How ai-dynamo/aisimulate works
AISimulate predicts LLM-serving performance and searches deployment configurations offline. The main workflow accepts typed deployment or search configuration, evaluates candidate behavior with the compiled performance engine and runtime data, then emits reports or concrete recommendations. Separate capabilities include deterministic workload replay, legacy deployment-artifact generation, and collection/publication of empirical performance data. The graph keeps these product branches distinct; links among unsampled subsystems are omitted rather than inferred from directory names.
Main entry points: engine.rs Replay engine component · engine.py Compiled engine · replayer.rs Replay orchestration · main.py Prediction and search
Components 25
User workflows
- Engine config
python/aisimulate/src/aisimulate/config/engine.py - Prediction and search
python/aisimulate/src/aisimulate/main.py - Recommendation
python/aisimulate/src/aisimulate/recommend.py - Prediction reports
python/aisimulate/src/aisimulate/output.py
Performance modeling
- Compiled engine
python/aisimulate/src/aisimulate_core/sdk/engine.py - Rust engine step
python/aisimulate/src/aisimulate_core/sdk/rust_engine_step.py - Operator models
python/aisimulate/src/aisimulate_core/sdk/operations/attention.py - Model families
python/aisimulate/src/aisimulate_core/sdk/deepseek_v41.py - Performance database
python/aisimulate/src/aisimulate_core/sdk/perf_database.py - Performance estimators
crates/core/src/perfmodel/fpm/estimator.rs
Workload replay
- Replay entry
crates/core/src/replay/engine.rs - Replay orchestration
crates/core/src/replay/replayer.rs - Replay engine component
crates/core/src/replay/components/engine.rs - Round-robin core
crates/core/src/replay/core/round_robin.rs - Disaggregated replay
crates/core/src/replay/disagg.rs - Replay protocol
crates/core/src/replay/protocol.rs - Scheduler contracts
crates/core/src/engine/generalized/contracts.rs - KV eviction policy
crates/core/src/engine/belady.rs - KV hashing
crates/core/src/engine/common/hashing.rs
Empirical data
- Performance collectors
python/aisimulate/collector/collect.py - Forward-pass collection
python/aisimulate/collector/fpm_forward/runner.py
Deployment generation
- Artifact generator
python/aisimulate/src/aisimulate/generator/rendering/engine.py - Configuration sweeper
python/aisimulate/src/aisimulate/sweeper/afd_engine.py - Compatibility workflows
python/aisimulate/src/aisimulate/legacy_cli/main.py
Other components
- Operator
Connections 24
- Operator → Prediction and search: runs
- Operator → Engine config: configures
- Prediction and search → Prediction reports: writes report
python/aisimulate/src/aisimulate/main.py - Prediction and search → Recommendation: searches candidates
python/aisimulate/src/aisimulate/main.py - Engine config → Compiled engine: configures
- Compiled engine → Rust engine step: executes through
python/aisimulate/src/aisimulate_core/sdk/engine.py - Rust engine step → Performance database: queries
python/aisimulate/src/aisimulate_core/sdk/rust_engine_step.py - Compiled engine → Operator models: uses
python/aisimulate/src/aisimulate_core/sdk/engine.py - Compiled engine → Model families: uses
python/aisimulate/src/aisimulate_core/sdk/engine.py - Operator → Replay entry: replays workload
- Replay entry → Replay orchestration: runs
crates/core/src/replay/engine.rs - Replay orchestration → Replay engine component: dispatches
- Replay orchestration → Round-robin core: uses
- Replay orchestration → Disaggregated replay: supports topology
- Replay engine component → Scheduler contracts: implements
crates/core/src/replay/components/engine.rs - Replay engine component → Replay protocol: accounts requests
crates/core/src/replay/components/engine.rs - Operator → Artifact generator: generates artifacts
- Operator → Compatibility workflows: runs workflows
- Compatibility workflows → Artifact generator: generates configs
python/aisimulate/src/aisimulate/legacy_cli/main.py - Operator → Configuration sweeper: searches configurations
- Performance collectors → Performance database: supplies measurements
- Forward-pass collection → Performance estimators: provides observations
- Replay engine component → KV eviction policy: uses eviction policy
- Replay engine component → KV hashing: uses cache keys