refactor(qwen): migrate generation to TextContinuation Task

[JiaxinD]

Background

Qwen advertises the legacy text_generation mode and cannot be accessed through the semantic TextContinuation SDK. This implements #1498.

Exit Criteria

• Qwen, Qwen2, Qwen3 and QwQ bundles expose text_continuation, with UTF-8 and token-ID inputs.
• Retain existing family graphs, tokenizer paths, sampling controls, native-KV and tensor-parallel implementation.
• Validate typed Config, linked C/C++ consumers and a real checkpoint before requesting review.

Implementation

• Bind the Qwen pipeline to ITextContinuation with family-owned Config parsing and vocabulary validation.
• Update family support, manifests, qualification candidates and E2E assertions to the new primary task.
• Add Config contract tests and C11/C++17 consumers for both input representations, owned results after model release and zero-token output.
• Rebuild Qwen bundles for the changed task. The public ABI and bundle format remain unchanged. Changes are confined to families/qwen.

Change categories

• Model or runtime behavior
• Public API
• ABI
• Bundle or artifact format
• Dependencies
• Documentation only
• CI or developer tooling

Validation

Commands and Results

Command	Result
cmake --build build-sm90 --parallel 8 --target trtmc trtmc_backend_trt trtmc_model_qwen test_qwen_task_config test_qwen_sampler test_qwen_tensor_names test_qwen_native_kv_cache	Passed, includes linked C/C++ SDK consumers
ctest --test-dir build-sm90 --output-on-failure -R '^(qwen_task_config|test_qwen_sampler|test_qwen_tensor_names|test_qwen_native_kv_cache)$'	4 passed, 0 skipped
python -m pytest families/qwen/tests/test_support.py families/qwen/tests/test_builder_policy.py families/qwen/tests/test_runtime_receipt.py families/qwen/tests/test_checkpoint_shape_contract.py -q	13 passed
python -m pytest 'families/qwen/tests/test_e2e.py::test_e2e[qwen3-0.6b-native-l0]' 'families/qwen/tests/test_e2e.py::test_e2e[qwen3-0.6b-fp16]' 'families/qwen/tests/test_e2e.py::test_e2e[qwen3-0.6b-fp16-native-kv-two-chunk-parity]' 'families/qwen/tests/test_e2e.py::test_e2e[qwen3-0.6b-topp]' -vv -s --basetemp=/work/e2e-temp --junitxml=/work/evidence/e2e.xml	4 passed, 0 skipped; HF-reference BF16/FP16, two-chunk receipt and sampling determinism; reference cases also pass C/C++ text/token/zero-token consumers

Earlier registry/impact checks, Ruff and SDK syntax checks passed; current worktree and git diff --check are clean. Current-head Stable/Dev Community CI pass; this is separate from the local GPU evidence.

Hardware, Environment, and Revisions

Head 9aa55f6ee9ccc7fde6715dd37e5c86a466431552, base 79f229243d517bf51db69f50b21c25d2d7e4b6fe. Qwen/Qwen3-0.6B revision c1899de289a04d12100db370d81485cdf75e47ca, cached offline. One H200, SM90, driver 595.71.05, Ubuntu 24.04, CUDA 13.3, TensorRT 11.1.0.106, PyTorch 2.12.0+cu130, Transformers 5.2.0. Image nvcr.io/nvidia/tensorrt:26.07-py3@sha256:b82db1abc23750ab0069abc99bbe4ea29138dbdc23ea39861199e2346638b48a.

E2E used TRTMC_E2E=1, TRTMC_BINARY=/work/source/build-sm90/trtmc, TRTMC_RUNTIME_ROOT=/work/source/build-sm90, TRTMC_NATIVE_BUILD_DIR=/work/source/build-sm90, CUDA_VISIBLE_DEVICES=0 and HF_HUB_OFFLINE=1. JUnit, checkpoint revision, binary hashes and family evidence were retained.

Not Run / Remaining Gaps

FP8, TP4, Qwen3-4B and the long-context regression were not run in this single-device validation. No throughput, latency or memory-equivalence claim. Protected Internal CI has not reported a premerge result for this PR.

Contributor Self-Review

• I have completed a self-review of this change.

Notes For Future Readers

Token-ID input bypasses chat templating and re-tokenization. Review Config, pipeline binding and consumers in that order. Rebuild bundles before use. Scoped native validation is complete; repository review capacity and protected CI remain separate requirements.

Risk level

• Low
• Medium
• High

The primary task changes SDK dispatch and requires rebuilding existing Qwen bundles.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

_{Comment @coderabbitai help to get the list of available commands.}