Workbench / Online

BUILD YOUR COMPUTE STACK

Start with the model. MuseBoard maps the memory and hardware. Settings are mirrored to the URL, so any configuration can be shared.

Workbench / Online

Model / Qwen 72BMode / EstimateUnits / GB (10⁹ B)

A / Model configuration

ModelAlibaba

Parameters (billions)

Editing the count switches to a custom model with an estimated architecture.

Precision0.5 B / param

Context length8,192 tokens

Batch sizeConcurrent sequences

Workload

Target hardwarePinned

Architecture / published config

Layers: 80
Hidden: 8,192
KV heads: 8 × 128
KV / token: 320 KB

B / Estimated VRAM

Inference

42.7GB

Serving: weights + KV cache + runtime overhead. Estimate — not a guarantee.

Model weights72.7B params × 0.5 B36.4 GB
KV cache8,192 tokens × batch 12.7 GB
Runtime overhead≈10% of weights3.6 GB

Memory headroom

Free after estimated load

Utilization

Target ≤ 90% of 48.0 GB

Decode ceiling

~17 tok/s

Theoretical, batch 1, bandwidth-bound

VRAM usage

42.7 GB / 48.0 GB

0 GB2 × 24 GB

Pinned configuration

2 × NVIDIA L4 24GB

Total VRAM: 48.0 GB
Topology: Single node · TP
Interconnect: PCIe 4.0
Board power: 144 W

Compatibility / 2 × L4

Inference
42.7 GB · 89% of 2 × L4
Ready
Fine-tuning
66.2 GB · Needs 4 × L4
Limited
Training
1,299 GB · Needs 64 × L4 (8 nodes)
Cluster

Alternative configurations

MuseBoard gives first-order estimates for planning conversations. It does not replace profiling, capacity testing or detailed ML infrastructure planning.