Purpose

Enable vision encoder CUDA graph capture for ModelRunnerV2 encoder-only instances. #49852 added encoder graph support to the ordinary V2 GPUModelRunner, but MMEncoderModelRunner overrides capture_model with an immediate return, so E-only instances do not capture those graphs.

Call the existing EncoderRunner capture path from the E-only runner, with GC handling, capture-memory accounting, and workspace locking. Reuse existing budget selection, replay, eager fallback, output copies, and EC transfer. The feature remains opt-in, independent of decoder graph mode, and disabled by enforce_eager.

Add E-only configuration/lifecycle/model tests, Mooncake graph-hit assertions, and CI configuration. The attention-padding zero-initialization workaround, its capture context, and its dedicated padding regression test have been removed. The production change is confined to MMEncoderModelRunner; this PR does not change attention kernels or default backend selection.

Known FA backend issue and backend validation

The separate NaN issue is tracked in #57136. With the workaround disabled, H20 / Qwen2.5-VL-3B / BF16 / head_dim=80 / TP=1 testing of nine encoder graph budgets sharing one pool produced:

Backend/version selection was inspected on the actual attention layers; there was no silent fallback. Each backend ran three small -> large -> small sequences, with nine graph hits and zero misses. Tail-injection tests covered capacity/active lengths 256/128 and 2048/1196, using fixed atol=rtol=0.016 against unpadded FP32 SDPA. FLASHINFER used its required offset metadata and interleaved QKV layout. All backends leave some output padding unwritten, but only FA3/FA4 propagated V-tail NaNs in these tests. The complete full-model propagation chain and interface-contract responsibility remain under investigation in #57136.

FLASHINFER and Triton did not reproduce this issue in the tested configuration; this is not a universal guarantee across models and shapes. FA3/FA4 remain affected in that configuration, and this PR does not fix or close #57136. All E-only model test cases (Qwen2.5-VL, Qwen3-VL, and Qwen3.5) select FLASHINFER. Mooncake CI selects TRITON_ATTN. Users can set --mm-encoder-attn-backend TRITON_ATTN or FLASHINFER when using encoder graphs on supported configurations; the integration script exposes E_MM_ENCODER_ATTN_BACKEND.

Environment: NVIDIA H20-3e, driver 580.105.08, Python 3.12, Torch 2.13.0+cu132; FLASHINFER 0.6.17, cuDNN package 9.20.0.48. Model revision: 66285546d2b821cf421d4f5eb2576359d3770cd3. Backend comparisons used 53a766f with the workaround disabled before worker initialization; the final change removes that workaround entirely. Existing precompiled extensions were reused. Multi-GPU, V1, RDMA, LoRA, and broad model-quality benchmarks were not evaluated in this follow-up. No performance improvement is claimed.

Test Plan

export PATH="$PWD/.venv/bin:$PATH"
export VLLM_USE_V2_MODEL_RUNNER=1
export VLLM_WORKER_MULTIPROC_METHOD=spawn
export OMP_NUM_THREADS=1

.venv/bin/python -m pytest \
  tests/kernels/attention/test_mha_attn.py \
  tests/v1/worker/test_encoder_runner.py \
  tests/v1/cudagraph/test_encoder_cudagraph.py \
  tests/v1/ec_connector/unit/test_worker_ec_connector.py -q

.venv/bin/python -m pytest \
  tests/models/multimodal/generation/test_vit_cudagraph.py \
  -k 'eonly and (qwen2_5_vl or qwen3_vl)' -v -s

Test Result

Code validation revision 05c72d0: pre-commit (including shellcheck and documentation checks) and manual mypy-3.12 passed. H20 rerun on this revision: combined attention/encoder/EC suites 132 passed, 14 skipped (27.85s; ROCm/AITER-only skips). Qwen2.5-VL and Qwen3-VL E-only tests using TRITON_ATTN: 2 passed (123.19s), covering images, video, budget fallback, and retained-output integrity. The backend comparisons above were completed without the workaround. The updated Mooncake CI backend configuration has not yet been run end-to-end on this revision; earlier EPD results with the workaround are not presented as validation of this final version.

The subsequent documentation-only commit 31993e3 removes all changes to docs/features/disagg_encoder.md; its pre-commit checks passed. Production code and tests are unchanged from the validated revision above.

Duplicate-work checks and human review

Open PR searches for E-only/encoder-only graphs, encoder cudagraph, and padding NaNs found no competing E-only capture implementation. #49852 is the merged infrastructure reused here. #55929 addresses input normalization; #52230 addresses V1 non-first PP ranks. The independent FA NaN issue is tracked in #57136 rather than fixed in this PR.

AI assistance (OpenAI Codex) was used for implementation, testing, and review. This PR remains a draft pending the human submitter's required review and test confirmation:

• Human submitter reviewed every changed line and understands the change end-to-end.
• Human submitter ran the relevant tests.

E-only model test backend follow-up

All three parameterized E-only model test cases now use FLASHINFER. The H20 FLASHINFER rerun described below covers Qwen2.5-VL; Qwen3-VL and Qwen3.5 await validation with this backend. Buildkite 89352 failed that case with TRITON_ATTN on H200 MIG (finite outputs, but maximum absolute difference 3.40625). On H20 at the same cc04b48 source revision, the isolated test passed with TRITON_ATTN and with a test-only FLASHINFER override: 1 passed in 88.82s and 57.55s respectively, both with maximum absolute error 0, seven graph hits and one fallback. These durations include startup and are not a backend performance comparison. Pre-commit passed for the committed backend change; the new CI result remains pending.