SyFI-ServingStudio
@SyFI-ServingStudio Organization
Public Skills
operate-gpu-spec
by SyFI-ServingStudio
Use when reading, querying, verifying, or extending the GPU spec catalog at gpu/spec.json — the per-GPU datacenter hardware table (mem/BW, dense TFLOPS by dtype, interconnect, nvl_domain_size, dollar_per_hour). Covers how to INTERPRET each field (TFLOPS are DENSE not sparse, interconnect BW is BIDIRECTIONAL, fp32 is CUDA-core not tensor, dollar_per_hour is the getdeploying on-demand "medium" rate, nvl_domain_size is a deployment param), how to SEARCH for a value (jq/python), how to VERIFY a value is right (sparse-doubling trap, one-way vs bidir, SXM vs PCIe vs NVL form-factor distinction, GB200 combined-for-2-GPU), what to do when a GPU/field is NOT in the table (web-search the fallback and clearly tell the user it is a searched result, not catalog-backed), and how to ADD/refresh a GPU. Only compute_capability reaches the simulator timing path (capability-keyed rules such as the FlashInfer all-reduce budget); modeled kernel timings come from measured kernels in profile.db, not from these peaks.
impl-wire-new-arch
by SyFI-ServingStudio
Use when wiring a just-built ServingStudio Sim L4 arch model into the shared dispatch sites so it is selectable by tag, predictable via timing-predict, and ready for a future worker — the integration step AFTER impl-compose-arch has produced the arch file. Covers arch/mod.rs re-exports, the arch/build.rs concrete builder + its uniform predictor dispatch (build_iter_model / build_attn_model / build_ffn_model), and the deployment build arm (worker pairing or explicit bail). The completion proof is a green timing-predict run on the new arch. Does NOT author the arch cost file / trait impl (that is impl-compose-arch) or build a worker.
operate-manage-jit-and-autotune-caches
by SyFI-ServingStudio
Use before restarting a serving engine or submitting an L1 profiling run, or when startup or profiling spends minutes compiling and tuning work an earlier process already did. Covers which caches persist, where they live, how to reuse them across restarts and worker processes, and when to invalidate. Not for the simulator's own kernel cost cache (that is impl-validate-kernel-cache).
dev-clean-reimplement-for-merge
by SyFI-ServingStudio
Use when a feature is substantially complete and accepted, but its development branch is too iterative or messy to merge directly. Reimplement the accepted behavior cleanly from current main while preserving equivalence and following each ServingStudio Sim layer's established patterns.
dev-create-worktree
by SyFI-ServingStudio
Use when the user asks to create, adopt, or set up a git worktree for ServingStudio Sim development, including converting an existing agent checkout into a wt-topic tree. Creates a sibling worktree from the current ServingStudio Sim checkout and provisions required working-copy artifacts. NOT for the parallel-writing-subagent isolation flow (that is dev-orchestrate-parallel-subagents).
impl-add-model-work-label
by SyFI-ServingStudio
Use when implementing independent model.work necessary-work support for a Transformer architecture or checkpoint so Optimality can produce R6/R7. Covers attention and FFN specs, heterogeneous LayerStack composition, exact architectures[0] registration, hand-derived parameter/FLOP goldens, semantic location maps, and exact-iteration verification. Not for simulator L1-L4 timing composition.
impl-wire-kernel-to-rust
by SyFI-ServingStudio
Use when implementing the Rust timing/cache wiring for a ServingStudio Sim L1 kernel after the orchestrator has validated the Python profiling handoff and selected the Rust Config/Input/cache shape. Covers KernelSpec implementation, sweep/enumerate/cache wiring, mod.rs exports, slot_input logging, Rust tests, and kernel-query grid smoke. Does not implement Python runners or choose the kernel semantics.
dev-file-design-review
by SyFI-ServingStudio
Use when asked to check, review, tidy, or validate a specific source file against repo design docs, behavioral correctness, naming quality, future-agent-facing comments, and readable code order. Produces either a focused patch plus validation results, or doc-backed findings/questions when code changes are not justified.
top-explore-models
by SyFI-ServingStudio
"Use as the top-level entry point when the user asks to understand a new Transformer model or Hugging Face checkpoint before ServingStudio Sim modeling work. Answers high-level model questions such as parameter scale, overall architecture, attention type, layer/head/hidden dimensions, context length, MoE layout, multimodal branches, and public architecture notes by collecting Hugging Face metadata/config/model-card evidence first. Route KV cache size/capacity calculations and attention/cache classification for MHA/GQA/MLA/DSA/GDN/custom mechanisms to dev-calculate-kv-cache-capacity. Route deeper Torch implementation, operation-list, forward-path, or kernel-semantics questions to dev-lookup-transformers-model."
