Skip to content

README.md

Latest commit

 

History

History
94 lines (72 loc) · 2.98 KB

README.md

File metadata and controls

94 lines (72 loc) · 2.98 KB

Tome

Distributed LLM inference and GRPO training infrastructure for Apple Silicon, featuring Paged KV Cache and Prefix Caching.

Features

  • Optimized for GRPO: Unified rollout-judge-train pipeline with multi-level prefix caching.
  • Paged KV Cache: 128-token block-based memory management with Copy-on-Write (CoW).
  • Prefix Caching: Radix-tree based KV reuse for shared rubrics and prompts.
  • Dual Model Architecture: Policy and Reference models live in unified memory for instant log-prob computation.
  • In-place Weight Updates: Merge LoRA adapters via RPC without reloads.
  • Distributed Scheduler: Rust-based orchestrator with prefix-aware routing.

Installation

uv sync

Running

Start the scheduler, inference node, and TUI in three terminals:

# Terminal 1: Scheduler (Rust HTTP server, routes requests to nodes)
./scripts/start_scheduler.sh

# Terminal 2: MLX inference node (gRPC server, runs Qwen3-0.6B by default)
./scripts/start_node.sh

# Terminal 3: TUI (interactive chat interface)
./scripts/start_tui.sh

GRPO Training iteration (example)

The GRPO API is split into Rollout (generates completions + ref logprobs) and Judge (scores completions):

# 1. Rollout
curl -X POST http://localhost:8080/v1/grpo/rollout \
  -H "Content-Type: application/json" \
  -d '{
    "batch_id": "iter-1",
    "prompts": [{"prompt_id": "p1", "prompt": "Write a poem about tensors."}],
    "group_size": 8,
    "max_tokens": 128
  }'

# 2. Judge
curl -X POST http://localhost:8080/v1/grpo/judge \
  -H "Content-Type: application/json" \
  -d '{
    "batch_id": "iter-1",
    "rubric": "Technical accuracy and rhyme.",
    "items": [{"item_id": "i1", "prompt": "Prompt: Poem about tensors. Completion: ..."}]
  }'

Tests & Benchmarks

# Run End-to-End GRPO pipeline test and benchmark
uv run tests/e2e_grpo.py

# Expected Output (approximate):
# --- BENCHMARK ---
# rollout + ref_lps Throughput: 136.0 tokens/s (aggregate)
# Judge Throughput: 4.3 items/s
uv run python3 mlx-impl/tests/test_kvcache.py
uv run python3 mlx-impl/tests/test_grpo.py

# Benchmark kernel and model performance
uv run python3 mlx-impl/benchmarks/benchmark_kernel.py
uv run python3 mlx-impl/benchmarks/benchmark.py

Project Structure

  • mlx-impl/node.py - gRPC inference node server with GRPO support.
  • mlx-impl/kvcache.py - Paged KV cache with CoW and Prefix Cache.
  • mlx-impl/model.py - Qwen3/Nanbeige model implementation in MLX.
  • mlx-impl/tests/ - Unit tests for MLX components.
  • mlx-impl/benchmarks/ - Performance and comparison benchmarks.
  • scheduler/ - Rust inference scheduler and prefix-aware router.
  • tests/e2e_grpo.py - System-wide end-to-end integration test.
  • scripts/ - Convenience scripts for starting components.
  • tui/ - Ratatui interactive chat interface.

Documentation