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.

Studio diffusion: keep the VRAM Budget reserve free in the decode-time residency release (#15048)

Daniel Han6792704cf094
diff --git a/studio/backend/core/inference/diffusion_memory.py b/studio/backend/core/inference/diffusion_memory.py
index a137937f884..030461b4bf3 100644
--- a/studio/backend/core/inference/diffusion_memory.py
+++ b/studio/backend/core/inference/diffusion_memory.py
@@ -4534,7 +4534,9 @@ def release_for_decode(
if dev.index is None:
dev = torch.device("cuda", torch.cuda.current_device())
need = decode_need_mib(lend, width = width, height = height, batch_size = batch_size)
- free = int(torch.cuda.mem_get_info(dev)[0])
+ free, total = (int(x) for x in torch.cuda.mem_get_info(dev))
+ # keep the VRAM Budget reserve free, like every other decode-time check
+ free -= user_budget_reserve_bytes(dev, total)
# only the cache an ordinary decode allocation can reuse: a CUDA graph pool's free blocks are not
cached = max(
0,
diff --git a/studio/backend/tests/test_diffusion_phase_peaks.py b/studio/backend/tests/test_diffusion_phase_peaks.py
index cfedc564ee5..6eed61d6f2b 100644
--- a/studio/backend/tests/test_diffusion_phase_peaks.py
+++ b/studio/backend/tests/test_diffusion_phase_peaks.py
@@ -410,6 +410,23 @@ class _GuardTorch:
self.cuda = cuda
+def test_decode_release_keeps_the_vram_budget_reserve_free(monkeypatch):
+ calls = []
+ monkeypatch.setattr(
+ dm,
+ "release_resident_groups",
+ lambda pipe, need, logger = None, reason = "", device = None, **kw: calls.append(need)
+ or (lambda: None),
+ )
+ pipe = types.SimpleNamespace()
+ pipe._unsloth_decode_lend = (4040, 5900, 1024 * 1024, False)
+ # 7500 MiB covers the decode, but not plus a 2 GiB reserve
+ monkeypatch.setattr(dm, "user_budget_reserve_bytes", lambda device, total: 2048 << 20)
+ monkeypatch.setitem(sys.modules, "torch", _GuardTorch(free_mib = 7500))
+ assert dm.release_for_decode(pipe, width = 1024, height = 1024, device = "cuda:0") is not None
+ assert calls == [6785 + dm._DECODE_SLACK_MIB - (7500 - 2048)]
+
+
def test_decode_streams_lent_encoders_only_when_free_memory_is_short(monkeypatch):
calls = []
monkeypatch.setattr(