[feat] [4/n] Improve API: refactor sampling param and merge with presets#1234
Merged
SolitaryThinker merged 8 commits intomainfrom Apr 16, 2026
Merged
[feat] [4/n] Improve API: refactor sampling param and merge with presets#1234SolitaryThinker merged 8 commits intomainfrom
SolitaryThinker merged 8 commits intomainfrom
Conversation
- Add fastvideo/api/profiles.py with PipelineProfile, ProfileStageSpec, and profile registry (register/get/validate) - Add pipeline-local profiles for all model families under fastvideo/pipelines/basic/*/profiles.py - Migrate Wan family: remove 14 SamplingParam subclasses, use default_profile in registry, profiles provide all defaults - Add _from_profile() to SamplingParam.from_pretrained() for profile-based default resolution - Add default_profile and model_family to ConfigInfo in registry - Add guidance_scale_2 to base SamplingParam - Break LingBotWorld cross-family inheritance - Update SSIM tests to use SamplingParam.from_pretrained() - Add comprehensive profile tests (test_profiles.py)
…m defaults Delete 12 SamplingParam subclass files (LTX2, Hunyuan, Hunyuan15, HYWorld, GameCraft, LingBotWorld, MatrixGame, GEN3C, Cosmos, Cosmos25, SD35, TurboDiffusion) and replace them with profile-based defaults. Each family's register_configs now sets sampling_param_cls=None and default_profile="<name>", so SamplingParam.from_pretrained() resolves defaults from the pipeline profile instead of subclass fields. Model-specific fields (LTX2 multi-modal CFG/STG, HYWorld attention masks, GameCraft camera/action control) promoted to base SamplingParam. LongCat entries also get default_profile for completeness.
Canonical location for SamplingParam is now fastvideo.api.sampling_param. All 48 import sites updated. Old configs/sample/base.py and __init__.py kept as backward-compatible re-export shims.
…swallowing, version sort - deepcopy profile defaults to prevent shared mutable list references between SamplingParam instances (e.g. ltx2_stg_blocks_video, sigmas) - narrow bare except Exception to (ValueError, RuntimeError) in _from_profile() so real errors aren't silently swallowed - use numeric-aware version sorting in get_profile() to handle multi-segment versions correctly (e.g. "2" > "10" string sort bug) - remove dead get_sampling_param_cls_for_name fallback from from_pretrained() since all sampling_param_cls are now None - update test_ltx2_registry.py for profile-based behavior
Delete the backward-compatible re-export shims in configs/sample/ and update all docs, tests, and CI config to point to the canonical location at fastvideo/api/sampling_param.py.
mergify
Bot
added
scope: training
Contributor
|
Contributor
Merge ProtectionsYour pull request matches the following merge protections and will not be merged until they are valid. 🔴 PR merge requirementsWaiting for:
This rule is failing.
🟢 📃 Configuration Change RequirementsWonderful, this rule succeeded.Mergify configuration change
|
SolitaryThinker
changed the title
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request introduces a centralized pipeline profile registry to manage inference presets and stage-based validation across model families, effectively replacing the legacy system of multiple SamplingParam subclasses. Key additions include the profile management system in fastvideo/api/profiles.py and the relocation of SamplingParam to fastvideo/api/sampling_param.py with integrated profile support. Review feedback highlights the need for more robust attribute validation in SamplingParam to prevent overwriting methods via hasattr, the correction of logger.exception usage, and a potential TypeError in the version sorting logic within the profile registry.
