thaw-vllm 0.2.1

Fast snapshot/restore for LLM inference. 17x faster cold starts, multi-GPU tensor parallel, KV cache snapshots. Supports vLLM and SGLang.

thaw

Fast snapshot/restore for LLM inference. Sub-second model hot-swap at 55 GB/s, 17x faster cold starts on 70B, multi-GPU tensor parallel, KV cache preservation.

thaw serve hot-swaps a Llama-3-8B-class model in 0.29 seconds at 55 GB/s — PCIe Gen5-saturating, bit-identical output. Cold-start Llama-3-70B on 2x A100 drops from 546s to 31.8s (17.2x). Rust+CUDA pipelined DMA, pinned-memory persistence, and KV cache snapshots that no other tool offers.

Benchmarks

Hot model swap (thaw serve, H100 SXM, Llama-3-8B-Instruct, 16 GB fp16):

Reload #	Time	Throughput	Backend
0 (cold, one-time pin)	6.40s	—	rust_pipelined_pinned_mmap
1	0.29s	55.0 GB/s	rust_pipelined_pinned_mmap
2	0.29s	55.1 GB/s	rust_pipelined_pinned_mmap
3	0.29s	55.1 GB/s	rust_pipelined_pinned_mmap
4	0.29s	55.1 GB/s	rust_pipelined_pinned_mmap

thaw serve pins the snapshot mmap once when a pool slot warms up (~6s for 16 GB — the one-time cudaHostRegister cost), then reuses that pinned buffer on every subsequent swap. Steady-state = pure PCIe Gen5 DMA at 86% of theoretical peak. Bit-identical output verified across reloads. Extrapolates to ~2.5s hot-swap for Llama-70B (140 GB), directly comparable to InferX's "sub-2s" claim. Bench: bench_slot_warm.py, correctness: bench_slot_warm_correctness.py.

Llama-3-70B-Instruct (141 GB fp16) on 2x A100 SXM 80GB — tensor parallel cold start:

Method	Time	Speedup
Normal vLLM cold start	546.5s	1x
thaw restore (TP=2)	31.8s	17.2x
Weight restore only	10.5s	6.74 GB/s per rank

Llama-3-8B-Instruct (16 GB fp16) — single GPU, H100 SXM:

Method	Time	Throughput	Speedup
Normal vLLM cold start	24.8s	—	1x
thaw restore (cold-cache NVMe)	2.6s	14.12 GB/s	9.7x
thaw restore (warm-cache)	2.5s	13.99 GB/s	9.9x
thaw freeze (D2H + NVMe write)	5.6s	2.88 GB/s	—

Cold-cache measurement verified with vmtouch -e (0% resident pages before restore, checked via mincore). fio parallel read on the same file confirms the NVMe ceiling — thaw's Rust reader saturates it. Reproducible across back-to-back runs. See docs/BENCHMARKS.md for methodology.

Freeze is currently slower than restore (2.88 GB/s vs 14 GB/s on the same hardware). Our restore path is chunked/O_DIRECT/write-combining pinned memory; the freeze path still allocates one pinned buffer per region and writes through the page cache. This is a Rust-side pipeline asymmetry, not a hardware limit — the rewrite to mirror restore's architecture is the next perf target. Tracked in plans/2026-04-17_pipelined-freeze-rewrite.md.

A pre-staged RAM path (mmap + cudaHostRegister) is implemented but gated off by default (THAW_ZEROCOPY_MMAP=1 to enable). cudaHostRegister is O(pages) — pinning a 16 GB mmap costs ~7s, which dominates one-shot restore. The path exists for thaw serve, where registration is amortized across many restores. Tracking proper amortization in a follow-up.

Agent fork — clone a running AI session (Llama-3-8B-Instruct, H100 SXM):

Operation	Time	Notes
Weight restore (warm-cache, post-freeze)	1.1s	14.79 GB/s — file was in page cache, see note below
KV cache restore	0.135s	65 blocks, 136 MB — prefill eliminated
Total restore (incl. vLLM init)	7.3s	vs 16s normal cold start
Fork 3 parallel completions	1.6s avg	All share 872-token cached prefix

The 14.79 GB/s weight restore here is a warm-cache measurement (the freeze that ran 5s earlier left the 16 GB file in Linux's page cache). The agent-fork flow is still the differentiator — no other tool restores KV cache at all. The KV restore number, the fork completion numbers, and the "skip prefill" claim all stand on their own regardless of where the weights came from.

All paths produce bit-identical inference output. KV cache restore preserves prefix cache across cold starts — new requests skip prefill entirely.

