Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Fix main's red CI: UI font tokens in the token preview, real-CUDA skip guards (#13936)

Daniel Han6d7afe0587da
diff --git a/studio/frontend/src/features/studio/sections/token-preview-dialog.tsx b/studio/frontend/src/features/studio/sections/token-preview-dialog.tsx
index 50724b4be81..670f806dfb2 100644
--- a/studio/frontend/src/features/studio/sections/token-preview-dialog.tsx
+++ b/studio/frontend/src/features/studio/sections/token-preview-dialog.tsx
@@ -66,7 +66,7 @@ const TokenStream = memo(function TokenStream({
showIds: boolean;
}) {
return (
- <div className="font-mono text-[12.5px] leading-6 whitespace-pre-wrap break-words">
+ <div className="font-mono text-ui-12p5 leading-6 whitespace-pre-wrap break-words">
{tokens.map(([id, text, trained], index) => (
<span
// biome-ignore lint/suspicious/noArrayIndexKey: tokens are positional
@@ -81,7 +81,7 @@ const TokenStream = memo(function TokenStream({
>
{visibleText(text)}
{showIds ? (
- <sub className="ml-0.5 text-[9px] opacity-60">{id}</sub>
+ <sub className="ml-0.5 text-ui-9 opacity-60">{id}</sub>
) : null}
</span>
))}
@@ -355,7 +355,7 @@ function TextPlayground({
value={text}
onChange={(event) => onTextChange(event.target.value)}
placeholder={t("studio.dataset.tokenPreview.textPlaceholder")}
- className="min-h-28 font-mono text-[12.5px]"
+ className="min-h-28 font-mono text-ui-12p5"
/>
<div className="flex flex-wrap items-center gap-2">
<p className="text-xs text-muted-foreground">
diff --git a/tests/test_fp8_rowwise_gemv_blackwell_gate.py b/tests/test_fp8_rowwise_gemv_blackwell_gate.py
index 93e6618f4b2..5eb1eaa52bb 100644
--- a/tests/test_fp8_rowwise_gemv_blackwell_gate.py
+++ b/tests/test_fp8_rowwise_gemv_blackwell_gate.py
@@ -8,6 +8,7 @@ import sys
from pathlib import Path
import pytest
+from real_accelerator import has_real_cuda
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
@@ -44,7 +45,7 @@ def test_gate(F, monkeypatch, capability, version, hip, expected):
F._fp8_gemv_miscompiles.cache_clear()
[email protected](not torch.cuda.is_available(), reason = "needs a GPU")
[email protected](not has_real_cuda(), reason = "needs a GPU")
@pytest.mark.parametrize("N, K", [(5000, 1536), (40000, 1024), (4096, 4096)])
def test_chosen_gemv_is_exact(F, N, K):
if F._fp8_kernel_unsupported(torch.empty(1, device = "cuda", dtype = torch.float8_e4m3fn)):
diff --git a/tests/test_norm_fast_launch.py b/tests/test_norm_fast_launch.py
index 8bf9c3829d9..070a0e61992 100644
--- a/tests/test_norm_fast_launch.py
+++ b/tests/test_norm_fast_launch.py
@@ -186,7 +186,7 @@ def test_launches_follow_the_current_stream(kind, width, gemma, train):
assert torch.equal(a, b)
[email protected](torch.cuda.device_count() < 2, reason = "needs two GPUs")
[email protected](not has_real_cuda() or torch.cuda.device_count() < 2, reason = "needs two GPUs")
@pytest.mark.parametrize("train", [False, True], ids = ["frozen", "trainable"])
@pytest.mark.parametrize("kind,width,gemma", CASES)
def test_second_device_with_first_current(kind, width, gemma, train):