Match a workload to the cheapest GPU
match_workloadRanks up to eight GPU configurations that can run a job at the lowest live price. Describe the job in query (an open-weight model to serve or fine-tune, a model size, or a GPU need) or with the structured fields. Each match gives the GPU count, the USD per hour for the whole configuration and a one-line reason; the answer also states the VRAM the job needs and how much cheaper the top match is than a hyperscaler running the same job. Use it to decide where to run a workload; use list_gpu_prices for the price of a named GPU. Open-weight models only: an API-only model such as GPT-4 returns no matches and a note. Ranked on price and fit alone: no provider pays for placement (each match still reports partner, a provider that pays FastGPU a referral). Hard limits (budget_usd_hr, max_total_usd with duration_hours, max_quote_age_minutes) drop every configuration that breaks them; limits reports what was dropped. Read-only: it does not rent, reserve or launch GPUs. No key required.
Tool call template
This tool declares no required parameters. Add optional parameters from the input schema as needed.
{
"id": 1,
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"arguments": {},
"name": "match_workload"
}
}Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| spot | No | Set true to include interruptible spot capacity for a cheaper rate. | |
| task | No | What the job does. | |
| model | No | Open model name to size against, e.g. "Llama 3 70B", "Qwen 72B", "Mixtral". | |
| query | No | Plain-language job, e.g. "cheapest to serve Llama 3 70B", "2x H100 for fine-tuning" or "a GPU with 80GB of memory". Provide this OR a structured spec below. | |
| region | No | Restrict to a data-residency region. | |
| vram_gb | No | Rough VRAM the job needs, in GB, if you already know it. | |
| params_b | No | Model size in billions of parameters when no exact model is named. | |
| reserved | No | Set true to include reserved / committed-term capacity for a lower rate. | |
| gpu_count | No | Exact positive GPU count. Overrides a count in query text. Returns no matches if no supported configuration fits; omit for automatic sizing. | |
| precision | No | Numeric precision to size the model at. | |
| weights_gb | No | Measured size of the weights as stored, in GB: the file you will load (for example a GGUF Q4_K_M build, which holds more bits per parameter than its nominal 4). Replaces parameters times bytes per parameter in the memory estimate. Used for a model sized by its parameter count; the answer's workload.weights_gb is null when it was not (an image or speech model sized as a whole pipeline). | |
| concurrency | No | For a serving or fine-tuning job: requests (or sequences) processed at once, 1 to 2048 (default 1). Sizes the KV cache, and for serving the tokens_per_sec and usd_per_million_tokens estimate: more requests at once share each step's weight reads. | |
| budget_usd_hr | No | Hard limit on the hourly price of the whole configuration, in USD. Configurations above it are not returned. | |
| max_total_usd | No | Hard limit on estimated_total_usd, the GPU rental cost of the whole job, in USD. Needs duration_hours. Configurations above it are not returned. | |
| context_tokens | No | Tokens of context each request holds, prompt and output together, 512 to 1000000 (default 4096). Sizes the KV cache and the per-token estimate. | |
| duration_hours | No | How long the job will run, in hours (fractions are fine). When sent, each match carries billed_hours and estimated_total_usd. | |
| active_params_b | No | For a mixture-of-experts model: the parameters it runs per generated token, in billions (Qwen3-30B-A3B runs about 3). Used for tokens_per_sec and usd_per_million_tokens; the memory estimate still holds every expert. Omit it to use FastGPU's figure for a model it knows, or what the Hugging Face repo's own config.json states (workload.active_params_source says which). | |
| engram_placement | No | For a model with Engram memory tables (DeepSeek-V4.1): gpu (default) holds them in GPU memory; host keeps them in the machine's RAM, as SGLang's --engram-config '{"placement":"cpu"}' does, so vram_required_gb leaves them out and workload.engram_gb is the RAM the machine needs for them. A token reads only a few of their rows, so the per-token estimate is the same either way. | |
| min_gpu_memory_gb | No | Memory each GPU card must have, in GB (e.g. 80 for 80GB-class cards such as the H100 or A100 80GB). Smaller cards are never returned. On its own it lists every card that size, cheapest first. | |
| max_quote_age_minutes | No | Hard limit on how old a price quote may be, in minutes. Configurations whose price was fetched from the provider longer ago are not returned. |
Output Schema
| Name | Required | Description | Default |
|---|---|---|---|
| hero | No | ||
| note | No | Why there are no matches, when there are none. | |
| as_of | No | When this answer was computed (ISO 8601). Quote ages are measured against it. | |
| count | Yes | ||
| error | No | Present when the request was not valid; says what to send instead. | |
| stale | No | ||
| limits | No | The hard limits this answer applied (null when one was not sent) and how many configurations each one left out. | |
| matches | Yes | ||
| workload | No | ||
| updated_at | No |