SGLang is a framework for running inference for large language models and multimodal models, meaning it processes inputs to produce model outputs such as text or other media. It is used to serve and accelerate open AI models and related workloads.
Borrowing it
Nothing to install: this file belongs to sgl-project/sglang. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/sgl-project/sglang/main/.claude/skills/compute-mamba-ratio/SKILL.mdgit clone --depth 1 https://github.com/sgl-project/sglangWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/sgl-project/sglang/compute-mamba-ratio)<a href="https://agentmods.dev/skills/sgl-project/sglang/compute-mamba-ratio"><img src="https://agentmods.dev/badge/skills/sgl-project/sglang/compute-mamba-ratio/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/sgl-project/sglang/compute-mamba-ratio"><img src="https://agentmods.dev/badge/skills/sgl-project/sglang/compute-mamba-ratio.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- Snyk pass
- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00088 | $0.03560 |
| Opus 5 | $0.00044 | $0.01780 |
| Sonnet 5 | $0.00018 | $0.00712 |
| Haiku 4.5 | $0.00009 | $0.00356 |
Grade A, and why
compute-mamba-ratio scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 10d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 138 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Optimal hybrid dual-pool ratio (--mamba-full-memory-ratio)
A hybrid model (attention layers + linear-attention layers — the recurrent-state family: Mamba/SSM, GDN, KDA, etc.) splits serving memory into two independently-budgeted pools, fixed once at startup:
- state pool (the linear-attention recurrent state) → caps concurrency (hard: whole slots, worst-case reserved, fail-loud)
- full-KV pool (attention KV) → caps context × concurrency (soft: paged, over-committable via retraction)
--mamba-full-memory-ratio r splits the post-weight budget: mamba_budget = rest · r/(1+r), i.e. mamba_budget : kv_budget = r. This skill picks the r (or the pin---max-mamba-cache-size alternative) at which neither pool bottlenecks first for the user's workload.
The formula
r* = (S + D) · token_equiv · dcp_size / L
token_equiv = state_bytes_per_slot / kv_bytes_per_token
L= average context length per request (input + output tokens)token_equiv= full-KV token-equivalent of one state slotS= state slots per running request (cache-strategy dependent, table below)D=--speculative-num-draft-tokens(0 if NOSPEC); each running req carriesDextra intermediate statesdcp_size=--dcp-size(1 without DCP). DCP shards the per-rank KV bydcp_size, so KV gets ~dcp_size×cheaper per request → the balance shifts that much toward the state pool.
r is dimensionless (just the split). To also predict the actual concurrency you need rest (below).
Inputs to collect from the user
L— average context (input + output) in tokens.- The two per-GPU byte constants — one of:
- (a) measured (preferred, exact) — from one boot log at any ratio:
Mamba Cache is allocated. ... ssm_state size X GBwithmax_mamba_cache_size: N→state_bytes_per_slot = X / NKV Cache is allocated. #tokens: M, KV size: Y GB→kv_bytes_per_token = Y / M
- (b) derived — model arch (linear-layer count + state dims
d_state/d_conv/heads/head_dim; attention type + KV dims: MLA latent dim, or GQAkv_heads·head_dim·layers) × the dtypes below.
- (a) measured (preferred, exact) — from one boot log at any ratio:
S— from--mamba-radix-cache-strategy, the overlap scheduler, andSGLANG_OPT_MAMBA_SKIP_DECODE_LOCK(table below).D—--speculative-num-draft-tokens(0 if NOSPEC; also 0 when ReplaySSM spec-verify is enabled — see caveats).dcp_size—--dcp-size(1 if no DCP). If DCP and spec with a replicated draft KV, apply the draft caveat (below).- KV dtype (bf16 / fp8) and ssm dtype (fp32 / bf16) — they set the two byte constants (see the
token_equiv2×2). - (only to also predict the clamp, not just
r)rest= per-GPU memory − weights at the chosen--mem-fraction-static. Readavail memafterLoad weight end, orMemory pool end. avail mem+ pool sizes, from the boot log.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 10d ago First seen · 138 lines · 88 tokens per session scan A 55827de7e89d
compute-mamba-ratio is a skill published in the GitHub repository sgl-project/sglang (35,682 stars, last pushed today), licensed Apache-2.0. It adds 88 tokens to every session and 3,560 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
llama-factory
Expert guidance for fine-tuning LLMs with LLaMA-Factory - WebUI no-code, 100+ models, 2/3/4/5/6/8-bit QLoRA, multimodal support.
groq-inference
Ultra-fast LLM inference on custom LPU hardware. OpenAI-compatible API at api.groq.com. Lowest latency in the industry (500-1000+ tok/s). Supports chat completions, vision, audio (Whisper STT + TTS), tool calling, JSON mode, and streaming. Free tier available. Inference only — no training.
fireworks-ai-inference
Fast inference and fine-tuning platform with serverless and on-demand GPU deployments. OpenAI-compatible API for chat completions, embeddings, function calling, vision, and structured output. Supports SFT, DPO, and RL fine-tuning. SOC2 + HIPAA compliant.
ort
ONNX Runtime in Rust via the ort crate (2.x): loading sessions, configuring CPU/CoreML/CUDA execution providers, tensor I/O with ndarray, async-safe spawnblocking wrapping, global thread-pool init, and debugging provider/opset issues.
minimind-learning
A Chinese-language learning assistant for MiniMind, a small language-model project, that records study notes and recognizes common machine-learning terms.
llama-factory
Expert guidance for fine-tuning LLMs with LLaMA-Factory - WebUI no-code, 100+ models, 2/3/4/5/6/8-bit QLoRA, multimodal support.