doppler-gpu 0.4.10 → 0.4.14
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/CHANGELOG.md +95 -0
- package/README.md +40 -1
- package/models/catalog.json +637 -32
- package/package.json +24 -3
- package/src/client/doppler-registry.js +6 -0
- package/src/client/doppler-registry.json +135 -13
- package/src/client/runtime/lora.d.ts +10 -0
- package/src/client/runtime/lora.js +69 -0
- package/src/client/runtime/model-manager.js +6 -167
- package/src/client/runtime/model-session.d.ts +3 -1
- package/src/client/runtime/model-session.js +6 -0
- package/src/client/wrap-pipeline-handle.js +7 -0
- package/src/config/conversion/amplify/amplify-120m-f16-af32.json +516 -0
- package/src/config/conversion/esm/esm2-t12-35m-ur50d-f32-af32.json +571 -0
- package/src/config/conversion/esmc/esmc-300m-f32-af32.json +342 -0
- package/src/config/conversion/nucleotide-transformer/nucleotide-transformer-v2-50m-f32-af32.json +388 -0
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +3 -2
- package/src/config/kernels/kernel-ref-digests.js +2 -1
- package/src/config/kernels/registry.json +73 -11
- package/src/config/merge.js +24 -0
- package/src/config/required-inference-fields-contract-check.js +2 -0
- package/src/config/schema/adapter.schema.d.ts +6 -1
- package/src/config/schema/adapter.schema.js +5 -0
- package/src/config/schema/converter.schema.d.ts +7 -0
- package/src/config/schema/converter.schema.js +15 -0
- package/src/config/schema/inference.schema.d.ts +8 -0
- package/src/config/schema/manifest.schema.d.ts +32 -1
- package/src/config/schema/manifest.schema.js +26 -0
- package/src/converter/conversion-plan.js +15 -1
- package/src/converter/core.js +5 -4
- package/src/converter/source-tensor-rules.d.ts +50 -0
- package/src/converter/source-tensor-rules.js +180 -0
- package/src/converter/source-tokenizer.d.ts +45 -0
- package/src/converter/source-tokenizer.js +155 -0
- package/src/experimental/adapters/artifact-contract.d.ts +83 -0
- package/src/experimental/adapters/artifact-contract.js +254 -0
- package/src/experimental/adapters/index.d.ts +25 -0
- package/src/experimental/adapters/index.js +12 -0
- package/src/experimental/adapters/lora-loader.js +24 -4
- package/src/experimental/bridge/trainer-artifact-bridge.d.ts +2 -1
- package/src/experimental/bridge/trainer-artifact-bridge.js +35 -13
- package/src/experimental/training/artifacts.d.ts +160 -0
- package/src/experimental/training/attention-backward.d.ts +64 -0
- package/src/experimental/training/attention-forward.d.ts +22 -0
- package/src/experimental/training/attention-forward.js +82 -0
- package/src/experimental/training/autograd.d.ts +76 -0
- package/src/experimental/training/checkpoint.d.ts +36 -0
- package/src/experimental/training/clip.d.ts +9 -0
- package/src/experimental/training/dataloader.d.ts +8 -0
- package/src/experimental/training/distillation/student-fixture.js +1 -0
- package/src/experimental/training/export.d.ts +41 -0
- package/src/experimental/training/lora-pipeline.d.ts +186 -0
- package/src/experimental/training/lora-pipeline.js +5 -0
- package/src/experimental/training/lora.d.ts +20 -0
- package/src/experimental/training/lora.js +4 -3
- package/src/experimental/training/loss-scaling.d.ts +21 -0
- package/src/experimental/training/loss.d.ts +10 -0
- package/src/experimental/training/native-qwen-lora.d.ts +106 -0
- package/src/experimental/training/native-qwen-lora.js +449 -0
- package/src/experimental/training/objectives/base.d.ts +58 -0
- package/src/experimental/training/optimizer.d.ts +22 -0
- package/src/experimental/training/optimizer.js +1 -0
- package/src/experimental/training/runner.d.ts +249 -0
- package/src/experimental/training/trainer.d.ts +89 -0
- package/src/experimental/training/workloads.d.ts +205 -0
- package/src/gpu/kernels/backward/adam.d.ts +1 -0
- package/src/gpu/kernels/backward/adam.js +6 -2
- package/src/gpu/kernels/backward/adam.wgsl +3 -2
- package/src/gpu/kernels/backward/matmul_backward_q4k.wgsl +90 -0
- package/src/gpu/kernels/backward/utils.js +18 -4
