@simulatte/doppler 0.1.6 → 0.1.7
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 +126 -0
- package/README.md +16 -23
- package/package.json +14 -1
- package/src/adapters/adapter-registry.js +12 -1
- package/src/adapters/lora-loader.js +23 -6
- package/src/bridge/extension-client.d.ts +5 -0
- package/src/bridge/extension-client.js +40 -0
- package/src/bridge/index.d.ts +2 -1
- package/src/bridge/index.js +6 -4
- package/src/browser/browser-converter.js +26 -1
- package/src/browser/file-picker.js +6 -0
- package/src/browser/safetensors-parser-browser.js +84 -1
- package/src/browser/shard-io-browser.js +2 -2
- package/src/browser/tensor-source-download.js +8 -2
- package/src/browser/tensor-source-http.d.ts +1 -0
- package/src/browser/tensor-source-http.js +5 -1
- package/src/client/doppler-api.browser.js +20 -4
- package/src/client/doppler-api.js +19 -3
- package/src/client/doppler-provider/generation.js +12 -0
- package/src/client/doppler-provider/model-manager.d.ts +10 -0
- package/src/client/doppler-provider/model-manager.js +91 -19
- package/src/client/doppler-provider/source-runtime.d.ts +2 -1
- package/src/client/doppler-provider/source-runtime.js +132 -13
- package/src/client/doppler-registry.json +8 -7
- package/src/config/backward-registry-loader.js +17 -2
- package/src/config/execution-v0-contract-check.js +113 -15
- package/src/config/kernel-path-contract-check.js +57 -29
- package/src/config/kernel-path-loader.js +5 -36
- package/src/config/kernels/kernel-ref-digests.js +1 -1
- package/src/config/kernels/registry.js +14 -1
- package/src/config/kernels/registry.json +7 -5
- package/src/config/loader.d.ts +1 -1
- package/src/config/loader.js +12 -2
- package/src/config/merge-contract-check.js +59 -4
- package/src/config/merge-helpers.js +128 -7
- package/src/config/merge.d.ts +1 -0
- package/src/config/merge.js +10 -0
- package/src/config/param-validator.js +47 -2
- package/src/config/presets/kernel-paths/{gemma2-q4k-dequant-f32a.json → gemma2-q4k-dequant-f32a-nosubgroups.json} +3 -3
- package/src/config/presets/kernel-paths/gemma3-f16-fused-f32a-online-streamingprefill.json +223 -0
- package/src/config/presets/kernel-paths/{gemma3-q4k-dequant-f32a.json → gemma3-q4k-dequant-f32a-nosubgroups.json} +3 -3
- package/src/config/presets/kernel-paths/registry.json +29 -8
- package/src/config/presets/models/gemma2.json +2 -2
- package/src/config/presets/runtime/experiments/bench/gemma3-bench-q4k.json +1 -1
- package/src/config/presets/runtime/experiments/debug/gemma3-debug-q4k.json +1 -1
- package/src/config/presets/runtime/experiments/verify/gemma3-verify.json +1 -1
- package/src/config/presets/runtime/kernels/dequant-f16-q4k.json +6 -13
- package/src/config/presets/runtime/kernels/dequant-f32-q4k.json +6 -13
- package/src/config/presets/runtime/kernels/embeddinggemma-q4k-dequant-f32a.json +37 -0
- package/src/config/presets/runtime/kernels/fused-q4k.json +6 -13
- package/src/config/presets/runtime/kernels/gemma2-q4k-dequant-f16a.json +33 -0
- package/src/config/presets/runtime/kernels/gemma2-q4k-dequant-f32a-nosubgroups.json +33 -0
- package/src/config/presets/runtime/kernels/gemma2-q4k-fused-f32a.json +33 -0
- package/src/config/presets/runtime/kernels/safe-q4k.json +6 -13
- package/src/config/presets/runtime/platform/metal-apple-q4k.json +1 -1
- package/src/config/runtime.js +6 -1
- package/src/config/schema/debug.schema.d.ts +5 -0
- package/src/config/schema/doppler.schema.js +16 -21
- package/src/config/schema/inference-defaults.schema.js +3 -3
- package/src/config/schema/kernel-path.schema.d.ts +5 -1
- package/src/config/schema/kernel-thresholds.schema.js +12 -4
- package/src/config/schema/manifest.schema.d.ts +2 -1
- package/src/config/schema/manifest.schema.js +16 -3
- package/src/config/training-defaults.js +30 -22
- package/src/converter/conversion-plan.js +94 -9
- package/src/converter/core.d.ts +7 -0
- package/src/converter/core.js +14 -9
- package/src/converter/execution-v0-manifest.js +4 -1
- package/src/converter/index.d.ts +1 -0
- package/src/converter/index.js +1 -0
- package/src/converter/manifest-inference.js +43 -12
