fix(moe): size mixed-GGUF expert banks from GGUF metadata

[zyy0212time-del]

Summary

bank_bytes_estimate() can now size mixed-GGUF expert banks (expert_quant="gguf") from model metadata, allowing the automatic pageable-GPU residency planner to account for them before loading. Allocation and estimation share the same GGUF expert-bank geometry helper.

Problem

For mixed-GGUF checkpoints such as qwen35moe / Laguna GGUF, bank_bytes_estimate() returned None, so the automatic planner had no usable bank size and skipped pre-load split-residency planning.

Raw GGUF tensor payload bytes are not an equivalent substitute. The expert banks use uniform flat [num_experts, stride] slots, with gate/up and down strides determined by the maximum participating quant geometry and aligned to 64 bytes. For example, a Q4_K down layer may occupy a slot sized for the larger Q6_K down geometry.

Root cause

The GGUF expert-bank geometry already existed in the qwen35moe and Laguna loaders, but the same calculation was duplicated locally and unavailable to bank_bytes_estimate().

The generic fixed-format sizing table therefore could not safely size the mixed-GGUF layout.

Fix

• Add a shared expert_bank_geometry(config) helper based on the existing GGUF row_bytes() block geometry and per-layer gguf_expert_types.

• Reuse that helper in both the qwen35moe and Laguna GGUF loaders.

• Add a "gguf" path to bank_bytes_estimate() using:

  num_moe_layers * num_experts * (gate_up_stride + down_stride)

• Return None for unknown quant types, incomplete metadata, or malformed type entries rather than producing a potentially unsafe estimate.

• Leave existing non-GGUF estimation paths unchanged.

Validation

Added 8 CPU/synthetic repository regression tests covering mixed global-max geometry, homogeneous GGUF layouts, exact sizing, unsupported/missing/malformed metadata, the existing non-GGUF path, and shared geometry use across both model loaders and the estimator.

For Ornith-1.5-35B-A3B-Abliterated-Q4_K_M:

• gate/up stride: 1,179,648 B/expert
• down stride: 860,160 B/expert
• logical bank size per layer: 522,190,848 B
• 40-layer total: 20,887,633,920 B
• estimator delta: 0

A native Windows / RTX 5060 Laptop E2E validation with a 15 GiB pin budget and no FT55_DEBUG_BANK_BYTES override selected 10 pageable + 30 pinned MoE layers, completed 60/60 host registrations with 14.589844 GiB cumulatively registered, reached /health = ready, and returned HTTP 200 with the expected "OK" smoke response.

This is a generic sizing fix rather than a Windows-specific change; no performance claims are made.

[zyy0212time-del]

Hi @lucaspirola — just a gentle ping on this stacked fix when you have a chance.

It is still scoped to the mixed-GGUF bank-sizing gap discovered while validating FlashML-org#196, with the regression tests and native-Windows E2E result included in the PR description.

Happy to rebase or retarget it if you would prefer to update FlashML-org#196 first. No rush — I mainly wanted to make sure it didn't get lost.