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 activation seeds keep part of the int8 DiT resident under a VRAM budget (#15985)

Daniel Han2aedb67ea08a
diff --git a/studio/backend/core/inference/diffusion_memory.py b/studio/backend/core/inference/diffusion_memory.py
index 50d813a8a75..13bb3e0b627 100644
--- a/studio/backend/core/inference/diffusion_memory.py
+++ b/studio/backend/core/inference/diffusion_memory.py
@@ -3344,18 +3344,24 @@ PARTIAL_RESIDENT_ENV = "UNSLOTH_DIFFUSION_PARTIAL_RESIDENT"
# FLUX.1 / Z-Image: their worst phase is the untiled 16-channel VAE decode (_MEASURED_IMAGE_ACTIVATION_MIB).
# MOSS-Image: 2905 MiB worst-phase peak fully streamed (B200 capped at 16 GiB).
# Qwen-Image-Edit-2511: 6046 MiB whole render (1024^2, one 1024^2 source, 40 steps, true CFG 3), int8 DiT + encoder.
+# Qwen-Image-2512 t2i: 6032 MiB whole render (1024^2, 20 steps, true CFG 4), int8 DiT + encoder, recorded on sm_120.
_MEASURED_IMAGE_PEAK_MIB: dict[str, int] = {
"qwen-image-2.1": 1849,
"flux.1": 2666,
"z-image": 2666,
"moss-image": 2976,
"qwen-image-edit": 6046,
+ "qwen-image": 6032,
}
# "0": Qwen-Image-Edit places from its learned peak only.
MEASURED_ACTIVATION_QWEN_EDIT_ENV = "UNSLOTH_DIFFUSION_MEASURED_ACTIVATION_QWEN_EDIT"
+# "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).
_MEASURED_STREAMED_IMAGE_PEAK_MIB: dict[str, int] = {
"qwen-image-edit": 3824,
+ "qwen-image": 3824,
}
MEASURED_ACTIVATION_STREAMED_ENV = "UNSLOTH_DIFFUSION_MEASURED_ACTIVATION_STREAMED"
_MEASURED_PEAK_SPEED_MODES = ("default", "max")
@@ -3533,6 +3539,10 @@ def measured_image_runtime_mib(
MEASURED_ACTIVATION_QWEN_EDIT_ENV
):
seed = None
+ if str(family or "").lower() == "qwen-image" and _env_off(
+ MEASURED_ACTIVATION_QWEN_IMAGE_ENV
+ ):
+ seed = None
if seed is None and phase_seeds:
phases = _MEASURED_IMAGE_ACTIVATION_MIB.get(str(family or "").lower())
if phases is not None:
diff --git a/studio/backend/tests/test_diffusion_qwen_image_seed.py b/studio/backend/tests/test_diffusion_qwen_image_seed.py
new file mode 100644
index 00000000000..1389e062830
--- /dev/null
+++ b/studio/backend/tests/test_diffusion_qwen_image_seed.py
@@ -0,0 +1,158 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Qwen-Image / Qwen-Image-2512 text-to-image placement seeds; CPU only, sizes from 16 / 12 GiB load logs."""
+
+from __future__ import annotations
+
+import types
+
+import pytest
+
+import core.inference.diffusion_memory as dm
+
+LOADED = {
+ "transformer": (32491, "dit"), # prequant cache line: 31.73 GiB
+ "text_encoder": (8559, "text_encoder"),
+ "vae": (242, "other"),
+}
+# "estimates={'model_dense_mib': 30603, 'companion_dense_mib': 9151, 'text_encoder_dense_mib': 8864, ...}"
+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
+BUDGETS = [(9550, 12288), (13638, 16376)]
+ENVS = (
+ "MEASURED_ACTIVATION_ENV",
+ "PARTIAL_RESIDENT_ENV",
+ "MEASURED_ACTIVATION_QWEN_EDIT_ENV",
+ "MEASURED_ACTIVATION_QWEN_IMAGE_ENV",
+ "MEASURED_ACTIVATION_STREAMED_ENV",
+ "STREAMED_RESIDENCY_ENV",
+ "RESIDENT_DIT_ENV",
+)
+
+
+def _clear(monkeypatch):
+ for env in ENVS:
+ name = getattr(dm, env, None)
+ if name:
+ monkeypatch.delenv(name, raising = False)
+ monkeypatch.setenv("UNSLOTH_DIFFUSION_LEARNED_PEAKS", "0")
+
+
+def _memory(budget_mib: int, total_mib: int) -> "dm.DeviceMemory":
+ reserve = dm._reserve_mib("discrete_vram", total_mib)
+ return dm.DeviceMemory("cuda", "cuda", "discrete_vram", budget_mib + reserve, total_mib)
+
+
+def _plan(budget_mib: int, total_mib: int):
