[Fix] dLLM: Preserve prompt tails and literal mask tokens across prefix reuse - #39592

Incremental review: Files changed against main.

dLLM prefix lookup can consume a partial prompt block and its synthetic masks when a cached prompt contains the same literal mask IDs. Literal mask IDs in a real prompt can also be denoised and emitted as generated tokens, causing KV accounting errors and a serving crash.

Bound cache lookup by complete real-token blocks and classify the request after matching its cache prefix. Carry an explicit committed-position mask into LowConfidence, JointThreshold and JointThresholdInDel so literal prompt IDs and previously emitted outputs stay immutable after retraction. Repeated denoising no longer subtracts the current block from the reported cached-token count. The change preserves autoregressive lookup and logprob boundaries.

The GPU/HTTP results below are historical qualification of the previous revision; they were not rerun for the October 10 refresh.

HTTP tests use JetLM/SDAR-1.7B-Chat-b32, BF16, TP=1, Triton, CUDA graphs disabled, block size 32, threshold 0.95, greedy sampling, 32 new tokens, and ignore_eos=true. Hardware: RTX PRO 4000 Blackwell 24 GB; Python 3.12, PyTorch 2.13.0+cu130, sglang-kernel 0.4.7. Each configuration runs 19 actual /generate requests. An observation-only wrapper records cache matches after Req.init_next_round_input.

The collision test first caches a 33-token prompt followed by 31 literal mask IDs, then submits the original 33-token prompt. Its output matches a clean-cache request and only 32 tokens are reused. Unit tests additionally exercise both JointThreshold implementations and both scheduling modes. These results do not qualify CUDA graph execution or multi-block prefill.

PYTHONPATH=python python -m pytest -q \
  test/registered/unit/mem_cache/test_dllm_fdfo_kv_reuse.py \
  test/registered/unit/dllm/algorithm/test_joint_threshold_indel.py \
  test/registered/unit/dllm/test_gemma4_renoise.py

Contributes to the prefix-cache / prompt-tail item in #39499. The post-cache phase ordering overlaps the prerequisite in #31586; this PR does not add its multi-block scheduler.

CI States

Latest PR Test (Base): ✅ Run #39056280279
Latest PR Test (Extra): ❌ Run #39056280293
Latest PR Test (AMD ROCm 10): ❌ Run #39056280358

October 10 review follow-up

Merged main c55790d38ba0c41170cec8993e7a85d5c687149c and addressed the remaining review items:

• Batch CPU-known committed lengths into the same pinned, nonblocking transfer as positions. LowConfidence uses torch.where to exclude committed positions. Algorithms read the declared ForwardBatch field directly, and test doubles provide it explicitly.
• Retraction phase selection and position masks include both prompt tokens and committed outputs, independently of KV residency.
• JointThresholdInDel uses the explicit immutable mask for state initialization and excludes those positions from mask counts and transfers, while retaining its contiguous-prefix fallback when metadata is absent.
• Existing registered CPU test files now contain regressions for a literal-mask cache collision, immutable literal prompts in synchronous/FDFO decoding, retracted outputs, repeated-pass cached-token accounting and InDel behavior.

All scoped pre-commit checks passed, including CI registry validation. CPU source execution passed 25 exact repository algorithm tests, committed-token phase/prefix controls over 72 prompt/output layouts, three actual position-branch controls and the repeated-pass cache-accounting guard. Negative controls reproduced the literal-prompt InDel failure; these tests also caught and corrected its mask-count/transfer path. This source-level run omits package initialization and native dependencies. Full registered-suite and HTTP reruns require the complete SGLang environment and were not executed in this refresh; no GPU performance or memory measurement is claimed.

Review fix commit: 0fbe7e46218ce8b728c72095f1332c3b98e6de19.