llama.cpp DeepSeek DSA seeks 768GB VRAM

This is less a launch than a correctness hunt for an inference kernel that only really proves itself under brutal benchmark conditions.

• –The 768 GB VRAM ask tells you this is well past normal workstation territory and into shared-cluster or proxy-runner territory.
• –lineage-bench is a smart choice here because it stresses reasoning behavior, not just token throughput.
• –The failed 8x RTX PRO 6000 run suggests the bottleneck is memory layout for indexer tensors, not just raw compute.
• –Comparing against prior sglang fp8 runs gives a useful cross-framework sanity check if the sparse-attention patch is doing the right thing.
• –If the sparse branch matches expected quality deltas, it would be a strong signal that llama.cpp can support DeepSeek V3.2 Speciale without silently flattening its behavior.