feat(server): reduce layer-split activation memory with backend precision policy

[weicj]

Summary

This PR reduces long-context target layer-split OOM risk by storing layer-split activation staging buffers in a backend-appropriate dtype instead of always using F32. The graph still casts back to F32 at ggml RMSNorm / norm-weight boundaries, so the lower-precision storage does not change those operator contracts.

The allocation reduced here is hidden_size * context_tokens * bytes_per_element * staging_buffer_count. Qwen35 and Laguna currently use two staging buffers; Gemma4 uses three because its per-layer input path also keeps an original embedding staging buffer.

At a 256K context cap, the staging allocation changes from:

• Qwen35 / Qwen3.6-27B / hidden=5120: 10240 MiB -> 5120 MiB.
• Laguna-XS.2 / hidden=2048: 4096 MiB -> 2048 MiB.
• Gemma4 31B / hidden=5376: 16128 MiB -> 8064 MiB.

Changes

• Add BackendActivationPolicy in common/backend_precision for target layer-split activation staging.
• Select activation dtype from backend architecture:
  • CUDA Ampere+ and HIP native-BF16 targets -> BF16.
  • CUDA tensor-core / GP100 and HIP gfx9/gfx10 targets without native BF16 -> F16.
  • Older or unknown CUDA/HIP targets -> F32.
• Add LUCEBOX_LAYER_SPLIT_ACT_TYPE=f32|f16|bf16 as an explicit local override.
• Extend shared layer-split activation buffers to allocate F32, F16, or BF16 tensors and upload F32 embeddings into the selected storage dtype.
• Add common/ggml_graph_precision.h for graph-side F32 casts.
• Make Qwen35, Gemma4, and Laguna RMSNorm / norm-weight graph boundaries F32-safe.
• Wire Qwen35, Gemma4, and Laguna target layer-split adapters through the shared activation policy.
• Add no-GPU unit coverage for the CUDA SM and HIP gfx dtype-selection tables.

Notes

• A real 16K Qwen3.6-27B Q4 HIP layer-split A/B on dual Pro VII reduced request-time peak VRAM from 10199/11224 MiB to 9896/10918 MiB, saving 303/306 MiB across the two cards. This matches the ~313 MiB expected from the sizing formula at 16K.
• Runtime smoke passed on dual Pro VII / gfx906 HIP target layer split with the default F16 activation path for Qwen3.6-27B Q4, Laguna-XS.2 Q4, and Gemma4 31B Q3.
• This PR is intended as a follow-up to refactor(server): share target layer-split runtime helpers #306 and feat(dflash): reduce feature mirror memory with dtype policy #309: refactor(server): share target layer-split runtime helpers #306 provides the shared layer-split runtime structure, while feat(dflash): reduce feature mirror memory with dtype policy #309 covers DFlash draft feature mirror storage.

[cubic-dev-ai]

4 issues found across 30 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>