[gty111]

Overview

Add native support for Cloudflare/clef-flash through /v1/systemone, reusing the Qwen3.5 backbone and the joint schema head introduced for Clef-Omni in #61064.

Claims

• Automatically detect and load the separate joint schema head, scoring choice, noul, and score questions in one pooling request.
• Support text/JSON, images, and videos sampled at 2 fps. Match the release's fixed prompt, media separators, schema spans, and 16K truncation.
• Share the head/pooler between Clef-Flash and Clef-Omni. Flash rejects audio inputs and ignores video soundtracks.

Validation

Compared the full checkpoint at revision 8b2e5fd17c09fd49bc2880b805d5323ec7ff4ff3 against the release's joint_schema_model.py in BF16. Ten model inputs covered invoice decisions, Unicode/empty instructions, a state truncated to 16K, one/two images, video, and image+video. Prompt token IDs and question/option spans matched the reference exactly in all ten cases.

Each mode ran the ten inputs once sequentially and twice at concurrency 4:

Mode	Requests	Highest-probability option matches	Max absolute probability difference
Eager	30	57/57	0.001508
Default compile	30	57/57	0.000863

The four invoice threshold labels were correct in every repetition (12/12 per mode). These are synthetic model correctness checks. Probability normalization and input-token counts also matched; output-token counts were zero.

The reference evaluation used encode_record, collate_records, and load_release_model(..., attn_implementation="sdpa"), comparing each question's softmax probabilities against /v1/systemone responses. Serve the model with:

vllm serve Cloudflare/clef-flash \
  --served-model-name clef-flash --dtype bfloat16 \
  --max-model-len 16384 --max-num-batched-tokens 512 --max-num-seqs 8
# Add --enforce-eager for the eager comparison.

Details

The checkpoint advertises Qwen3_5ForConditionalGeneration and supplies joint_head_config.json plus joint_head.safetensors. Architecture conversion detects that head, and the Flash wrapper uses the output embedding table for lexical option features, matching the reference implementation.

The existing Clef head/pooler is moved unchanged into clef_joint_schema.py. Flash shares the joint schema strategy's typed answers and span/truncation logic, with its own media rendering for adjacent placeholders and a single trailing newline.

Duplicate checks found no open Clef-Flash support PR. #61064 supports Clef-Omni; the other open Clef-related PRs concern API/frontend behavior.

Model comparisons cover BF16 with TP=1. Multi-GPU configurations, quantization, and the full Clef checkpoint were not evaluated.

AI assistance (Codex) was used to implement and validate this model support.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

[mergify]

Documentation preview: https://vllm--61308.org.readthedocs.build/en/61308/

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[Isotr0py]

/ci run --allow-stale

[github-actions]

✅ Triggered Buildkite CI #94363 for commit 2591dfdc36be.

⚠️ This PR is 5 commits behind upstream main. Running CI at your own risk because --allow-stale was requested; outdated CI configuration may cause failures. Before merging, merge or rebase onto the latest main, then rerun /ci run on the latest PR commit.

[Isotr0py]

/ci run --allow-stale

[github-actions]

✅ Triggered Buildkite CI #94373 for commit feba86ccfe27.

⚠️ This PR is 6 commits behind upstream main. Running CI at your own risk because --allow-stale was requested; outdated CI configuration may cause failures. Before merging, merge or rebase onto the latest main, then rerun /ci run on the latest PR commit.

[vllm-agent]

CI selector (shadow): 27 test steps (60 jobs) instead of 68 (100 jobs)

Shadow mode: this changes nothing about what CI runs. It shows what the evidence-based selector would pick for this PR, next to today's rules. How it works.

Feedback welcome: reply here if it would skip a step this change needs, or runs something unrelated.

steps (jobs)	Today's rules	Selector	Would skip	Would add
NVIDIA, CPU and others	68 (100)	27 (60)	46 (53)	5 (13)
AMD mirrors	62 (89)	18 (41)	47 (54)	3 (6)

Selector would run (27)

• arm-cpu-test ×3
• basic-models-test-other-cpu
• basic-models-tests-extra-initialization ×14
• basic-models-tests-initialization
• basic-models-tests-other
• cpu-language-generation-and-pooling-model-tests ×3
• cpu-multi-modal-model-tests-n ×4
• cpu-multimodal-config
• cpu-params-env-tokenizers-parser
• cpu-reasoning-renderers
• distributed-model-tests-2-gpus ×3
• engine
• entrypoints-integration-api-server-generate
• entrypoints-integration-speech_to_text
• language-models-tests-extra-standard ×2
• language-models-tests-granite-l4-compatibility
• language-models-tests-hybrid ×2
• language-models-tests-standard
• multi-modal-models-standard-1-qwen2
• multi-modal-models-standard-4-other-whisper
• multi-modal-processor ×4
• multi-modal-processor-cpu ×4
• pipeline-context-parallelism-4-gpus
• quantization ×4
• rayexecutorv2-4-gpus
• v1-metrics-lmeval
• v1-others-cpu