More GPUs and models

GPU	Model	Normal	thaw	Speedup
2x A100 SXM 80GB	Llama-3-70B (TP=2)	546.5s	31.8s	17.2x
H100 SXM 80GB	Llama-3-8B	24.8s	2.6s	9.7x
RTX PRO 6000 (Blackwell)	Llama-3-8B	28.6s	3.2s	8.9x
RTX A6000	Llama-3-8B	73.2s	5.8s	12.6x

Larger models show bigger speedups because weight loading dominates more of the total cold start time.

How it works

Normal vLLM cold start:
  Download weights → deserialize safetensors → copy to GPU → init KV cache → ready
  [======================================] 24.8s

thaw restore (cold-cache NVMe):
  Dummy init → parallel NVMe read → pipelined DMA to GPU
  [====] 2.6s

Freeze captures all GPU state into binary snapshots — model weights (.thaw) and KV cache blocks (.thawkv).

Restore initializes vLLM with dummy weights (fast — no disk I/O), then overwrites them from the snapshot using double-buffered pipelined DMA through pinned host memory. Two CUDA streams overlap PCIe transfers with disk reads. KV cache blocks are restored separately with their prefix cache hash mappings, so new requests immediately get cache hits.

Three restore modes:

• Disk: reads snapshot from NVMe with O_DIRECT, bypassing the kernel page cache. Throughput limited by NVMe bandwidth — on H100 SXM NVMe this hits 14 GB/s with the Rust pipelined path, saturating the drive.
• Pre-staged RAM: snapshot already in memory (tmpfs, shared memory, or mmapped with page cache warm). The full zero-copy path (mmap + cudaHostRegister) is implemented behind THAW_ZEROCOPY_MMAP=1, but the one-time registration cost makes it a win only when amortized across many restores.
• Slot-warm hot-swap (thaw serve): when a pool slot warms up, thaw serve pins the snapshot mmap once (~6s cudaHostRegister for 16 GB) and persists the pinned handle on the slot. Every subsequent model swap into that slot reuses the pinned buffer and runs as pure PCIe DMA — 0.29s at 55 GB/s for an 8B model on H100 SXM, saturating Gen5 PCIe.

KV cache snapshots capture the prefix-cached blocks that vLLM retains after generation. On restore, block data is DMA'd back to GPU and the prefix cache hash table is reconstructed. Requests with matching prefixes skip prefill — the most expensive part of inference.

Architecture

thaw/
  crates/
    thaw-core/       Rust. File format, region tables, I/O. No CUDA dep.
    thaw-cuda-sys/   Rust. FFI bindings to CUDA runtime (cudaMallocHost,
                     cudaMemcpyAsync, streams). Built via build.rs.
    thaw-runtime/    Rust. Orchestration: freeze/restore pipelines, double-
                     buffered DMA, O_DIRECT, MockCuda for Mac testing.
    thaw-py/         Rust. PyO3 bindings exposing pipelined freeze/restore
                     to Python. Builds a native .so via maturin.
    thaw-cli/        Rust. GPU benchmark binary.
  python/
    thaw_vllm/       Python package (pip install thaw-vllm).
      snapshot.py    Freeze/restore weights, Rust backend fallback.
      kv_snapshot.py KV cache freeze/restore.
      loader.py      vLLM ModelLoader: load_format="thaw".
      pool.py        Engine pool: pre-warmed slots, model hot-swap, OpenAI API.
      server.py      Single-engine OpenAI-compatible API server.
      cli.py         CLI: thaw freeze, thaw serve, thaw info.
    vllm_demo.py     End-to-end benchmark: normal vs thaw cold start.
    kv_cache_demo.py KV cache snapshot/restore demo with correctness test.
  demos/
    agent_fork.py    Agent fork demo: clone session, fork parallel completions.

Testing on Mac, shipping on GPU. The CudaBackend trait abstracts all GPU operations. MockCuda (a HashMap-backed fake) lets 48 runtime tests run on any machine. The cuda feature flag activates real GPU paths only when needed.

Quick start

pip install thaw-vllm[all]

This installs the Python package, FastAPI server, and pre-built Rust+CUDA native extension. No Rust toolchain needed.

