Release 0.12.0

docs/ACCELERATION.md

@@ -198,8 +198,9 @@ Compile a component once it is fully configured - the graph captures final dtype
 
 Typical gains are 1.3-1.5x on diffusion transformers, and compilation stacks with the caches above. Notes:
 
-- **First run pays the compile cost.** The [server](SERVER.md)'s persistent worker keeps compiled pipelines loaded between runs, so the cost is paid once per session rather than once per generation. A new session still retraces, but the kernels it built come back from the on-disk cache, which dw keeps across reboots (`TORCHINDUCTOR_CACHE_DIR`, see [Environment Defaults](#environment-defaults)).
+- **First run pays the compile cost.** The [server](SERVER.md)'s persistent worker keeps compiled pipelines loaded between runs, so the cost is paid once per session rather than once per generation. A new session still retraces, but the kernels it built come back from the on-disk cache, which dw keeps across reboots (`TORCHINDUCTOR_CACHE_DIR`, see [Environment Defaults](#environment-defaults)). SDNQ's `triton.jit`/autotune kernels are cached separately by Triton itself (`TRITON_CACHE_DIR`, else `$TRITON_HOME/.triton/cache`, else `~/.triton/cache`); dw sets no `TRITON_*` variable and leaves that cache where it is, shared with any other Triton user on the machine. Both caches' sizes show in `get_server_info`'s `caches` and `GET /api/caches`, and `DELETE /api/caches/{name}` empties one ([SERVER.md](SERVER.md)); the next first step of each shape pays the compile or re-autotune again (100-300 s per shape on MiniMax-H3 for Triton).
 - **Pin the attention backend** on a compiled component (`"attention_backend"` in the same `components` entry) rather than using the pipeline-level per-call context manager, which forces recompiles.
+- **Quantize-on-load time is a separate cost.** An SDNQ on-the-fly component re-reads and re-quantizes its bf16 weights on every cold load; `DW_QUANTIZED_CACHE=1` saves the quantized result and reads it back instead - see [QUANTIZATION.md](QUANTIZATION.md#cached-on-the-fly-dw_quantized_cache).
 - **Composes with offloading**: apply `group_offload` and `compile` on the same component and the offload hooks are installed first, as required. Skipped with a warning on MPS.
 - **Don't combine `fullgraph` with a `cache`**: the cache hooks decide skip-or-compute per step, a data-dependent branch diffusers wraps in `torch.compiler.disable` - it needs the graph break that `fullgraph: true` forbids. Compile with the default (partial) graph mode when a cache is active.
 - **TorchAO quantization needs compile to be fast** - see [QUANTIZATION.md](QUANTIZATION.md#torchao).
@@ -295,7 +296,7 @@ Set automatically at import unless already present in the environment (export yo
 | `PYTORCH_CUDA_ALLOC_CONF` | `expandable_segments:True` | Lets the CUDA allocator grow segments instead of fragmenting fixed-size ones. Multi-step workflows churn differently-shaped allocations (generate, upscale, interpolate); fragmentation is what OOMs a card that nominally has room. |
 | `HF_ENABLE_PARALLEL_LOADING` | `true` | Loads sharded checkpoints in parallel - faster cold starts. |
 | `PYTORCH_MPS_HIGH_WATERMARK_RATIO` | `0.0` | MPS only - use all available unified memory. |
-| `TORCHINDUCTOR_CACHE_DIR` | `$XDG_CACHE_HOME/dw/torchinductor` (`~/.cache/dw/torchinductor`) | Keeps `torch.compile`'s on-disk cache across reboots - Inductor's own default is under `/tmp`. A new worker still retraces, but a cache hit skips codegen and kernel builds, the bulk of a cold compile. Entries are per shape, so a new size or frame count compiles once more. |
+| `TORCHINDUCTOR_CACHE_DIR` | `$XDG_CACHE_HOME/dw/torchinductor` (`~/.cache/dw/torchinductor`) | Keeps `torch.compile`'s on-disk cache across reboots - Inductor's own default is under `/tmp`. A new worker still retraces, but a cache hit skips codegen and kernel builds, the bulk of a cold compile. Entries are per shape, so a new size or frame count compiles once more. Listed as `inductor` by `GET /api/caches` and cleared with `DELETE /api/caches/inductor`. Triton's cache is not set by dw and stays at its own default (`~/.triton/cache`) - see the compile notes above. |
 
 For faster model downloads, optionally `pip install hf_transfer` and set `HF_HUB_ENABLE_HF_TRANSFER=1`. Not enabled automatically - it bypasses the Python HTTP stack and breaks some proxy setups.
 

