#3103 batch-groups: STRATA_SPLIT_MTP_BATCH beside STRATA_PIPED_CONCURRENT (opt-in STRATA_PIPED_CONCURRENT_SMB=1)

Beschreibung

## Summary

`STRATA_PIPED_CONCURRENT=1` is off beside `STRATA_SPLIT_MTP_BATCH=1` (`STRATA_PIPED_CONCURRENT is off beside STRATA_SPLIT_MTP_BATCH`), because the batched draft K/V (`Prefill::draft_kv`) would allocate on the read's thread. This adds an opt-in, `STRATA_PIPED_CONCURRENT_SMB=1`, that removes the allocation and lets the two run together. Off (the default), nothing changes.

## What allocates

Reading `Prefill::draft_kv`:

- its row buffers are carved from the prompt path's scratch region (`m.region`): no allocation;
- its two uploads (the bounds, 16 B; the tokens, `nb` x 4 B) go through `up_ring`, and are smaller than what `reserve_uploads` already reserves before the read (`T` x `K` x 4 B);
- the Q8_0 projections run through `mmq::Context` (llama.cpp's MMQ): the context is created at first use, and llama.cpp takes its stream-k fixup buffer from the context's pool, which grows at the first larger request.

So the MMQ pool is the one thing that can allocate. On devices whose MMQ configs do not use stream-k it never does (in the ggml we build with, all 254 RDNA2 configs have `stream_k = false`); on NVIDIA (Volta and newer) and CDNA it can.

## The change

- `mmq::Context::reserve_fixup(bytes)`: takes `bytes` from the pool and gives them back, so a later run that needs at most that much allocates nothing.
- `Prefill::reserve_draft_kv()`: creates the MMQ context and grows the pool to the largest fixup, one 128 x 128 tile of floats per SM (`nsm` x 128 x 128 x 4 B).
- In `generate.cpp`, with `STRATA_SPLIT_MTP_BATCH=1` and `STRATA_PIPED_CONCURRENT_SMB=1`, it is called on the drafter's stage before each concurrent read, beside the existing `mtp.reserve_prefill` and `sp.reserve_uploads`, with nothing in flight. It logs once: `STRATA_PIPED_CONCURRENT_SMB=1: the batched draft K/V (STRATA_SPLIT_MTP_BATCH) runs beside the concurrent read; its MMQ pool is grown before each read`.
- `ConcurrentReadSetup::split_mtp_batch_reserved`: `concurrent_read_off` no longer names `STRATA_SPLIT_MTP_BATCH` when it is set. `batch_read_test` covers the new case (checked that it fails when the new assertion is flipped).

## Measured

2x RX 6900 XT (gfx1030), ROCm 10.0, Linux. Flash-Next GSQ-RCO IQ3_S, 262144 context, `--kv int8 --kv-resident 32768`, layer split with part of the experts in RAM, `--spec 3 --spec-min-p 0.7`, `"parallel": 2`, `--batch-groups 2 --batch-mtp`, greedy. This branch's binary, a fresh engine per arm. Both arms have `STRATA_PIPED_CONCURRENT=1`; "with" adds `STRATA_SPLIT_MTP_BATCH=1 STRATA_PIPED_CONCURRENT_SMB=1`.

**Stability:** 9 arms, 0 stalls. In the three full runs (two with, one without), each arm read a 64K prompt while a short request arrived (4 give-ways per arm, short TTFT 8.0-8.1 s), two streams, a hand-off, and a needle test across the yield point (6 of 6 correct).

**Prompt read**, 3 interleaved pairs (each arm's outputs are token-identical run to run, so these are the same requests):

| | 4K | 32K | 128K |
|---|---:|---:|---:|
| without | 951-954 | 1,723-1,724 | 1,990-1,991 |
| with | 990-991 (**+4.0%**) | 1,754 (**+1.8%**) | 2,006-2,008 (**+0.9%**) |

**The cost, which is `STRATA_SPLIT_MTP_BATCH` itself and not this change:** the draft layer's K/V come from Q8_1 x Q8_0 MMQ instead of the drafter's own mmvq, so the drafts move (as `docs/MULTI_GPU.md` says). On the same requests, drafts accepted fell from 87.5% to 85.3% (and 84.5% to 82.4% in the full runs, 6 of 6 comparisons lower), and decode was about 3% lower at 4K and 32K, unchanged at 128K. On our rig the two roughly cancel, so we leave it off. On rigs where the prompt read is a larger share, or where `STRATA_SPLIT_MTP_BATCH` gains more, it may be worth it.

## Not tested

The path this guards: **no NVIDIA card here**, so the stream-k fixup allocation that this change pre-grows has not been exercised. On gfx1030 the MMQ never takes from the pool, so our runs show the two coexisting, not that the reservation is sufficient on a stream-k device. A run on CUDA with `STRATA_PIPED_CONCURRENT=1 STRATA_SPLIT_MTP_BATCH=1 STRATA_PIPED_CONCURRENT_SMB=1` and a few long reads beside a decoding slot would settle it.

---
Developed with an AI coding assistant; every number above was measured on 2x RX 6900 XT (gfx1030), ROCm 10.0, Linux.