Freeze a model, then serve it:

# Llama models are gated — authenticate with HuggingFace first
huggingface-cli login

# Step 1: Freeze model weights to a snapshot
thaw freeze --model meta-llama/Llama-3.1-8B-Instruct --output weights.thaw

# Step 2: Serve with pre-warmed engine pool
thaw serve --model meta-llama/Llama-3.1-8B-Instruct --snapshot weights.thaw

That's it. You now have an OpenAI-compatible API at http://localhost:8000/v1:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "meta-llama/Llama-3.1-8B-Instruct",
       "messages": [{"role": "user", "content": "Hello!"}],
       "max_tokens": 64}'

How thaw serve works

thaw serve is PgBouncer for GPU inference. It keeps vLLM engines pre-initialized with dummy weights, then DMA-swaps real model weights from a snapshot on demand. First swap into a slot pays the one-time cudaHostRegister pin cost (~6s for 16 GB); every subsequent swap runs at 55 GB/s (0.29s for 8B, ~2.5s for 70B) — that's the pinned mmap reused through PCIe Gen5 DMA without ever leaving the slot.

• OpenAI-compatible API — /v1/completions, /v1/chat/completions, streaming via SSE
• Model affinity — requests for an already-loaded model have zero swap cost
• Hot model registration — register new snapshots at runtime via /admin/snapshots
• Pool status — monitor slots, loaded models, and utilization via /admin/pool

# Multi-model pool with 2 warm slots
thaw serve --model meta-llama/Llama-3.1-8B-Instruct \
  --snapshot base.thaw \
  --pool-size 2 \
  --register finetune-v2=/snapshots/v2.thaw

# The model field in each request selects which snapshot to serve
curl localhost:8000/v1/completions -d '{"model": "finetune-v2", "prompt": "..."}'

Python API

import thaw_vllm
from vllm import LLM, SamplingParams

# Freeze: save model weights to a snapshot
llm = LLM(model="meta-llama/Meta-Llama-3-8B", dtype="float16", enforce_eager=True)
thaw_vllm.freeze_model_pipelined(model, "/path/to/weights.thaw")

# Restore: two lines, 9.2x faster cold start
llm = thaw_vllm.load("meta-llama/Meta-Llama-3-8B", "/path/to/weights.thaw")

Or use load_format="thaw" directly with vLLM:

import thaw_vllm  # registers the loader
llm = LLM(model="meta-llama/Meta-Llama-3-8B",
          load_format="thaw",
          model_loader_extra_config={"snapshot": "/path/to/weights.thaw"})

Multi-GPU — tensor parallel with per-rank snapshots:

# Freeze: each GPU saves its shard
llm = LLM(model="meta-llama/Meta-Llama-3-70B-Instruct", tensor_parallel_size=2, ...)
thaw_vllm.freeze_model_tp(llm, "/path/to/weights.thaw")
# Creates: weights.thaw (rank 0), weights.rank1.thaw (rank 1)

# Restore: 17.2x faster than normal cold start
llm = thaw_vllm.load("meta-llama/Meta-Llama-3-70B-Instruct", "/path/to/weights.thaw",
                      tensor_parallel_size=2)

Cloud storage (S3) — load snapshots directly from S3 URIs (install with pip install thaw-vllm[cloud]):

# Freeze once, upload to S3, restore anywhere
llm = thaw_vllm.load("meta-llama/Meta-Llama-3-8B",
                     "s3://my-bucket/llama-3-8b.thaw")

First call downloads to ~/.cache/thaw/snapshots/ (override with THAW_CACHE_DIR); subsequent calls hit the local cache. For TP, per-rank files live at s3://bucket/weights.thaw and s3://bucket/weights.rank1.thaw — thaw derives the per-rank URIs automatically. AWS credentials come from the standard boto3 chain (env vars, ~/.aws/credentials, IAM role).

SGLang — same API, class-passthrough loader (install with pip install thaw-vllm[sglang]):

import sglang
from thaw_sglang import ThawSGLangModelLoader

engine = sglang.Engine(
    model_path="meta-llama/Meta-Llama-3-8B",
    load_format=ThawSGLangModelLoader,
    model_loader_extra_config={"snapshot": "/path/to/weights.thaw"},
    dtype="float16",
)

TP works automatically — each SGLang worker loads its own rank-specific snapshot. Freeze via thaw freeze --engine sglang ... or ThawSGLangFreezeLoader. Note: vLLM and SGLang cannot coexist in one env (torch version conflict) — use separate pods.

Agent fork demo — clone a running AI session, fork parallel completions:

python demos/agent_fork.py --snapshot weights.thaw
python demos/agent_fork.py --snapshot weights.thaw --full-cycle  # destroy + restore

CLI reference

thaw freeze --model meta-llama/Meta-Llama-3-8B --output weights.thaw
thaw serve  --model meta-llama/Meta-Llama-3-8B --snapshot weights.thaw [--pool-size N] [--register NAME=PATH]
thaw info   weights.thaw

Troubleshooting

hf-xet download crash — Some versions of huggingface_hub ship with an hf-xet backend that can crash during large model downloads. If you see RuntimeError: Data processing error: File reconstruction error, set:

export HF_HUB_DISABLE_XET=1

Disk space — pip install thaw-vllm[all] plus a 8B model snapshot needs ~50 GB. Use at least 100 GB container disk on cloud providers.

