Support Blackwell GPUs and MIG slices for structure inference

[DimaMolod]

Summary

Makes the new Blackwell hardware actually usable for structure inference: one real bug fix for MIG
slices, plus documentation/routing for both Blackwell node types.

Companion to KosinskiLab/AlphaPulldown#629 (container side).

The bug: unified memory is silently disabled on MIG slices

structure_inference_xla_mem_fraction: auto computes host RAM / GPU VRAM from
nvidia-smi --query-gpu=memory.total. On a MIG slice that query reports the parent card, not
the slice — 32623 MiB for a 16 GB 1g.16gb instance. The fraction therefore comes out roughly
halved, which switches off host spill on exactly the small slices that need it most: borderline
complexes OOM instead of spilling to host RAM.

Fixed by reading the slice's own profile from nvidia-smi -L when CUDA_VISIBLE_DEVICES is a MIG
UUID, falling back to --query-gpu on whole cards.

Measured on EMBL gpu-el8 with 32000 MiB host RAM:

GPU	gpu_mem_mb	fraction	JAX actually sees
RTX PRO 4500 MIG 1g.16gb	16384 (was 32623)	1.953 (was 0.981)	~16032 MiB
RTX PRO 6000 Blackwell 96 GB	97887	0.327	~97252 MiB
RTX 3090	24576	1.302	~24126 MiB

Whole cards are byte-identical before/after — this only affects MIG nodes.

Docs / routing

• Blackwell documented as verified rather than expected: AlphaPulldown 2.5.0 containers run on
  RTX PRO 6000 (gpu50-53) and on the RTX PRO 4500 16 GB MIG slices (gpu60-68) for both AF3 and
  AF2, with confidence scores matching A100/H100/3090. Only pre-2.5.0 images fail there, at the
  first kernel compilation with ptxas too old / UNIMPLEMENTED — so excluding those nodes is an
  old-image workaround, not a hardware statement.
• Added both node groups to the structure_inference_gpu_tiers example: a 16 GB tier (monomers and
  small complexes) and a 96 GB tier (large ones).
• Noted that a plain gpu:1 gres already lands on a MIG slice (SLURM sets
  CUDA_VISIBLE_DEVICES=MIG-<uuid>), so no slurm_gres change is needed to use them.

Testing

• Fraction logic run on real hardware — MIG slice, RTX PRO 6000, and RTX 3090 — producing the table
  above; non-MIG output unchanged.
• The edited shell: block verified to expand under Snakemake's formatter (brace escaping) and to
  produce valid bash (bash -n).

🤖 Generated with Claude Code