fix: reuse conditioning cache for reference-free Fixed V3.9 runs
Reuse the existing bounded prompt-conditioning cache only for safe reference-free Fixed groups. Preserve Reference, List/Timeline, terminal-pair, Sampling and Run Storage paths; add 22 regression cases and bilingual update guidance. Local validation: 1558 CPU passed, 1 skipped; eight successful GPU generations. Post-fix same-session warm total medians: V3.8 273.579s vs V3.9 273.779s, with overlapping ranges. This scoped result is not a universal speed or GPU-kernel claim. Official workflows/ZIPs and historical Releases are unchanged.
This commit is contained in:
+6
-1
@@ -15,7 +15,8 @@ from ..compatibility import accelerator_summary, check_comfy_h3_runtime
|
||||
from ..constants import (
|
||||
DIAGNOSTICS_FULL, DIAGNOSTICS_OFF, DIAGNOSTICS_OPTIONS, FPS,
|
||||
CONTINUUM_ACTUAL_PREFIX_STEPS, SEAM_CORRECTION_AUTO, SEAM_CORRECTION_OFF,
|
||||
SEAM_CORRECTION_OPTIONS, V2_CONTINUITY_AUTO, normalize_diagnostics_mode,
|
||||
SEAM_CORRECTION_OPTIONS, V2_CONTINUITY_AUTO, normalize_diagnostics_mode,
|
||||
PROMPT_MODE_FIXED,
|
||||
)
|
||||
from ..continuation import POLICY_REPLACE, prepare_conditioning
|
||||
from ..driving_audio import (
|
||||
@@ -1014,6 +1015,10 @@ def _run_runtime_internal(*,model:Any,clip:Any,video_vae:Any,audio_vae:Any,sampl
|
||||
timeline_video_assets=timeline_video_assets,
|
||||
reference_encode_cache=reference_encode_cache,
|
||||
cache_event=reference_encode_cache_events.append,
|
||||
prompt_conditioning_cache=bool(
|
||||
prompt_conditioning_cache and plan["mode"] == PROMPT_MODE_FIXED
|
||||
),
|
||||
prompt_cache_event=prompt_cache_events.append,
|
||||
expected_group_contract=(reference_storage_plan.group_for(sequence_index+1)
|
||||
if reference_storage_plan is not None else None),
|
||||
first_frame_hash=assets.first_frame_hash,
|
||||
|
||||
Reference in New Issue
Block a user