Placement

When you ask the fleet for a host, hwLedger ranks every candidate with a single blended score and returns the top N.

Formula

$$\text{rank}=0.7\cdot \text{fit\_score}+0.3\cdot \text{cost\_score}$$

• fit_score ∈ [0, 1] — how comfortably the plan fits in the host's free VRAM. free_bytes / plan.total_bytes, clamped to 1.0.
• cost_score ∈ [0, 1] — 1 - (hourly_usd / max_hourly_in_pool). Local hosts with hourly_usd = 0 get cost_score = 1.0.

The 0.7/0.3 split biases toward fit: a host that barely fits is penalised more than one that costs a bit more.

Worked example

Plan needs 62 GB for DeepSeek-V3 at seq 32 768, users 2.

Host	Free VRAM	$/hr	fit	cost	rank
Local M2 Max	80 GB	0.00	1.00	1.00	1.000
RunPod A100-80	78 GB	1.89	1.00	0.37	0.811
Vast.ai 2×A6000	92 GB	1.20	1.00	0.60	0.880
Lambda H100	75 GB	2.99	1.00	0.00	0.700

Local wins outright. Drop the local host and Vast.ai edges out RunPod because of cheaper hourly cost at equal fit.

Tie-breaks

When two hosts score within 0.01:

1. Prefer local over rented (kind == Local).
2. Prefer lower latency (last_seen_rtt_ms).
3. Prefer higher free VRAM headroom (absolute bytes, not ratio).
4. Stable sort by host_id as a deterministic final tiebreaker.

Configuration

The weights live in hwledger-server/config.toml:

[placement]
fit_weight = 0.7
cost_weight = 0.3
min_free_headroom_bytes = 1073741824   # 1 GiB slack

Set min_free_headroom_bytes to force a safety margin — any host with less free VRAM than plan.total_bytes + headroom is filtered before ranking.

Related

• Fleet overview
• Cloud rentals
• Server crate