doppler-gpu 0.4.11 → 0.4.15

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 (75) hide show
  1. package/CHANGELOG.md +70 -0
  2. package/README.md +79 -45
  3. package/models/catalog.json +29 -29
  4. package/package.json +12 -3
  5. package/src/client/doppler-registry.json +9 -9
  6. package/src/client/runtime/lora.js +5 -0
  7. package/src/client/runtime/model-session.d.ts +69 -1
  8. package/src/client/runtime/model-session.js +231 -0
  9. package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +3 -2
  10. package/src/config/kernels/kernel-ref-digests.js +2 -1
  11. package/src/config/kernels/registry.json +21 -2
  12. package/src/config/schema/adapter.schema.d.ts +6 -1
  13. package/src/config/schema/adapter.schema.js +5 -0
  14. package/src/experimental/adapters/artifact-contract.d.ts +83 -0
  15. package/src/experimental/adapters/artifact-contract.js +254 -0
  16. package/src/experimental/adapters/index.d.ts +25 -0
  17. package/src/experimental/adapters/index.js +12 -0
  18. package/src/experimental/adapters/lora-loader.js +9 -4
  19. package/src/experimental/training/artifacts.d.ts +160 -0
  20. package/src/experimental/training/attention-backward.d.ts +64 -0
  21. package/src/experimental/training/attention-forward.d.ts +22 -0
  22. package/src/experimental/training/attention-forward.js +82 -0
  23. package/src/experimental/training/autograd.d.ts +76 -0
  24. package/src/experimental/training/checkpoint.d.ts +36 -0
  25. package/src/experimental/training/clip.d.ts +9 -0
  26. package/src/experimental/training/dataloader.d.ts +8 -0
  27. package/src/experimental/training/distillation/student-fixture.js +1 -0
  28. package/src/experimental/training/export.d.ts +41 -0
  29. package/src/experimental/training/lora-pipeline.d.ts +186 -0
  30. package/src/experimental/training/lora-pipeline.js +5 -0
  31. package/src/experimental/training/lora.d.ts +20 -0
  32. package/src/experimental/training/lora.js +4 -3
  33. package/src/experimental/training/loss-scaling.d.ts +21 -0
  34. package/src/experimental/training/loss.d.ts +10 -0
  35. package/src/experimental/training/native-qwen-lora.d.ts +106 -0
  36. package/src/experimental/training/native-qwen-lora.js +449 -0
  37. package/src/experimental/training/objectives/base.d.ts +58 -0
  38. package/src/experimental/training/optimizer.d.ts +22 -0
  39. package/src/experimental/training/optimizer.js +1 -0
  40. package/src/experimental/training/runner.d.ts +249 -0
  41. package/src/experimental/training/trainer.d.ts +89 -0
  42. package/src/experimental/training/workloads.d.ts +205 -0
  43. package/src/gpu/kernels/backward/adam.d.ts +1 -0
  44. package/src/gpu/kernels/backward/adam.js +6 -2
  45. package/src/gpu/kernels/backward/adam.wgsl +3 -2
  46. package/src/gpu/kernels/backward/matmul_backward_q4k.wgsl +90 -0
  47. package/src/gpu/kernels/backward/utils.js +18 -4
  48. package/src/inference/pipelines/text/ffn/dense.js +16 -0
  49. package/src/inference/pipelines/text/generator.d.ts +5 -1
  50. package/src/inference/pipelines/text/generator.js +106 -0
  51. package/src/inference/pipelines/text/linear-attention.d.ts +2 -0
  52. package/src/inference/pipelines/text/linear-attention.js +77 -6
  53. package/src/inference/pipelines/text/lora-types.d.ts +6 -1
  54. package/src/inference/pipelines/text/lora-types.js +5 -0
  55. package/src/inference/pipelines/text/ops.js +1 -0
  56. package/src/inference/pipelines/text/types.d.ts +28 -0
  57. package/src/inference/pipelines/text.d.ts +7 -2
  58. package/src/inference/pipelines/text.js +4 -0
  59. package/src/models/diffusiongemma.js +1 -1
  60. package/src/models/embeddinggemma.js +1 -1
  61. package/src/models/gemma3.js +1 -1
  62. package/src/models/gemma4.js +1 -1
  63. package/src/models/qwen3.js +1 -1
  64. package/src/tooling/hf-registry-utils.js +1 -1
  65. package/src/tooling/node-browser-command-runner.d.ts +16 -0
  66. package/src/tooling/node-browser-command-runner.js +1 -1
  67. package/src/training-browser.d.ts +19 -0
  68. package/src/training-browser.js +13 -0
  69. package/src/training.d.ts +137 -0
  70. package/src/training.js +219 -0
  71. package/src/version.js +1 -1
  72. package/tests/kernels/browser/kernel-suite.js +9 -1
  73. package/tests/kernels/browser/test-page.js +4 -2
  74. package/tests/kernels/reference/adam.js +4 -2
  75. package/NOTICE +0 -5
package/CHANGELOG.md CHANGED
@@ -8,6 +8,76 @@ docs so the `0.1.x` line has one conventional npm-visible history surface.
8
8
 
