Will It Fit?
Estimate single-GPU text inference VRAM across Transformers and vLLM with a compact, explainable breakdown.
Qwen 3.5 9B · vLLM · Official BF16 checkpoint
Required GPU VRAM (0.9 budget)
24.7 GB
Core estimate: 22.2 GB. Against RTX 4090 24GB, this leaves 1.1 GB of headroom.
At 4,096 tokens, the estimated max concurrency is 5 concurrent requests.
Quick read
Selected GPU
Runtime
KV cache dtype
Context length
Current concurrency
Max concurrency @ context
Class
Bandwidth
Nominal VRAM
Core estimate
Required GPU VRAM (0.9 budget)
Headroom
Breakdown
Where the memory goes
Weights
10B resident parameters at 2.00 bytes each. Calibrated from the official checkpoint profile.
20.0 GB
KV cache
Concurrency 1, context 4,096, 8 KV-bearing layers, 4 KV heads, BF16 cache storage.
0.1 GB
Linear attention state
Concurrency 1, 24 linear-attention layers, static recurrent state, and short-convolution buffers. This term stays flat as context grows.
0.1 GB
Runtime / safety overhead
Conservative buffer for allocator fragmentation, kernels, and runtime scratch space.
2.0 GB
Weights = parameter count × bytes per parameter.
KV cache grows with context length, KV-bearing layers, concurrent requests, and the selected KV cache dtype.
Hybrid Qwen3.5 layers also keep a static linear-attention state, and runtime presets can inflate the required card VRAM beyond the core estimate.
Show the substituted formulas
Weights
parameter count × bytes per weight
10B × 2.00 = 20.0 GB
Qwen documents Qwen3.5-9B for Transformers and vLLM.
KV cache
batch × effective KV tokens across attention layers × 2 × KV heads × head dim × bytes per KV element
1 × 32,768 × 2 × 4 × 256 × 2 = 0.1 GB
Only the attention-bearing layers contribute KV cache in this hybrid stack, and BF16 controls the bytes per stored KV element.
Linear state
batch × linear layers × (recurrent state + short-conv buffers) × state bytes
1 × 24 × 557,056 × 4.00 = 0.1 GB
Hybrid Qwen3.5 layers keep a static recurrent state plus q/k/v short-convolution buffers. The published configs keep that state in float32, so it is modeled separately from the BF16 weight dtype.
Overhead
max(1.5 GB, 10% of weights + KV cache + linear state)
max(1.5 GB, 10% of 20.2 GB) = 2.0 GB
This leaves room for runtime buffers instead of claiming an unrealistically exact fit.
Tips
What to change next
This technically fits, but the headroom is thin. Leave extra space for kernels and runtime buffers.
Model
Selected model
Qwen 3.5 9B
10B dense • 262,144 context • 4 KV heads
Architecture
Total params
Active params
Layers
Hidden size
Attention heads
KV heads
KV-bearing layers
Context length
Modality
License
Assumptions
The calculator works in raw bytes, displays decimal GB, and keeps both the core tensor footprint and the runtime-adjusted card requirement explicit instead of pretending every engine uses the full card the same way.