Add or verify model support in SGLang's Rust processor so chat requests produce the same prompt text and token IDs as Python. Use reference-generated test cases to find differences in tools, reasoning settings, conversation history, and errors.
Verify file splits, function moves, module extractions, and renames by reproducing changes from a base commit and comparing files byte for byte. Generate repeatable proof scripts and reports that expose changes beyond relocation.
Add a CUDA/C++ GPU operation to sgl-kernel using AOT, meaning compilation before use. Follow the implementation through PyTorch registration, Python API exposure, SGLang integration, correctness tests, and benchmarks that account for cache reuse.
Monitor selected GitHub Actions checks on an SGLang pull request until they pass on its latest commit. Diagnose failures, push small validated fixes, and retry only failed jobs within defined limits.
Investigate SGLang runs that freeze across multiple GPUs. It uses stack traces, communication logs, GPU memory dumps, and comparisons between processes to locate state differences, trace their causes, and guide fixes.
Organize SGLang GPU operations when adding or moving code. It guides API placement, operation registration, test organization, and migration checks to preserve existing behavior, references, and continuous integration coverage.
Analyze LLM serving traces to identify GPU time usage, opportunities to run operations concurrently, and known patterns for combining operations. Produce three consistent tables from saved traces or live captures on supported servers.
Add a diffusion model or Diffusers generation pipeline to SGLang. Examine reference code, choose a suitable pipeline structure, reuse existing components, and implement model settings and processing stages, with guidance for later multi-GPU support.
Develop lightweight CUDA operations for SGLang using JIT, which compiles code on first use. A tensor-scaling example explains implementation conventions, input validation, memory access helpers, and when to choose this approach.
Collect and review SGLang startup logs to identify unnecessary output and trace its source. Get saved logs and specific cleanup proposals while preserving useful diagnostics, with runtime changes left for user approval.
Calculate how to divide serving memory between recurrent state and KV cache, two memory pools used by hybrid models. Use startup logs and workload settings to recommend a ratio, estimate concurrency, and identify the limiting pool.
Add, rename, or review SGLang environment variables using consistent definitions, types, names, and access methods. The skill also guides migration from old names and temporary test overrides that restore the original values afterward.
Review existing tests and CI, the automated checks run during development, for redundant coverage, misplaced tests, and opportunities to reduce runtime or cost. Receive file-specific recommendations based on documented improvement patterns, with uncertain findings separated.
Understand and modify SGLang's CI pipelines, including test stage order, start conditions, early stopping after failures, and workload splitting. Use the guide to add jobs and investigate failed, skipped, or cancelled checks.
Write and register SGLang unit and end-to-end tests, from choosing models and file locations to reusing server fixtures. Follow project rules for meaningful assertions, CI registration, dependency mocking, and safe server cleanup after tests.
Measure denoising latency in SGLang diffusion models and identify slow operations with profiler traces. Compare performance and output quality on the same GPUs, check existing optimizations, and gather evidence before starting low-level kernel work.
Choose SGLang Diffusion settings for a specific model, GPU, and memory budget. Compare generation speed, memory use, and output quality to select command-line options, including CPU offloading, multiple GPUs, caching, and reduced precision.
Convert an existing SGLang cookbook page to shared templates driven by model configuration. Preserve its deployment commands and interactive choices, transfer eligible benchmark measurements, and check equivalence, links, and rendered behavior.
Integrate or upgrade cache-dit in SGLang diffusion, the image generation system. Follow guidance for resolving code conflicts, adding cache settings, testing feature combinations, and comparing generated images using quality measurements and logs.
Investigate CUDA crashes in SGLang using logs and inputs saved before execution. Inspect tensor shapes, data types, and numerical anomalies, with guidance for separate process logs and additional CUDA debugging tools.
Guide changes and reviews for SGLang’s Scheduler, TokenizerManager, and ModelRunner classes. Keep initialization in small, replaceable helpers and separate computation from coordination so large classes retain clear responsibilities and explicit dependencies.
Add a model page to the SGLang Cookbook using shared templates. Organize verified launch settings and measured benchmarks, create configuration files and documentation, and update navigation, vendor cards, and page validation.
Apply consistent names to SGLang speculative decoding code, which verifies draft tokens to generate text. Distinguish tokens with or without the bonus token, counts, counters, and arrays when adding identifiers, renaming code, or reviewing changes.
Guide changes to SGLang runtime configuration and shared process state. Identify where resolved settings, runtime flags, resources, and per-forward state belong, and use the documented access patterns and checks when developing or testing.
Is this your account? Link GitHub to bring it onto your creator page (Japanese)