WebGPU + ONNX provider: Qwen 3 0.6B in-browser (7.4.0)#66

Open

esokullu wants to merge 33 commits into

mainfrom

webgpu

Commits on May 22, 2026

WebGPU + ONNX provider (Qwen 3 0.6B, in-browser, no server needed)

esokullu
and
claude
committed
chore: bump version 7.3.1 → 7.4.0

esokullu
and
claude
committed
Merge branch 'main' into webgpu
esokullu
authored
Update offscreen.html
esokullu
authored
WebGPU: download progress indicator

esokullu
and
claude
committed
Vendor @huggingface/transformers 4.2.0 (WebGPU + ONNX runtime)

esokullu
and
claude
committed
version up
esokullu
committed
WebGPU: fix bare-specifier import (vendor ort.webgpu.bundle, patch

esokullu
and
claude
committed
WebGPU: vendor UNMINIFIED builds (Web Store policy + readability)

esokullu
and
claude
committed
WebGPU: vendor onnxruntime-common, patch second bare specifier

esokullu
and
claude
committed
WebGPU: vendor asyncify WASM variant for CPU fallback ops

esokullu
and
claude
committed
WebGPU: switch default dtype from q4 to q4f16

esokullu
and
claude
committed
WebGPU: document fp16 fallback when q4f16 kernel overflows

esokullu
and
claude
committed
WebGPU: surface fallback-adapter / no-GPU case in Test Connection

esokullu
and
claude
committed
WebGPU: force .jsep wasm variant so WebGPU EP actually engages

esokullu
and
claude
committed
WebGPU: enable cross-origin isolation for SharedArrayBuffer

esokullu
and
claude
committed
WebGPU: address PR #66 codex review (tool-call streaming + cache key)

esokullu
and
claude
committed
WebGPU: disable wasm-cache for chrome-extension scheme

esokullu
and
claude
committed
WebGPU: move inference to a dedicated Worker + upgrade to Qwen 3.5 0.8B

esokullu
and
claude
committed
title change
esokullu
committed
WebGPU: keep outputs on GPU to avoid mapAsync OOM

esokullu
and
claude
committed

Commits on May 26, 2026

Merge branch 'main' into webgpu
esokullu
committed
9.0.0

esokullu
and
claude
committed
9.0.2
esokullu
committed
Fix WebGPU CPU tensor access error in worker pipeline
esokullu
committed
Handle WebGPU OrtRun buffer download/CPU data failures with retry mode
esokullu
committed
Add robust fallback for WebGPU buffer map failures
esokullu
committed
Improve WebGPU fallback errors and switch default ONNX model to Gemma
esokullu
committed
Reset WebGPU mode when WASM kernel init fails
esokullu
committed
Retry WebGPU unaligned-access failures with fp16 dtype
esokullu
committed
Skip WASM fallback for quantized WebGPU models
esokullu
committed
Retry quantized WebGPU map failures with fp16 before aborting
esokullu
committed
improvements
esokullu
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

WebGPU + ONNX provider: Qwen 3 0.6B in-browser (7.4.0)#66

WebGPU + ONNX provider: Qwen 3 0.6B in-browser (7.4.0)#66
esokullu wants to merge 33 commits into
mainfrom
webgpu

Commits on May 22, 2026

WebGPU + ONNX provider (Qwen 3 0.6B, in-browser, no server needed)

chore: bump version 7.3.1 → 7.4.0

Merge branch 'main' into webgpu

Update offscreen.html

WebGPU: download progress indicator

Vendor @huggingface/transformers 4.2.0 (WebGPU + ONNX runtime)

version up

WebGPU: fix bare-specifier import (vendor ort.webgpu.bundle, patch

WebGPU: vendor UNMINIFIED builds (Web Store policy + readability)

WebGPU: vendor onnxruntime-common, patch second bare specifier

WebGPU: vendor asyncify WASM variant for CPU fallback ops

WebGPU: switch default dtype from q4 to q4f16

WebGPU: document fp16 fallback when q4f16 kernel overflows

WebGPU: surface fallback-adapter / no-GPU case in Test Connection

WebGPU: force .jsep wasm variant so WebGPU EP actually engages

WebGPU: enable cross-origin isolation for SharedArrayBuffer

WebGPU: address PR #66 codex review (tool-call streaming + cache key)

WebGPU: disable wasm-cache for chrome-extension scheme

WebGPU: move inference to a dedicated Worker + upgrade to Qwen 3.5 0.8B

title change

WebGPU: keep outputs on GPU to avoid mapAsync OOM

Commits on May 26, 2026

Merge branch 'main' into webgpu

9.0.0

9.0.2

Fix WebGPU CPU tensor access error in worker pipeline

Handle WebGPU OrtRun buffer download/CPU data failures with retry mode

Add robust fallback for WebGPU buffer map failures

Improve WebGPU fallback errors and switch default ONNX model to Gemma

Reset WebGPU mode when WASM kernel init fails

Retry WebGPU unaligned-access failures with fp16 dtype

Skip WASM fallback for quantized WebGPU models

Retry quantized WebGPU map failures with fp16 before aborting

improvements