9
9
  ## [Unreleased]
10
10
 
11
+ ## [0.4.15] - 2026-07-24
12
+
13
+ ### Added
14
+
15
+ - Added `generateWithEvidence()` to loaded model handles. The browser-safe
16
+ result binds generated token IDs, transcript, resolved generation config,
17
+ runtime profile, WebGPU backend identity, and execution-plan identity with
18
+ canonical SHA-256 hashes.
19
+
20
+ ## [0.4.14] - 2026-07-19
21
+
22
+ ### Added
23
+
24
+ - Added native browser, Node, and Bun LoRA training for the final Qwen 3.5
25
+ 0.8B `down_proj` over frozen row-wise Q4K base weights. The bounded native
26
+ path uses completion-masked causal-LM loss, AdamW, deterministic
27
+ checkpoint/resume, and runtime-loadable safetensors/RDRR export.
28
+ - Added Q4K frozen-weight input gradients for the native LoRA suffix and a
29
+ production inference capture contract for the final feed-forward activation.
30
+
31
+ ### Fixed
32
+
33
+ - Replaced the Qwen capability gate that incorrectly treated frozen Q4K base
34
+ weights as preventing native LoRA optimization. Other Qwen targets remain
35
+ unavailable on the native backend until their full backward graphs exist.
36
+ - Added explicit provider teardown to the validated Bun operator sequence;
37
+ model unload alone does not release Dawn's host-owned callback pump.
38
+
39
+ ### Release status
40
+
41
+ - `0.4.13` was tagged but never published. Its tarball is an invalid release
42
+ candidate because it rejects the supported native Qwen target. `0.4.14`
43
+ supersedes it; npm `latest` must not be updated with the `0.4.13` tarball.
44
+
45
+ ## [0.4.13] - 2026-07-19
46
+
47
+ ### Added
48
+
49
+ - Added the public `doppler-gpu/training` npm entry point for capability-gated
50
+ SFT/LoRA workloads, WebGPU autograd and optimization primitives, checkpoint
51
+ evaluation, and runtime-compatible safetensors adapter export.
52
+ - Added explicit `webgpu_native` and `external` training-backend selection.
53
+ Packed Q4K Gemma and Qwen workloads fail closed on the native backend and
54
+ require an injected or workload-declared external trainer.
55
+
56
+ ### Changed
57
+
58
+ - Added installed-package import, type, and capability smokes for the training
59
+ entry point so repository-only deep imports cannot masquerade as npm support.
60
+
61
+ ## [0.4.12] - 2026-07-19
62
+
63
+ ### Added
64
+
65
+ - Added LoRA activation for Qwen 3.5 linear-attention projections, including
66
+ `in_proj_a`, `in_proj_b`, `in_proj_qkv`, `in_proj_z`, and `out_proj` PEFT
67
+ targets, with deterministic intermediate-buffer cleanup.
68
+ - Added immutable adapter-custody catalog validation for preserved WGSL Repair
69
+ artifacts and their exact base-model, tokenizer, conversion, and source
70
+ lineage.
71
+
72
+ ### Changed
73
+
74
+ - Corrected the Qwen 3.5 0.8B chat contract to use the Qwen template with
75
+ thinking disabled, and published the refreshed complete RDRR artifact at an
76
+ immutable hosted revision.
77
+ - Reconciled WGSL Writer V3 campaign records with explicit generation budgets
78
+ and finite confirmation metrics; this does not promote general shader
79
+ authorship.
80
+
11
81
  ## [0.4.11] - 2026-07-19
12
82
 
13
83
  ### Added
package/README.md CHANGED
@@ -1,3 +1,7 @@
1
+ <p align="center">
2
+ <img src="assets/doppler-logo.svg" alt="doppler" width="190" />
3
+ </p>
4
+
1
5
  # doppler-gpu
2
6
 
3
7
  [![Build](https://img.shields.io/github/actions/workflow/status/clocksmith/doppler/check-green.yml?branch=main&label=build)](https://github.com/clocksmith/doppler/actions/workflows/check-green.yml)
@@ -6,15 +10,16 @@
6
10
  [![PRs welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](https://github.com/clocksmith/doppler/pulls)
7
11
 
8
12
  Doppler turns inference engineering into a verifiable search problem. Optimizers
9
- edit RDRR manifests, execution plans, and kernels; parity and benchmark gates
13
+ edit JSON manifests, execution plans, and WGSL kernels; parity and benchmark gates
10
14
  accept or reject each candidate against correctness and speed.
11
15
 
12
- Today engineers use these contracts and gates to tune a JavaScript/WGSL WebGPU
16
+ These contracts and gates are used by operators to tune a configurable and pure JavaScript and WGSL sourced WebGPU
13
17
  runtime for supported
14
18
  [RDRR artifacts](https://github.com/clocksmith/doppler/blob/main/docs/rdrr-format.md)
15
- in browser and Node. Doppler runs text generation, embeddings, and reranking
16
- locally, with CLI and OpenAI-compatible server entry points. Bun WebGPU support
17
- is experimental.
19
+ in browser, Node, and Bun. Doppler runs text generation, embeddings, and reranking
20
+ locally, with CLI and OpenAI-compatible server entry points.
21
+
22
+ **Read this first:** [getting started](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md) → [Root API](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md) → [open source evidence](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md)
18
23
 
19
24
  **[Try the live demo](https://d4da.com/doppler)** | **[npm](https://www.npmjs.com/package/doppler-gpu)** | **[docs](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)**
20
25
 
