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 MiniMax-H3: keep the whole denoiser forward compilable with the render text cache (follow-up to #17203) (#17226)

Daniel Han83fd83f68dbc
diff --git a/studio/backend/core/inference/video_minimax_h3_forward.py b/studio/backend/core/inference/video_minimax_h3_forward.py
index ee778071f99..cbd46dcedc1 100644
--- a/studio/backend/core/inference/video_minimax_h3_forward.py
+++ b/studio/backend/core/inference/video_minimax_h3_forward.py
@@ -7,6 +7,8 @@
inference tensor has no version counter, so an in-place edit inside a render is not seen
(``UNSLOTH_H3_REFINER_CACHE=0`` disables).
* the packed sequence filled in place with ``index_copy_`` (``UNSLOTH_H3_INPLACE_PACK=0`` disables).
+
+Under a whole-forward ``torch.compile`` the stock forward runs (regional compile keeps this forward eager).
"""
from __future__ import annotations
@@ -58,6 +60,12 @@ def _forward_sha(cls: type) -> Optional[str]:
return sha
+def _compiling() -> bool:
+ import torch
+ is_compiling = getattr(getattr(torch, "compiler", None), "is_compiling", None)
+ return bool(is_compiling is not None and is_compiling())
+
+
def _stats(module: Any) -> dict:
stats = module.__dict__.get(STATS_ATTR)
if stats is None:
@@ -215,9 +223,14 @@ def _forward(
) -> Any:
import torch
- use_cache = self.__dict__.get(_SCOPE_ATTR) is not None and refiner_cache_enabled()
extra = {}
- if torch.is_grad_enabled() or not (use_cache or inplace_pack_enabled()):
+ if _compiling():
+ # cache key (is_inference, data_ptr) and stats counters cause graph breaks / recompiles when traced
+ fn = type(self).forward
+ elif torch.is_grad_enabled() or not (
+ (self.__dict__.get(_SCOPE_ATTR) is not None and refiner_cache_enabled())
+ or inplace_pack_enabled()
+ ):
_stats(self)["stock"] += 1
fn = type(self).forward
else:
diff --git a/studio/backend/tests/test_h3_forward.py b/studio/backend/tests/test_h3_forward.py
index 11d1a9ba50f..73648c286f9 100644
--- a/studio/backend/tests/test_h3_forward.py
+++ b/studio/backend/tests/test_h3_forward.py
@@ -311,6 +311,52 @@ def test_regional_compile_adds_no_graphs(monkeypatch):
assert patched <= stock
+def _two_renders(model, fn):
+ outs = []
+ layout = _layout()
+ for seed in (3, 4):
+ enc = _enc(seed)
+ with torch.no_grad(), h3f.h3_render_scope(_Pipe(model)):
+ for i in range(3):
+ outs.append(fn(**_step_inputs(i, enc), **layout))
+ return outs
+
+
+@needs_mirror
+def test_regional_compile_still_uses_the_text_cache(monkeypatch):
+ monkeypatch.delenv(h3f.REFINER_CACHE_ENV, raising = False)
+ monkeypatch.delenv(h3f.INPLACE_PACK_ENV, raising = False)
+ stock = _two_renders(_model(), _model())
+ torch._dynamo.reset()
+ m = _model()
+ h3f.install_h3_forward(m)
+ m.compile_repeated_blocks(backend = "eager", fullgraph = True, dynamic = False)
+ _assert_same(stock, _two_renders(m, m))
+ stats = m.__dict__[h3f.STATS_ATTR]
+ assert stats["text_misses"] == 2 and stats["text_hits"] == 4 and stats["inplace_packs"] == 6
+
+
+@needs_mirror
[email protected]("fullgraph", [True, False])
+def test_whole_forward_compile_one_graph_across_renders(monkeypatch, fullgraph):
+ from torch._dynamo.testing import CompileCounter
+ from torch._dynamo.utils import counters
+
+ monkeypatch.delenv(h3f.REFINER_CACHE_ENV, raising = False)
+ monkeypatch.delenv(h3f.INPLACE_PACK_ENV, raising = False)
+ stock = _two_renders(_model(), _model())
+ torch._dynamo.reset()
+ counters.clear()
+ m = _model()
+ assert h3f.install_h3_forward(m) is True
+ cnt = CompileCounter()
+ compiled = torch.compile(m, backend = cnt, fullgraph = fullgraph, dynamic = False)
+ _assert_same(stock, _two_renders(m, compiled))
+ assert cnt.frame_count == 1
+ assert sum(counters["graph_break"].values()) == 0
+ torch._dynamo.reset()
+
+
@needs_mirror
def test_lora_scale_is_part_of_the_key(monkeypatch):
monkeypatch.delenv(h3f.REFINER_CACHE_ENV, raising = False)