KV Cache Compaction RL

RL training with mid-generation KV cache compaction using Attention Matching. This enables learning over long effective contexts (8K+ tokens) within fixed KV memory budgets (2K), by compressing the cache between generation segments.

Built on PRIME-RL for async RL and vLLM for inference.

How It Works

The CompactionWorker drives model forward passes inside vLLM's collective_rpc, bypassing the scheduler for full control over the KV cache. Generation proceeds until the KV cache fills a budget (max_kv_len), then:

1. Select top-k keys by attention importance (C1)
2. Solve least-squares for replacement values (C2)
3. Optionally compute NNLS beta bias for partition function correction
4. Inject [prompt | C1/C2 | suffix] back into paged blocks
5. Continue generating until max_total_tokens or EOS

By default, suffix queries (real attention patterns from generated tokens) score key importance, with forced indices passed to the trainer for deterministic replay.

With compute_beta=true, an additive per-key bias corrects the softmax partition function mismatch after compaction. This uses contiguous KV mirrors (BetaState) alongside vLLM's paged cache, with SDPA decode replacing FlashAttention for the bias-corrected path.

Budget Injection

Budget injection tells the model how many tokens remain. Two modes:

• With compaction (inject_budget_message=true): injects a user message after each compaction event
• Without compaction (inject_only=true): injects every inject_budget_every tokens during standard generation (no KV compression)

Injected tokens get completion_mask=0 so they don't contribute to loss. Logprobs are 0.0 at inject positions.

Setup

git clone https://github.com/HyperPotatoNeo/attention-matching-rl.git
cd attention-matching-rl
uv sync --all-extras

Inference

Single-GPU server

uv run inference @ configs/compaction/qwen3_4b_serve_tp1.toml --server.port 8000

Test compaction generation

import requests

resp = requests.post("http://localhost:8000/compact_generate", json={
    "prompt_ids": tokenizer.encode("Solve this problem..."),
    "max_kv_len": 2048,
    "max_total_tokens": 8192,
    "compact_target_ratio": 0.25,
    "compact_window": 1024,
    "n_compacts": 99,
    "compute_beta": True,
    "temperature": 0.6,
})
result = resp.json()
print(result["final_text"])
print(f"Tokens: {result['diagnostics']['total_tokens']}, "
      f"Compactions: {len(result['diagnostics']['compaction_events'])}")

Evaluate on rg-mix-env

# Start 4 TP=1 servers
bash scripts/start_4servers.sh

# Run compaction eval
python scripts/eval_rg_mix.py --mode compaction --n 100 \
    --max-kv-len 2048 --max-total-tokens 8192 \
    --n-compacts 99 --compact-ratio 0.25 --compact-window 1024

# Run baseline eval (no compaction)
python scripts/eval_rg_mix.py --mode baseline --n 100

Training

Training uses 2 nodes with a 4+4 layout:

• Node 1 (4 GPUs): 4 independent TP=1 compaction servers (ports 8000-8003)
• Node 2 (4 GPUs): FSDP2 trainer (4 GPUs) + orchestrator (CPU)

Weight broadcast via Lustre filesystem. Containers use --net=host for cross-node communication.

Launch

sbatch -A m5017 -C "gpu&hbm80g" --qos=premium --time 48:00:00 \
    --gpus-per-node 4 --nodes=2 ~/compaction_suffix_queries.sh

The launch script handles container setup, config resolution (replacing __INFERENCE_NODE__ placeholder), inference server health checks, and trainer startup.

Training parameters

• Full fine-tune (no LoRA), lr=1e-6, AdamW with CPU offload
• KV budget: max_kv_len=2048, compact_window=1024, ratio=0.25, max_total_tokens=8192
• batch_size=256, rollouts_per_example=8, seq_len=9216
• Checkpoints every 50 steps, auto-resume with resume_step = -1

Key Files

Configs

Docs

• Implementation details — algorithm, beta correction, CUDA graphs
• Speed optimizations — batching, graph capture, profiling

Citation

Based on Attention Matching:

@article{zweiger2025attention,
  title={Attention Matching: an Attention Decomposition Framework for Efficient KV Cache Compression},
  author={Zweiger, Adam},
  journal={arXiv preprint arXiv:2602.16284},
  year={2025}
}