@@ -30,16 +35,16 @@ Doppler classifies artifacts by what they consume and produce. This is
30
35
  separate from lineage (`family`), runtime implementation (`modelType`), and
31
36
  artifact-size tier.
32
37
 
33
- | Type | Input → output | Runtime-verified / cataloged | Representative lanes |
34
- | --- | --- | --- | --- |
35
- | Text generators | text → text | 12 / 14 | gemma-3-1b-it-q4k-ehf16-af32<br>gemma-3-270m-it-f16-af32<br>gemma-3-270m-it-q4k-ehf16-af32<br>+11 more |
36
- | Multimodal generators | audio + image + text → text | 3 / 3 | gemma-4-e2b-it-q4k-ehf16-af16-int4ple<br>gemma-4-e2b-it-q4k-ehf16-af32<br>gemma-4-e2b-it-q4k-ehf16-af32-int4ple |
37
- | Diffusion language models | text → text | 0 / 1 | diffusiongemma-26b-a4b-it-q4k-ehf16-af16 |
38
- | Translation specialists | text → text | 2 / 2 | translategemma-4b-1b-enes-q4k-ehf16-af32<br>translategemma-4b-it-q4k-ehf16-af32 |
39
- | Language embedders | text → pooled-embedding | 2 / 2 | google-embeddinggemma-300m-q4k-ehf16-af32<br>qwen-3-embedding-0-6b-q4k-ehf16-af32 |
40
- | Rerankers | text-pair → relevance-score | 2 / 2 | qwen-3-reranker-0-6b-f16-af32<br>qwen-3-reranker-0-6b-q4k-ehf16-af32 |
41
- | Protein encoders | protein-sequence → pooled-embedding + token-embedding + token-logits | 3 / 3 | amplify-120m-f16-af32<br>esm2-t12-35m-ur50d-f32-af32<br>esmc-300m-f32-af32 |
42
- | Nucleotide encoders | dna-sequence → pooled-embedding + token-embedding | 1 / 1 | nucleotide-transformer-v2-50m-f32-af32 |
38
+ | Type | Verified / Cataloged | What it does |
39
+ | --- | --- | --- |
40
+ | Text generators | 12 / 14 | `text → text` |
41
+ | Multimodal generators | 3 / 3 | `audio + image + text → text` |
42
+ | Diffusion language models | 0 / 1 | `text → text` |
43
+ | Translation specialists | 2 / 2 | `text → text` |
44
+ | Language embedders | 2 / 2 | `text → embedding` |
45
+ | Rerankers | 2 / 2 | `text-pair → relevance-score` |
46
+ | Protein encoders | 3 / 3 | `sequence → embedding + token outputs` |
47
+ | Nucleotide encoders | 1 / 1 | `dna-seq → embedding` |
43
48
 
44
49
  The [full model-support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md)
45
50
  lists every lane and its lifecycle evidence. Classification says what an
@@ -49,10 +54,29 @@ modality.
49
54
 
50
55
  <!-- model-type-clusters:end -->
51
56
 
57
+ ## LoRA: runtime loading and training
58
+
59
+ Doppler supports:
60
+
61
+ - SafeTensors LoRA loading and hot-swap at runtime.
62
+ - Experimental SFT/LoRA training in Node, Bun, and browser via `doppler-gpu/training`.
63
+ - Native packed-Q4K LoRA for `qwen-3-5-0-8b-q4k-ehaf16` in `webgpu_native` (frozen base, LoRA A/B updates only).
64
+ - External backends for other packed-Q4K training targets.
65
+
66
+ ```bash
67
+ npx doppler-gpu lora --config ./workload.json --surface node
68
+ ```
69
+
70
+ Active adapters are registered in [models/adapters/catalog.json](models/adapters/catalog.json).
71
+ The training API, workload schema, and adapter families live in the
72
+ [LoRA Format specification](docs/lora-format.md),
73
+ [Training handbook](docs/training-handbook.md), and [Training API](docs/api/training.md).
74
+
52
75
  ## Evidence
53
76
 
54
- Doppler has higher steady-state inference throughput than Transformers.js in each
55
- of the comparable browser WebGPU results indexed below.
77
+ Doppler has higher steady-state inference throughput than Transformers.js in
78
+ accepted, throughput-comparable browser WebGPU comparisons indexed below where
79
+ the throughput gate passes.
56
80
 
57
81
  ![Metal and Vulkan browser WebGPU throughput distributions](https://raw.githubusercontent.com/clocksmith/doppler/main/assets/doppler-webgpu-evidence.svg)
58
82
 
@@ -65,33 +89,29 @@ receipt, and the [methodology](https://github.com/clocksmith/doppler/blob/main/d
65
89
 
66
90
  ## Why these lanes are faster
67
91
 
68
- Doppler authors the WGSL path and pins it in RDRR. Runtime profiles pin session
69
- cadence. The measured wins come from different mechanisms in different phases:
92
+ Doppler speedups come from specific, explicit runtime changes:
70
93
 
