fix(ds4): use monolithic backend for single-device launches, not only for HIP

[Graffioh]

Summary

• route DeepSeek 4 single-device launches through DeepSeek4Backend on both CUDA and HIP
• reserve LayerSplitBackend for explicit multi-device placement or a configured remote target shard
• preserve automatic CUDA/Halo layer splitting whenever remote target-shard IPC is enabled

Why

CUDA launches previously always selected DeepSeek4LayerSplitAdapter. Without an explicit or remote split, its automatic sizing capped the local shard at 42 layers, so a single H200 executed a virtual [0, 42) + [42, 43) split on the same GPU.

That virtual boundary adds split orchestration overhead and prevents the full-model fused decode path, whose eligibility requires the same range to start at layer 0 and own the output layer. A single local device should instead use the monolithic backend; devices that cannot fit the full model can use its existing hybrid-expert fallback.

Expected behavior

Configuration	Backend	Behavior
H200/Halo single	DeepSeek4Backend	Full monolithic
3090 single	DeepSeek4Backend	Full attempt → hybrid fallback
3090 + Halo IPC	LayerSplitBackend	Auto-split CUDA/Halo
Multi-GPU explicit	LayerSplitBackend	Layer split

Validation

• git diff --check
• Run on H200 both single and virtual split (based on fix(ds4): correct layer-split HC handoff and per-device state #507)

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}

[davide221]

Rebased onto current main ae98774 and resolved the backend_factory.cpp conflict at head 6e5b149. The resolution preserves current DS4 prefill validation/config propagation and applies the intended routing rule: single local targets use DeepSeek4Backend; explicit or remote splits use LayerSplitBackend.

Lucebox4 validation (exact pushed head 6e5b149):

• git diff --check: PASS
• CUDA 13.3 / sm_86 Release: dflash_server and test_deepseek4_unit built successfully
• CUDA deepseek4_unit: PASS
• ROCm 7.2.2 / gfx1151 Release: dflash_server and test_deepseek4_unit built successfully
• HIP deepseek4_unit: PASS
• Model-backed RTX 3090 generation using DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf: PASS
  • DeepSeek4Backend selected (no virtual layer split)
  • expected 81.7 GiB full CUDA allocation failure cleanly fell back to hybrid placement
  • hybrid initialized with 2322 hot / 8686 cold experts
  • real chat request returned HTTP 200 and exactly OK
  • 5.75 s prefill; 4.4 tok/s decode; 6.20 s total

The isolated test server was stopped after validation. GitHub CI was approved and is running.