digital-garbageandClaude Sonnet 5 1e4553e658 fix: H3 attention decoupling was scaling K magnitude, which barely moves softmax ...

Live 0-1 strength sweep came back flat -- confirmed the mechanism, not just the
value, was wrong. H3's Q/K get RMSNorm'd before attention, so scaling a key
toward zero pushes its logit toward the *mean* of an already near-zero-centered
distribution, not toward excluded. Switched to what actually suppresses a
softmax entry: an additive negative bias on the targeted logits, still passed
through the same per-query-group split (still no S^2 mask, still same total
FLOPs -- the bias is a per-key vector broadcast over the group's queries).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

2026-09-05 02:23:30 +03:00

..

_comfy_stubs.py

feat(lora): report weights that match by name but not by shape

2026-08-22 23:18:36 +03:00

conftest.py

test: kill the 22 chronic failures — they were stub collisions, not redundant tests

2026-08-03 11:48:58 +03:00

test_alg_guidance.py

feat: schedules for every sampler, a usable frames field, ALG on any sampler

2026-08-11 03:53:11 +03:00

test_bachvid_kv.py

KV-Lock Phase 2: variance-gated scheduler drives BachVid injection

2026-06-21 17:14:06 +03:00

test_batched_phrase_vectors.py

perf(studio): classify every uncertain phrase in one encode instead of one each

2026-08-25 13:36:23 +03:00

test_block_influence.py

fix: rank blocks by share of movement, not by a ratio depth decides

2026-09-04 18:10:55 +03:00

test_block_repeat_last_steps.py

feat: h3_block_repeat_last_steps goes negative for 'apply early, then stop'

2026-09-04 23:01:27 +03:00

test_category_vector_cache.py

perf(studio): embed the category descriptions once per encoder, not once per run

2026-08-25 13:34:08 +03:00

test_companion_conditioning.py

fix(sampler): a wired reference conditioning is not a second scene

2026-08-25 14:06:52 +03:00

test_conditioning_payload_length.py

feat(log): one line per feature, state then reason

2026-08-26 23:10:27 +03:00

test_context_windows.py

fix: don't reject projects saved with the old context_window_schedule names

2026-07-31 18:10:10 +03:00

test_cut_opening.py

fix: cut_opening_frames cuts decoded pixels, not latents (noisy first frame)

2026-08-11 23:47:51 +03:00

test_detail_probe.py

feat: span loop — 31,32..40,31,32..40 instead of 31,31,32,32..40,40

2026-09-04 19:56:38 +03:00

test_detailing.py

fix: drop guide keyframes on a resized second pass; unify the two ALG switches

2026-08-11 23:39:47 +03:00

test_diagnostics.py

feat: FUNPACK_FAULTHANDLER for crashes and hangs

2026-08-23 04:35:41 +03:00

test_distilled_flow_ancestral.py

fix(sampler): Distilled Flow's s_noise is rectified-flow ancestral now

2026-08-25 17:53:45 +03:00

test_distilled_flow_correction_window.py

fix(sampler): final_correction_steps counts steps that can be corrected

2026-08-25 17:45:13 +03:00

test_dynashift.py

feat: re-enable DynaShift on H3, add a positive-pull bank alongside

2026-09-03 23:22:35 +03:00

test_encode_tally.py

perf(studio): count text-encoder passes by purpose, per run

2026-08-25 13:32:44 +03:00

test_encoder_family.py

fix(studio): right tokenizer family, and never the network first

2026-08-23 04:32:44 +03:00

test_finite_sample_guard.py

fix: the H3 v2a hint only fires when JoyAI audio memory is on; guard the inputs too

2026-08-04 22:48:15 +03:00

test_funpack_log.py

maint: remove the chunk-causal lane

2026-08-26 23:39:32 +03:00

test_gguf_support.py

revert(gguf): back to the load path from 6f04642

2026-08-23 08:12:43 +03:00

test_h3_av_decouple.py

fix: H3 attention decoupling was scaling K magnitude, which barely moves softmax

2026-09-05 02:23:30 +03:00

test_h3_final_layer.py

fix(h3): picture detail ran backwards — H3's final-layer scale is negative

2026-08-27 00:38:46 +03:00

test_h3_interior_pins.py

feat(h3): region-locked anchors, and the taste direction lands after the refiner

2026-08-25 21:48:28 +03:00

test_h3_latent_upscaler.py

feat(detailing): load MiniMax H3's latent upscaler natively

2026-08-21 00:03:43 +03:00

test_h3_region_lock.py

feat(h3): region-locked anchors, and the taste direction lands after the refiner

2026-08-25 21:48:28 +03:00

test_h3_repr_steering_multiblock.py

feat: REINS steers multiple blocks at once, block picker covers all 50

2026-09-05 00:03:56 +03:00

test_h3_repr_steering.py

fix: sweep gallery ratings were single-slot, not per-item -- rebuilt on independent slots

2026-09-05 01:22:23 +03:00

test_h3_token_weights.py

feat: variability dial on H3 phrase emphasis

2026-09-03 16:58:34 +03:00

test_h3_weight_dtype_warning.py

feat(loaders): say it when H3 is loaded at a precision that breaks it

2026-08-27 00:16:43 +03:00

test_host_memory.py

feat: cap ComfyUI's pinned memory without a launch flag

2026-08-23 04:17:41 +03:00

test_identity_transfer.py

Identity transfer: support ComfyUI's new RoPE matrix layout

2026-07-29 17:28:35 +03:00

test_joyai_memory.py

editor: JoyAI-Echo paired audio memory + v2a coupling (phase 2)

