vh
6a8582936e
Merge + quantize path for turning the Gemma-4 26B-A4B ERP/RP LoRA into a
servable NVFP4A16 seat, plus a playbook entry for the defect found while
validating it.
The landmine (playbook §3.15): a `targets=["Linear"]` NVFP4 recipe silently
misses every MoE expert on this architecture. Gemma-4 stores each layer's 128
experts as two fused 3-D nn.Parameter tensors, not nn.Linear modules, so the
recipe resolves 205 of 427 modules and ZERO experts — 22.84 B params, 88.5% of
the model, left in BF16 with no warning. This is the same blind spot that
killed QLoRA here via bitsandbytes; the tool changed, the checkpoint layout did
not.
before linearize_moe: 427 Linears, 205 targeted, experts 0
after linearize_moe: 11,947 Linears, 11,725 targeted, experts 11,520
(30 layers x 128 experts x 3 projections)
llmcompressor's linearize_moe unfuses them; no registration needed because
Gemma-4 satisfies FusedExpertsProtocol structurally. Caught by an §4.1 dry run
that asserts the expert count before any GPU spend, which is now the documented
requirement rather than an optional step.
Scheme is NVFP4A16, deviating from the playbook's mixed-W4A4 default on
measured grounds: brokkr-smithy-dev benched the W4A4 quant of this checkpoint
at 12% on contradiction detection with CoT off against gen's 81%, the signature
of 4-bit input activations on a reasoning-dense task, and W4A4 KLD degrades
2-4x past ~10k ctx on sm_120. This is a 16,384-ctx RP seat. Marlin's prefill
cost is accepted.
Two further silent-failure guards, both from prior hard-won lessons:
- the merged model ships the UPSTREAM chat template, not the trainee base's
stale 365-line one, because training rendered through upstream and the
mismatch would present as a tuning failure
- calibration reads the run's own encode cache rather than re-tokenizing, which
sidesteps §3.14 (a fast tokenizer mutated by truncation=True and persisted by
save_pretrained clamps every prompt forever)
Merge-then-quantize rather than LoRA hot-swap, since hot-swap onto NVFP4 was a
silent no-op on vLLM 0.24.0 (#47639). merge_lora.py asserts sampled target
weights actually changed, so an inert adapter cannot ship as a tune.
docs/
Navigation map for the documentation tree. New session? Read
orientation.md first — it's the narrative overview
of the fleet, backup architecture, governing principles, and gotchas,
and it points at everything else.
Tree
docs/
├── orientation.md # start here — fleet overview + where-to-look guide
├── runbooks/ # ops runbooks (recovery, deployment phases)
│ ├── disaster-recovery.md
│ ├── nh3-prune-ritual.md
│ └── pbs-deployment.md
└── pfi/ # PFI-specific reference (services, models, VMs)
├── docker-stack.md
├── model-list.md
├── proxmox-vms.md
├── recommended-model-settings.md
├── vm-102-matrix-appservice.md
└── vm-102-matrix-synapse.md
What goes where
runbooks/— step-by-step ops procedures. Anything you'd reach for during an incident or while standing up new infrastructure. Examples: disaster recovery (blast-radius tiers + restoration steps), PBS deployment (9-phase rollout). New runbook → new file here.pfi/— PFI-specific reference material that's too narrow for the top-level CLAUDE.md but doesn't change incident response. AI model inventory, recommended inference settings, Matrix bridge config, Proxmox VM map. New stable reference → new file here.- Top-level (
docs/orientation.md,docs/README.md) — narrative guides about the workspace itself, not about specific infra.
Cross-references
- Fleet topology + servers table: top-level
CLAUDE.md. - Open work + recent milestones: top-level
STATUS.md. - Durable cross-session facts:
~/.claude/projects/-home-lkraven-development-eshpfi-management/memory/.
Conventions
- Markdown, GitHub-flavored. CommonMark renders fine in most viewers.
- File names are lowercase-kebab-case, descriptive. No dates in filenames — git history covers that.
- One topic per file. If a file grows past ~500 lines, look for a natural split before adding more.
- No checked-in binaries or checksums. Build/release artifacts belong
in a build pipeline or
tools/, notdocs/.