Search Vast.ai GPU Offers

Help the user find the best GPU offers on Vast.ai.

User Request

$ARGUMENTS

Instructions

1. Understand requirements: Key dimensions to determine:

   • GPU model (RTX 4090, A100, H100, etc.)
   • Number of GPUs
   • VRAM requirements
   • Budget ($/hr)
   • Region preferences
   • Reliability needs
   • Pricing type (on-demand vs spot/interruptible vs reserved)

2. Build the search query from the mapping below.

3. Run the search and present results clearly.

4. Help interpret: Highlight best options based on user priorities.

Query Construction

User says	Query field
"4090", "RTX 4090"	gpu_name=RTX_4090
"A100 80GB"	gpu_name=A100 gpu_ram>=80
"H100"	gpu_name=H100
"4 GPUs", "multi-GPU"	num_gpus>=4
"at least 48GB VRAM"	gpu_total_ram>=48
"under $1/hr", "cheap"	dph_total<1.0
"reliable"	reliability>0.95
"US only"	geolocation=US
"datacenter"	datacenter=true
"spot", "interruptible"	add -b flag
"reserved"	add -r flag
"NVLink"	bw_nvlink>0
"fast internet"	inet_down>500

Always include reliability>0.9 unless the user wants cheaper unreliable machines.

Default sort: -o 'dph_total' (cheapest first).

Examples

# Cheap single 4090
vastai search offers 'gpu_name=RTX_4090 num_gpus=1 reliability>0.9' -o 'dph_total'

# 8x H100 for large training
vastai search offers 'gpu_name=H100 num_gpus=8 reliability>0.95' -o 'dph_total'

# Any GPU with 80GB+ VRAM in US datacenters
vastai search offers 'gpu_total_ram>=80 geolocation=US datacenter=true reliability>0.95' -o 'dph_total'

# Budget multi-GPU fine-tuning
vastai search offers 'num_gpus>=2 gpu_ram>=24 dph_total<1.5 reliability>0.9' -o 'dph_total'

# Spot pricing for H100s
vastai search offers 'gpu_name=H100 num_gpus>=4' -b -o 'dph_total'

Presenting Results

After running the search:

1. Summarize top 3-5 options in a clear table with: Offer ID, GPU, # GPUs, VRAM, $/hr, Reliability, Location
2. Highlight the offer ID prominently — user needs it to launch
3. Note trade-offs (price vs reliability vs perf)
4. Mention spot pricing if it would save significantly
5. If no results, suggest relaxing constraints