71
- | Lever | Affected phase | Measured receipt |
72
- | --- | --- | --- |
73
- | Fused Q4_K projection and FFN work removes separate dispatches | Text decode | [Qwen 3.5 0.8B Metal](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260709T154633.json) |
74
- | Fixed `head_dim=128` attention avoids the generic prefill path | Retrieval attention | [embedding](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/embedding_compare_qwen-3-embedding-0-6b-q4k-ehf16-af32_20260709T180853.json), [reranking](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/rerank_compare_qwen-3-reranker-0-6b-q4k-ehf16-af32_20260709T192830.json) |
75
- | Batch-four decode amortizes one readback across four tokens | Submit and map waits | [Qwen 3.5 2B Vulkan](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260707T161623.json) |
76
- | INT4-PLE and Q4_K layouts lower projection traffic | Text decode | [Gemma 4 Vulkan](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260707T170557.json) |
94
+ - **Fusion:** projection + FFN path reductions reduce kernel count.
95
+ - **Attention/layout tuning:** lowers prefill overhead.
96
+ - **Batched decode:** amortizes readback overhead.
97
+ - **Packed kernels:** Q4_K + INT4-PLE reduce projection traffic.
77
98
 
78
- The Gemma 4 comparison uses its declared product-format output policy. It does
79
- not claim exact greedy token parity.
99
+ Evidence receipts:
80
100
 
81
- Parity and comparable benchmark gates decide what is retained. The
82
- [challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md)
83
- records accepted and rejected trials, and each verdict's receipt binds it to a
84
- model, profile, workload, and device; the
85
- [negative-results findings](https://github.com/clocksmith/doppler/blob/main/docs/developer-guides/16-kernel-performance-optimization.md#what-we-ruled-out)
86
- keep the dead ends.
101
+ - [Qwen 3.5 0.8B (Metal)](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260709T154633.json)
102
+ - [Qwen embedding + rerank](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/embedding_compare_qwen-3-embedding-0-6b-q4k-ehf16-af32_20260709T180853.json), [Qwen rerank](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/rerank_compare_qwen-3-reranker-0-6b-q4k-ehf16-af32_20260709T192830.json)
103
+ - [Qwen 3.5 2B (Vulkan)](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260707T161623.json)
104
+ - [Gemma 4 (Vulkan)](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260707T170557.json)
105
+
106
+ Retention is gated: a lane moves forward only when both parity and benchmark
107
+ receipts pass in the
108
+ [challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md).
87
109
 
88
110
  ### Long-term direction
89
111
 
90
- Humans run this loop today. The WGSL-distillation workload is experimental.
91
- Automated kernel generation and autotuning are directions, not shipped product
92
- paths. Doppler aims to reduce the manual steps between a new checkpoint or GPU
93
- and a verified runtime lane. Automated proposals will face the same parity and
94
- benchmark gates.
112
+ Humans run this loop today. WGSL-distillation is experimental.
113
+ Automation in kernel generation/autotuning is a roadmap item, not shipped yet.
114
+ New checkpoints and GPUs still go through the same parity and benchmark gate flow.
95
115
 
96
116
  ## How it works
97
117
 
@@ -197,8 +217,22 @@ for await (const token of model.generate('Describe WebGPU briefly')) {
197
217
 