docs/ARCHITECTURE.md

@@ -28,6 +28,7 @@ one to open.
 | Device translation | `dw/pipeline_processors/placement.py`: `place_component` | A device is translated by `resolve_device()` before anything reads its backend, so the MPS accommodations also fire for a translated device. | `tests/test_device_portability.py::TestPlacement::test_a_translated_device_gets_its_backends_offload_downgrade` |
 | On-demand wrappers | `dw/pipeline_processors/placement.py`: `apply_on_demand_placement` | The on-demand wrappers keep the wrapped signature, because H3's denoiser reads `signature(transformer.forward)`. | `tests/test_modular_pipeline.py::TestOnDemandResidency::test_the_wrapped_signature_survives` |
 | On-demand and offload | `dw/pipeline_processors/placement.py`: `place_component`, `has_component_group_offload` | On-demand residency and `group_offload` cannot both be set on one component, and either one stops the whole pipeline from being moved to the device. | `tests/test_modular_pipeline.py::TestOnDemandResidency::test_group_offload_and_on_demand_together_are_rejected`, `tests/test_modular_pipeline.py::TestOffloadDevice::test_component_group_offload_is_not_moved_to_device` |
+| Quantized component cache | `dw/quantized_cache.py`: `load_components`, `from_pretrained`, `cache_key` | A key that can't pin the weights isn't cached, the cache never fails a job, and the save comes before pruning, so an entry is the whole component. | `tests/test_quantized_cache.py` |
 | LoRAs | `dw/pipeline_processors/adapters.py`: `active_loras`, `load_loras` | A `loras` entry whose `model_name` is null is switched off. | `tests/test_lora_disable.py::TestLoadLoras::test_an_all_null_entry_loads_nothing` |
 | Step progress | `dw/pipeline_processors/progress.py`: `reported_progress_bars` | A pipeline with no step callback reports its denoise steps through its progress bars. | `tests/test_modular_progress.py::test_every_advance_of_the_bar_is_reported` |
 | Community pipelines | `dw/community_pipelines/` (e.g. `pipeline_ltx2_refine.py`: `LTX2RefinePipeline`), `dw/introspection.py`: `community_pipelines`, `load_allowed_class` | A pipeline dw ships is named by its dotted path (`dw.community_pipelines.<module>.<Class>`), loads untrusted because it is under `dw`, and is listed by `list_pipelines` and resolved by `get_pipeline_signature` - only the modules in that directory, never a caller-chosen import. `LTX2RefinePipeline` VAE-encodes `video` at `width` x `height` and passes it unnormalized as `latents`, because `LTX2Pipeline.prepare_latents` normalizes 5-D latents itself. | `tests/test_ltx2_refine_pipeline.py::TestIntrospection::test_only_the_shipped_modules_resolve`, `tests/test_ltx2_refine_pipeline.py::TestEncode::test_the_latents_are_left_unnormalized_for_the_parent` |
@@ -125,7 +126,7 @@ one to open.
 | Opening media | `dw/media.py` | This is the only module that calls `av.open`. It imports no torch and no step types. | `tests/test_media_layering.py::test_av_open_appears_only_in_media`, `tests/test_media_layering.py::test_media_imports_no_step_types` (no torch: —) |
 | Signal processing | `dw/dsp.py` | It measures and transforms waveforms, decides nothing, and imports nothing from `dw`. | `tests/test_media_layering.py::test_dsp_imports_nothing_from_dw` |
 | Assessment probes | `dw/tasks/assess.py`, `dw/assessment_rules.py` | Exactly three registered commands are assessment probes. Each measures a finished file, and every rule names a real probe field. | `tests/test_assessment_rules.py::TestRuleProbesAreRealCommands::test_assessment_is_exactly_the_three_probes`, `tests/test_assessment_rules.py::TestRuleProbesAreRealCommands::test_every_rule_names_a_registered_json_command` |
