Summary

• Widen the multi-step draft CUDA-graph cuda_graph_kv_indices buffer by the missing topk factor (max_bs * max_context_len -> max_bs * topk * max_context_len), matching the eager init_forward_metadata sizing.
• Add an always-on size invariant in common_template so an undersized row fails fast instead of silently corrupting memory.

Bug

• generate_draft_decode_kv_indices packs topk per-branch sequences (prefix + draft) into each row, so a row needs topk * seq_len entries; the CUDA-graph buffer omitted the topk factor that the eager path already has.
• For speculative_eagle_topk > 1, once seq_lens_sum * topk > max_bs * max_context_len the kernel writes past the row into adjacent graph-pool memory. The corruption is usually silent (wrong-but-in-vocab draft tokens); intermittently the stomped int32 indices land on the int64 draft token list, producing out-of-vocab ids that crash the target verify embedding gather (vectorized_gather_kernel index-out-of-bounds).
• The undersizing is about the topk packing, not page layout; demonstrated at page_size = 1 + topk = 8 (see Repro). Addresses a flashinfer topk > 1 draft coredump tracked in CUDA Coredump Tracker #26340.

Fix

• Size the CUDA-graph buffer with the topk factor (eager-equivalent).
• Invariant: assert seq_lens_sum * topk + bs * num_steps <= row_width host-side before the kernel launch -- a cheap always-on check that turns the silent overflow into a deterministic, clearly-messaged failure and guards future sizing regressions.

Repro

Isolates the bug to this buffer (flashinfer EAGLE topk>1, page_size=1 concurrent burst): https://gist.github.com/hnyls2002/3d20584c6d2f2af97fc9ff0e80472780

With this PR the burst passes (buffer sized max_bs * topk * max_context_len). Revert the * topk in init_cuda_graph_state (keeping the size invariant) and the first batch trips the invariant deterministically:

AssertionError: EAGLE draft kv_indices row too small: need 84280 but row width is 32768 (topk=8, num_seqs=7, seq_lens_sum=10500, num_steps=5)

page_size=1 keeps the repro on the draft kv_indices buffer; a single request cannot overflow it, hence the concurrent burst.

CI States

Latest PR Test (Base): 🚫 Run #27002608851
Latest PR Test (Extra): ❌ Run #27002608748