ds4 TP with --ssd-streaming on Metal for GLM and V4.1

Facts

• "TP with streaming" here means two Macs running `ds4 --tensor-parallel` with `--ssd-streaming` for GLM 5.x or DeepSeek V4.1. The ds4 docs give no command for it, so every statement below comes from the source. [source]
• Because `tp_shard = tp.role != NONE && !ssd_streaming`, a streaming TP run skips every step guarded by `tp_shard`: the contiguous routed-expert half mapping, the sharded-bytes memory accounting, the GLM ownership-kernel validation and the wired-limit lazy-paging warning. [source]
• The GLM check `glm_tp_validate_ownership_kernels` (error text "GLM tensor parallelism lacks ownership-aware kernels for routed expert type %u in layer %u") runs only under `tp_shard`, so a streaming GLM TP run never reaches it. [source]
• The Metal warning "iogpu.wired_limit_mb is 0 -- TP expert shard will page lazily" is also inside the `tp_shard` path, so the documented `sudo sysctl iogpu.wired_limit_mb=120000` step is about the resident shard (about 97 GiB of sharded span views in the source comment), not about streaming. [source]
• Memory accounting returns the sharded byte count only `if (!ssd_streaming && g_tp_shard_model_bytes != 0)`, so a streaming run accounts memory by the normal streaming rules. [source]
• The Metal static-weights lock (keeps non-routed weights out of competition with streamed experts in the file cache) is conditioned on `ssd_streaming && !load_slice && !tp_shard`, so it stays active in a streaming TP run; `DS4_METAL_DISABLE_STREAMING_STATIC_LOCK` turns it off and the message "Metal SSD static weights remain pageable to preserve runtime headroom" says the lock was skipped. [source]
• For V4.1, the TP fast decode path (`ds41_decode_island` with the shared-owner split) is taken only when `tp_world == 2` and not `streaming`; with streaming the code falls back to the ordinary `ds41_graph_layer`. `DS4_METAL_DISABLE_V41_TP_SHARED_OWNER` disables the fast path in any mode. [source]
• The Metal TP batched-MoE path for 2 to 6 rows is gated on `!g->ssd_streaming` and on an M5-class GPU, with `DS4_METAL_DISABLE_TP_BATCH_MOE` as the off switch. [source]
• The ds4 docs list tested two-Mac TP setups that do not use the streaming flag: V4 Flash Q4/MXFP4, GLM 5.3 Flash Q4, GLM 5.2 IQ2_XXS, and V4.1 Flash Q2 "with disk-only Engram tables". [source]
• For the two-Spark CUDA TP section the docs say "Do not add `--ssd-streaming` or `--cuda-tensor-parallel`: those select different memory/execution modes". That sentence is about CUDA network TP; the Mac section has no such sentence. [source]
• The docs tell the operator not to treat repeated handshake or RDMA timeouts as a pass just because a retry works, and to keep both logs. [source]
• Do not probe the waiting coordinator with `curl` or `nc`; it may treat the connection as a worker handshake. [source]
• Is the combination supported? The Metal validator accepts it and the GLM gate schedule has a streaming branch; the docs show no command and no benchmark; the CUDA section forbids it. Both sides are real. [source]
• How routed experts are divided between the two ranks in a streaming run, since the contiguous-half shard map is off. The code read does not show it. [source]
• Whether the V4.1 fall-back to `ds41_graph_layer` under streaming makes TP slower than single-Mac streaming. No benchmark exists. [source]
• A streaming TP run skips the `tp_shard` steps: half-expert mapping, sharded-bytes accounting, GLM ownership validation and the lazy-paging warning. [source]
• The GLM ownership-aware kernel validation does not run when `--ssd-streaming` is set. [source]
• The `iogpu.wired_limit_mb` lazy-paging warning belongs to the resident TP shard path, not to streaming. [source]
• The Metal static-weights lock still applies in a streaming TP run. [source]
• ds4's V4.1 TP shared-owner decode path is disabled when streaming is on. [source]
• ds4's Metal TP batched-MoE path (2 to 6 rows, M5 GPUs) is disabled when streaming is on. [source]
• The ds4 docs list V4.1 Flash Q2 on two 128 GB Macs with disk-only Engram tables as a TP setup, without the streaming flag. [source]
• The ds4 two-Spark CUDA section forbids `--ssd-streaming` with TP; the Mac section does not say so. [source]
• Probing a waiting ds4 TP coordinator with `curl` or `nc` can be read as a worker handshake. [source]