talkie-1930-13b-it-ONNX

Talkie 1930 13B IT ONNX WebGPU

Unofficial community q4f16/q8 ONNX + WebGPU quantization of lewtun/talkie-1930-13b-it-hf for browser inference with Transformers.js tokenizer/chat-template handling and direct ONNX Runtime WebGPU cached decoding.

• Live browser demo: https://scasella.github.io/talkie-quant-webgpu/
• GitHub runner and export scripts: https://github.com/scasella/talkie-quant-webgpu
• Source model revision: 6311dedf518470856a8503f2080bb4b54fcb3323
• Default browser dtype: q4f16
• Fallback dtype: q8
• Stop token IDs: [65535, 65536]
• Best measured browser speed: 3.60 tok/s reported rolling token latency on a MacBook Pro with M4 Pro and 24 GB unified memory

This model keeps the source tokenizer, chat template, generation config, and Apache-2.0 license metadata. It is not an official Talkie release. The ONNX artifacts validate outside the browser. The fast cached q4f16 path has also passed a Chrome/WebGPU smoke on a MacBook Pro with M4 Pro and 24 GB unified memory using direct ONNX Runtime WebGPU.

Badges And Discovery

The model-card metadata is set so the Hugging Face page badges and filters match what this repo actually ships:

• License: apache-2.0, matching the source model metadata and repo license.
• Language: en, because Talkie targets English text.
• Library: transformers.js, because the browser runner uses Transformers.js for tokenizer/chat-template handling.
• Task: text-generation, because the artifact is a decoder-only chat/text generation model even though the app uses a custom manual loop instead of stock pipeline("text-generation").
• Base model: lewtun/talkie-1930-13b-it-hf, so Hugging Face links and filters this as a quantized derivative of the source model.
• Discovery tags: webgpu, onnx, onnxruntime-web, kv-cache, browser-ai, client-side-inference, edge-ai, apple-silicon, macos, q4f16, and q8.

Quantization At A Glance

The default browser artifact is the fast cached q4f16 ONNX file: about 13.0 GB across 55 smaller external-data chunks. It quantizes query/key attention projections while leaving value projections unquantized, preserving top-1 agreement on the smoke prompt and avoiding a full prompt re-run on every generated token.

The q4f16 files are larger than a theoretical pure 4-bit checkpoint because ONNX stores scales, metadata, and unquantized tensors. The cached files trade first-load size for faster steady per-token decoding.

Performance Journey

This repo started with the browser as the acceptance test. The first validated path was the smaller full-sequence q4f16 graph: it loaded in Chrome/WebGPU and generated real text, but every new token reran the whole accumulated prompt. On the MacBook Pro M4 Pro / 24 GB unified-memory smoke, that measured about 0.61 tok/s.

The current default is the result of a few browser-specific turns:

• KV-cache ONNX export so decode can feed only the newest token after prefill.
• A custom Talkie generation loop because this is not a stock Transformers.js causal-LM architecture.
• Direct onnxruntime-web/webgpu execution for the cached graph, while keeping Transformers.js for tokenizer and chat-template handling.
• Manual input_ids, position_ids, and past_key_values.* feeds, with present.* cache tensors kept on GPU where possible and only logits copied back for CPU sampling.
• Browser graph optimization set to disabled after default optimization hit std::bad_alloc during session creation.
• A page-level Hugging Face fetch limiter. A service-worker-only queue still let too many long model fetches accumulate and hit browser request timeouts.
• Partial top-k sampling and throttled UI updates to keep JavaScript overhead from dominating decode.

The quantization path also had a failed candidate: quantizing all attention projections was faster-looking on paper but failed reference top-5 validation. The published fast q4f16 artifact quantizes query/key projections and leaves the value projection unquantized. That preserved top-1 agreement on the smoke prompt and produced the default onnx/model_kv_fast_q4f16.onnx.

Current browser smoke result: kv-cache / ort-direct generated 16 non-NUL words at about 3.60 tok/s reported rolling token latency and 3.58 tok/s p50 token latency after the default warmup, about 6x the original full-sequence steady token rate. Cold load is still the big caveat. A fetch-concurrency sweep found fetches=4 was the best stable default on this machine: without warmup, that run measured about 530.6s to Ready and 17.4s TTFT. With the tiny cached-graph warmup enabled by default, the same path measured about 472.4s to Ready and 1.1s TTFT in the latest run. The additive model_kv_cold64_q4f16.onnx candidate improved the no-warmup Ready time to about 484.2s with about 19.0s TTFT, but that is only a modest partial win, so it remains an opt-in candidate rather than replacing the default. A later static-compatible compression experiment added gzip companions for the fast q4f16 external-data chunks. That reduced transfer from about 12.97 GB to 8.85 GB. The best compressed cold-start run used compressed=1, warmup=0, and fetches=8, and measured about 316.6s Ready, 17.3s TTFT, and 3.61 tok/s. That improves cold load materially, but it still misses the 2x cold-start target, so it is documented as opt-in.