Gated models — Llama models require HuggingFace authentication. Run huggingface-cli login before freeze/serve.

Building from source (alternative to pre-built wheels)

If you need to build the Rust+CUDA backend yourself (e.g., custom CUDA version):

git clone https://github.com/thaw-ai/thaw.git && cd thaw
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
source "$HOME/.cargo/env"
pip install "maturin[patchelf]" vllm
maturin build --release --features cuda -m crates/thaw-py/Cargo.toml -o /tmp/wheels
pip install /tmp/wheels/*.whl
pip install -e ".[serve]"

Competitive landscape

The model loading space is active. Here's how thaw compares:

Project	Approach	Throughput	Limitations
thaw	Pipelined DMA, pinned memory, O_DIRECT + KV cache snapshot	6.7-14.8 GB/s per GPU	—
fastsafetensors (IBM)	GDS + 4x NVMe RAID0	26.4 GB/s	Requires GDS setup + RAID hardware
NVIDIA Model Streamer	Multi-threaded concurrent streaming	~2 GB/s (single SSD)	NVIDIA-maintained, less flexible
CoreWeave Tensorizer	HTTP/S3 streaming + deserialization	~4.6 GB/s local	Tied to CoreWeave ecosystem
vLLM Sleep Mode	Offload to CPU RAM, reload	0.26-3s	Not a cold start — requires prior warm load
Modal GPU Snapshots	CUDA checkpoint/restore API	~10x reduction	Alpha. Doesn't help with large model weight loading
InferX	GPU runtime snapshotting	Claims 2s for 70B	No public code or benchmarks

thaw's differentiation:

1. KV cache snapshot/restore — nobody else does this. Preserves prefix cache across cold starts, eliminates prefill. Enables agent forking, session migration, warm handoff.
2. Single NVMe performance — most deployments don't have RAID0. thaw already matches or beats multi-threaded alternatives on one drive.
3. No special hardware — no GDS, no RAID, no driver patches. Works on any CUDA 12+ GPU.

See docs/LANDSCAPE.md for detailed analysis.

Roadmap

• Weight snapshot/restore (pure Python path)
• Rust+CUDA pipelined freeze/restore (double-buffered DMA, O_DIRECT)
• RAM-backed restore path (mmap + chunked pinned staging; zero-copy mmap variant gated behind THAW_ZEROCOPY_MMAP for thaw serve)
• PyO3 bindings + vLLM integration shim
• H100 / A6000 / Blackwell benchmarks
• KV cache snapshot/restore — the moat (freeze/restore prefix-cached blocks, verified on Llama-3-8B)
• pip install thaw-vllm + CLI (thaw freeze, thaw serve, thaw info)
• load_format="thaw" — native vLLM ModelLoader integration
• OpenAI-compatible API server (thaw serve)
• Streaming support in API server (SSE, OpenAI-compatible)
• Agent fork demo — clone a running AI session, fork parallel completions from shared KV cache (full-cycle: 14.79 GB/s restore, 0.135s KV restore on H100 SXM)
• Multi-GPU / tensor parallel — 17.2x speedup on Llama-3-70B with 2x A100 (TP=2), bit-exact correctness verified
• Engine pool (thaw serve) — pre-warmed vLLM engines with hot model swapping, OpenAI-compatible API, multi-model serving
• Pre-built native wheels — pip install thaw-vllm[all], no Rust toolchain needed
• SGLang integration — class-passthrough loader, freeze + restore, validated on H100 TP=2 (5.0 GB/s)
• Slot-warm hot-swap — persistent cudaHostRegister per pool slot, 0.29s / 55 GB/s model swap on H100 SXM (thaw serve)
• Cloud snapshot storage (S3) — thaw freeze --output s3://... and thaw serve --snapshot s3://..., validated H100 SXM 2026-04-17 (15 GiB freeze+upload in 5.6s, 229s single-stream S3 download — ranged-GET crate is next)
• Pipelined-freeze parity with restore — rewrite freeze to use chunked WC-pinned buffers + O_DIRECT, target 10-14 GB/s (3-4x current)
• Rust thaw-cloud crate — concurrent ranged GETs for S3 restore at NIC line-rate
• GPUDirect Storage support

Design

Full technical architecture, file format spec, and rationale: DESIGN.md

License

MIT