vllm - 💡(How to fix) Fix [Bug]: AssertionError at kv_cache_utils.py:1042 — dense draft model + hybrid-attention main (DeltaNet+SWA) fails in unify_kv_cache_spec_page_size

Summary

Pairing a dense draft model (LocoOperator/LocoOperator-4B) with a hybrid-attention main model (Qwen/Qwen3-Coder-Next-80B-A3B, DeltaNet + sliding-window attention) fails at engine init with a bare AssertionError in unify_kv_cache_spec_page_size — the dense draft's page_size_bytes doesn't equal the hybrid main's max_page_size.

Stack trace (live, deterministic across 3 boot attempts)

File "/usr/local/lib/python3.12/site-packages/vllm/v1/worker/gpu_model_runner.py", line 5951, in profile_cudagraph_memory
    self._init_minimal_kv_cache_for_profiling()
File "/usr/local/lib/python3.12/site-packages/vllm/v1/worker/gpu_model_runner.py", line 5870, in _init_minimal_kv_cache_for_profiling
    kv_cache_groups = get_kv_cache_groups(self.vllm_config, kv_cache_spec)
File "/usr/local/lib/python3.12/site-packages/vllm/v1/core/kv_cache_utils.py", line 1654, in get_kv_cache_groups
    kv_cache_spec = unify_kv_cache_spec_page_size(kv_cache_spec)
File "/usr/local/lib/python3.12/site-packages/vllm/v1/core/kv_cache_utils.py", line 1042, in unify_kv_cache_spec_page_size
    assert new_spec.page_size_bytes == max_page_size
AssertionError

Reproduction

• Main: Qwen/Qwen3-Coder-Next-80B-A3B-NVFP4 (DeltaNet + SWA hybrid attention, MoE 3B-active / 80B total)
• Draft: LocoOperator/LocoOperator-4B (4B dense)
• Args added to an otherwise-working spec:

  --speculative-config '{"model": "/draft", "num_speculative_tokens": 3}'

• Hardware: 1× RTX PRO 6000 Blackwell (SM_120, 96 GiB), TP=1, --kv-cache-dtype fp8_e4m3
• Image: vllm/vllm-openai:v0.20.2-cu129-ubuntu2404

Failure reproduces deterministically — 3 boot attempts ~80s apart, identical traceback, weights load to ~60% before the engine hits the assertion.

Secondary observation

The assertion at kv_cache_utils.py:1042 has no error message attached (assert x == y with no , f"..." clause). Operators get a bare AssertionError with no expected-vs-actual values. Even just assert new_spec.page_size_bytes == max_page_size, f"draft spec page_size_bytes={new_spec.page_size_bytes} != main max_page_size={max_page_size}" would save significant git-archeology time.

Known historical working config

A fork tracking commit a29a754d1b (early 2026) had this exact pairing serving correctly and producing ~2× decode TPS lift on the same hardware class. The regression appears to be downstream of that point — the validation was either added or tightened.

Ask

1. Is dense-draft + hybrid-attention-main a supported pairing? It worked in the past.
2. If yes — what config are we missing?
3. If no — is there a tracked RFC for adding support, and can we contribute a minimal repro?

Related

• RFC #42082 — Standardize KV-cache Layouts
• Closed RFC #31634 — Decouple page_size_bytes calculation in AttentionSpec for TPU/RPA

Reproduction host is a Blackwell fleet running Qwen3-Coder-Next-80B-A3B as the AGENTIC_CODE primary — restoring this spec-decode path would meaningfully reduce per-user decode latency.