+ plan = dm.plan_diffusion_memory(
+ target = types.SimpleNamespace(supports_model_cpu_offload = True),
+ device_memory = _memory(budget_mib, total_mib),
+ runtime_headroom_mib = dm.estimate_image_runtime_mib(
+ width = None, height = None, family = "qwen-image"
+ ),
+ requested_mode = None,
+ **FLAT,
+ )
+ if plan.offload_policy == dm.OFFLOAD_MODEL:
+ plan = dm.torchao_streaming_plan(plan)
+ return plan
+
+
+def pipe(monkeypatch):
+ _clear(monkeypatch)
+ monkeypatch.setattr(dm, "_loaded_component_mib", lambda p: dict(LOADED), raising = False)
+ monkeypatch.setattr(dm, "_pipe_denoisers_hold_torchao", lambda p: True)
+ monkeypatch.setattr(dm, "_stream_window_mib", lambda p: WINDOW)
+ return object()
+
+
+def _refine(
+ pipe,
+ plan,
+ family = "qwen-image",
+ speed = "default",
+):
+ return dm.refine_plan_from_loaded_weights(pipe, plan, family = family, speed_mode = speed)
+
+
+def _wide_room(memory, headroom: int) -> int:
+ return (
+ int(memory.free_mib)
+ - dm._resident_dit_slack_mib(memory)
+ - LOADED["vae"][0]
+ - headroom
+ - WINDOW
+ )
+
+
+def test_seed_values(monkeypatch):
+ _clear(monkeypatch)
+ m = dm.measured_image_runtime_mib
+ assert m("qwen-image", "default") == RESIDENT_HEADROOM
+ assert m("qwen-image", "max") == RESIDENT_HEADROOM
+ assert m("qwen-image", "default", streamed = True) == STREAMED_HEADROOM
+ assert m("qwen-image", "default", width = 1328, height = 1328) > RESIDENT_HEADROOM
+ assert m("qwen-image", "off") is None
+ assert m("qwen-image", "off", dense_transformer_mib = 38968) == 5120
+ # 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
+
+
[email protected]("budget,total", BUDGETS)
+def test_budget_keeps_part_of_the_denoiser_resident(pipe, budget, total):
+ plan = _plan(budget, total)
+ assert plan.stream_transformer and plan.offload_policy in (
+ dm.OFFLOAD_GROUP,
+ dm.OFFLOAD_STREAMING,
+ )
+ assert not plan.resident_transformer_mib
+ new = _refine(pipe, plan)
+ memory = plan.device_memory
+ resident = new.resident_transformer_mib
+ assert resident and resident == _wide_room(memory, STREAMED_HEADROOM)
+ assert resident < LOADED["transformer"][0]
+ assert new.estimates["measured_streamed_headroom_mib"] == STREAMED_HEADROOM
+ 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(
+ memory.free_mib
+ )
+ assert new.offload_policy == plan.offload_policy
+ assert new.stream_text_encoders == plan.stream_text_encoders
+
+
[email protected]("budget,total", BUDGETS)
+def test_kill_switch_restores_the_all_streamed_plan(pipe, monkeypatch, budget, total):
+ monkeypatch.setenv(dm.MEASURED_ACTIVATION_QWEN_IMAGE_ENV, "0")
+ assert dm.measured_image_runtime_mib("qwen-image", "default") is None
+ assert dm.measured_image_runtime_mib("qwen-image", "default", streamed = True) is None
+ plan = _plan(budget, total)
+ assert _refine(pipe, plan) is plan
+ assert dm.measured_image_runtime_mib("qwen-image-edit", "default") == RESIDENT_HEADROOM
+
+
+def test_streamed_kill_switch_falls_back_to_the_resident_seed(pipe, monkeypatch):
+ monkeypatch.setenv(dm.MEASURED_ACTIVATION_STREAMED_ENV, "0")
+ plan = _plan(13638, 16376)
+ new = _refine(pipe, plan)
+ assert new.resident_transformer_mib == _wide_room(plan.device_memory, RESIDENT_HEADROOM)
+ assert "measured_streamed_headroom_mib" not in new.estimates
+
+
[email protected]("speed", ["off", "eager"])
+def test_unseeded_tiers_keep_the_flat_plan(pipe, speed):
+ plan = _plan(13638, 16376)
+ assert _refine(pipe, plan, speed = speed) is plan
+
+
+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