emberserve
From-scratch LLM inference engine: continuous batching, paged KV cache, prefix caching, hand-written Triton paged attention, CUDA graphs, OpenAI-compatible API. Benchmarked against vLLM on an A100
What XERJ indexed
Snapshot
The whole repository at commit 440137e37e (2026-09-30). main is 34 commits ahead.
Selected work
“cold start C, Qwen3-8B, fresh-trace sweeps”. 26 files, 6,913 lines in results.
Licence
The repository
LICENSE file checked at the pinned commit; the author agreed to inclusion.
The XERJ corpus
The author agreed by email on 2026-10-24. Same terms as the repository. Answers name the repo, the commit and the file:line.
selected for the XERJ Corpus Hublicence clearedsuggested use: adapt with attribution
continuous-batchingcuda-graphskv-cachellamallm-inferencepaged-attentionpytorchqwen
Example questions
- How does emberserve implement continuous-batching?
- Where does emberserve handle “cold start C, Qwen3-8B, fresh-trace sweeps”?
- What does results do in emberserve?
Inside the indexed corpus
- 409 files, 7.6 MB in total, about 18.9 KB per file.
- The selected commit changed 26 of those files (6.4%).
- Languages by size: Python 97.8%, Shell 1.7%, Slim 0.2%, Roff 0.2%, Dockerfile 0.2%.
- On GitHub since 2026-09-27, last push 2026-10-02, 211 commits on main.
- The pin is 34 commits behind main; the corpus answers from the pinned code, not the newest.
Licence on GitHub and in the corpus
GitHub reported MIT for ashwinsreedhar28/emberserve on 2026-10-29, and the LICENSE file at commit 440137e37e was reviewed as MIT. They match, so answers from this corpus carry the same terms as the live repository.
Contributors
Ashwin Sreedhar (@ashwinsreedhar28) has 1 selected repository on the hub.
Using XERJ with emberserve
- RAG without a vector database: one hybrid index The hub files it under AI, prompts and agents.
- Best grep alternative for AI coding agents (2026)
- Ask questions about a GitHub repo without cloning it
- Index a reference library for a coding agent
The prompt
Paste it into your AI agent. It clones ashwinsreedhar28/emberserve at the pinned commit and indexes it on your machine.
Search the GitHub repository ashwinsreedhar28/emberserve with XERJ (https://hub.xerj.org/corpus/ashwinsreedhar28--emberserve). Run these in a shell, in order, and check each result before the next: 1. xerj --help (success: it prints usage). If xerj is missing: curl -fsSL https://xerj.org/get | sh 2. git clone https://github.com/ashwinsreedhar28/emberserve emberserve && git -C emberserve checkout 440137e37e95ae294f88a6d50b72148fe1b267ce 3. Start a local node and leave it running: xerj --insecure --data-dir ./xerj-data 4. xerj autoindex ./emberserve --url http://127.0.0.1:9200 --prefix emberserve --progress plain 5. xerj autoindex map --url http://127.0.0.1:9200 Then answer my question from the indexed files and cite the ax_path of every hit. Query syntax: https://xerj.org/llms.txt Licence: MIT. Keep the copyright notice and credit the authors.
Related
Ashwin SreedharAI, prompts and agents corporaaiderdynalm-project/dynalmp95labs/ifameta-pytorch/tritonparse
Stats: GitHub API, fetched 2026-10-29. Snapshot: the git tree at the pinned commit. Selected work and consent: the XERJ registry.