198
218
  // One-shot
199
219
  const text = await model.generateText('Explain WebGPU in one sentence');
220
+
221
+ // One-shot with a hash-bound execution receipt
222
+ const evidence = await model.generateWithEvidence('Explain WebGPU in one sentence', {
223
+ maxTokens: 64,
224
+ temperature: 0,
225
+ });
226
+ console.log(evidence.outputText, evidence.transcriptHash);
227
+
228
+ await model.unload();
200
229
  ```
201
230
 
231
+ `generateWithEvidence()` returns the text and token IDs together with canonical
232
+ SHA-256 hashes for the transcript, resolved generation config, runtime profile,
233
+ WebGPU backend, and execution-plan identity. The receipt records what ran; it
234
+ does not by itself establish output correctness.
235
+
202
236
  ### OpenAI-compatible server
203
237
 
204
238
  ```bash
@@ -216,16 +250,16 @@ const response = await client.chat.completions.create({
216
250
  });
217
251
  ```
218
252
 
219
- Registry IDs resolve to hosted RDRR artifacts from `Clocksmith/rdrr` by default. See the [Root API guide](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md).
253
+ Registry IDs resolve to hosted RDRR artifacts from `clocksmith/rdrr` by default. See the [Root API guide](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md).
220
254
 
221
255
  ## Start here
222
256
 
223
- | Reader | Entry points |
224
- | --- | --- |
225
- | Application developers | [Getting started](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md), [Root API](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md), and the [OpenAI-compatible server](#openai-compatible-server) |
226
- | Model integrators | [RDRR format](https://github.com/clocksmith/doppler/blob/main/docs/rdrr-format.md), [model support](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md), and [Program Bundles](https://github.com/clocksmith/doppler/blob/main/docs/integration/program-bundle.md) |
227
- | Runtime and kernel engineers | [Architecture](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md), [kernel optimization](https://github.com/clocksmith/doppler/blob/main/docs/developer-guides/16-kernel-performance-optimization.md), and the [challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md) |
228
- | Evidence reviewers | [Competition scoreboard](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md), [benchmark methodology](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md), and [release matrix](https://github.com/clocksmith/doppler/blob/main/docs/release-matrix.md) |
257
+ ### Pick a path
258
+
259
+ - **Application builders:** start with [getting started](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md), [Root API](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md), and the [OpenAI-compatible server](#openai-compatible-server).
260
+ - **Integrators:** use the [RDRR format](https://github.com/clocksmith/doppler/blob/main/docs/rdrr-format.md), [support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md), and [Program Bundles](https://github.com/clocksmith/doppler/blob/main/docs/integration/program-bundle.md).
261
+ - **Kernel/inference engineers:** follow [architecture](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md), [performance optimization](https://github.com/clocksmith/doppler/blob/main/docs/developer-guides/16-kernel-performance-optimization.md), and the [challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md).
262
+ - **Evidence reviewers:** review the [scoreboard](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md), [methodology](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md), and [release matrix](https://github.com/clocksmith/doppler/blob/main/docs/release-matrix.md).
229
263
 
230
264
  The [docs index](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)
231
265
  owns the complete model, subsystem, API, architecture, and operator inventory.
@@ -35,8 +35,8 @@
35
35
  "format": "rdrr"
36
36
  },
37
37
  "hf": {
38
- "repoId": "Clocksmith/rdrr",
39
- "revision": "7e194b4a6824a8b71cbd0eaa511d9f2c7e1c0129",
38
+ "repoId": "clocksmith/rdrr",
39
+ "revision": "a8591b20bce7c22d75becde1315482e76ff85fc9",
40
40
  "path": "models/gemma-3-270m-it-q4k-ehf16-af32"
41
41
  },
42
42
  "baseUrl": null,
@@ -125,7 +125,7 @@
125
125
  "format": "rdrr"
126
126
  },
127
127
  "hf": {
128
- "repoId": "Clocksmith/rdrr",
128
+ "repoId": "clocksmith/rdrr",
129
129
  "revision": null,
130
130
  "path": "models/gemma-3-270m-it-f16-af32"
131
131
  },
@@ -197,7 +197,7 @@
197
197
  "format": "rdrr"
198
198
  },
199
199
  "hf": {
200
- "repoId": "Clocksmith/rdrr",
200
+ "repoId": "clocksmith/rdrr",
201
201
