WaveCut/Bonsai-web-GGUF

Bonsai Web GGUF

Browser-ready, deterministic shards for the WaveCut/bonsai-webgpu-chat Space. The files are the pinned Prism ML Bonsai Q1_0 GGUF releases, re-sharded with llama-gguf-split --split-max-size 512M so every browser download stays below the current wllama single-file limit.

Contents

Tier	Architecture	Download	Shards	CPU fallback
1_7b	Qwen3 dense	248,302,336 bytes	1	yes
4b	Qwen3 dense	572,270,816 bytes	2	yes
8b	Qwen3 dense	1,158,654,816 bytes	3	yes
27b	Qwen3.5 hybrid	3,803,453,440 bytes	8	no

The repository also carries dflash/Qwen3.6-27B-DFlash-Q8_0.gguf (1,849,481,440 bytes) as a native correctness reference only. It is not part of a tier download and is not loaded by the browser Space.

manifest.json is the machine-readable contract: immutable source provenance, ordered shard paths, bytes, SHA-256, context defaults, WebGPU limits, chat template evidence, and the default-off DFlash research status. SHA256SUMS is the independent shard checksum list.

Source provenance

Tier	Source repository	Pinned revision	Original SHA-256
1_7b	prism-ml/Bonsai-1.7B-gguf	210a9e99f79cb184909d49595906526eb2b3dd9a	3d7c6c90dd98717a203adb22d5eacd2581850e40aa5327e144b97766cae5f7e3
4b	prism-ml/Bonsai-4B-gguf	78f2c2bacd0904ffaba24b4873ed975e5818354a	4524b3f997f0f06444e568d1f26e2efd69effa3218c7ad3047432fb171e42168
8b	prism-ml/Bonsai-8B-gguf	48516770dd04643643e9f9019a2a349cf26c5dbd	284a335aa3fb2ced3b1b01fcb40b08aa783e3b70832767f0dd2e3fdfa134bd54
27b	prism-ml/Bonsai-27B-gguf	0cf7e3d21581b169b4df1de8bf01316000e2fbb7	17ef842e47450caeb8eaa3ebfbbab5d2f2278b62b79be107985fb69a2f819aa0

The shard container checksums differ from each original monolithic file because GGUF splitting adds split metadata. No quantization or tensor conversion is performed.

The first shard of every tier was loaded through llama.cpp's automatic multi-file resolver and compared with its monolithic source under greedy sampling. Token IDs and text match exactly: 64/64 tokens for 1.7B and 27B, and 16/16 tokens for 4B and 8B. The shorter middle-tier checks are load and reassembly guards, not quality evaluations.

Loading with wllama

Pass the ordered URLs from manifest.json to loadModelFromUrl:

await wllama.loadModelFromUrl(model.files.map((file) => file.url), {
  n_ctx: model.defaultContext,
  n_gpu_layers: 99999,
  flash_attn: false,
  useCache: true,
});

The 27B tier is WebGPU-only in the Space. The application must reject it when adapter limits, storage quota, or full-graph placement are insufficient; it must not silently fall back to CPU.

Runtime evidence

The release gate uses pinned llama.cpp 00fa7cb284cbf133fc426733bd64238a3588a33e and Dawn 18eb229ef5f707c1464cc581252e7603c73a3ef0. On Apple M2 Max, all required Q1_0 and Qwen3.5 hybrid WebGPU operation tests passed. Full-graph 1.7B and 27B inference requires the token embedding tensor to remain on WebGPU; Flash Attention stays disabled because quantized K/V combinations are not fully supported.

DFlash

DFlash remains inspection-only and default-off in the browser. The manifest pins the BF16 source and the bundled Q8_0 GGUF reference. Shape, complete tokenizer equality, target taps [2, 17, 32, 47, 62], conversion, and a 32-token temperature-0 native Bonsai-27B parity smoke pass.

The current custom wllama build contains the DFlash llama.cpp architecture, but its TypeScript/worker/glue load path does not yet expose spec_type=draft-dflash. The Q8_0 fc.weight is 139,264,000 bytes, above the 128 MiB baseline WebGPU storage-buffer binding limit. Browser activation, acceptance measurement, and a positive decode-speed result therefore remain hard release gates. The reference must not be presented as a browser-ready drafter.

License

The Bonsai model files are published under Apache License 2.0. The DFlash reference is MIT-licensed. The source license is included as LICENSE and tier/DFlash attribution is preserved under notices/.