ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Wire the hosted HiDream I1 int8/fp8 checkpoints

Browse source 
Gate-validated: all 28 per-case pairs pass per scheme (LPIPS suite means 0.291
int8 / 0.278 fp8, in the 50-step trajectory-divergence band; CLIP delta means
0.007-0.008), and the int8 checkpoint is verified bit-identical to on-the-fly
quantize across all 1615 state dict tensors (1073 quantized, max abs diff 0.0).
Uploaded to unsloth/HiDream-I1-Full-FP8.
This commit is contained in:
Daniel Han 2026-07-18 05:19:36 +00:00
commit 930d29be98
2 changed files with 17 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

7
studio/backend/core/inference/diffusion_families.py
View file

@ -363,6 +363,13 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (
# need the same TE4 assembly for tiny demand, so no GGUF artifact is wired yet.
DiffusionFamily(
name = "hidream-i1",
# Hosted checkpoints: 28/28 per-case gate pairs per scheme (LPIPS suite means 0.291
# int8 / 0.278 fp8, the 50-step trajectory band); int8 verified bit-identical to
# on-the-fly quantize across all 1615 state dict tensors.
prequant_repos = (
("int8", "unsloth/HiDream-I1-Full-FP8"),
("fp8", "unsloth/HiDream-I1-Full-FP8"),
),
pipeline_class = "HiDreamImagePipeline",
transformer_class = "HiDreamImageTransformer2DModel",
base_repo = "HiDream-ai/HiDream-I1-Full",

10
studio/backend/tests/test_diffusion_more_families.py
View file

@ -296,6 +296,16 @@ def test_ideogram4_memory_table_counts_both_dits():
assert text_encoders_gb > 5.0
def test_hidream_prequant_wiring():
# Hosted int8/fp8 checkpoints (28/28 per-case gate pairs per scheme; int8 verified
# bit-identical to on-the-fly quantize) serve the family default base.
from core.inference.diffusion_families import family_prequant_repo
fam = detect_family("HiDream-ai/HiDream-I1-Full")
for scheme in ("int8", "fp8"):
assert family_prequant_repo(fam, scheme) == "unsloth/HiDream-I1-Full-FP8"
def test_hidream_quant_schemes_not_denied_and_no_extra_excludes():
# Measured on a B200 (outputs/hidream_smoke): int8 and fp8 both engage and render
# cleanly, including a 2-3 token prompt on int8 -- the routed MoE expert Linears