메뉴
Modal.com knowledge for answering user questions about serverless GPU compute, cost comparison with Daytona, and Python-based sandbox workflows. Use when the user asks about GPU access, ML training, Modal pricing, serverless sandboxes, or wants to compare cloud compute options for their Rust/Python projects.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | modal |
| description | Modal.com knowledge for answering user questions about serverless GPU compute, cost comparison with Daytona, and Python-based sandbox workflows. Use when the user asks about GPU access, ML training, Modal pricing, serverless sandboxes, or wants to compare cloud compute options for their Rust/Python projects. |
Modal is a serverless cloud platform optimized for compute-heavy workloads. Unlike Daytona (full Docker containers), Modal runs functions in lightweight containers with per-second billing. It excels at GPU workloads (T4, A100, H100) and Python-based data processing.
The user's project (specgen) may use Modal for:
export MODAL_TOKEN_ID=***
export MODAL_TOKEN_SECRET=*** Use sandbox crate
cargo run -p specgen-sandbox -- spin --provider modal \
--repo https://github.com/bl1nk-bot/specbl1z.git \
--branch main --name my-sandbox
cargo run -p specgen-sandbox -- exec --id <sandbox-id> \
--cmd "cargo build --workspace --release"
cargo run -p specgen-sandbox -- delete --id <sandbox-id>
import modal
app = modal.App("specgen-build")
image = modal.Image.debian_slim().run_commands([
"curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y",
"git clone https://github.com/bl1nk-bot/specbl1z.git /workspace",
])
@app.function(image=image, gpu="T4")
def build():
import subprocess
subprocess.run(["cargo", "build", "--workspace", "--release"], cwd="/workspace")
$30/month free credits on Starter plan. You only pay for actual compute cycles used — idle containers cost nothing.
| GPU | Per Second | Per Hour (approx) |
|---|---|---|
| Nvidia B200 | $0.001736 | $6.25 |
| Nvidia H200 | $0.001261 | $4.54 |
| Nvidia H100 | $0.001097 | $3.95 |
| Nvidia RTX PRO 6000 | $0.000842 | $3.03 |
| Nvidia A100 80GB | $0.000694 | $2.50 |
| Nvidia A100 40GB | $0.000583 | $2.10 |
| Nvidia L40S | $0.000542 | $1.95 |
| Nvidia A10 | $0.000306 | $1.10 |
| Nvidia L4 | $0.000222 | $0.80 |
| Nvidia T4 | $0.000164 | $0.59 |
CPU: $0.125 per physical core per hour (1 physical core ≈ 2 vCPU) Memory: $0.00000222 per GiB per second Volumes: $0.09 per GiB per month (1 TiB free)
Plans: Starter $0 + compute (3 seats, 100 containers, $30 free credits), Enterprise for unlimited seats + SSO + HIPAA.
Modal.com with T4 GPU at $0.70/hr, per-second billing, $30 free credits. For a small model that trains in 2 hours: $1.40 (or $0 with free credits).
Other options:
| Condition | Use Daytona | Use Modal |
|---|---|---|
| CPU (1 vCPU, 1hr) | $0.0504 + RAM $0.0162/GiB | per-cycle, idle=free |
| GPU cheapest | H100 $3.95/hr | T4 $0.59/hr |
| GPU best | RTX PRO 6000 $3.03/hr | H100 $3.95/hr, B200 $6.25/hr |
| Free credits | $200 one-time | $30/month |
| Billing | per-second wall-clock | per CPU cycle (idle = $0) |
| Android target | Yes ($0.0504/vCPU/hr) | No |
| Full Docker | Yes | Lightweight container |
| Best for | CI/CD, cross-platform | ML training, burst compute |
Recommendation:
Modal bills per CPU cycle, not wall-clock. Idle time costs $0. Active CPU: $0.125 per physical core per hour.
Example: a 10-hour continuous CPU build using 1 core = ~$1.25 Example: 1 hour T4 GPU training = $0.59
Daytona bills per wall-clock second. Example: 10 hours CPU (1 vCPU + 1GiB RAM) = $0.504 + $0.162 = $0.67
Bottom line: Daytona is predictable per-hour. Modal can be cheaper if your workload has idle periods.
export MODAL_TOKEN_ID=***export MODAL_TOKEN_SECRET=***4. Test: cargo run -p specgen-sandbox -- spin --provider modal --repo `| Error | Cause | Fix |
|---|---|---|
MODAL_TOKEN_ID not set | No credentials | Create token at https://modal.com/settings |
sandbox creation failed | Invalid image or quota | Check image name; verify quota on dashboard |
exec exit code != 0 | Command failed | Read output; may need apt install in sandbox |
out of memory | GPU OOM | Reduce batch size or use larger GPU (A100) |
timeout after 1hr | Modal default limit | Use @app.function(timeout=3600*3) for 3hr |
pip install modal (primary)specgen-sandbox cratehttps://api.modal.com/v1Basic base64(token_id:token_secret)Systematic pre-merge verification workflow for Rust monorepo branches. Check compilation, tests, clippy, grep leftovers, then fix → commit → push → PR. Use when user says "check work done in this branch" or before merging any PR.
Dependency lock and Cargo.lock conflict policy for Rust workspaces. Covers when to update, when to regenerate, manual conflict handling, and merge/sync guidelines to prevent recurring Cargo.lock merge conflicts.
Audit and fix documentation (README, ARCHITECTURE, CHANGELOG) against current codebase while waiting for CI/bot reviews on a PR. Fix factual drift, dead references, and add changelog entries — then commit docs to the same branch.
Daytona sandbox knowledge for answering user questions about remote development environments. Use when the user asks about Daytona, sandboxes, remote execution, cross-platform builds, or needs to run code on cloud infrastructure instead of locally. Provides API details, pricing, troubleshooting, and CLI commands.
Mandatory project quality workflow for the specgen Rust monorepo. Enforces zero-warning policy, formatting standards, and test compliance. Trigger before any commit, pull request, or completion of a coding task. Also trigger when the user says "check quality", "run QA", or "is it ready?".
Comprehensive toolkit for detecting and eliminating "AI slop" - generic, low-quality AI-generated patterns in natural language, code, and design. Use when reviewing or improving content quality, preventing generic AI patterns, cleaning up existing content, or enforcing quality standards in writing, code, or design work. Trigger on first run in any new codebase, after AI-assisted coding sessions, or when user says "clean up", "review quality", "check slop", or "anti-slop".