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: Qwen-Image text-to-image streamed placement uses its own activation figure under a VRAM budget (#16095)

Daniel Han08b323912901
diff --git a/studio/backend/core/inference/diffusion_memory.py b/studio/backend/core/inference/diffusion_memory.py
index 13bb3e0b627..cdedea2d4be 100644
--- a/studio/backend/core/inference/diffusion_memory.py
+++ b/studio/backend/core/inference/diffusion_memory.py
@@ -3358,12 +3358,15 @@ MEASURED_ACTIVATION_QWEN_EDIT_ENV = "UNSLOTH_DIFFUSION_MEASURED_ACTIVATION_QWEN_
# "0": Qwen-Image / Qwen-Image-2512 text-to-image places from its learned peak only.
MEASURED_ACTIVATION_QWEN_IMAGE_ENV = "UNSLOTH_DIFFUSION_MEASURED_ACTIVATION_QWEN_IMAGE"
# Streamed tier: peak reserved above resident weights AND the prefetch window (the resident seed already pads it).
-# Qwen-Image t2i takes the Edit figure: same DiT class and VAE, half the image tokens (no source latents).
+# Qwen-Image t2i has its own figure: the borrowed Edit figure (3824) left 3.5 GiB unused at a 16 GiB budget.
_MEASURED_STREAMED_IMAGE_PEAK_MIB: dict[str, int] = {
"qwen-image-edit": 3824,
- "qwen-image": 3824,
+ "qwen-image": 2560,
}
MEASURED_ACTIVATION_STREAMED_ENV = "UNSLOTH_DIFFUSION_MEASURED_ACTIVATION_STREAMED"
+# "0": Qwen-Image t2i streams with the borrowed Edit figure again.
+QWEN_IMAGE_STREAMED_SEED_ENV = "UNSLOTH_DIFFUSION_QWEN_IMAGE_STREAMED_SEED"
+_QWEN_IMAGE_BORROWED_STREAMED_MIB = 3824
_MEASURED_PEAK_SPEED_MODES = ("default", "max")
_MEASURED_PEAK_MARGIN = 1.15
_MEASURED_PEAK_ROUND_MIB = 256
@@ -3535,6 +3538,8 @@ def measured_image_runtime_mib(
seed = _MEASURED_IMAGE_PEAK_MIB.get(str(family or "").lower())
if streamed and phase is None and not _env_off(MEASURED_ACTIVATION_STREAMED_ENV):
seed = _MEASURED_STREAMED_IMAGE_PEAK_MIB.get(str(family or "").lower(), seed)
+ if str(family or "").lower() == "qwen-image" and _env_off(QWEN_IMAGE_STREAMED_SEED_ENV):
+ seed = _QWEN_IMAGE_BORROWED_STREAMED_MIB
if str(family or "").lower() == "qwen-image-edit" and _env_off(
MEASURED_ACTIVATION_QWEN_EDIT_ENV
):
diff --git a/studio/backend/tests/test_diffusion_qwen_image_seed.py b/studio/backend/tests/test_diffusion_qwen_image_seed.py
index 1389e062830..7e7976a05b1 100644
--- a/studio/backend/tests/test_diffusion_qwen_image_seed.py
+++ b/studio/backend/tests/test_diffusion_qwen_image_seed.py
@@ -20,7 +20,10 @@ LOADED = {
FLAT = dict(model_dense_mib = 30603, companion_dense_mib = 9151, text_encoder_dense_mib = 8864)
WINDOW = 1620 # "streams 60 offload groups with event-fenced prefetch (2 ahead, 1620 MiB window)"
RESIDENT_HEADROOM = 7168 # seed x 1.15, rounded up to 256 MiB
-STREAMED_HEADROOM = 4608
+STREAMED_HEADROOM = (
+ 3072 # Qwen-Image t2i's own streamed figure (2560) x 1.15, rounded up to 256 MiB
+)
+EDIT_STREAMED_HEADROOM = 4608 # the Qwen-Image-Edit figure (3824), which t2i borrowed before
BUDGETS = [(9550, 12288), (13638, 16376)]
ENVS = (
"MEASURED_ACTIVATION_ENV",
@@ -30,6 +33,7 @@ ENVS = (
"MEASURED_ACTIVATION_STREAMED_ENV",
"STREAMED_RESIDENCY_ENV",
"RESIDENT_DIT_ENV",
+ "QWEN_IMAGE_STREAMED_SEED_ENV",
)
@@ -101,7 +105,7 @@ def test_seed_values(monkeypatch):
# GGUF / packed denoisers (no static seed) still answer from a learned peak alone
assert m("qwen-image", "default", static_seed = False) is None
assert m("qwen-image-edit", "default") == RESIDENT_HEADROOM
- assert m("qwen-image-edit", "default", streamed = True) == STREAMED_HEADROOM
+ assert m("qwen-image-edit", "default", streamed = True) == EDIT_STREAMED_HEADROOM
@pytest.mark.parametrize("budget,total", BUDGETS)
