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: read the hosted LTX-2.5 int8 checkpoint when the load's own plan offloads the denoiser (#17143)

Daniel Han5cafbf2b1156
diff --git a/studio/backend/core/inference/video.py b/studio/backend/core/inference/video.py
index 8c038c63e28..9ef8587db6a 100644
--- a/studio/backend/core/inference/video.py
+++ b/studio/backend/core/inference/video.py
@@ -4409,6 +4409,19 @@ def _auto_offload_scheme(requested: Optional[str]) -> Optional[str]:
return TQ_INT8 if scheme == TQ_AUTO else scheme
+# "0" disables reading the hosted native file when the load's own plan offloads a seeded denoiser.
+OFFLOAD_HOSTED_READ_ENV = "UNSLOTH_VIDEO_OFFLOAD_HOSTED_READ"
+
+
+def _offload_hosted_read_enabled() -> bool:
+ return (os.environ.get(OFFLOAD_HOSTED_READ_ENV) or "").strip().lower() not in (
+ "0",
+ "off",
+ "false",
+ "no",
+ )
+
+
def _video_native_hosted_scheme(fam: Any, scheme: Optional[str], target: Any) -> Optional[str]:
"""``int8`` when an offloaded load of ``scheme`` builds the W8A8 ConvRot native DiT that the family's hosted int8
checkpoint already holds (``diffusion_native_hosted``), else None. The load re-checks the file; never raises."""
@@ -9519,6 +9532,16 @@ class VideoBackend:
denoiser_seed_scheme,
_video_plan_label(plan),
)
+ if _video_native_hosted_planned is None and _offload_hosted_read_enabled():
+ # The probe planned a resident seed; read the hosted native build rather than quantising dense shards.
+ _video_native_hosted_planned = _video_native_hosted_scheme(
+ fam, denoiser_seed_scheme, target
+ )
+ if _video_native_hosted_planned is not None:
+ logger.info(
+ "video.denoiser_prequant: the hosted %s checkpoint is read as offload-safe native Linears",
+ _video_native_hosted_planned,
+ )
denoiser_seed_scheme = None
denoiser_seed_gb = None
plan, bf16_plan, quant_replanned = _plan_for_te_scale(te_scale, log = False)
diff --git a/studio/backend/tests/test_video_backend.py b/studio/backend/tests/test_video_backend.py
index 4489ca8426b..124b50be7b0 100644
--- a/studio/backend/tests/test_video_backend.py
+++ b/studio/backend/tests/test_video_backend.py
@@ -10689,6 +10689,44 @@ def test_seeding_is_skipped_entirely_under_offload(fake_runtime, monkeypatch):
assert "moves the DiT" in status["resolved"]["transformer_quant"]["reason"]
[email protected]("switch", [None, "0"])
+def test_a_seed_declined_by_the_loads_own_offload_plan_reads_the_hosted_native_build(
+ fake_runtime, monkeypatch, switch
+):
+ """Probe prices the seed resident but the load's plan offloads it (LTX-2.5 sidecar encoder): read the hosted file."""
+ import core.inference.video as video_mod
+
+ if switch is not None:
+ monkeypatch.setenv(video_mod.OFFLOAD_HOSTED_READ_ENV, switch)
+ monkeypatch.setattr(video_mod, "dense_transformer_supported", lambda target: True)
+ monkeypatch.setattr(video_mod, "quantize_transformer", lambda *a, **k: "int8")
+ _stub_denoiser_seed(monkeypatch, scheme = "int8", plan_scheme = "int8")
+ real_plan = video_mod.plan_diffusion_memory
+ monkeypatch.setattr(
+ video_mod,
+ "plan_diffusion_memory",
+ lambda **kwargs: dataclasses.replace(real_plan(**kwargs), offload_policy = "model"),
+ )
+ _stub_apply_memory_plan(monkeypatch, video_mod)
+ monkeypatch.setattr(
+ video_mod, "_video_native_hosted_scheme", lambda fam, scheme, target: "int8"
+ )
+ monkeypatch.setattr(video_mod, "_video_native_cache_spec", lambda fam, **kw: {"scheme": "int8"})
+ monkeypatch.setattr(video_mod, "_read_native_quant_from_cache", lambda *a, **k: {})
+ monkeypatch.setattr(video_mod, "_video_native_stream_build", lambda *a, **k: {})
+ hosted_reads = []
+ monkeypatch.setattr(
+ video_mod, "_read_native_quant_from_hosted", lambda *a, **k: hosted_reads.append(k) or {}
+ )
+
+ # The fake target lacks a native int8 kernel, so the load refuses after the read.
+ with contextlib.suppress(RuntimeError):
+ VideoBackend().load_pipeline(
+ "Wan-AI/Wan2.2-T2V-A14B-Diffusers", model_kind = "pipeline", transformer_quant = "int8"
+ )
+ assert len(hosted_reads) == (0 if switch == "0" else 1)
+
+
def _plans_for(monkeypatch, video_mod):
"""Every ``model_dense_mib`` the memory planner was asked to judge, in order."""
real_plan = video_mod.plan_diffusion_memory