Profile
Agents sit on your GPUs and flag the true bottleneck: kernel, batch, decode, or HBM.
Product
A mill for inference silicon.
Dielet is an inference foundry. Performance engineers are agents that rewrite kernels. You buy the endpoint that those agents verified on your mix — serverless Turbo or a dedicated die.
Search
Candidates include fused kernels, FP8 passes, and new shard maps. Each is measured.
Verify
A candidate ships only if outputs match and the clock is lower on your mix.
Serve
OpenAI-compatible endpoints. Dedicated capacity in under a day.
Search space
What the mill actually tries.
Nothing ships because it looks faster on a blog. A candidate has to beat your recorded mix and keep the tokens.
Kernels and decode
Fused attention, KV layouts, and decode paths measured against the sequence lengths you actually serve.
Batching and shards
Batch shapes, FP8 passes, and shard maps. The mill keeps the map that won on your GPUs, not a generic recipe.
Weights
Open models on the rack.
Bring a public checkpoint or a LoRA. Named families we mill first:
- Llama 3.1 70B
- Llama 3.1 8B
- DeepSeek V3
- DeepSeek R1
- Mistral Large
Serving
Two ways to take the die.
Hosted Turbo
Shared, optimized endpoints. Playground tokens on Spark; Forge when you want SSO and a 99.9% target.
Dedicated Die
Reserved H100 or H200, private VPC, 24-hour cutover. Fab if the mill has to run on-prem.
Teams
Who it is for
Builders
Keep your chat client. Point it at a Dielet base URL after the mill prints a verified die.
Operators
GPU-seconds, versions, and spend sit next to the same objects your engineers debug.
Security
SSO, audit logs, and regional storage. Ujjwal Kumar Singh signs the DPA.