[AMD] Add algo-config option in Quark

[poganesh]

• Adds an algo_configs parameter to the QuarkQuantization pass (torch path), letting a recipe pass a custom quantization algorithm config (e.g. AWQ scaling_layers / model_decoder_layers) through to LLMTemplate.get_config(algo_configs=...).

• This is needed for architectures that Quark 0.12.post1 does not ship a built-in algorithm config for, e.g. qwen3, which is absent from Quark's AWQ_MAP and otherwise raises NotImplementedError: No built-in awq configuration is available for the 'qwen3' architecture. With this change, the recipe can supply the config (including scaling_layers: [] for auto-detection), unblocking AWQ quantization for such models.

• The parameter is optional and defaults to None; existing recipes and architectures with built-in configs are unaffected.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Pull request overview

This PR extends Olive’s Quark Torch quantization pass to accept user-supplied algorithm configuration (“algo_configs”) from a recipe and forward it into Quark’s LLMTemplate.get_config(...), enabling AWQ quantization for model architectures that Quark doesn’t yet ship built-in configs for (e.g., qwen3).

Changes:

• Add an optional algo_configs pass parameter to QuarkQuantization (torch path) to accept per-algorithm config dictionaries from recipes.
• In the torch quantization runner, translate recipe-provided AWQ config dicts into Quark’s AWQConfig objects and pass them into template.get_config(algo_configs=...).
• Remove the explicit processor=None argument from the calibration dataloader call.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
olive/passes/quark_quantizer/torch/quark_torch_quantization.py	Builds algo_configs (currently AWQ-only) from recipe JSON into AWQConfig and forwards it to LLMTemplate.get_config; removes processor=None from the calibration dataloader call.
olive/passes/quark_quantizer/quark_quantization.py	Adds the optional algo_configs PassConfigParam for the torch pathway so recipes can supply algorithm configs.

[poganesh]

@thpereir, @thiagocrepaldi - Please help review the quark changes.

[thiagocrepaldi]

Code itself looks OK. Are you planning to add tests? any extra documentation planned, with maybe a real usage example?

[poganesh]

Code itself looks OK. Are you planning to add tests? any extra documentation planned, with maybe a real usage example?

@thiagocrepaldi

• The new Quark 0.12 torch path had no existing tests (only the ONNX path in test_quark_onnx_quantization.py), so I added a new test_quark_torch_quantization.py covering the algo_configs handling introduced in the PR.

• The primary use case is olive-recipes (e.g. qwen3, which Quark has no built-in AWQ config for). detailed config semantics are deferred to the Quark docs, in the olive-recipes README.