[OpenVINO] Support Falcon-H1-0.5B-Instruct with task text-generation

[popovaan]

Description

Enabled Falcon-H1 (hybrid Mamba+attention, FalconH1ForCausalLM) for OpenVINO GenAI. Root cause: explicit PagedAttention request failed with 'Model references undeclared parameters: beam_idx' because SDPAToPagedAttention cannot rewrite the recurrent Mamba conv/ssm state. Fix: in all three LLMPipeline constructors, catch the specific PA model-construction failure (is_paged_attention_model_construction_failure) and fall back to the stateful backend, stripping scheduler_config. model-checker EXIT=0: GenAI llm_bench 1st=2.34ms 2nd=1.67ms/tok tput=597 tok/s; optimum similarity 1.0 and GenAI similarity 1.0 (threshold 0.95). Verified no regression: pytest test_llm_pipeline.py -k linear -> 32 passed, 6 skipped (qwen3-next STATEFUL+PAGED_ATTENTION). Docs updated with FalconH1ForCausalLM entry (npm build passed). Test wiring added to LINEAR_ATTENTION_MODELS_LIST but left commented: no non-degenerate tiny-random FalconH1 fixture is published; public community fixtures (tiny-random/falcon-h1, yujiepan/falcon-h1-tiny-random) segfault due to degenerate configs.

Reproduce generation

import openvino_genai

pipe = openvino_genai.LLMPipeline("output_dir", "CPU")
print(pipe.generate("What is the capital of France?", max_new_tokens=10))

Validation

• Validation model: Tiny model (/home/openvino_bot/.local/share/openvino-model-agent/requests/issue-39/repository/workspace/tiny_falcon_h1)
• WWB Optimum similarity: 1.0
• WWB GenAI similarity: 1.0
• First-token latency: 2.34
• Throughput: 597.08

Final real-model validation

• Estimated FP16 size: Not available
• Available RAM at validation: 122.39 GiB
• 70% RAM limit: 85.67 GiB
• Status: Skipped
• Reason: Could not estimate FP16 model size: No module named 'huggingface_hub'

Related model-support PRs

• Optimum Intel: [OpenVINO] Support Falcon-H1-0.5B-Instruct with task text-generation huggingface/optimum-intel#1908

Checklist

• This PR follows the GenAI contributing guidelines.
• Tests have been updated or added to cover the new code.
• This PR fully addresses the model-support work.
• Documentation was updated, or its absence is explained in the validation handoff.

[Copilot]

Pull request overview

This PR enables Falcon-H1 (hybrid Mamba+attention) usage in OpenVINO GenAI text-generation by making LLMPipeline robust to PagedAttention/ContinuousBatching construction failures that occur when SDPAToPagedAttention cannot rewrite recurrent Mamba state. It also updates supported-model documentation and leaves test-model wiring notes for future regression coverage once a stable tiny-random fixture exists.

Changes:

• Added utils::is_paged_attention_model_construction_failure() to detect the specific “undeclared/unregistered beam_idx” incompatibility during PagedAttention model construction.
• Updated all three LLMPipeline constructors to catch that specific failure and fall back to the stateful backend (dropping scheduler_config).
• Added FalconH1ForCausalLM to the supported LLM architectures list in site docs; updated test model list notes.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/python_tests/data/models.py	Adds an explanatory (commented) placeholder for a future Falcon-H1 tiny-random regression fixture.
src/cpp/src/utils.hpp	Declares a new helper for detecting PA model-construction incompatibility from exceptions.
src/cpp/src/utils.cpp	Implements the PA incompatibility detector using message substring matching.
src/cpp/src/llm/pipeline.cpp	Adds explicit-PA constructor fallback logic to stateful backend when the model cannot be converted to PagedAttention.
site/docs/supported-models/_components/llm-models-table/models.ts	Documents Falcon-H1 architectures/models as supported.

Suppressed comments (2)

src/cpp/src/llm/pipeline.cpp:319

• The comment says the constructor will re-throw when Continuous Batching (PagedAttention) is invoked explicitly, but the new logic also falls back to the stateful backend for known model-construction incompatibilities. Updating the comment will keep it consistent with the actual behavior.

    } else if (utils::explicitly_requires_paged_attention(user_properties)) {
        // If CB is invoked explicitly, create CB adapter as is and re-throw in case if internal issues
        auto [device_properties, scheduler_config] = utils::extract_scheduler_config(properties, utils::get_latency_oriented_scheduler_config());

src/cpp/src/llm/pipeline.cpp:383

• The comment says the constructor will re-throw when Continuous Batching (PagedAttention) is invoked explicitly, but the new logic also falls back to the stateful backend for known model-construction incompatibilities. Updating the comment will keep it consistent with the actual behavior.

    } else if (utils::explicitly_requires_paged_attention(user_properties)) {
        // If CB is invoked explicitly, create CB adapter as is and re-throw in case if internal issues
        auto [device_properties, scheduler_config] = utils::extract_scheduler_config(properties, utils::get_latency_oriented_scheduler_config());