Summary

• Mamba / hybrid-linear models (GDN — Qwen3-Next; Mamba2 — Nemotron-H) with --speculative-eagle-topk > 1 previously fell back to the spec v1 worker. This runs them on spec v2, reusing the tree-drafting machinery from Spec v2 tree drafting (topk>1) with page_size>1 #26972.

Changes

Spec v2 mamba tree verify

• eagle_worker_v2.py: compute the tree-aware last_correct_step_indices by gathering each request's last accepted node from accept_index, replacing the chain-only accept_lens - 1; drop the topk != 1 guard (and the now-unused verify_input arg).
• speculative_hook.py: remove the mamba + topk > 1 -> spec v1 fallback.

Hybrid KV pool

• HybridLinearKVPool now sets enable_kv_cache_copy (matching the MHA / SWA / FP4 pools), so the topk > 1 tree path's move_kv_cache (accepted-branch KV compaction) is initialized for hybrid models. Without it the tree path hit KV copy not initialized.

Guards & docs

• LightningAttentionBackend (seg_la): raise NotImplementedError on topk > 1 — seg_la has no parent-indices plumbing and verifies a draft tree as a chain, so tree verify is genuinely unsupported there; fail fast instead of silently mis-decoding.
• Correct the stale "Mamba2 tree verify is structurally unsupported" comments in the attention unit-test suite — the production SSM kernel does consume the parent-indices plumbing; the unit test skips topk > 1 only for lack of a tree-aware reference.

Validation

• TestQwen3NextMTPTopk (Qwen3-Next-80B, NEXTN, topk=4) now runs on spec v2.
• H200, topk=4, accept length ~4: Qwen3-Next-80B gsm8k 0.96, Nemotron-H-120B gsm8k 0.965.

CI States

Latest PR Test (Base): ❌ Run #27088568107
Latest PR Test (Extra): ✅ Run #27088568046