2026-06-24 19:04:32 +03:00

test_ksampler_schedule.py

feat: schedules for every sampler, a usable frames field, ALG on any sampler

2026-08-11 03:53:11 +03:00

test_loaders.py

revert(gguf): back to the load path from 6f04642

2026-08-23 08:12:43 +03:00

test_logs_do_not_quote_the_prompt.py

fix(h3): the emphasis line reports counts, never the phrases themselves

2026-08-25 14:42:05 +03:00

test_loop_style.py

Loop roll v2: content-only roll — guides and audio memory stay pinned

2026-07-15 21:55:22 +03:00

test_lora_shape_mismatch.py

maint: remove the chunk-causal lane

2026-08-26 23:39:32 +03:00

test_ltx25_compat.py

fix: two LTX-2.5 assumptions that would have failed silently

2026-08-11 22:36:38 +03:00

test_mask_safe_attention.py

test: kill the 22 chronic failures — they were stub collisions, not redundant tests

2026-08-03 11:48:58 +03:00

test_minimax_h3_sampler.py

maint: remove the mid-scene guide

2026-08-27 14:32:20 +03:00

test_minimax_h3_studio.py

fix(studio): the wire is the instruction — drop the ownership toggle

2026-08-23 09:12:22 +03:00

test_minimax_h3.py

fix(lora): a file that matched nothing says what it looked like, and is not "Active"

2026-08-26 23:20:31 +03:00

test_negative_erase.py

fix(studio): cap the negative_erase renorm gain

2026-08-23 03:52:45 +03:00

test_per_block_any_model.py

feat(log): one line per feature, state then reason

2026-08-26 23:10:27 +03:00

test_project_editor_settings.py

feat: the project remembers your Editor settings, so a new rental doesn't reset them

2026-08-06 19:41:18 +03:00

test_prompt_enhancer.py

fix(studio): the enhancer generated twice on every Editor run

2026-08-26 17:48:07 +03:00

test_pulse_temporal.py

Add Pulse temporal style with per-scene ease-down segments

2026-06-10 16:50:19 +03:00

test_quality_sharpness_ksampler.py

feat(samplers): quality sharpness on any KSampler

2026-08-23 00:08:57 +03:00

test_reference_conditioning_ownership.py

feat(log): one line per feature, state then reason

2026-08-26 23:10:27 +03:00

test_reference_row_protection.py

fix(h3): the tokenizer's labels are not the user's prompt

2026-08-23 09:55:21 +03:00

test_refinement_keys.py

feat: disable every rating-driven mechanism on H3 except phrase emphasis and REINS

2026-09-03 18:46:30 +03:00

test_run_phase.py

fix: don't reject projects saved with the old context_window_schedule names

2026-07-31 18:10:10 +03:00

test_scene_chain_sampler.py

maint: remove the mid-scene guide

2026-08-27 14:32:20 +03:00

test_scene_wrapper_hygiene.py

fix(sampler): sweep leaked hooks below the blocks; log a hook census per run

2026-08-20 19:14:24 +03:00

test_score_slider.py

sampler: contrastive-pair FreeSliders — minus pole from learned bad_dir

2026-06-24 17:14:25 +03:00

test_second_pass_layout_conds.py

fix: drop guide keyframes on a resized second pass; unify the two ALG switches

2026-08-11 23:39:47 +03:00

test_second_pass_multiscene.py

fix: multi-scene chains work with a resolution-changing second pass

2026-08-12 00:05:52 +03:00

test_second_pass.py

feat(samplers): optional second_pass_sampler

2026-08-23 03:25:10 +03:00

test_settings_card_noise.py

maint: finish prompt repair, and label bounded attention untested

2026-08-27 14:41:10 +03:00

test_shortcut_revolver.py

Shortcut revolver: no-repeat cycling through multi-replacement shortcuts

2026-07-06 05:26:21 +03:00

test_sla_attention.py

refactor(loaders): SLA composes with the attention backend instead of replacing it

2026-08-20 23:36:01 +03:00

test_stage_timing.py

feat(editor): Settings > Appearance gets "Show timeline on hover"

2026-08-25 14:32:11 +03:00

test_state_io_tally.py

perf(studio): report what the refinement key costs, and what is making it big

2026-08-25 13:48:35 +03:00

test_studio_wasted_encodes.py

perf(studio): don't locate phrase ranges for a patch that already declined

2026-08-25 13:26:43 +03:00

test_tensor_sidecar.py

perf(studio): keep the key's tensors in a binary sidecar, not as base64 in its JSON

2026-08-25 13:54:01 +03:00

test_timeline_shortcut_safety.py

fix(editor): a stray letter could split a clip from a window you were typing in

2026-08-27 00:32:17 +03:00

test_trajectory_guidance.py

fix: trajectory guidance never trained — the heads were empty every run

2026-09-03 14:41:00 +03:00

test_trajectory_probe.py

feat: a way to throw the learned state away and start clean

2026-09-03 14:47:26 +03:00

test_v2_category_alignment.py

fix: H3 scale ratings were silently discarded and replaced with the default

2026-09-03 18:56:28 +03:00

test_value_function_device.py

diag(vf): say where the value functions trained, and how big they are

2026-08-25 14:34:26 +03:00