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: offloaded native int8 denoisers release their safetensors file once pinned (LTX-2.5 holds ~23 GiB less host mapping) (#17065)

Daniel Han3c1fbf2a0cff
diff --git a/studio/backend/core/inference/diffusion_native_cache.py b/studio/backend/core/inference/diffusion_native_cache.py
index 0d03a1e715e..6f5f40c4f4e 100644
--- a/studio/backend/core/inference/diffusion_native_cache.py
+++ b/studio/backend/core/inference/diffusion_native_cache.py
@@ -501,6 +501,10 @@ def load(
scale = scale,
),
)
+ from .prequant_safetensors import _unviewed
+
+ # No-copy aliases: a view's base survives offload's ``.data =`` re-point and keeps the whole file mapped.
+ state = {key: _unviewed(value) for key, value in state.items()}
model.load_state_dict(state, strict = True, assign = True)
del state
if any(t.device.type == "meta" for t in (*model.parameters(), *model.buffers())):
diff --git a/studio/backend/core/inference/diffusion_native_hosted.py b/studio/backend/core/inference/diffusion_native_hosted.py
index e519c5a4589..d672b9c988c 100644
--- a/studio/backend/core/inference/diffusion_native_hosted.py
+++ b/studio/backend/core/inference/diffusion_native_hosted.py
@@ -213,6 +213,10 @@ def load_native_from_hosted(
)
state[fqn + ".weight_q"] = codes
state[fqn + ".weight_scale"] = scale.view(torch.int32)
+ from .prequant_safetensors import _unviewed
+
+ # No-copy aliases: a view's base survives offload's ``.data =`` re-point and keeps the whole file mapped.
+ state = {key: _unviewed(value) for key, value in state.items()}
model.load_state_dict(state, strict = True, assign = True)
del state
if any(t.device.type == "meta" for t in (*model.parameters(), *model.buffers())):
diff --git a/studio/backend/tests/test_video_native_hosted.py b/studio/backend/tests/test_video_native_hosted.py
index 5732ef9cdce..1fa7b168058 100644
--- a/studio/backend/tests/test_video_native_hosted.py
+++ b/studio/backend/tests/test_video_native_hosted.py
@@ -7,7 +7,9 @@ file, verified against its recorded fingerprint, and refused (dense build) whene
from __future__ import annotations
+import gc
import json
+import os
import sys
import types
from pathlib import Path
@@ -288,3 +290,26 @@ def test_only_an_opted_in_family_offload_int8_takes_the_hosted_file(monkeypatch)
monkeypatch.setattr(V, "native_int8_act", lambda target: True)
monkeypatch.setenv(NH.NATIVE_HOSTED_ENV, "0")
assert V._video_native_hosted_scheme(ltx25, "int8", _target()) is None
+
+
+def _file_mapped(path) -> bool:
+ real = os.path.realpath(str(path))
+ with open("/proc/self/maps") as maps:
+ return any(line.rstrip("\n").endswith(real) for line in maps)
+
+
+def test_offload_repoint_releases_the_hosted_file(tmp_path):
+ # Group offload re-points each weight at its pinned copy; a view's base would keep the file mapped.
+ path, cfg_dir, _built = _write_hosted(tmp_path)
+ model = _load(path, cfg_dir)
+ assert model is not None
+ tensors = [*model.parameters(), *model.buffers()]
+ assert all(t._base is None for t in tensors)
+ if not os.path.exists("/proc/self/maps"):
+ return
+ assert _file_mapped(path)
+ for t in tensors:
+ t.data = t.data.clone()
+ del t, tensors
+ gc.collect()
+ assert not _file_mapped(path)
diff --git a/studio/backend/tests/test_video_native_prequant_cache.py b/studio/backend/tests/test_video_native_prequant_cache.py
index e6ac8af8631..3541fb45aef 100644
--- a/studio/backend/tests/test_video_native_prequant_cache.py
+++ b/studio/backend/tests/test_video_native_prequant_cache.py
@@ -357,3 +357,34 @@ def test_hosted_int8_is_not_the_offload_build():
assert not int8_convrot_enabled(FAMILY)
assert _spec()["rot_group"] == 256
+
+
+def _file_mapped(path) -> bool:
+ real = os.path.realpath(str(path))
+ with open("/proc/self/maps") as maps:
+ return any(line.rstrip("\n").endswith(real) for line in maps)
+
+
+def test_offload_repoint_releases_the_cache_entry(base_dir):
+ import gc
+
+ fam = _family()
+ spec = _spec()
+ fresh = _fresh_build(base_dir, "transformer")
+ assert NC.save(fresh, fam, spec, base_repo = base_dir, component = "transformer") is not None
+ del fresh
+ path = NC.lookup(fam, spec, base_repo = base_dir, component = "transformer")
+ cached = NC.load(
+ path, diffusers.WanTransformer3DModel, base_dir, component = "transformer", spec = spec
+ )
+ assert cached is not None
+ tensors = [*cached.parameters(), *cached.buffers()]
+ assert all(t._base is None for t in tensors)
+ if not os.path.exists("/proc/self/maps"):
+ return
+ assert _file_mapped(path)
+ for t in tensors:
+ t.data = t.data.clone()
+ del t, tensors
+ gc.collect()
+ assert not _file_mapped(path)