Catalog memory components and the vLLM weight-offload split law (tier model, part 1 of 2)

[gunbai-bot]

What this is

Part 1 of 2 of the memory-tier model for the GH200 lane (node adhoc-e2818997-28d). It is split out of #13321 because the dashboard reviewer timed out three times on the combined diff (reviews 76027, 76033, 76057), each run ending at persona load. The parent asked for this split. #13321 keeps the arm_memory_fit migration and shrinks to that alone once this lands.

extdeps.systems.types: a system's memory components, addressable.

• Before this change, catalog rows carried each memory's facts but gave no way to name "this memory of this system".
• SystemMemoryComponent is the system's catalog name plus MemoryFacts plus GpuMemoryAddressability. Addressability takes one of three forms:
  • GpuAndCpuShareOnePool
  • GpuAddressesDirectly
  • GpuAddressesAcrossCoherentLink { per_direction }
• Each row shape has its own projection, derived from fields that already exist (no new figures):
  • integrated_system_memory_components gives the GB10's one pool.
  • coherent_superchip_memory_components gives the GH200's HBM, plus its LPDDR5X across the C2C link at the row's own link rate.
• A component's identity is the system name plus how the GPU reaches that memory. One system never reaches two memories the same way, so the pair names exactly one component.

extdeps.vllm.weight_offload: the engine's split law, cited at vLLM 8d09804c8.

• vLLM's weight offloaders move whole parameters of decoder layers: make_layers calls wrap_modules.
• UVA keeps no device copy.
• Prefetch keeps prefetch_step device copies of each distinct offloaded parameter. The source is PrefetchOffloader.post_init, which builds StaticBufferPool(slot_capacity=prefetch_step).
• The row states the condition the UVA law depends on: the MoE backend must not re-lay-out weights after loading. That holds for the TRITON / VLLM_CUTLASS arms of convert_to_fp8_moe_kernel_format. It does not hold for the DeepGEMM, FlashInfer, Marlin or AITER arms, which allocate new tensors on the device. The launch has to read this back; the row does not assume it.

Repairs after the side-chat REQUEST_CHANGES at 3303271

1. Zero-copy UVA is an established realization, not a configured backend.
   • VllmWeightOffloadSplitLaw is now sole_constructor. Its only producer is vllm_weight_offload_split_standing.
   • For UVA, that producer reads three facts: is_uva_available(), VLLM_WEIGHT_OFFLOADING_DISABLE_UVA, and the realized VllmFp8MoeBackend. The backend is a closed coproduct of every Fp8MoeBackend arm, and each arm carries its convert_to_fp8_moe_kernel_format relayout fact.
   • It returns one of three standings:
     • an unread fact gives WeightOffloadSplitUnestablished, naming each obligation;
     • a relayout backend gives WeightOffloadSplitRefused;
     • UVA unavailable or disabled gives the functional_call fallback law, with one device copy;
     • only a full read with an in-place backend gives UvaZeroCopyView, with zero copies.
   • VllmPrefetchStep is sole_constructor behind vllm_prefetch_step, which refuses zero (upstream OffloadConfig.validate_offload_config).
   • Reds in test.claim.vllm_weight_offload_witness: UVA unavailable, UVA disabled, a relayout backend (DeepGEMM, Marlin), unread facts, prefetch step zero.
2. Memory-component identity cannot alias.
   • SystemMemoryComponent is sole_constructor. Only the row-shape projections build it.
   • system_memory_component_same compares the whole fact: system, capacity, memory kind, and the complete addressability, including the coherent link's per-direction rate.
   • Reds in test.claim.system_memory_component_witness: a catalog row edited to the same system name with different GPU memory, and one with a different link rate. Neither aliases the original component.

Consumers (DESIGN §3c)

This PR is a declared frontier. Nothing in it is consumed in this diff. The named consumer is #13321, the second half of this split:

• gunbc.spark.arm_memory_fit takes its memory tiers from these components. It joins them against the host's catalog memory, uses addressability to decide which tiers the engine envelope governs, and gives the Spark a one-tier subject through integrated_system_memory_components(nvidia_dgx_spark_4tb_catalog).
• checkpoint_tier_split consumes VllmWeightOffloadSplitLaw.

The trigger is #13321 landing. If it does not land, both modules here are dangling and should be reverted.

🤖 Generated with Claude Code