOCATION)
         2. user-specified variant (VLLM_PRECOMPILED_WHEEL_VARIANT) from nightly repo
-           or auto-detected CUDA variant based on system (torch, nvidia-smi)
-        3. the default variant from nightly repo
+           or CUDA variant selected from VLLM_MAIN_CUDA_VERSION, torch, or nvidia-smi
 
         If downloading from the nightly repo, the commit can be specified via
         VLLM_PRECOMPILED_WHEEL_COMMIT; otherwise, the head commit in the main branch
@@ -924,28 +923,21 @@ class precompiled_wheel_utils:
                 )
                 commit = precompiled_wheel_utils.get_base_commit_in_main_branch()
             print(f"Using precompiled wheel commit {commit} with variant {variant}")
-            try_default = False
-            wheels, repo_url, download_filename = None, None, None
+            download_filename = None
             try:
                 wheels, repo_url = precompiled_wheel_utils.fetch_metadata_for_variant(
                     commit, variant
                 )
             except Exception as e:
-                logger.warning(
-                    "Failed to fetch precompiled wheel metadata for variant %s: %s",
-                    variant,
-                    e,
-                )
-                try_default = True  # try outside handler to keep the stacktrace simple
-            if try_default:
-                print("Trying the default variant from remote")
-                wheels, repo_url = precompiled_wheel_utils.fetch_metadata_for_variant(
-                    commit, None
-                )
-                # if this also fails, then we have nothing more to try / cache
-            assert wheels is not None and repo_url is not None, (
-                "Failed to fetch precompiled wheel metadata"
-            )
+                raise RuntimeError(
+                    "Failed to fetch precompiled wheel metadata for CUDA "
+                    f"variant {variant!r} at commit {commit}. The "
+                    "root/default variant is not used as a fallback because "
+                    "its CUDA compatibility with the selected variant cannot "
+                    "be verified. Provide a compatible wheel with "
+                    "VLLM_PRECOMPILED_WHEEL_LOCATION or disable "
+                    "VLLM_USE_PRECOMPILED to build from source."
+                ) from e
             # The metadata.json has the following format:
             # see .buildkite/scripts/generate-nightly-index.py for details
             """[{