  "revision": "95f0b29ec73dea70394c6bcfa8407bc6796df6c9",
202
202
  "path": "models/google-embeddinggemma-300m-q4k-ehf16-af32"
203
203
  },
@@ -275,7 +275,7 @@
275
275
  "format": "rdrr"
276
276
  },
277
277
  "hf": {
278
- "repoId": "Clocksmith/rdrr",
278
+ "repoId": "clocksmith/rdrr",
279
279
  "revision": "7c3d30e300bcb02cbd68fb0db3eee64fbf738f99",
280
280
  "path": "models/gemma-3-1b-it-q4k-ehf16-af32"
281
281
  },
@@ -357,7 +357,7 @@
357
357
  "format": "rdrr"
358
358
  },
359
359
  "hf": {
360
- "repoId": "Clocksmith/rdrr",
360
+ "repoId": "clocksmith/rdrr",
361
361
  "revision": "2070777c77047d54e6eae105f6dcb1891cf6f21a",
362
362
  "path": "models/gemma-4-e2b-it-q4k-ehf16-af32"
363
363
  },
@@ -438,7 +438,7 @@
438
438
  "format": "rdrr"
439
439
  },
440
440
  "hf": {
441
- "repoId": "Clocksmith/rdrr",
441
+ "repoId": "clocksmith/rdrr",
442
442
  "path": "models/gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
443
443
  "revision": "16256bf16dc8f92b8fe5105b07628cef91587f0a"
444
444
  },
@@ -524,7 +524,7 @@
524
524
  "format": "rdrr"
525
525
  },
526
526
  "hf": {
527
- "repoId": "Clocksmith/rdrr",
527
+ "repoId": "clocksmith/rdrr",
528
528
  "revision": "2d326649ae3d04b83323de0e6247e1e42fb1d9af",
529
529
  "path": "models/gemma-4-e2b-it-q4k-ehf16-af16-int4ple"
530
530
  },
@@ -602,7 +602,7 @@
602
602
  "format": "rdrr"
603
603
  },
604
604
  "hf": {
605
- "repoId": "Clocksmith/rdrr",
605
+ "repoId": "clocksmith/rdrr",
606
606
  "revision": null,
607
607
  "path": "models/gemma-4-12b-it-text-q4k-ehf16-af32"
608
608
  },
@@ -675,7 +675,7 @@
675
675
  "format": "rdrr"
676
676
  },
677
677
  "hf": {
678
- "repoId": "Clocksmith/rdrr",
678
+ "repoId": "clocksmith/rdrr",
679
679
  "revision": null,
680
680
  "path": "models/gemma-4-12b-it-text-q4k-ehf16-af16"
681
681
  },
@@ -749,7 +749,7 @@
749
749
  "format": "rdrr"
750
750
  },
751
751
  "hf": {
752
- "repoId": "Clocksmith/rdrr",
752
+ "repoId": "clocksmith/rdrr",
753
753
  "revision": null,
754
754
  "path": "models/gemma-4-12b-it-text-w4a16-ct-ehf16-af16"
755
755
  },
@@ -881,7 +881,7 @@
881
881
  "format": "rdrr"
882
882
  },
883
883
  "hf": {
884
- "repoId": "Clocksmith/rdrr",
884
+ "repoId": "clocksmith/rdrr",
885
885
  "revision": null,
886
886
  "path": "models/gemma-4-31b-it-text-q4k-ehf16-af32"
887
887
  },
@@ -959,7 +959,7 @@
959
959
  "format": "rdrr"
960
960
  },
961
961
  "hf": {
962
- "repoId": "Clocksmith/rdrr",
962
+ "repoId": "clocksmith/rdrr",
963
963
  "revision": null,
964
964
  "path": "models/gemma-4-31b-it-text-q4k-ehf16-af16"
965
965
  },
@@ -1033,7 +1033,7 @@
1033
1033
  "format": "rdrr"
1034
1034
  },
1035
1035
  "hf": {
1036
- "repoId": "Clocksmith/rdrr",
1036
+ "repoId": "clocksmith/rdrr",
1037
1037
  "path": "models/translategemma-4b-1b-enes-q4k-ehf16-af32",
1038
1038
  "revision": "1e2c047357f07dfe45a8effc2fa915acd73999ee"
1039
1039
  },
@@ -1111,7 +1111,7 @@
1111
1111
  "format": "rdrr"
1112
1112
  },
1113
1113
  "hf": {
1114
- "repoId": "Clocksmith/rdrr",
1114
+ "repoId": "clocksmith/rdrr",
1115
1115
  "revision": "6fc46049882e961a57d1690ba1ffde21677d001a",
1116
1116
  "path": "models/translategemma-4b-it-q4k-ehf16-af32"
1117
1117
  },
@@ -1189,7 +1189,7 @@
1189
1189
  "format": "rdrr"
1190
1190
  },
1191
1191
  "hf": {
1192
- "repoId": "Clocksmith/rdrr",
1192
+ "repoId": "clocksmith/rdrr",
1193
1193
  "revision": "049000f49325dca7db2ed2c9de2c8881bd0f4603",
1194
1194
  "path": "models/qwen-3-embedding-0-6b-q4k-ehf16-af32"
1195
1195
  },
@@ -1277,7 +1277,7 @@
1277
1277
  "format": "rdrr"
1278
1278
  },
1279
1279
  "hf": {
1280
- "repoId": "Clocksmith/rdrr",
1280
+ "repoId": "clocksmith/rdrr",
1281
1281
  "revision": "f86fe245b9bbc275cd69af46b1d45d47ea685a55",
1282
1282
  "path": "models/qwen-3-reranker-0-6b-q4k-ehf16-af32"
1283
1283
  },
@@ -1375,7 +1375,7 @@
1375
1375
  "format": "rdrr"
1376
1376
  },
1377
1377
  "hf": {
1378
- "repoId": "Clocksmith/rdrr",
1378
+ "repoId": "clocksmith/rdrr",
1379
1379
  "revision": "cc1fafff8cda609372b608cd92e487f6a2c32bc8",
1380
1380
  "path": "models/qwen-3-reranker-0-6b-f16-af32"
1381
1381
  },
@@ -1535,8 +1535,8 @@
1535
1535
  "format": "rdrr"
1536
1536
  },
1537
1537
  "hf": {
1538
- "repoId": "Clocksmith/rdrr",
1539
- "revision": "f58f1d0b58641c84e7ea50d13fea0dd4dc91389a",
1538
+ "repoId": "clocksmith/rdrr",
1539
+ "revision": "80d7716270b6371d541de979eff3370edaf34e13",
1540
1540
  "path": "models/qwen-3-5-0-8b-q4k-ehaf16"
1541
1541
  },
1542
1542
  "baseUrl": null,
@@ -1569,7 +1569,7 @@
1569
1569
  "result": "pass",
1570
1570
  "lastVerifiedAt": "2026-06-15",
1571
1571
  "source": "local-node-webgpu-verify",
1572