Collaborator
Author
|
/test full |
…tage_schemas (#1234) Address PR #1234 review feedback. Rename the profile registry to "preset" throughout: - PipelineProfile → InferencePreset, ProfileStageSpec → PresetStageSpec - fastvideo/api/profiles.py → presets.py; 13 pipeline profiles.py → presets.py - Registry API, ConfigInfo.default_profile, PipelineSelection fields renamed Simplify version: str → int; drop the _version_key helper (every registered version is "1", plain int comparison suffices). Rename InferencePreset.stages → stage_schemas to clarify it is validation metadata; the real execution DAG lives in each pipeline's create_pipeline_stages(). Misc review fixes: - logger.exception → logger.warning in SamplingParam.update() - Collapse double config lookup via get_preset_selection() - Fix "image-to-prompt" typo on wan_fun_1_3b_inp - Loosen test_total_profile_count (== 37 → >= 37) - Revert CRLF whitespace churn in test_lingbot_similarity.py - Remove unused _clear_registry
SolitaryThinker
added a commit
that referenced
this pull request
Apr 16, 2026
…tage_schemas (#1234) Address PR #1234 review feedback. Rename the profile registry to "preset" throughout: - PipelineProfile → InferencePreset, ProfileStageSpec → PresetStageSpec - fastvideo/api/profiles.py → presets.py; 13 pipeline profiles.py → presets.py - Registry API, ConfigInfo.default_profile, PipelineSelection fields renamed Simplify version: str → int; drop the _version_key helper (every registered version is "1", plain int comparison suffices). Rename InferencePreset.stages → stage_schemas to clarify it is validation metadata; the real execution DAG lives in each pipeline's create_pipeline_stages(). Misc review fixes: - logger.exception → logger.warning in SamplingParam.update() - Collapse double config lookup via get_preset_selection() - Fix "image-to-prompt" typo on wan_fun_1_3b_inp - Loosen test_total_profile_count (== 37 → >= 37) - Revert CRLF whitespace churn in test_lingbot_similarity.py - Remove unused _clear_registry Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…ields (#1234) hasattr-based validation let callers overwrite methods (e.g. update itself) or other non-field attributes. Switch update() and _from_preset() to check dataclasses.fields instead, and use logger.error with a clearer "has no field" message on unknown keys.
SolitaryThinker
added a commit
that referenced
this pull request
Apr 16, 2026
…ields (#1234) hasattr-based validation let callers overwrite methods (e.g. update itself) or other non-field attributes. Switch update() and _from_preset() to check dataclasses.fields instead, and use logger.error with a clearer "has no field" message on unknown keys. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Collaborator
Author
|
/merge |
SolitaryThinker
force-pushed
the
will/api_4
branch
from
6254d6a to
fbfa43c
Compare
SolitaryThinker
changed the title
SolitaryThinker
changed the title
SolitaryThinker
added a commit
that referenced
this pull request
Apr 22, 2026
PR 4 (#1234) moved SamplingParam defaults into preset.defaults dicts. Four entries (matrixgame × 1, turbodiffusion × 3) set 'negative_prompt': None, which overrides SamplingParam.negative_prompt (typed str) with None. When the CFG branch in text_encoding.py:81 runs, 'assert isinstance(batch.negative_prompt, str)' crashes. Concrete repro: the TurboWan2.2-I2V-A14B-Diffusers SSIM test on Modal L40S fails with AssertionError at text_encoding.py:81 before any generation happens. CFG is entered despite guidance_scale=1.0 because of a separate bug — LTX-2-specific class defaults on shared SamplingParam (ltx2_cfg_scale_video=3.0, ltx2_cfg_scale_audio=7.0) force do_classifier_free_guidance=True in ForwardBatch.__post_init__ for any non-LTX2 model that doesn't explicitly override them. That deeper issue is tracked in PR plan.md § 'LTX2-specific defaults leak onto shared SamplingParam' — this commit only fixes the type-contract violation that surfaces it. Fix: change 'negative_prompt': None -> 'negative_prompt': ''. Runtime-equivalent (both mean 'no negative prompt') but keeps the str invariant so the CFG branch no longer crashes. Regression guard: new test_presets.py::TestPresetDefaultTypes asserts no preset sets negative_prompt=None across all registered presets.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Purpose
PR 4 of the API improvement series. Introduces a profile registry — named inference presets that replace the 20+
SamplingParam subclasses scattered across fastvideo/configs/sample/ — and moves the canonical SamplingParam to
fastvideo/api/sampling_param.py.
Continues from #1226 (PR 3), #1220 (PR 2), #1218 (PR 1).
Changes
register_profile() / get_profile() with typed validation of stage overrides
hunyuan, hunyuan15, hyworld, gamecraft, lingbotworld, longcat, matrixgame, gen3c, cosmos, sd35, turbodiffusion) —
37 profiles total
HYWorld, GameCraft, GEN3C, LongCat) promoted to the base class; from_pretrained() → _from_profile() chain for
profile-based defaults
model_family + default_profile; added _register_profiles(), get_model_family(), get_default_profile()
,cosmos,cosmos2_5,sd35,turbodiffusion}.py); configs/sample/{base,init}.py are backward-compatible re-export
shims
except Exception to (ValueError, RuntimeError), numeric-aware version sorting
Checklist
pre-commit run --all-filesand fixed all issuesFor model/pipeline changes, also check: