@simulatte/doppler 0.1.1 → 0.1.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +19 -20
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -10,7 +10,7 @@ Inference and training on raw WebGPU. Pure JS + WGSL.
|
|
|
10
10
|
npm install @simulatte/doppler
|
|
11
11
|
```
|
|
12
12
|
|
|
13
|
-
## Quick
|
|
13
|
+
## Quick start
|
|
14
14
|
|
|
15
15
|
```js
|
|
16
16
|
import { doppler } from '@simulatte/doppler';
|
|
@@ -26,30 +26,14 @@ Tokens stream from a native `AsyncGenerator`. See [more examples](#more-examples
|
|
|
26
26
|
|
|
27
27
|
## Why Doppler
|
|
28
28
|
|
|
29
|
-
**JS → WGSL → WebGPU.**
|
|
29
|
+
**JS → WGSL → WebGPU.** Direct JavaScript orchestration into native WebGPU kernels, avoiding ONNX runtimes, WASM blobs, and bridge layers.
|
|
30
30
|
|
|
31
|
-
**`for await` streaming.**
|
|
31
|
+
**`for await` streaming.** Generation uses a native `AsyncGenerator` that fits normal app control flow.
|
|
32
32
|
|
|
33
33
|
**LoRA hot-swap.** Swap adapters at runtime without reloading the base model.
|
|
34
34
|
|
|
35
35
|
**Independent model instances.** Run multiple models concurrently. Each owns its pipeline, buffers, and KV cache.
|
|
36
36
|
|
|
37
|
-
## Under the Hood
|
|
38
|
-
|
|
39
|
-
- Sharded weight loading via OPFS. Gigabytes into VRAM without blocking the main thread.
|
|
40
|
-
- Quantized inference: Q4K, Q8, F16. Real models on consumer GPUs.
|
|
41
|
-
- Kernel hot-swap between prefill and decode paths.
|
|
42
|
-
- Config-driven runtime. Presets, kernel path selection, and sampling are policy, not code.
|
|
43
|
-
- Reproducible benchmarks with deterministic knobs and auditable kernel traces.
|
|
44
|
-
|
|
45
|
-
## Browser Support
|
|
46
|
-
|
|
47
|
-
- Chrome / Edge 113+ (WebGPU required)
|
|
48
|
-
- Firefox (behind flag, WebGPU support varies)
|
|
49
|
-
- Safari (WebGPU support in progress)
|
|
50
|
-
|
|
51
|
-
---
|
|
52
|
-
|
|
53
37
|
## Evidence
|
|
54
38
|
|
|
55
39
|
|
|
@@ -58,7 +42,15 @@ Snapshot artifacts:
|
|
|
58
42
|
- [g3-1b-p064-d064-t0-k1.compare.json](benchmarks/vendors/fixtures/g3-1b-p064-d064-t0-k1.compare.json)
|
|
59
43
|
- [lfm2-5-1-2b-p064-d064-t0-k1.compare.json](benchmarks/vendors/fixtures/lfm2-5-1-2b-p064-d064-t0-k1.compare.json)
|
|
60
44
|
|
|
61
|
-
##
|
|
45
|
+
## Under the hood
|
|
46
|
+
|
|
47
|
+
- Sharded weight loading via OPFS moves multi-GB weights into VRAM without blocking the main thread.
|
|
48
|
+
- Quantized inference paths (Q4K, Q8, F16) support practical model sizes on consumer GPUs.
|
|
49
|
+
- Kernel hot-swap between prefill and decode paths.
|
|
50
|
+
- Config-driven runtime keeps presets, kernel-path selection, and sampling explicit.
|
|
51
|
+
- Reproducible benchmarks expose deterministic knobs and auditable kernel traces.
|
|
52
|
+
|
|
53
|
+
## More examples
|
|
62
54
|
|
|
63
55
|
```js
|
|
64
56
|
// Non-streaming
|
|
@@ -85,6 +77,13 @@ for await (const token of doppler('Hello', { model: 'gemma-3-1b' })) {
|
|
|
85
77
|
- Runtime config contract: [docs/config.md](docs/config.md)
|
|
86
78
|
- Architecture: [docs/architecture.md](docs/architecture.md)
|
|
87
79
|
|
|
80
|
+
## Environment requirements
|
|
81
|
+
|
|
82
|
+
- WebGPU-capable browser runtime is required.
|
|
83
|
+
- Chrome / Edge 113+ supported.
|
|
84
|
+
- Firefox support varies (typically behind a flag).
|
|
85
|
+
- Safari support is evolving.
|
|
86
|
+
|
|
88
87
|
## License
|
|
89
88
|
|
|
90
89
|
Apache License 2.0 (`Apache-2.0`). See [LICENSE](LICENSE) and [NOTICE](NOTICE).
|