Native Qwen3-Reranker CausalLM support in RerankCalculatorOV

[ambeckley]

Summary

• Adds native support for Qwen3-Reranker models (0.6B, 8B, all sizes) using CausalLM architecture, exported with --task text-generation
• Auto-detects Qwen3 via model_type in config.json — no changes needed for existing reranker models
• Applies server-side chat template formatting and CausalLM graph postprocessing (yes/no logit extraction via PrePostProcessor), so clients use the standard /v3/rerank API with no workarounds

Motivation

Qwen3-Reranker models use CausalLM architecture, not cross-encoder text-classification. The current workaround (#3578) requires community-modified seq-cls model. This PR enables all official Qwen3-Reranker model sizes to work natively through OVMS without client modifications. It is still backwards compatibility with tomaarsen/Qwen3-Reranker-*-seq-cls models.

Changes

src/rerank/rerank_servable.hpp

• Added isQwen3, hasPositionIds, hasBeamIdx detection flags
• Override applyPrePostProcessing() to:
  • Parse config.json for model_type: "qwen3"
  • Detect position_ids and beam_idx model inputs
  • Check output dimensionality (3D = CausalLM, 2D = text-classification with warning)
  • Look up yes/no token IDs via tokenizer
  • Build PrePostProcessor graph: Slice last token → Squeeze → Gather yes/no logits → Subtract (yes - no), producing [batch, 1] output compatible with existing sigmoid scoring

src/rerank/rerank_calculator_ov.cc

• Added Qwen3 chat template input formatting in PrepareInputsForRerankModel()
• Compute position_ids from attention mask for CausalLM models
• Zero-fill beam_idx for CausalLM models
• Guard token_type_ids creation with !isQwen3 check (CausalLM uses position_ids, not token_type_ids, as 3rd input)

Model Export

Models must be exported with --task text-generation (not the default text-classification):

optimum-cli export openvino --model Qwen/Qwen3-Reranker-8B --task text-generation --weight-format int8 Qwen3-Reranker-8B-causal-int8-ov

The default text-classification export produces a model with an untrained random classification head that outputs garbage scores.

Test plan

• Tested with Qwen3-Reranker-0.6B (int8) on CPU — correct relevance scores
• Tested with Qwen3-Reranker-8B (int8) on CPU and Intel Arc GPU — correct relevance scores
• Verified existing non-Qwen3 reranker models are unaffected (isQwen3 = false, no code path changes)
• CI/unit tests (to be added if maintainers request)