@@ -121,7 +125,7 @@ def test_budget_keeps_part_of_the_denoiser_resident(pipe, budget, total):
assert new.estimates["measured_runtime_headroom_mib"] == RESIDENT_HEADROOM
# the encode streams the resident groups back out (the 16 GiB Edit overrun)
assert new.estimates["encode_resident_transformer_mib"] == 0
- assert resident + 3824 + WINDOW + LOADED["vae"][0] + dm._resident_dit_slack_mib(memory) <= int(
+ assert resident + 2560 + WINDOW + LOADED["vae"][0] + dm._resident_dit_slack_mib(memory) <= int(
memory.free_mib
)
assert new.offload_policy == plan.offload_policy
@@ -156,3 +160,61 @@ def test_dense_bf16_denoiser_is_not_placed_from_the_int8_seed(pipe, monkeypatch)
monkeypatch.setattr(dm, "_pipe_denoisers_hold_torchao", lambda p: False)
plan = _plan(13638, 16376)
assert _refine(pipe, plan) is plan
+
+
+def test_t2i_streamed_figure_is_its_own_and_below_the_edit_figure(monkeypatch):
+ _clear(monkeypatch)
+ assert dm._MEASURED_STREAMED_IMAGE_PEAK_MIB["qwen-image"] == 2560
+ assert dm._MEASURED_STREAMED_IMAGE_PEAK_MIB["qwen-image-edit"] == 3824
+ m = dm.measured_image_runtime_mib
+ assert m("qwen-image", "default", streamed = True, width = 1328, height = 1328) > STREAMED_HEADROOM
+ assert m("qwen-image", "default", streamed = True, width = 1328, height = 1328) < m(
+ "qwen-image-edit", "default", streamed = True, width = 1328, height = 1328
+ )
+ assert m("qwen-image", "default") == RESIDENT_HEADROOM
+
+
+def test_t2i_streamed_kill_switch_restores_the_borrowed_figure(monkeypatch):
+ _clear(monkeypatch)
+ monkeypatch.setenv(dm.QWEN_IMAGE_STREAMED_SEED_ENV, "0")
+ m = dm.measured_image_runtime_mib
+ assert m("qwen-image", "default", streamed = True) == EDIT_STREAMED_HEADROOM
+ assert m("qwen-image", "default") == RESIDENT_HEADROOM
+ assert m("qwen-image-edit", "default", streamed = True) == EDIT_STREAMED_HEADROOM
+ monkeypatch.setenv(dm.MEASURED_ACTIVATION_STREAMED_ENV, "0")
+ assert m("qwen-image", "default", streamed = True) == RESIDENT_HEADROOM
+ monkeypatch.setenv(dm.MEASURED_ACTIVATION_QWEN_IMAGE_ENV, "0")
+ assert m("qwen-image", "default", streamed = True) is None
+
+
[email protected]("budget,total", BUDGETS)
+def test_t2i_figure_keeps_more_resident_than_the_borrowed_one(pipe, monkeypatch, budget, total):
+ plan = _plan(budget, total)
+ new = _refine(pipe, plan)
+ monkeypatch.setenv(dm.QWEN_IMAGE_STREAMED_SEED_ENV, "0")
+ old = _refine(pipe, plan)
+ memory = plan.device_memory
+ # kill switch == the placement before this change, byte for byte
+ assert old.resident_transformer_mib == _wide_room(memory, EDIT_STREAMED_HEADROOM)
+ assert old.estimates["measured_streamed_headroom_mib"] == EDIT_STREAMED_HEADROOM
+ gained = new.resident_transformer_mib - old.resident_transformer_mib
+ assert gained == EDIT_STREAMED_HEADROOM - STREAMED_HEADROOM
+ assert new.resident_transformer_mib + STREAMED_HEADROOM + WINDOW + LOADED["vae"][
+ 0
+ ] + dm._resident_dit_slack_mib(memory) == int(memory.free_mib)
+ assert new.estimates["encode_resident_transformer_mib"] == 0
+
+
+def test_learned_peak_above_the_t2i_figure_still_raises_it(monkeypatch):
+ _clear(monkeypatch)
+ import core.inference.diffusion_measured_peaks as mp
+
+ monkeypatch.setattr(mp, "lookup", lambda **kw: 4000)
+ assert dm.measured_image_runtime_mib("qwen-image", "default", quant = "q", streamed = True) == 4608
+ monkeypatch.setattr(
+ mp, "lookup", lambda **kw: 1907
+ ) # the learned 16 GiB render: never lowers the figure
+ assert (
+ dm.measured_image_runtime_mib("qwen-image", "default", quant = "q", streamed = True)
+ == STREAMED_HEADROOM
+ )