- package/src/inference/pipelines/text/attention/attn-config.js +2 -0
- package/src/inference/pipelines/text/attention/projections.js +161 -24
- package/src/inference/pipelines/text/attention/record.js +67 -9
- package/src/inference/pipelines/text/attention/run.js +60 -7
- package/src/inference/pipelines/text/attention/types.js +18 -9
- package/src/inference/pipelines/text/config.d.ts +5 -0
- package/src/inference/pipelines/text/config.js +91 -0
- package/src/inference/pipelines/text/ffn/dense.js +205 -2
- package/src/inference/pipelines/text/ffn/standard.js +50 -3
- package/src/inference/pipelines/text/generator-helpers.js +11 -1
- package/src/inference/pipelines/text/generator-runtime.js +63 -15
- package/src/inference/pipelines/text/generator-steps.js +3 -1
- package/src/inference/pipelines/text/generator.d.ts +5 -1
- package/src/inference/pipelines/text/generator.js +196 -3
- package/src/inference/pipelines/text/init-qkv-fusion.js +85 -1
- package/src/inference/pipelines/text/init.d.ts +2 -0
- package/src/inference/pipelines/text/init.js +2 -0
- package/src/inference/pipelines/text/layer.js +6 -0
- package/src/inference/pipelines/text/linear-attention.d.ts +2 -0
- package/src/inference/pipelines/text/linear-attention.js +77 -6
- package/src/inference/pipelines/text/logits/cpu.d.ts +6 -0
- package/src/inference/pipelines/text/logits/cpu.js +22 -0
- package/src/inference/pipelines/text/logits/index.d.ts +1 -1
- package/src/inference/pipelines/text/logits/index.js +85 -11
- package/src/inference/pipelines/text/logits/types.d.ts +4 -0
- package/src/inference/pipelines/text/lora-types.d.ts +6 -1
- package/src/inference/pipelines/text/lora-types.js +5 -0
- package/src/inference/pipelines/text/ops.js +33 -0
- package/src/inference/pipelines/text/types.d.ts +66 -0
- package/src/inference/pipelines/text/weights.d.ts +10 -0
- package/src/inference/pipelines/text/weights.js +33 -1
- package/src/inference/pipelines/text.d.ts +9 -2
- package/src/inference/pipelines/text.js +116 -0
- package/src/inference/tokenizers/behavior-flags.js +5 -1
- package/src/inference/tokenizers/bundled.js +99 -8
- package/src/loader/doppler-loader.d.ts +2 -0
- package/src/loader/doppler-loader.js +9 -0
- package/src/loader/final-weights-loader.d.ts +8 -0
- package/src/loader/final-weights-loader.js +42 -0
- package/src/loader/layer-loader.d.ts +6 -0
- package/src/loader/layer-loader.js +102 -3
- package/src/loader/loader-types.d.ts +15 -2
- package/src/models/diffusiongemma.js +1 -1
- package/src/models/embeddinggemma.js +1 -1
- package/src/models/gemma3.js +1 -1
- package/src/models/gemma4.js +1 -1
- package/src/models/qwen3.js +1 -1
- package/src/tooling/hf-registry-utils.js +1 -1
- package/src/tooling/node-browser-command-runner.d.ts +16 -0
- package/src/tooling/node-browser-command-runner.js +1 -1
- package/src/tooling/node-converter.js +33 -1
- package/src/tooling/node-webgpu.js +85 -2
- package/src/training-browser.d.ts +19 -0
- package/src/training-browser.js +13 -0
- package/src/training.d.ts +137 -0
- package/src/training.js +219 -0
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +9 -1
- package/tests/kernels/browser/test-page.js +4 -2
- package/tests/kernels/reference/adam.js +4 -2
- package/NOTICE +0 -5
package/CHANGELOG.md
CHANGED
|
@@ -8,6 +8,101 @@ docs so the `0.1.x` line has one conventional npm-visible history surface.
|
|
|
8
8
|
|
|
9
9
|
## [Unreleased]
|
|
10
10
|
|
|
11
|
+
## [0.4.14] - 2026-07-19
|
|
12
|
+
|
|
13
|
+
### Added
|
|
14
|
+
|
|
15
|
+
- Added native browser, Node, and Bun LoRA training for the final Qwen 3.5
|
|
16
|
+
0.8B `down_proj` over frozen row-wise Q4K base weights. The bounded native
|
|
17
|
+
path uses completion-masked causal-LM loss, AdamW, deterministic
|
|
18
|
+
checkpoint/resume, and runtime-loadable safetensors/RDRR export.
|
|
19
|
+
- Added Q4K frozen-weight input gradients for the native LoRA suffix and a
|
|
20
|
+
production inference capture contract for the final feed-forward activation.
