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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Studio diffusion: arm Z-Image's prompt-length inputs for an int8 / fp8 compile (#14715)
Daniel Han725455e1ae59
diff --git a/studio/backend/core/inference/diffusion_dynamic_text.py b/studio/backend/core/inference/diffusion_dynamic_text.pyindex b7a4143de46..849cd062f4d 100644--- a/studio/backend/core/inference/diffusion_dynamic_text.py+++ b/studio/backend/core/inference/diffusion_dynamic_text.py@@ -121,11 +121,13 @@ def prompt_length_varies(transformer: Any) -> bool:return isinstance(PROMPT_LENGTH_POLICY.get(cls.__name__), tuple)-# What the compile arms (Z-Image is classified above but not armed here). FLUX.1 int8 / fp8 (torchao weights compile-# with automatic dynamic) recompiled both blocks on the first prompt past 256 T5 tokens and again on the first new-# resolution (31 s and 60 s on a loaded B200); armed, the first compile covers every text length and image size.+# What the compile arms. int8 / fp8 FLUX.1 and Z-Image (torchao weights compile with automatic dynamic) recompiled on+# new text lengths: FLUX.1 both blocks on the first prompt past 256 T5 tokens and again on the first new resolution (31 s+# and 60 s on a loaded B200); Z-Image's shared block on the caption, image and unified lengths (12 recompiles in the+# first render) and again on the first longer prompt. Armed, the first compile covers every text length and image size._FAMILY_SOURCES: dict[str, tuple[str, ...]] = {"FluxTransformer2DModel": _FLUX1_SOURCES,+ "ZImageTransformer2DModel": _ZIMAGE_SOURCES,"Flux3Transformer2DModel": _FLUX3_SOURCES,"QwenImage21Transformer2DModel": _QWEN_IMAGE_21_SOURCES,"QwenImageTransformer2DModel": _QWEN_IMAGE_SOURCES,diff --git a/studio/backend/tests/test_diffusion_dynamic_text.py b/studio/backend/tests/test_diffusion_dynamic_text.pyindex 7df70cf4ff1..c5d257a433d 100644--- a/studio/backend/tests/test_diffusion_dynamic_text.py+++ b/studio/backend/tests/test_diffusion_dynamic_text.py@@ -543,6 +543,52 @@ def test_slice_identity_fix_is_probe_gated_and_reversible(monkeypatch):assert SliceVariable.__dict__.get("get_real_python_backed_value") is stock+class ZImageTransformer2DModel(torch.nn.Module):+ """Z-Image's one block class runs the caption refiner (caption tokens), the noise refiner (image tokens) and the+ main layers (caption + image), positionally as (x, attn_mask, freqs_cis); the caption is padded to 32-token buckets."""++ def __init__(self):+ super().__init__()+ self.block = torch.compile(self._block, backend = "eager", dynamic = None)++ @staticmethod+ def _block(x, attn_mask, freqs_cis):+ return (x * freqs_cis).masked_fill(~attn_mask[..., None], 0.0) + x++ def forward(self, caption_len, image_len):+ def run(n):+ x = torch.randn(1, n, 8)+ return self.block(x, torch.ones(1, n, dtype = torch.bool), torch.randn(1, n, 8))++ run(caption_len)+ run(image_len)+ return run(caption_len + image_len)+++def test_zimage_new_caption_bucket_reuses_the_first_graphs():+ """An automatic-dynamic Z-Image recompiled its block for each new length (caption, image, unified) inside the+ first render, then again on the first longer prompt; armed, the first compile covers every length."""+ from torch._dynamo.utils import counters++ def graphs_per_render(install):+ torch._dynamo.reset()+ counters.clear()+ m = ZImageTransformer2DModel()+ if install:+ assert dt.install(m)+ seen = []+ for caption_len, image_len in ((32, 64), (32, 64), (96, 64), (32, 48)):+ before = counters["stats"]["unique_graphs"]+ m(caption_len, image_len)+ seen.append(counters["stats"]["unique_graphs"] - before)+ dt.uninstall(m)+ torch._dynamo.reset()+ return seen++ assert graphs_per_render(False)[0] > 1+ assert graphs_per_render(True) == [1, 0, 0, 0]++class FluxTransformer2DModel(torch.nn.Module):"""FLUX.1's two compiled blocks: the double block sees the text stream beside the image tokens, the single blockconcatenates them, and the rotary tables span text + image."""diff --git a/studio/backend/tests/test_diffusion_prompt_length_policy.py b/studio/backend/tests/test_diffusion_prompt_length_policy.pyindex d3cfd0856ce..5b40d3e2b9b 100644--- a/studio/backend/tests/test_diffusion_prompt_length_policy.py+++ b/studio/backend/tests/test_diffusion_prompt_length_policy.py@@ -62,9 +62,10 @@ def test_prompt_length_varies_reads_the_policy():def test_the_compile_arms_only_its_own_sources():- # Z-Image is classified but compiles as before+ # every classified family is armedassert set(dt._FAMILY_SOURCES) == {"FluxTransformer2DModel",+ "ZImageTransformer2DModel","Flux3Transformer2DModel","QwenImage21Transformer2DModel","QwenImageTransformer2DModel",