[JiaxinD]

Background

Qwen needs semantic TextContinuation support in #1498 while preserving streaming and generation capacity. Its embedding comparator also hardcodes the 0.6B output dimension. The integration retains upstream's native conversation history support.

Exit Criteria

Expose typed text/token-ID continuation, preserve streaming and conversation behavior, validate published embedding profiles and reject invalid inputs. Larger embedding execution remains outside this PR's qualification matrix.

Implementation

• Bind continuation with family-owned Config and an exclusive generation lease. TP models do not advertise streaming.
• Preserve cancellation, ChatML, vocabulary checks, zero-token continuation and owned C11/C++17 results.
• Retain upstream conversation bindings and history rendering. Share Qwen configuration validation, exclude single-turn template switches from conversations and retain their positive token cap. Continuation and streaming retain zero-token support.
• Accept official 4B/8B embedding topology and dimensions 2560/4096 in CPU contracts and the reference comparator.
• Keep unqualified 4B/8B E2E manifests removed per maintainer feedback. Retain pinned CPU fixtures and rejection tests.
• Preserve upstream benchmark-profile removals, S1 Mini ownership and core provenance validation. Reject prebuilt bundles advertising legacy text_generation; pass system prompts unchanged to text SDK consumers while token-ID inputs bypass preprocessing.

Changes remain under families/qwen. Rebuild legacy generation bundles for the semantic primary task. Shared API/ABI definitions and bundle format are unchanged by this PR. The upstream merge preserves published history.

Change categories

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

Validation

Commands and Results

Executed check	Result
pytest families/qwen/tests core/builder/tests/test_model_support.py -q	118 passed, 16 skipped
Architecture, impact, metadata and public source hygiene suites	99 passed
Fresh public SDK/API/loader/bundle sources and actual family CMake targets, selected CTest	Seven native CPU contracts passed; C11/C++17 consumers compiled
Conversation integration regression on mechanical merge, then repaired parser	Task pipeline failed before repair; all seven contracts passed after repair
Model inventory, impact, legal headers, clang-format and git diff --check	Passed

Conversation tests verify binding/schema, exact once-only history rendering and pre-engine rejection boundaries. CPU engine/state doubles do not establish generated output correctness.

Hardware, Environment, and Revisions

Head 31f1b53d30a5afc315f0d0ceda2de8a9b78a071b integrates NVIDIA/main 8194f406fb0f3b89d81d719566852e71f1a34f16 into prior head 1404788d78fddc82adebe58dc9090fb6e4aa6185. Windows Python3.13 CPU, WSL Ubuntu GCC13.3, C11/C++17 and CUDA12.6 headers/runtime; CUDA_VISIBLE_DEVICES empty. Pinned fixtures: Qwen/Qwen3-Embedding-4B@5cf2132abc99cad020ac570b19d031efec650f2b and Qwen/Qwen3-Embedding-8B@1d8ad4ca9b3dd8059ad90a75d4983776a23d44af.

Not Run / Remaining Gaps

Current-head TensorRT/GPU inference, protected CI, larger embeddings, FP8/TP4 and performance are unvalidated. Radix authentication expired. Earlier head 1404788d passed public CPU/Stable/Dev; those results do not qualify this merge. Fifteen Python skips are unselected GPU E2Es; one Windows native build probe is skipped, with native CMake validation separately completed in WSL.

Contributor Self-Review

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

Reviewed provenance/task boundaries, conversation integration, configuration, streaming lifecycle and consumers.

Notes For Future Readers

S1 Mini criteria and checkpoint pins are preserved. Removed manifests can return after qualification. Ready conversion remains subject to the three-PR author capacity policy; protected execution remains required.

Risk level

• Low
• Medium
• High

Primary-task and streaming integration need protected execution; CPU tests do not establish GPU parity or performance equivalence.

[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

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

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

[JiaxinD]

@yifeif-nv Could you rerun Dev Community CI for #1558 (eba91686) and #1583 (27214f4c) after #1595? Their previous attempts stopped during GPU provisioning before model tests.
The manual qualification of #1413's current head passed; these two unchanged heads still need that validation.

[yifeif-nv]

@yifeif-nv Could you rerun Dev Community CI for #1558 (eba91686) and #1583 (27214f4c) after #1595? Their previous attempts stopped during GPU provisioning before model tests. The manual qualification of #1413's current head passed; these two unchanged heads still need that validation.

Hey, I've triggered the internal CI for you. The dev CI is the community GPU CI that is still in development. It is not gating the merging. So once the internal CI is passed, I will merge the PR. Thanks.

[JiaxinD]

@yifeif-nv Thanks for triggering Internal CI. It failed on eba91686 with details withheld; could you share the failing test name and a sanitized error excerpt so I can fix it?

[yifeif-nv]

Hey @JiaxinD I think we'll need to remove
families/qwen/tests/manifests/qwen3-embedding-4b.json
families/qwen/tests/manifests/qwen3-embedding-8b.json

for this PR. Then you can locally verify it via

python -m pytest \
  families/qwen/tests/test_embedding_profiles.py \
  families/qwen/tests/test_embedding_e2e_dimensions.py -q

python -m tools.model_ci validate
python tools/test_impact.py --validate

This is due to the private CI require any ckpt declared to be tested in L0 and both ckpt are current declared as premerge = false. We can add in the test in a later time

[JiaxinD]

Removed the 4B/8B manifests and integrated upstream at cb6291bb. CPU profile coverage remains; 170 Python tests and seven native contracts passed locally.
@yifeif-nv Could you rerun Internal CI for this head once Community CPU completes?

[JiaxinD]

@yifeif-nv Community CPU, Stable and Dev pass on bc52bbfd. Could you trigger Internal CI for this current head?

[JiaxinD]

@yifeif-nv Community CPU, Stable and Dev passed for 1404788. Please trigger protected CI for this head.

[JiaxinD]

Integrated upstream conversation history in 31f1b53, retaining the family-owned Config and streaming lease. Seven native CPU contracts and 118 Python tests passed; GPU/protected qualification remains pending.

[JiaxinD]

@yifeif-nv Community CPU passed for 31f1b53. Please trigger protected CI for this head.