|
|
21
|
+
|
|
22
|
+
### Fixed
|
|
23
|
+
|
|
24
|
+
- Replaced the Qwen capability gate that incorrectly treated frozen Q4K base
|
|
25
|
+
weights as preventing native LoRA optimization. Other Qwen targets remain
|
|
26
|
+
unavailable on the native backend until their full backward graphs exist.
|
|
27
|
+
- Added explicit provider teardown to the validated Bun operator sequence;
|
|
28
|
+
model unload alone does not release Dawn's host-owned callback pump.
|
|
29
|
+
|
|
30
|
+
### Release status
|
|
31
|
+
|
|
32
|
+
- `0.4.13` was tagged but never published. Its tarball is an invalid release
|
|
33
|
+
candidate because it rejects the supported native Qwen target. `0.4.14`
|
|
34
|
+
supersedes it; npm `latest` must not be updated with the `0.4.13` tarball.
|
|
35
|
+
|
|
36
|
+
## [0.4.13] - 2026-07-19
|
|
37
|
+
|
|
38
|
+
### Added
|
|
39
|
+
|
|
40
|
+
- Added the public `doppler-gpu/training` npm entry point for capability-gated
|
|
41
|
+
SFT/LoRA workloads, WebGPU autograd and optimization primitives, checkpoint
|
|
42
|
+
evaluation, and runtime-compatible safetensors adapter export.
|
|
43
|
+
- Added explicit `webgpu_native` and `external` training-backend selection.
|
|
44
|
+
Packed Q4K Gemma and Qwen workloads fail closed on the native backend and
|
|
45
|
+
require an injected or workload-declared external trainer.
|
|
46
|
+
|
|
47
|
+
### Changed
|
|
48
|
+
|
|
49
|
+
- Added installed-package import, type, and capability smokes for the training
|
|
50
|
+
entry point so repository-only deep imports cannot masquerade as npm support.
|
|
51
|
+
|
|
52
|
+
## [0.4.12] - 2026-07-19
|
|
53
|
+
|
|
54
|
+
### Added
|
|
55
|
+
|
|
56
|
+
- Added LoRA activation for Qwen 3.5 linear-attention projections, including
|
|
57
|
+
`in_proj_a`, `in_proj_b`, `in_proj_qkv`, `in_proj_z`, and `out_proj` PEFT
|
|
58
|
+
targets, with deterministic intermediate-buffer cleanup.
|
|
59
|
+
- Added immutable adapter-custody catalog validation for preserved WGSL Repair
|
|
60
|
+
artifacts and their exact base-model, tokenizer, conversion, and source
|
|
61
|
+
lineage.
|
|
62
|
+
|
|
63
|
+
### Changed
|
|
64
|
+
|
|
65
|
+
- Corrected the Qwen 3.5 0.8B chat contract to use the Qwen template with
|
|
66
|
+
thinking disabled, and published the refreshed complete RDRR artifact at an
|
|
67
|
+
immutable hosted revision.
|
|
68
|
+
- Reconciled WGSL Writer V3 campaign records with explicit generation budgets
|
|
69
|
+
and finite confirmation metrics; this does not promote general shader
|
|
70
|
+
authorship.
|
|
71
|
+
|
|
72
|
+
## [0.4.11] - 2026-07-19
|
|
73
|
+
|
|
74
|
+
### Added
|
|
75
|
+
|
|
76
|
+
- Added manifest-first conversion and WebGPU inference support for four
|
|
77
|
+
biological sequence encoders: ESM-2 35M, Nucleotide Transformer v2 50M,
|
|
78
|
+
AMPLIFY 120M, and ESMC 300M.
|
|
79
|
+
- Published hosted RDRR registry lanes and Node/WebGPU qualification evidence
|
|
80
|
+
for tokenizer identity, finite outputs, reference parity, and synthetic LoRA
|
|
81
|
+
activation/unload behavior. This evidence does not claim downstream
|
|
82
|
+
biological-task quality or trained-adapter quality.
|
|
83
|
+
- Added source tokenizer and tensor-mapping contracts for importing sequence
|
|
84
|
+
checkpoints without runtime family detection.
|
|
85
|
+
- Added a model taxonomy for domain, task, architecture role, inputs, and
|
|
86
|
+
outputs, and exposed that classification through the quickstart registry.
|
|
87
|
+
- Added the `tinker_peft_browser_adapter` trainer-artifact bridge profile with
|
|
88
|
+
explicit Gamma selection authority and browser-parity requirements.
