How semianalysisai/inferencex works
InferenceX is a multi-project inference research platform: its end-to-end tooling launches serving benchmarks, gathers and analyzes results, and supports evaluations; CollectiveX benchmarks communication, while OperatorX benchmarks individual operators across hardware backends. The graph preserves those distinct product branches and their principal stages. Some subsystems are represented by their source directories because their implementation was not sampled; those branches are documented capabilities, not claims that their internal calls were inspected.
Main entry points: context.py Launch context · base.py Launch backend · main.py Operator runner · __init__.py Cluster records
Components 25
Benchmark orchestration
- Configuration
inferencex-e2e/infx/config.py - Benchmark matrix
inferencex-e2e/infx/matrix - Cluster records
inferencex-e2e/infx/clusters/__init__.py - Launch request
inferencex-e2e/infx/launch/request.py - Launch policy
inferencex-e2e/infx/launch/policy.py - Launch context
inferencex-e2e/infx/launch/context.py - Launch backend
inferencex-e2e/infx/launch/backends/base.py - Cluster scheduler
inferencex-e2e/infx/clusters/slurm.py
Serving and analysis
- Serving benchmark
inferencex-e2e/infx/bench_serving/benchmark_serving.py - Inference server
- Result collection
inferencex-e2e/infx/results - Agentic metrics
inferencex-e2e/infx/results/agentic/backends - Result analysis
inferencex-e2e/infx/results
Evaluation workflows
- Evaluations
inferencex-e2e/infx/evals - Benchmark workflows
inferencex-e2e/infx/workflows
CollectiveX
- Collective sweep matrix
collectivex/sweep_matrix.py - Collective runtime
collectivex/runtime - Bandwidth summary
collectivex/summarize.py
OperatorX
- Operator runner
operatorx/main.py - Runtime snapshot
operatorx/runtime.py - Operation registry
operatorx/core/op_registry.py - Operation model
operatorx/core/op.py - Hardware backends
operatorx/runners - Run results
operatorx/core/result.py
Other components
- Researcher
Connections 29
- Researcher → Configuration: configures
README.md - Configuration → Benchmark matrix: configures
- Benchmark matrix → Cluster records: reads topology
inferencex-e2e/infx/clusters/__init__.py - Launch context → Cluster records: uses
inferencex-e2e/infx/launch/context.py - Launch context → Launch backend: uses
inferencex-e2e/infx/launch/context.py - Launch context → Launch request: uses
inferencex-e2e/infx/launch/context.py - Launch context → Launch policy: uses
inferencex-e2e/infx/launch/context.py - Launch backend → Cluster records: uses
inferencex-e2e/infx/launch/backends/base.py - Launch backend → Launch request: uses
inferencex-e2e/infx/launch/backends/base.py - Launch backend → Cluster scheduler: dispatches
- Launch policy → Cluster records: reads
inferencex-e2e/infx/launch/policy.py - Launch policy → Launch request: uses
inferencex-e2e/infx/launch/policy.py - Launch context → Serving benchmark: launches
- Serving benchmark → Inference server: sends requests
- Inference server → Result collection: produces metrics
- Result collection → Agentic metrics: adapts metrics
inferencex-e2e/infx/results/agentic/backends/vllm.py - Result collection → Result analysis: feeds analysis
- Benchmark workflows → Benchmark matrix: coordinates runs
inferencex-e2e/infx/workflows/recover_failed_ingest.py - Benchmark workflows → Evaluations: coordinates
- Researcher → Collective sweep matrix: selects sweep
- Collective sweep matrix → Collective runtime: creates shards
- Collective runtime → Bandwidth summary: summarizes
- Bandwidth summary → Collective runtime: imports bandwidth
- Researcher → Operator runner: runs benchmarks
README.md - Operator runner → Runtime snapshot: snapshots runtime
operatorx/main.py - Operator runner → Hardware backends: runs operations
operatorx/main.py - Operator runner → Run results: records results
- Hardware backends → Operation model: executes
operatorx/runners/amd/backends/torch.py - Operation registry → Operation model: registers
operatorx/core/op_registry.py