dev-explore-kernel
by SyFI-ServingStudio
Use to search the kernel/serving ecosystem (vLLM, SGLang, TensorRT-LLM, FlashInfer, flash-attn, cutlass) for how an operation is really implemented — the exact call path, the launch granularity (one fused kernel or several), and the constraints. Read-only. Reports findings and, when a source is chosen to wrap, a concrete wrapper plan for impl-register-kernel. Does NOT decide the op's ServingStudio Sim kernel home (that is top-split-model-into-kernels) and does not edit code or register KernelProfilerSpec rows.
operate-profile-sim-speed
by SyFI-ServingStudio
Use when asked to profile, perf, or speed up the ServingStudio Sim simulator wallclock (how fast a run executes), not the modeled cluster throughput and not L1 kernel profiling. Covers building a symbol-rich release binary, capturing a perf record of a representative run, and reading the per-thread / flat breakdown to find the bottleneck.
add-analyzer-subject
by SyFI-ServingStudio
Use when adding a new ServingStudio Sim analyzer subject (a metric) end to end — the Rust compute side (a report + payload JSON built from a run's parquet via DataFusion) and the Python render side (a matplotlib PNG drawn from the payload). Covers the flat registry, the report/payload contract, the Applies gate, keeping a large run within the speed budget, and — emphatically — reusing the shared session / io / cdf infra and the Python common style / figure / layout helpers instead of re-implementing them. NOT for running an existing analyzer (that is run-moesim-analyzer) or adding an L1 kernel (that is top-add-kernel).
dev-lookup-transformers-model
by SyFI-ServingStudio
"Use as the lower-level code-exploration step for a Transformer model or checkpoint when high-level Hugging Face config/model-card/public-source evidence is insufficient. Trigger for local Transformers or Torch implementation questions such as operation lists, exact forward paths, class/function locations, QKV/RoPE/cache/MLP/MoE semantics, inherited model code, kernel-relevant behavior, or the smallest Torch reference. Usually follows top-explore-models; this dev lookup skill does not edit code."
top-compose-real-framework-from-sim
by SyFI-ServingStudio
Build or optimize a real serving framework through repeated simulation, one measured trial, and the Align framework to ServingStudio Sim side of operate-run-alignment.
operate-use-analyzer
by SyFI-ServingStudio
Use when selecting, reading, comparing, interpreting, or citing an existing ServingStudio Analyzer result, including simulation sweeps/singletons, timing predictions, kernel profiles, and kernel measurements. Makes Analyzer the numerical authority, selects an exact resource from explicit UI context, a just-created managed result, or the recent catalog, and copies exact citation tokens returned beside result values. Not for producing the result itself; use the matching operate-run- or operate-profile- skill first.
operate-profile-serving-run
by SyFI-ServingStudio
Capture and attribute a bounded nsys profile of a real LLM-serving process. Covers vLLM, SGLang, and framework-owned engines. Not for simulator wallclock, L1 profile DB rows, or the full alignment pipeline.
dev-present-changes-for-review
by SyFI-ServingStudio
Guide a user-paced review of a branch or diff: inventory first, then one coherent change at a time. Not a correctness or design audit of one file.
dev-compose-kernel
by SyFI-ServingStudio
Develop or integrate a real serving kernel, fused boundary, or communication path. Not for ServingStudio Sim L1 cost kernels.
operate-align-moe-kernel
by SyFI-ServingStudio
Diagnose alignment of one existing fused MoE kernel through exact-input, compressed-popularity, and full-run comparisons. Not for whole-model MoE attribution, routing-model design, or adding a new kernel backend.
dev-llm-serving
by SyFI-ServingStudio
Implement or review code in a real LLM or multimodal serving framework using the routed reference library. Under a parent workflow, follows its frozen trial.
dev-build-serving-repetitive-unit
by SyFI-ServingStudio
Build a reduced-depth model that keeps the production execution path for kernel and layer-timing work. Not for scheduling, capacity, or E2E results.
operate-compare-serving-performance
by SyFI-ServingStudio
Compare two real serving engines by end-to-end performance, then compare broad GPU-time regions without per-iteration alignment. Not for simulator alignment or an A/B within one implementation.
impl-compose-op
by SyFI-ServingStudio
Use when implementing an L2 op under simulator/src/op/ after top-split-model-into-kernels assigned an op to "reuse an existing kernel" (atomic) or "new compound op", and the L1 kernels it wraps already build. Covers the live code contract — atomic Op<K> (one Op::new line, no file) vs a compound op file with Config/Input/Self + build/compile/eval over the CostTree. Follows the code, not design.md's retired init_ops/dry_run_init_ops/lookup/ LookupResult. NOT for L1 kernels (impl-register-kernel / impl-wire-kernel-to-rust) or L3 worklets (impl-compose-worklet).