-| Script check | `dw/tasks/script_check.py`: `check`; `dw/script_lines.py`: `parse_lines`, `parse_shots`, `shot_names_error` | `check_script` aligns a take's heard words to its expected lines and guards out words heard over silence and repetition loops (Whisper's hallucination over music). It builds its findings with `assessment_rules.finding()` and reads the guard's floor and window (`DEAD_AIR_FLOOR_DBFS`, `DEAD_AIR_WINDOW`) rather than copying them. It is not an assessment probe and adds nothing to `RULES`. Its literal `lines` is refused at validation through `task_problems` (`script_lines_errors`), which calls the same `parse_lines` the task runs, and a literal `shots` through the same `parse_shots` and `shot_names_error`; the parsers live in `dw/script_lines.py`, not the task, so that `task_problems` reaches them without importing `script_check` (the `import_cycles` ratchet in `scripts/arch_metrics.py`). Its shot map comes from `assess.resolve_shots`, the one the probes use, never a copy; it places each shot on the soundtrack with `assess.sample_span`, decides repeated shot names with `shots.duplicate_shot_names`, and asks `locations.is_http_url` whether the take is a URL. | `tests/test_script_check.py::TestRegistration::test_registered_as_json_command_without_assessment_flag`, `tests/test_script_check.py::TestEmptyLines::test_words_only_over_silence_are_discarded_not_found`, `tests/test_script_check.py::TestLinesErrors::test_literal_with_a_colon_is_not_a_reference` |
+| Script check | `dw/tasks/script_check.py`: `check`; `dw/script_lines.py`: `parse_lines`, `parse_shots`, `shot_names_error` | `check_script` aligns a take's heard words to its expected lines and guards out words heard over silence and repetition loops (Whisper's hallucination over music). It builds its findings with `assessment_rules.finding()` and reads the guard's floor and window (`DEAD_AIR_FLOOR_DBFS`, `DEAD_AIR_WINDOW`) rather than copying them; the loudest-window measure is `dsp.loudest_window_dbfs` (`script_check.word_level` wraps it), which `transcribe_audio`'s `suspect: "silent"` mark reads with the same floor and window; the repetition-loop detector (`script_lines.looped_texts`, `REPEAT_RUN_MIN`, `REPEAT_MAX_PERIOD`) lives in `dw/script_lines.py` and backs both `check_script`'s `repetition` discard and `transcribe_audio`'s `suspect: "repetition"`, and `script_lines.suspect_runs`/`SUSPECT_LEFT_OUT` are what `plan_cuts` groups and leaves out suspect chunks by - so neither task imports the other. It is not an assessment probe and adds nothing to `RULES`. Its literal `lines` is refused at validation through `task_problems` (`script_lines_errors`), which calls the same `parse_lines` the task runs, and a literal `shots` through the same `parse_shots` and `shot_names_error`; the parsers live in `dw/script_lines.py`, not the task, so that `task_problems` reaches them without importing `script_check` (the `import_cycles` ratchet in `scripts/arch_metrics.py`). Its shot map comes from `assess.resolve_shots`, the one the probes use, never a copy; it places each shot on the soundtrack with `assess.sample_span`, decides repeated shot names with `shots.duplicate_shot_names`, and asks `locations.is_http_url` whether the take is a URL. | `tests/test_script_check.py::TestRegistration::test_registered_as_json_command_without_assessment_flag`, `tests/test_script_check.py::TestEmptyLines::test_words_only_over_silence_are_discarded_not_found`, `tests/test_script_check.py::TestLinesErrors::test_literal_with_a_colon_is_not_a_reference` |
 | Audio sync against a source slice | `dw/tasks/measure_sync.py`: `measure_sync`; `dw/dsp.py`: `envelope_lag` | `measure_sync` cross-correlates the onset envelopes of a shot's audio and its source slice and answers the lag and a 0-1 confidence. A positive lag means the audio is LATE (its events fall after the reference's), negative early; `lag_seconds` is None for a silent or flat track. It builds its findings (`sync_unmeasurable`, `sync_low_confidence`, `audio_out_of_sync`) with `assessment_rules.finding()` over module-local rule dicts and adds nothing to `RULES`; it reads the silence floor from `beats.SILENT_DBFS`. `envelope_lag` imports nothing from `dw`. | `tests/test_measure_sync.py::TestLag::test_late_audio_is_a_positive_lag`, `tests/test_measure_sync.py::TestLag::test_early_audio_is_a_negative_lag`, `tests/test_measure_sync.py::TestUnmeasurable::test_a_silent_reference_has_no_lag` |
 
