dflash : run DFlash2 under split mode tensor

[Piggidragon]

Overview

A DFlash2 draft aborts on the first decode under --split-mode tensor.

DFlash2 uses the target's output.weight. Under split mode tensor that weight is split over the vocab, so the draft logits are too. The selector then runs top_k and get_rows on them, and the meta backend cannot run either op on a vocab split:

ggml-backend-meta.cpp:543: GGML_ASSERT(src_ss[0].axis != GGML_BACKEND_SPLIT_AXIS_0) failed

The backend has no all-gather. The target has the same problem and turns off backend sampling under split mode tensor. This PR does the same for the selector: it copies the logits to the CPU for the candidate pick, and the scheduler gathers the slices on the way. Only the top_k and get_rows on the logits run on the CPU. The rest of the selector stays tensor-parallel.

The CPU copy exposes a second bug. The scheduler keeps a view in the split where it sits, not where its data lives. A reshape of the CPU copy can land in a meta split, and ggml_backend_meta_buffer_simple_tensor aborts on it. The meta backend already skips one such case, a view of a host leaf (s_copy_main). This PR extends that check to any view whose data is outside the meta buffers, and lets that view close the last subgraph, as #66 does for the host cache.

Pinning the drafter with -devd CUDA0 does not work around the bug. The draft then runs layer split and cannot read the target's output.weight from the meta buffer:

ggml-backend.cpp:941: pre-allocated tensor (output.weight) in a buffer (Meta()) that cannot run the operation (NONE)

Testing

Qwen3.8-27B-UD-Q4_K_XL target, Qwen3.8-27B-DFlash2-Q4_K_M draft, RTX 4070 + RTX 3060, -sm tensor -ts 0.5,0.5 -np 2 -c 32768, KV on device, --spec-draft-n-max 7:

	result
llama/dev	aborts in warmup (assert above)
this PR	76 tok/s, draft acceptance 89.5% (94/105) and 91.2% (83/91) over two runs
this PR, -sm layer	67 tok/s, acceptance 94.3%, unchanged path

test-llama-archs -a qwen35 / llama / qwen3next: the Meta rows pass, the same as on llama/dev.

Not covered here:

• Host-resident draft KV (--kv-gpu-layers-draft, or -nkvo) under split mode tensor does not crash on llama/dev with this PR, but the draft acceptance drops to 0.2%. With ggml-meta : split a host-resident KV cache by head #66 and the other stacked split-mode-tensor PRs on top, the same setup reaches 91-94%. So this belongs to those PRs, not this one.
• With host KV on the target, llama/dev aborts in warmup before the draft loads (i_start == cgraph->n_nodes). ggml-meta : split a host-resident KV cache by head #66 fixes that.
• On the local stack with those PRs, the full config at 262k context runs at 68 tok/s, 94% acceptance.

The subgraph hunk touches the same loop as #66, so whichever PR merges second needs a small rebase.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - implemented by an agent on my instruction, see the Assisted-by: commit trailers.

🤖 Generated with Claude Code

https://claude.ai/code/session_01W8hAxqSiBjFwK9CdBaYCut