- "notes": "Browser/Node release evidence is package-visible at reports/release-claims/qwen-3-5-0-8b-q4k-ehaf16/2026-05-10T02-22-04.891Z.json. The hosted Clocksmith/rdrr artifact is approved for quickstart resolution with explicit stop-token metadata.",
1572
+ "notes": "Browser/Node release evidence is package-visible at reports/release-claims/qwen-3-5-0-8b-q4k-ehaf16/2026-05-10T02-22-04.891Z.json. The hosted clocksmith/rdrr artifact is approved for quickstart resolution with explicit stop-token metadata.",
1573
1573
  "contracts": {
1574
1574
  "executionContractOk": true
1575
1575
  }
@@ -1637,7 +1637,7 @@
1637
1637
  "format": "rdrr"
1638
1638
  },
1639
1639
  "hf": {
1640
- "repoId": "Clocksmith/rdrr",
1640
+ "repoId": "clocksmith/rdrr",
1641
1641
  "revision": "977d145bf2478a7fb542e6aca65030585620ca60",
1642
1642
  "path": "models/qwen-3-5-2b-q4k-ehaf16"
1643
1643
  },
@@ -1671,7 +1671,7 @@
1671
1671
  "result": "pass",
1672
1672
  "lastVerifiedAt": "2026-06-15",
1673
1673
  "source": "local-node-webgpu-verify",
1674
- "notes": "Browser/Node release evidence is package-visible at reports/release-claims/qwen-3-5-2b-q4k-ehaf16/2026-05-03T02-33-21.397Z.json. The execution-v1 lane keeps the fixed fused-Q4 decode path and q4 LM-head/prefill kernels; the hosted Clocksmith/rdrr artifact is approved for quickstart resolution with explicit stop-token metadata.",
1674
+ "notes": "Browser/Node release evidence is package-visible at reports/release-claims/qwen-3-5-2b-q4k-ehaf16/2026-05-03T02-33-21.397Z.json. The execution-v1 lane keeps the fixed fused-Q4 decode path and q4 LM-head/prefill kernels; the hosted clocksmith/rdrr artifact is approved for quickstart resolution with explicit stop-token metadata.",
1675
1675
  "contracts": {
1676
1676
  "executionContractOk": true
1677
1677
  }
@@ -1720,7 +1720,7 @@
1720
1720
  "format": "rdrr"
1721
1721
  },
1722
1722
  "hf": {
1723
- "repoId": "Clocksmith/rdrr",
1723
+ "repoId": "clocksmith/rdrr",
1724
1724
  "revision": "b402f6f27837857d51636da5f78c12bcd47e2a03",
1725
1725
  "path": "models/qwen-3-6-27b-q4k-ehaf16"
1726
1726
  },
@@ -1798,7 +1798,7 @@
1798
1798
  "format": "rdrr"
1799
1799
  },
1800
1800
  "hf": {
1801
- "repoId": "Clocksmith/rdrr",
1801
+ "repoId": "clocksmith/rdrr",
1802
1802
  "revision": "3dee21b3b12d65ac7fef9b24cbf759cacc953a67",
1803
1803
  "path": "models/qwen-3-6-27b-q4k-eaf16"
1804
1804
  },
@@ -1932,7 +1932,7 @@
1932
1932
  "format": "rdrr"
1933
1933
  },
1934
1934
  "hf": {
1935
- "repoId": "Clocksmith/rdrr",
1935
+ "repoId": "clocksmith/rdrr",
1936
1936
  "path": "models/amplify-120m-f16-af32",
1937
1937
  "revision": "990de206946a0972cc92e9dbcc9726399d257cac"
1938
1938
  },
@@ -2004,7 +2004,7 @@
2004
2004
  "format": "rdrr"
2005
2005
  },
2006
2006
  "hf": {
2007
- "repoId": "Clocksmith/rdrr",
2007
+ "repoId": "clocksmith/rdrr",
2008
2008
  "path": "models/esm2-t12-35m-ur50d-f32-af32",
2009
2009
  "revision": "feea324bae6c06c9d129ce39e188a75478ccdbbc"
2010
2010
  },
@@ -2076,7 +2076,7 @@
2076
2076
  "format": "rdrr"
2077
2077
  },
2078
2078
  "hf": {
2079
- "repoId": "Clocksmith/rdrr",
2079
+ "repoId": "clocksmith/rdrr",
2080
2080
  "path": "models/esmc-300m-f32-af32",
2081
2081
  "revision": "8354795dc544798ca7a4419bf00f844648f4b6f0"
2082
2082
  },
@@ -2148,7 +2148,7 @@
2148
2148
  "format": "rdrr"
2149
2149
  },
2150
2150
  "hf": {
2151
- "repoId": "Clocksmith/rdrr",
2151
+ "repoId": "clocksmith/rdrr",
2152
2152
  "path": "models/nucleotide-transformer-v2-50m-f32-af32",
2153
2153
  "revision": "a5703ef6971cc883438a393d73f2f54dccf94976"
2154
2154
  },
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "doppler-gpu",
3
- "version": "0.4.11",
3
+ "version": "0.4.15",
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",
@@ -30,6 +30,8 @@
30
30
  "program-bundle:parity": "node tools/check-program-bundle-parity.js examples/program-bundles/gemma-3-270m-it-q4k-ehf16-af32.program-bundle.json",
31
31
  "artifact-identity:inventory": "node tools/artifact-identity-inventory.js",
32
32
  "artifact-identity:check": "node tools/artifact-identity-inventory.js --check",
33
+ "adapter-artifacts:check": "node tools/check-adapter-artifact-catalog.js",
34
+ "adapter-artifacts:verify-local": "node tools/check-adapter-artifact-catalog.js --verify-local-bytes",
33
35
  "ci:check": "npm run test:ci && npm run kernels:check && npm run check:green",
34
36
  "ci:contracts:check": "node tools/check-contract-artifacts.js --with-lean --lean-require-manifest-match",
35
37
  "ci:diffusion:contract": "node tools/ci-diffusion-contract-gates.js",
@@ -87,7 +89,7 @@
87
89
  "goals:check": "node tools/check-goal-completion.js",
88
90