Would skip (today's rules run them) (46)

• amd-lm-eval-small-models-harness
• amd-qwen3-next-mtp-async-eplb-accuracy
• ascend-npu-test
• async-engine-inputs-utils-worker
• basic-correctness ×2
• basic-correctness-cpu-offload
• basic-correctness-cumem
• basic-correctness-prefetch-offload
• basic-correctness-sleep-mode
• benchmarks-cli-test
• cpu-tool-parsers
• entrypoints-integration-api-server ×4
• entrypoints-integration-api-server-openai-chat_completion
• entrypoints-integration-api-server-openai-completion
• entrypoints-integration-llm
• entrypoints-integration-multimodal
• entrypoints-integration-pooling
• entrypoints-integration-responses-api
• entrypoints-unit-tests
• examples
• fusion-e2e-quick-h100
• fusion-e2e-tp2-b200
• fusion-e2e-tp2-quick-h100
• jit-monitor-no-runtime-jit
• kernels-fla-ops-test-b200
• kernels-mhc-test-b200
• kernels-root-misc-test-b200
• metrics-tracing-2-gpus
• model-executor
• multi-modal-models-standard-2-qwen3-gemma
• multi-modal-models-standard-3-llava-qwen2-vl
• pytorch-compilation-dynamic-shapes
• pytorch-compilation-passes-unit-tests
• pytorch-compilation-unit-tests
• pytorch-compilation-unit-tests-h100
• pytorch-fullgraph-cudagraph-l4-compatibility
• pytorch-fullgraph-test
• regression
• rl-entrypoints-tests
• v1-core
• v1-executor-worker
• v1-kv-connectors ×4
• v1-kv-offload
• v1-logits-oracle
• v1-sample
• v1-spec-decode

Would add (today's rules do not run them) (5)

• arm-cpu-test ×3 (code map)
• cpu-multi-modal-model-tests-n ×4 (code map)
• engine (code map)
• quantization ×4 (code map)
• rayexecutorv2-4-gpus (code map)

AMD mirrors: would skip (47)

• async-engine-inputs-utils-worker
• basic-correctness ×2
• basic-correctness-cpu-offload
• basic-correctness-cumem
• basic-correctness-prefetch-offload
• basic-correctness-sleep-mode
• benchmarks-cli-test
• entrypoints-integration-api-server ×4
• entrypoints-integration-api-server-openai-chat_completion
• entrypoints-integration-api-server-openai-completion
• entrypoints-integration-llm
• entrypoints-integration-multimodal
• entrypoints-integration-pooling
• entrypoints-integration-responses-api
• entrypoints-unit-tests
• examples
• extract-hidden-states-integration-2-gpus
• fusion-e2e-quick-h100
• fusion-e2e-tp2-ar-rms-config-sweep-h100
• fusion-e2e-tp2-b200
• fusion-e2e-tp2-quick-h100
• jit-monitor-no-runtime-jit
• kernels-fla-ops-test-b200
• kernels-mhc-test-b200
• kernels-root-misc-test-b200
• lm-eval-small-models
• metrics-tracing-2-gpus
• model-executor
• mrcr-eval-small-models
• multi-modal-models-standard-2-qwen3-gemma
• multi-modal-models-standard-3-llava-qwen2-vl
• platform-tests
• pytorch-compilation-dynamic-shapes
• pytorch-compilation-passes-unit-tests
• pytorch-compilation-unit-tests
• pytorch-compilation-unit-tests-h100
• pytorch-fullgraph-cudagraph-l4-compatibility
• pytorch-fullgraph-test
• regression
• rl-entrypoints-tests
• v1-core
• v1-executor-worker
• v1-kv-connectors ×4
• v1-kv-offload
• v1-logits-oracle
• v1-sample
• v1-spec-decode

AMD mirrors: would add (3)

• engine (code map)
• quantization ×4 (code map)
• rayexecutorv2-4-gpus (code map)

_{11 changed files · base c6566bb9f6 · head feba86ccfe · Python record: build 93039 at 1e5d0ea888 · kernel record: table fe341a5 (build 94315), map fe341a5 · not counted: 11 build steps, 5 A100 steps the generator no longer emits, 12 optional steps the selector would also run}