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

Keep transformer_quant tri-state through pre-eviction validation

Browse source
This commit is contained in:
Daniel Han 2026-07-05 02:14:31 +00:00
commit 68735819cd
1 changed files with 4 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

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

@ -927,8 +927,9 @@ class DiffusionBackend:
kind = resolve_model_kind(gguf_filename, model_kind)
# Validate every mode string that can raise NOW, before this load evicts the
# previous pipeline below: their first in-line uses all sit past _unload_locked,
# where a bad request would cost the user their working model.
transformer_quant = normalize_transformer_quant(transformer_quant)
# where a bad request would cost the user their working model. Validate-only for
# transformer_quant: the raw value keeps the unset/auto vs explicit-off tri-state.
normalize_transformer_quant(transformer_quant)
normalize_speed_mode(speed_mode)
normalize_attention_backend(attention_backend)
normalize_transformer_cache(transformer_cache)
@ -1018,7 +1019,7 @@ class DiffusionBackend:
quant_plan = None
if (
kind == "gguf"
and transformer_quant is not None # normalized above, pre-eviction
and normalize_transformer_quant(transformer_quant) is not None
and dense_transformer_supported(target)
and plan.offload_policy != OFFLOAD_NONE
):