  "goals:report": "node tools/render-goal-completion-summary.js",
89
91
  "commands:surface:check": "node tools/check-command-surface-contract.js",
90
- "check:green": "npm run agents:verify && npm run goals:check && npm run public:boundaries:check && npm run package:smoke && npm run commands:surface:check && npm run api:docs:check && npm run config:single-source:check && npm run capability-policy:check && npm run imports:check:browser && npm run pending:check && npm run exports:parity:check && npm run artifact:contract:check && npm run claims:evidence:check && npm run support:subsystems:check && npm run policy:schemas:check",
92
+ "check:green": "npm run agents:verify && npm run goals:check && npm run public:boundaries:check && npm run package:smoke && npm run commands:surface:check && npm run api:docs:check && npm run config:single-source:check && npm run capability-policy:check && npm run imports:check:browser && npm run pending:check && npm run exports:parity:check && npm run artifact:contract:check && npm run adapter-artifacts:check && npm run claims:evidence:check && npm run support:subsystems:check && npm run policy:schemas:check",
91
93
  "verify:model": "node src/cli/doppler-cli.js verify",
92
94
  "quickstart:sync": "node tools/sync-quickstart-registry.js",
93
95
  "quickstart:check": "node tools/sync-quickstart-registry.js --check",
@@ -285,6 +287,14 @@
285
287
  "types": "./src/generation/index.d.ts",
286
288
  "import": "./src/generation/index.js"
287
289
  },
290
+ "./training": {
291
+ "browser": {
292
+ "types": "./src/training-browser.d.ts",
293
+ "default": "./src/training-browser.js"
294
+ },
295
+ "types": "./src/training.d.ts",
296
+ "import": "./src/training.js"
297
+ },
288
298
  "./diffusion": {
289
299
  "types": "./src/experimental/diffusion/index.d.ts",
290
300
  "import": "./src/experimental/diffusion/index.js"
@@ -338,7 +348,6 @@
338
348
  "README.md",
339
349
  "CHANGELOG.md",
340
350
  "LICENSE",
341
- "NOTICE",
342
351
  "tools/convert-safetensors-node.js"
343
352
  ],
344
353
  "devDependencies": {
@@ -39,8 +39,8 @@
39
39
  "label": "Text generators"
40
40
  },
41
41
  "hf": {
42
- "repoId": "Clocksmith/rdrr",
43
- "revision": "7e194b4a6824a8b71cbd0eaa511d9f2c7e1c0129",
42
+ "repoId": "clocksmith/rdrr",
43
+ "revision": "a8591b20bce7c22d75becde1315482e76ff85fc9",
44
44
  "path": "models/gemma-3-270m-it-q4k-ehf16-af32"
45
45
  },
46
46
  "vendorBenchmark": {
@@ -91,7 +91,7 @@
91
91
  "label": "Language embedders"
92
92
  },
93
93
  "hf": {
94
- "repoId": "Clocksmith/rdrr",
94
+ "repoId": "clocksmith/rdrr",
95
95
  "revision": "95f0b29ec73dea70394c6bcfa8407bc6796df6c9",
96
96
  "path": "models/google-embeddinggemma-300m-q4k-ehf16-af32"
97
97
  },
@@ -136,7 +136,7 @@
136
136
  "label": "Text generators"
137
137
  },
138
138
  "hf": {
139
- "repoId": "Clocksmith/rdrr",
139
+ "repoId": "clocksmith/rdrr",
140
140
  "revision": "7c3d30e300bcb02cbd68fb0db3eee64fbf738f99",
141
141
  "path": "models/gemma-3-1b-it-q4k-ehf16-af32"
142
142
  },
@@ -183,7 +183,7 @@
183
183
  "label": "Multimodal generators"
184
184
  },
185
185
  "hf": {
186
- "repoId": "Clocksmith/rdrr",
186
+ "repoId": "clocksmith/rdrr",
187
187
  "revision": "2070777c77047d54e6eae105f6dcb1891cf6f21a",
188
188
  "path": "models/gemma-4-e2b-it-q4k-ehf16-af32"
189
189
  },
@@ -230,7 +230,7 @@
230
230
  "label": "Multimodal generators"
231
231
  },
232
232
  "hf": {
233
- "repoId": "Clocksmith/rdrr",
233
+ "repoId": "clocksmith/rdrr",
234
234
  "revision": "16256bf16dc8f92b8fe5105b07628cef91587f0a",
235
235
  "path": "models/gemma-4-e2b-it-q4k-ehf16-af32-int4ple"
236
236
  },
@@ -276,8 +276,8 @@
276
276
  "label": "Text generators"
277
277
  },
278
278
  "hf": {
279
- "repoId": "Clocksmith/rdrr",
280
- "revision": "f58f1d0b58641c84e7ea50d13fea0dd4dc91389a",
279
+ "repoId": "clocksmith/rdrr",
280
+ "revision": "80d7716270b6371d541de979eff3370edaf34e13",
281
281
  "path": "models/qwen-3-5-0-8b-q4k-ehaf16"
282
282
  },
283
283
  "vendorBenchmark": {
@@ -322,7 +322,7 @@
322
322
  "label": "Text generators"
323
323
  },
324
324
  "hf": {
325
- "repoId": "Clocksmith/rdrr",
325
+ "repoId": "clocksmith/rdrr",
326
326
  "revision": "977d145bf2478a7fb542e6aca65030585620ca60",
327
327
  "path": "models/qwen-3-5-2b-q4k-ehaf16"
328
328
  },
@@ -13,6 +13,11 @@ const LORA_TARGET_WEIGHT_KEYS = Object.freeze({
13
13
  up_proj: 'up',
14
14
  down_proj: 'down',
15
15
  gate_up_proj: 'gateUp',
16
+ in_proj_a: 'linearInProjA',
17
+ in_proj_b: 'linearInProjB',
18
+ in_proj_qkv: 'qkvProj',
19
+ in_proj_z: 'linearInProjZ',
20
+ out_proj: 'oProj',
16
21
  });
17
22
 
18
23
  async function getExperimentalLoRAModule() {