[feat] [3/n] Improve API: extend support to cli

[SolitaryThinker]

Motivation

The previous CLI surface mirrored every SamplingParam / FastVideoArgs field
as a flat argparse flag (~80+ flags on generate alone). This made it hard to
keep CLI parity with the config schema, created brittle translation code, and
required per-model shell scripts full of duplicate flag lists. The nested config
approach lets users maintain one YAML per model and override individual values
on the command line without needing to know the full flag set.

Purpose

• Replaces the flat argparse-flag inference CLI with a config-driven model.
  fastvideo generate and fastvideo serve now accept only --config <path>
  plus optional dotted overrides (e.g. --request.sampling.seed 42).

• Config-only CLI: Both subcommands take a single --config pointing to a
  nested YAML/JSON file with generator: and request: (or server:)
  top-level sections. All other CLI input uses dotted paths that map directly
  into the config tree.

• Request metadata tracking: New request_metadata.py module tracks which
  GenerationRequest fields were explicitly provided by the user vs filled by
  schema defaults. Uses a 3-way merge (raw config + baseline snapshot + current
  state) with __setattr__-based dirty-path recording so that post-load
  mutations (e.g. config.request.sampling.seed = 7) are correctly reflected
  when translating to the legacy SamplingParam.

• Typed config builders: New inference_config.py provides
  build_generate_run_config() and build_serve_config() with validation
  (prompt source checks, num_gpus > 0, dotted-override prefix enforcement).

• Shell scripts → YAML configs: 12 per-model shell scripts replaced with
  15 declarative YAML configs plus a single run.sh launcher.

• FlexibleArgumentParser: Extended with parse_known_args override and
  deferred config loading so dotted overrides pass through to the config
  builders instead of being consumed by argparse.

• Dashed-key normalization: --request.output.output-path is automatically
  normalized to request.output.output_path in the override parser.

Changes

Test Plan

# Commands you ran

Test Results

Test output

# Paste output here

Checklist

• I ran pre-commit run --all-files and fixed all issues
• I added or updated tests for my changes
• I updated documentation if needed
• I considered GPU memory impact of my changes

For model/pipeline changes, also check:

• I verified SSIM regression tests pass
• I updated the support matrix if adding a new model

[mergify]

⚠️ PR title format required

Your PR title must start with a type tag in brackets. Examples:

• [feat] Add new model support
• [bugfix] Fix VAE tiling corruption
• [refactor] Restructure training pipeline
• [perf] Optimize attention kernel
• [ci] Update test infrastructure
• [docs] Add inference guide
• [misc] Clean up configs
• [new-model] Port Flux2 to FastVideo

Valid tags: feat, feature, bugfix, fix, refactor, perf, ci, doc, docs, misc, chore, kernel, new-model

Please update your PR title and the merge protection check will pass automatically.

[mergify]

Merge Protections

Your pull request matches the following merge protections and will not be merged until they are valid.

🔴 PR merge requirements

Waiting for:

• #approved-reviews-by>=1
• check-success=fastcheck-passed
• check-success=full-suite-passed

This rule is failing.

• #approved-reviews-by>=1
• check-success=fastcheck-passed
• check-success=full-suite-passed
• check-success~=pre-commit
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model)\]

[gemini-code-assist]

Code Review

This pull request refactors the FastVideo CLI to be "config-first," moving from a flat argument structure to a nested YAML/JSON configuration with dotted-path overrides. This change simplifies CLI usage by consolidating parameters into structured config files and introduces a new request_metadata module to track explicitly provided fields. The old CLI scripts have been replaced with new YAML config files and a generic run.sh wrapper. Documentation has been updated to reflect these changes. Review comments highlight a potential architectural concern regarding the global side effects of __setattr__ patching for request tracking and suggest improving error messages for users transitioning from the old flat config format.

[gemini-code-assist]

Patching __setattr__ on the class level has global side effects for all instances of these dataclasses within the process. While this is necessary for the current tracking implementation, it should be noted that any part of the codebase using these types will now incur the overhead of the tracking logic, even if they are not part of a tracked request. Consider if a more localized approach (e.g., a wrapper or a specific subclass) could achieve the same goal without global monkey-patching.

[gemini-code-assist]

The check for the generator key is used to distinguish between the new nested schema and the legacy flat schema. It might be beneficial to provide a more descriptive error message or a migration hint if a flat config is detected, as users transitioning from older versions might be confused by the requirement for a top-level mapping.

[Eigensystem]

/merge