feature: cross-block no-spec fallback for variable-length spec decoding proposers

[rebel-wonsubkim]

🚀 Summary of Changes

What does this PR do? What feature, fix, or improvement does it bring?

Motivation

Query backfill (#604) keeps every decode step at a single (batch, num_spec+1) shape by prepending past tokens. But variable-length proposers
(ngram/suffix) can need backfill that crosses a KV block boundary: right after entering a new block, a short draft would pull past tokens from the
previous block, which the single-block decode path can't express → silent KV corruption. (Fixed-length proposers never hit this — num_spec+1 <=
block_size keeps their backfill in-block.)

Key idea

• Detect cross-block in the scheduler (desired_slide > tokens_used_in_block) and fall back to no-spec (query_len=1) for that step.
• Cross-DP collective: any rank tripping it → every rank drops to no-spec via the step_no_spec_required OR-reduce, so all ranks keep the same decode
  shape.
• Compile the no-spec graph in warmup. A decode graph's compile key has two axes — input_ids shape and max_pads_across_dp size (num_padded). The warmup
  dummy drives the same runtime path (RBLNSchedulerOutput(step_no_spec_required=True)) so num_padded matches too; setting query_len=1 alone would still
  hot-path recompile at runtime.
• Two runtime asserts guard the invariant: no backfill window may cross a block boundary, and the no-spec scrub must clear all slide + force
  query_len=1.

Verification

Full model (MiniMax-M2.5, DP4/EP4, ngram num_spec=3), isl=512/osl=2048: cross-block no-spec fired 322× organically, with 0 recompile / 0 hot-path / 0
assert, outputs coherent.

▎ Full motivation & key idea are kept in the rbln_scheduler.py module docstring.

host tensor - 검증 완료

RBLN_WEIGHT_FREE=0 VLLM_RBLN_USE_DEVICE_TENSOR=0 VLLM_RBLN_SAMPLER=0 \
vllm serve MiniMaxAI/MiniMax-M2.5 --port 8000 --data-parallel-size 4 --enable-expert-parallel \
--max-model-len 196608 --block-size 1024 --enable-chunked-prefill \
--max-num-batched-tokens 512 --max-num-seqs 8 --gpu-memory-utilization 0.8 --trust-remote-code \
--speculative-config '{"method": "ngram", "num_speculative_tokens": 3, "prompt_lookup_max": 5, "prompt_lookup_min": 2}'`

device tensor - 검증 중

RBLN_WEIGHT_FREE=1 VLLM_RBLN_USE_DEVICE_TENSOR=1 VLLM_RBLN_SAMPLER=0 \
vllm serve MiniMaxAI/MiniMax-M2.5 --port 8000 --data-parallel-size 4 --enable-expert-parallel \
--max-model-len 196608 --block-size 1024 --enable-chunked-prefill \
--max-num-batched-tokens 512 --max-num-seqs 8 --gpu-memory-utilization 0.8 --trust-remote-code \
--speculative-config '{"method": "ngram", "num_speculative_tokens": 3, "prompt_lookup_max": 5, "prompt_lookup_min": 2}'`

vllm bench serve \
    --model MiniMaxAI/MiniMax-M2.5 --backend vllm \
    --base-url http://127.0.0.1:8000 --endpoint /v1/completions \
    --dataset-name random --random-input-len 512 --random-output-len 2048 \
    --num-prompts 320 --max-concurrency 32 --request-rate 4 \
    --percentile-metrics ttft,tpot,itl,e2el --temperature 0

---

📌 Related Issues / Tickets

• Resolves #
• Related to feature(spec_dec): implement spec decode backfill for fixed length drafts (full_spec only) #604

---

✅ Type of Change

• 🚀 Release (release)
• ✨ Feature (feature)
• 🧠 Model support (model)
• 🧬 Core engine changes (core)
• 🛠 Bug fix (fix)
• ⚙️ Performance improvement (perf)
• 🔁 Refactor or code cleanup (refactor)
• 📄 Documentation (docs)
• ❓ Other (other): please describe

---

🧪 How to Test

1. Run ...
2. Verify output: ...
3. Edge case tested: ...

---

📸 Screenshots / Logs (if applicable)

---

📋 Checklist

• PR title follows Conventional Commits format
• This PR is linked to an existing issue
• The test method is described, and the expected result is clearly stated
• Relevant documentation has been updated (if applicable)

---

💬 Notes

---

[codecov]

Codecov Report

❌ Patch coverage is 34.88372% with 28 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
vllm_rbln/v1/worker/rbln_model_runner.py	6.89%	26 Missing and 1 partial ⚠️
vllm_rbln/v1/core/rbln_scheduler.py	92.85%	0 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!