[cwschroeder]

This adds opt-in parallel sparse attention and a second experimental two-Mac MoE partition for DeepSeek-V4.1-Flash-1017 EXL3 with SSD Engram. The current TCP pair remains slower than one Mac. The branch is based on upstream d1018b318fb51981c4a18f12e40269a4d3ee9713 and preserves its backend-neutral family files. The Metal path reads the shared configuration through a validated f32/slice view; its GPU loader, model and cache have separate capability files.

On two M3 Ultra Macs with 256 GB each, runtime cdfa5203 measured:

Configuration	Decode forwards/s	Repetitions
One Mac, parallel attention	4.8619	Median of 3 on the second Mac
TCP pair, MoE output columns	3.0992	1 per rank
TCP pair, whole routed slots	3.6003	Median of 3 per rank

Each generation has 128 completion tokens but 127 decode forwards. The routed-slot partition improves this TCP decode sample by about 16% over output columns, while both ranks still hold the complete backbone and attention. It provides no RAM reduction. The pair uses ordinary TCP over the TB5 interfaces, with both data and ACK sockets observed on bridge0/en3; it does not open an RDMA device.

TF_DSV41_PARALLEL_ATTENTION=1 scores independent keys and output dimensions concurrently while preserving each reduction order. The earlier attention-only prototype 61ab890f increased the single-Mac three-run median from 4.0519 to 4.8595 forwards/s. The current runtime repeats that result with the explicit 256-thread reservation. The separate teacher prefill in the earlier prototype became slightly slower, so this is not a claim that every workload improves.

TF_DSV41_TP_SLOTS=1 assigns slots 0/2/4 to rank 0 and 1/3/5 to rank 1, duplicates the shared expert, exchanges completed routed rows once per layer and restores the original six-slot combine order. Measured generation exchanges fall from 15,489 to 5,249, corresponding to 121 versus 41 per forward plus startup identity. Both switches default to the predecessor paths and must agree across the pair.

All six routed-slot generation token files match the frozen original. Teacher268 logits match the original SHA on both ranks with both partitions. One-token teacher chunks match the same original logits on both ranks. The interleaved two-lane test ends at positions 134/134 with zero bit disagreements against solo execution. All model phases pass their execution I/O gate, with zero page-ins. Both original Qwen services reply 4 after restoration; the original reranker returns the correct synthetic ordering. Test processes, ports and locks are clear.

Validation completed locally uses Zig 0.17.0 and -Dcpu=apple_m3. Both Macs pass the DeepSeek GPU, EXL3 GPU, reference and planted numerical control tests. The new attention passes its kernel-contract gates and actual M3 launch limits. The global limits check remains red with twelve missing M1/M2 Flash-Next declarations, reproduced on the unchanged upstream base with matching findings. Linux runs every applicable No GPU workflow command and Lean successfully; the full Zig format check and all five original encoding/template fixtures pass.

The complete desktop suite is 819/821 on both candidate and unchanged base, with the same sandbox-Git assertion and settings child timeout; their app trees are identical. Typecheck, mobile verification and whitespace pass on both. These local failures remain open findings. Remote CI is pending on the updated head and is separate from the local results above.

This stays a draft. Whole-model FP64/noise-floor qualification, the original-checkpoint anchor matrix, MTP, long context and full Rust-normalizer parity remain open under #847. The new native binary builds, but its HTTP serving behavior has not been requalified at cdfa5203; the existing native HTTP/cancellation/peer-loss receipts belong to b27e39af. Native Metal serving retains the explicit TF_EXPERIMENTAL=deepseek admission gate. No merge or production binary promotion is part of this patch.

Pinned data: current runner and slot receipts, earlier attention comparison, and historical native HTTP receipt.

[cwschroeder]

Closing this draft. The two-Mac partition decodes slower than one Mac: 3.60 against 4.86 forwards/s.

What the receipts measure:

• Wall time minus GPU time is 5.6 s for 127 forwards, about 44 ms a token. Each token waits for 41 TCP exchanges, and nettop shows a 1.0 to 1.2 ms round trip on the link.
• GPU time per rank rose from 25.4 to 29.7 s, although each rank computes only half the routed slots.

What we suspect but have not isolated: the code drains the command buffer twice per layer, once so the CPU can pick slots from the router output and once before the exchange. We do not yet know how much of the GPU gap that explains. A faster link alone would leave that gap.

We will rework the pair on our fork after speed-up mode's pattern (one command buffer, GPU-posted sequences and polled flags over MCDMA), and come back only with a receipt that beats one Mac.

The parallel sparse attention, the bit-identical part that gained speed, is now its own PR onto the family line: akol1#4. The receipts here stay pinned at 8c6e65ac.