 ## Security and architecture guardrails
@@ -152,6 +153,7 @@ one to open.
 | Worker pool dispatch | `dw/server/jobs.py`: `JobManager._next_dispatch`; `dw/server/pool.py`: `WorkerSlot`, `check_fits`, `largest_ceiling_gb`; `dw/vram_estimate.py`: `required_vram_gb`; `dw/server/admission.py`: `pool_capacity` | `--devices` runs one worker per card; a queued job goes to a free card that fits what it needs, a job that fits no free card keeps its place while a smaller one behind it takes the card, a job bigger than every card is refused at submit, and cancel or a crash touches only the worker running that job. On a backend no `cost` entry describes, a declared `vram_estimate` is held at admission to the pool's largest card, so a job the first card cannot hold is not refused when a larger card can hold it. | `tests/test_worker_pool.py::TestConcurrency::test_two_workers_run_two_jobs_at_once`, `tests/test_worker_pool.py::TestBackfill::test_a_job_too_big_for_the_free_card_waits_while_a_smaller_one_takes_it`, `tests/test_worker_pool.py::TestFitAtSubmit::test_a_job_too_big_for_every_card_is_refused_naming_the_largest`, `tests/test_worker_pool.py::TestCancel::test_cancel_reaches_the_worker_running_the_job`, `tests/test_worker_pool.py::TestCrashIsolation::test_a_crashed_worker_fails_only_its_own_job`, `tests/test_worker_pool.py::TestOomRanking::test_the_later_started_worker_is_ranked_first`, `tests/test_worker_pool.py::TestFitAtSubmit::test_a_job_only_the_larger_card_holds_is_admitted` |
 | Card affinity | `dw/server/pool.py`: `choose_slot`; `dw/server/jobs.py`: `JobManager.route`; `dw/worker_protocol.py`: `workflow_identity` | Among the free cards a job fits it goes to the rerun's original card, else the card whose worker last ran its workflow identity (`workflow_identity`, the one rule the worker caches by too), else one with no worker running, else the first in `--devices` order; a cache probe and `plan.estimate` (`priced_for`) ask about the card `route` names, and memory is read and cleared per card. | `tests/test_worker_pool.py::TestIdentityAffinity::test_a_second_run_lands_on_the_card_that_ran_the_workflow_first`, `tests/test_worker_pool.py::TestRerunAffinity::test_a_rerun_lands_on_the_original_card_when_it_is_free`, `tests/test_worker_pool.py::TestProbeRouting::test_the_probe_goes_to_the_worker_the_job_would_be_routed_to`, `tests/test_worker_pool.py::TestMemoryPerCard::test_a_named_card_running_a_job_is_refused_though_the_other_is_idle` |
 | Per-card memory and cache | `dw/server/worker_memory.py`: `memory_status`, `clear_memory`, `probe_cache` | Each card keeps its own last memory reading; an idle card's worker is asked live, a busy one answers with the run's last report and says why it is not live, and a clear leaves a card running a job alone. Every request takes the card's lock with a bounded wait, and a card's lock is never taken while the JobManager's lock is held. | `tests/test_worker_pool.py::TestMemoryPerCard::test_a_named_card_running_a_job_is_refused_though_the_other_is_idle`, `tests/test_worker_pool.py::TestProbeRouting::test_the_probe_goes_to_the_worker_the_job_would_be_routed_to`, `tests/test_server.py::test_memory_says_why_a_reading_is_not_the_worker_s` |
+| dw disk caches | `dw/disk_caches.py` | Members `inductor`, `triton`, `quantized`. Clear only a root carrying the `.dw-cache` marker, never while a job is running or queued (`quantized` also not during a model download); an entry delete takes only a 64-hex id, checked before any path is built; clear renames into `.clearing-*` then deletes. | `tests/test_disk_caches.py` |
 | Job-to-run link | `dw/server/job_record.py`, `dw/server/jobs.py`: `JobManager.realized` | A job records `run_id`, `run_dir` and `run_version`, and `JobManager.realized` reads that run's `workflow.json`, confined to the output root. | `tests/test_server_jobs.py::test_realized_reads_the_file_the_run_wrote`, `tests/test_server_jobs.py::test_realized_refuses_a_run_dir_that_escapes_the_output_root` |
 | Job history | `dw/server/job_history.py`: `JobHistory` | Finished jobs persist in `jobs.sqlite`, in WAL mode, so the Jobs view survives a restart. | `tests/test_server.py::test_job_history_survives_restart_and_reruns`, `tests/test_job_history_wal.py::test_the_jobs_database_uses_wal_mode` |
 | HTTP security | `dw/server/http_security.py`: `install_middleware`, `query_token_ok` | Every API route needs the token, and only a route marked `query_token_ok` takes it as `?token=`. | `tests/test_security_auth.py::TestTheTokenGate::test_every_api_spelling_needs_the_token`, `tests/test_security_auth.py::TestTheTokenGate::test_the_query_token_is_refused_where_it_is_not_allowed` |