@simulatte/doppler 0.1.0 → 0.1.1

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.
Files changed (2) hide show
  1. package/README.md +21 -16
  2. package/package.json +3 -5
package/README.md CHANGED
@@ -1,6 +1,6 @@
1
1
  # @simulatte/doppler
2
2
 
3
- Browser-native inference engine for local AI workloads.
3
+ Inference and training on raw WebGPU. Pure JS + WGSL.
4
4
 
5
5
  **[Live Demo](https://d4da.com)** · **[npm](https://www.npmjs.com/package/@simulatte/doppler)** · **[simulatte.world](https://simulatte.world)**
6
6
 
@@ -22,19 +22,25 @@ for await (const token of model.generate('Hello, world')) {
22
22
  }
23
23
  ```
24
24
 
25
- That's it. Streaming is the default. See [more examples](#more-examples) below or the full [API contract](docs/doppler-api-contract.md).
25
+ Tokens stream from a native `AsyncGenerator`. See [more examples](#more-examples) below or the full [API contract](docs/doppler-api-contract.md).
26
26
 
27
- ## Features
27
+ ## Why Doppler
28
28
 
29
- - WebGPU-accelerated inference in browser and Node — no WASM bridge
30
- - Native `for await` streaming — not callbacks
31
- - Sharded weight loading via OPFS
32
- - LoRA adapter hot-swap at runtime
33
- - Quantized model support (Q4K, Q8, F16)
34
- - Multi-model with independent instances
35
- - Kernel hot-swap (prefill/decode paths)
36
- - Reproducible benchmark tooling
37
- - Auditable kernel execution tracing
29
+ **JS → WGSL → WebGPU.** One hop to the GPU. No ONNX runtime, no WASM blob, no bridge layer.
30
+
31
+ **`for await` streaming.** Not callbacks. Not a `TextStreamer` class. A loop.
32
+
33
+ **LoRA hot-swap.** Swap adapters at runtime without reloading the base model.
34
+
35
+ **Independent model instances.** Run multiple models concurrently. Each owns its pipeline, buffers, and KV cache.
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.
38
44
 
39
45
  ## Browser Support
40
46
 
@@ -46,12 +52,11 @@ That's it. Streaming is the default. See [more examples](#more-examples) below o
46
52
 
47
53
  ## Evidence
48
54
 
49
- Lower is better, comparing per-phase latency by workload.
50
-
51
55
  ![Phase-latency comparison on one workload across models](benchmarks/vendors/results/compare_1b_multi-workload_favorable_phases.svg)
52
56
 
53
- Snapshot artifact:
54
- - [g3-p064-d064-t0-k1.apple-m3pro.compare.json](benchmarks/vendors/fixtures/g3-p064-d064-t0-k1.apple-m3pro.compare.json)
57
+ Snapshot artifacts:
58
+ - [g3-1b-p064-d064-t0-k1.compare.json](benchmarks/vendors/fixtures/g3-1b-p064-d064-t0-k1.compare.json)
59
+ - [lfm2-5-1-2b-p064-d064-t0-k1.compare.json](benchmarks/vendors/fixtures/lfm2-5-1-2b-p064-d064-t0-k1.compare.json)
55
60
 
56
61
  ## More Examples
57
62
 
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@simulatte/doppler",
3
- "version": "0.1.0",
3
+ "version": "0.1.1",
4
4
  "description": "Browser-native WebGPU inference engine for local intent and inference loops",
5
5
  "main": "src/index.js",
6
6
  "types": "src/index.d.ts",
@@ -134,11 +134,9 @@
134
134
  "tools/convert-safetensors-node.js"
135
135
  ],
136
136
  "devDependencies": {
137
+ "@huggingface/transformers": "^3.8.1",
137
138
  "jest": "^30.2.0",
139
+ "onnxruntime-web": "^1.24.1",
138
140
  "playwright": "^1.58.2"
139
- },
140
- "dependencies": {
141
- "@huggingface/transformers": "^3.8.1",
142
- "onnxruntime-web": "^1.24.1"
143
141
  }
144
142
  }