Files

The config.json includes the Transformers.js external-data chunk map used by the browser loader.

Browser Use

The easiest path is the hosted demo:

https://scasella.github.io/talkie-quant-webgpu/

For local development:

git clone https://github.com/scasella/talkie-quant-webgpu.git
cd talkie-quant-webgpu
npm install
npm run dev

The app defaults to cached q4f16. The smaller full-sequence q4f16 artifact remains available as a fallback.

The app also warms a tiny cached decode path after session creation so the first visible reply token arrives faster. Opt out with ?warmup=0 when benchmarking raw TTFT.

Opt into gzip-compressed external data for the best measured cold-start tradeoff:

https://scasella.github.io/talkie-quant-webgpu/?compressed=1&warmup=0&fetches=8

Opt into the smaller cold-start candidate without changing the default:

https://scasella.github.io/talkie-quant-webgpu/?revision=main&q4file=model_kv_cold64_q4f16.onnx&fetches=4

Transformers.js Notes

Talkie has a custom model_type, so stock pipeline("text-generation") is not used here. The browser runner formats messages with the shipped chat template, tokenizes them, runs an explicit manual generation loop, samples on the CPU, suppresses token 0, and stops on token IDs 65535 or 65536.

Minimal tokenizer and fallback-model loading sketch:

import { AutoModel, AutoTokenizer } from "@huggingface/transformers";

const repo = "scasella91/talkie-1930-13b-it-ONNX";
const tokenizer = await AutoTokenizer.from_pretrained(repo);
const model = await AutoModel.from_pretrained(repo, {
  device: "webgpu",
  dtype: "q4f16"
});

Use the GitHub runner for the complete manual generation loop, including direct ONNX Runtime loading, limited concurrent model fetches, and manual past_key_values.* cache management for the cached files.

Validation

• Hub artifact validation confirms tokenizer/config files, ONNX files, and all q4/q8 external-data chunks, including the additive fast cached q4 artifact.
• The additive fast cached q4f16 artifact validated top-1 against the PyTorch full-sequence wrapper with value projection left unquantized.
• The cached q8 fallback validated against the same reference on the CPU provider and is kept as a browser/WebGPU fallback.
• Chrome on a MacBook Pro with M4 Pro and 24 GB unified memory loaded cached q4f16 with cache=0, opt=disabled, fetches=4, and the default warmup, then generated 16 non-NUL words with kv-cache / ort-direct, about 3.60 tok/s reported rolling latency, and about 3.58 tok/s p50 token latency.
• Cold load remains slow: the default warmup run took about 472.4s to Ready, but TTFT fell to about 1.1s. The best no-warmup fetch-concurrency run was about 530.6s Ready / 17.4s TTFT / 3.69 tok/s.
• The additive model_kv_cold64_q4f16.onnx candidate validated and loaded, but only improved the same smoke to about 484.2s Ready / 19.0s TTFT, so it is documented as opt-in rather than default.
• The gzip external-data opt-in for model_kv_fast_q4f16.onnx validated in the browser at 316.6s Ready / 17.3s TTFT / 3.61 tok/s with compressed=1, warmup=0, fetches=8, cache=0, and opt=disabled. It improves cold load by about 40% versus the original 528.7s Ready baseline, but it still misses the <=265s Ready target, so it is not the default.

Known Limitations

• First load is large and slow because the model is split across external-data chunks.
• Browser cache writes are disabled by default to reduce duplicate large allocations during cold load.
• q4f16 keeps q4 weights but uses float32 runtime tensors in the current browser artifact for WebGPU stability.
• Cached q4f16 is faster after the first token, but first load and first-token latency are still slow.
• The displayed tok/sec is a rolling token-latency rate, not a cold-start average.
• The compressed opt-in reduces network transfer, but the browser still materializes decompressed ONNX external data for WebGPU session creation.
• The older full-sequence artifacts remain slower fallbacks.

Attribution

Talkie was developed by Alec Radford, Nick Levine, and David Duvenaud. This ONNX repo builds on the Hugging Face Transformers-format conversion by lewtun/talkie-1930-13b-it-hf and the original Talkie project at https://github.com/talkie-lm/talkie.