[Core][SM70] Decouple the Mamba state grid from the KV block size

[yangzhuxinyzx]

What

The long-prefill chunk is min(max_num_scheduled_tokens, mamba_state_block_size), and the 75T Q8000 dense prefill route only dispatches for a chunk in [8000, 8192]. So the Mamba recurrent-state grid -- not the token budget -- decided whether the fast prefill path was reachable at all. Page-size unification scaled that grid together with the KV block size, which tied the fast path to a 4096 block and forced a choice between fast prefill and KV capacity.

This honours an explicit --mamba-block-size in align mode, and stops unify_kv_cache_spec_page_size multiplying a MambaSpec's block_size by the page-unification ratio when one is given. The physical page is still padded to max_page_size, and align-mode state memory is page_size_bytes * (2 + num_speculative_blocks), which does not depend on block_size -- so the change is memory-neutral. Both edits are inert unless the flag is passed.

Result

Qwen3.8-27B NVFP4, TP4, fp8_e4m3 KV, DFlash2, unique-salt 100K prompt:

--block-size	--mamba-block-size	grid	chunk	75T	@100k	KV pool
4096 (shipped)	-	8192	8192	yes	3426	818,142
2048	-	4096	4096	no	2954	1,058,133
2048	8192	8192	8192	yes	3318-3336	1,058,133-1,108,650

The default path is unchanged, re-measured with --block-size 4096 and no flag: grid 8192, chunk 8192, 75T dispatch, 858,310 KV tokens, 3272 tok/s.

Rejected: --block-size 1648

1648 is the smallest legal block and measured the largest pool (1,112,091) and the fastest prefill (3551 tok/s), but it does not divide 8192, so no prefix length is simultaneously block-aligned (KV blocks) and grid-aligned (recurrent state). Result: prefix_cache_queries_total = 812,040 against hits_total = 0.0, with an identical repeated 232K-token prompt still taking 86.5 s. The alignment assertion is retained for that reason; the in-tree example in scheduler.py obeys the same rule (816 = 51 x 16).

Note on prefix-cache reuse

Reuse is quantised by the grid, not the block: at grid 8192 a byte-identical prompt below 8192 tokens gets no reuse at all. Since chunk <= grid, the reuse quantisation, the chunk size and the grid are one and the same knob, so the smaller block buys KV capacity only and not finer reuse. That is inherent to align-mode hybrid scheduling, and is documented rather than worked around.

Evidence

• docs/design/sm70_mamba_state_grid_decoupling.md -- derivation, raw probe output, reuse-granularity table, adopted default.
• docs/design/sm70_v100_migration_control.md -- control entry.
• scripts/serve_qwen38_27b_nvfp4_v100.sh -- adopted default launcher for this model.

Verification

Re-tested on the committed revision, launched through the canonical script: 100K prompt at 3318 tok/s, 75T dispatch, 1,108,650 KV tokens, prefix_cache_hits_total = 32,768 at 8192 granularity, and cache-hit greedy outputs identical to the cold outputs at every size tested.