|
|
89
|
+
|
|
90
|
+
### Changed
|
|
91
|
+
|
|
92
|
+
- Extended the text pipeline and loader contracts for bidirectional sequence
|
|
93
|
+
encoders, including model-declared normalization and logits bias, QK
|
|
94
|
+
normalization shape, token-level outputs, and sequence embedding results.
|
|
95
|
+
- Expanded WGSL Writer V3 corpus, semantic-oracle, training, and evaluation
|
|
96
|
+
contracts after the four reference packages passed AMD/Vulkan Chromium
|
|
97
|
+
execution qualification. General shader authorship remains unproven.
|
|
98
|
+
- Rebuilt kernel reachability metadata for the current 301 registered
|
|
99
|
+
variants.
|
|
100
|
+
|
|
101
|
+
### Fixed
|
|
102
|
+
|
|
103
|
+
- Required WGSL Writer V3 evaluation receipts to disclose output-generation
|
|
104
|
+
budgets so compared lanes cannot hide unequal generation work.
|
|
105
|
+
|
|
11
106
|
## [0.4.10] - 2026-07-18
|
|
12
107
|
|
|
13
108
|
### Fixed
|
package/README.md
CHANGED
|
@@ -22,6 +22,45 @@ is experimental.
|
|
|
22
22
|
npx doppler-gpu
|
|
23
23
|
```
|
|
24
24
|
|
|
25
|
+
<!-- model-type-clusters:start -->
|
|
26
|
+
|
|
27
|
+
## Supported RDRR model types
|
|
28
|
+
|
|
29
|
+
Doppler classifies artifacts by what they consume and produce. This is
|
|
30
|
+
separate from lineage (`family`), runtime implementation (`modelType`), and
|
|
31
|
+
artifact-size tier.
|
|
32
|
+
|
|
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 |
|
|
43
|
+
|
|
44
|
+
The [full model-support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md)
|
|
45
|
+
lists every lane and its lifecycle evidence. Classification says what an
|
|
46
|
+
artifact is shaped to do; only lifecycle receipts establish what is
|
|
47
|
+
verified, and a runtime pass does not by itself qualify every declared input
|
|
48
|
+
modality.
|
|
49
|
+
|
|
50
|
+
<!-- model-type-clusters:end -->
|
|
51
|
+
|
|
52
|
+
## Supported LoRA adapters (PEFT)
|
|
53
|
+
|
|
54
|
+
Doppler supports runtime loading and hot-swapping of LoRA adapters (in SafeTensors format) bound to specific base models.
|
|
55
|
+
|
|
56
|
+
| Adapter ID | Base Model ID | HF Repository | Lifecycle & Purpose |
|
|
57
|
+
| --- | --- | --- | --- |
|
|
58
|
+
| `doppler-wgsl-qwen35-9b-v12-external20-seed11` | `qwen-3-5-9b-f16-af32` | [clocksmith/lora](https://huggingface.co/clocksmith/lora/blob/5662f8f7e3dbcf1e2ec353b5a04d0d4f2f5254ce/adapters/seed11/adapter_model.safetensors) | Preserved portable adapter (v12) |
|
|
59
|
+
| `doppler-wgsl-qwen35-9b-v12-external20-seed29` | `qwen-3-5-9b-f16-af32` | [clocksmith/lora](https://huggingface.co/clocksmith/lora/blob/5662f8f7e3dbcf1e2ec353b5a04d0d4f2f5254ce/adapters/seed29/adapter_model.safetensors) | Candidate adapter for repair validation (v12) |
|
|
60
|
+
| `doppler-wgsl-qwen35-9b-v12-external20-seed47` | `qwen-3-5-9b-f16-af32` | [clocksmith/lora](https://huggingface.co/clocksmith/lora/blob/5662f8f7e3dbcf1e2ec353b5a04d0d4f2f5254ce/adapters/seed47/adapter_model.safetensors) | Preserved portable adapter (v12) |
|
|
61
|
+
|
|
62
|
+
Active adapters are registered in [models/adapters/catalog.json](models/adapters/catalog.json). See the [LoRA Format specification](docs/lora-format.md) for details on the adapter manifest schema.
|
|
63
|
+
|
|
25
64
|
## Evidence
|
|
26
65
|
|
|
27
66
|
Doppler has higher steady-state inference throughput than Transformers.js in each
|
|
@@ -189,7 +228,7 @@ const response = await client.chat.completions.create({
|
|
|
189
228
|
});
|
|
190
229
|
```
|
|
191
230
|
|
|
192
|
-
Registry IDs resolve to hosted RDRR artifacts from `
|
|
231
|
+
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).
|
|
193
232
|
|
|
194
233
|
## Start here
|
|
195
234
|
|