- package/src/converter/parsers/diffusion.js +0 -3
- package/src/converter/quantization-info.js +35 -15
- package/src/converter/shard-packer.d.ts +1 -1
- package/src/converter/shard-packer.js +4 -1
- package/src/debug/config.js +123 -11
- package/src/debug/signals.js +7 -1
- package/src/debug/tensor.d.ts +2 -0
- package/src/debug/tensor.js +13 -2
- package/src/distribution/p2p-control-plane.js +52 -12
- package/src/distribution/p2p-observability.js +43 -7
- package/src/distribution/p2p-webrtc-browser.js +20 -0
- package/src/distribution/shard-delivery.js +77 -26
- package/src/formats/gguf/types.js +33 -16
- package/src/formats/rdrr/groups.d.ts +12 -4
- package/src/formats/rdrr/groups.js +3 -6
- package/src/formats/rdrr/parsing.js +39 -2
- package/src/formats/rdrr/types.d.ts +2 -1
- package/src/gpu/command-recorder.js +86 -61
- package/src/gpu/device.d.ts +1 -0
- package/src/gpu/device.js +73 -19
- package/src/gpu/kernel-tuner/benchmarks.js +326 -316
- package/src/gpu/kernel-tuner/cache.js +71 -4
- package/src/gpu/kernel-tuner/tuner.js +22 -4
- package/src/gpu/kernels/attention.js +15 -34
- package/src/gpu/kernels/backward/adam.js +62 -58
- package/src/gpu/kernels/backward/attention_backward.js +257 -169
- package/src/gpu/kernels/backward/conv2d_backward.js +14 -1
- package/src/gpu/kernels/cast.js +191 -149
- package/src/gpu/kernels/check-stop.js +33 -44
- package/src/gpu/kernels/conv2d.js +27 -17
- package/src/gpu/kernels/cross_entropy_loss.js +21 -15
- package/src/gpu/kernels/depthwise_conv2d.js +36 -26
- package/src/gpu/kernels/dequant.js +178 -126
- package/src/gpu/kernels/energy.d.ts +3 -21
- package/src/gpu/kernels/energy.js +111 -88
- package/src/gpu/kernels/feature-check.js +1 -1
- package/src/gpu/kernels/fused_ffn.js +84 -65
- package/src/gpu/kernels/fused_matmul_residual.js +56 -33
- package/src/gpu/kernels/fused_matmul_rmsnorm.js +62 -45
- package/src/gpu/kernels/gather.js +33 -15
- package/src/gpu/kernels/gelu.js +19 -11
- package/src/gpu/kernels/grouped_pointwise_conv2d.js +33 -23
- package/src/gpu/kernels/groupnorm.js +34 -23
- package/src/gpu/kernels/kv-quantize.js +5 -2
- package/src/gpu/kernels/layernorm.js +35 -19
- package/src/gpu/kernels/logit-merge.js +5 -3
- package/src/gpu/kernels/matmul.js +58 -39
- package/src/gpu/kernels/modulate.js +23 -15
- package/src/gpu/kernels/moe.js +221 -175
- package/src/gpu/kernels/pixel_shuffle.js +22 -14
- package/src/gpu/kernels/relu.js +18 -10
- package/src/gpu/kernels/repeat_channels.js +25 -17
- package/src/gpu/kernels/residual.js +37 -27
- package/src/gpu/kernels/rmsnorm.js +57 -41
- package/src/gpu/kernels/rope.js +3 -0
- package/src/gpu/kernels/sample.js +27 -38
- package/src/gpu/kernels/sana_linear_attention.js +18 -10
- package/src/gpu/kernels/scale.js +18 -11
- package/src/gpu/kernels/shader-cache.js +4 -2
- package/src/gpu/kernels/silu.js +120 -72
- package/src/gpu/kernels/softmax.js +44 -25
- package/src/gpu/kernels/split_qkv.js +23 -13
- package/src/gpu/kernels/transpose.js +18 -10
- package/src/gpu/kernels/transpose.wgsl +5 -3
- package/src/gpu/kernels/upsample2d.js +21 -13
- package/src/gpu/kernels/utils.js +20 -13
- package/src/gpu/partitioned-buffer-pool.js +10 -2
- package/src/gpu/perf-guards.js +2 -9
- package/src/gpu/profiler.js +27 -22
- package/src/gpu/readback-utils.d.ts +16 -0
- package/src/gpu/readback-utils.js +41 -0
- package/src/gpu/submit-tracker.js +13 -0
- package/src/gpu/uniform-cache.d.ts +1 -0
- package/src/gpu/uniform-cache.js +30 -9
- package/src/hotswap/intent-bundle.js +6 -0
- package/src/hotswap/manifest.d.ts +10 -1
- package/src/hotswap/manifest.js +12 -2
- package/src/hotswap/runtime.js +30 -8
- package/src/index-browser.d.ts +44 -0
- package/src/index-browser.js +14 -0
- package/src/inference/browser-harness-contract-helpers.d.ts +5 -0
- package/src/inference/browser-harness-contract-helpers.js +28 -0
- package/src/inference/browser-harness-diffusion-energy-suites.d.ts +2 -0
- package/src/inference/browser-harness-diffusion-energy-suites.js +269 -0
- package/src/inference/browser-harness-model-helpers.d.ts +16 -0
- package/src/inference/browser-harness-model-helpers.js +217 -0
- package/src/inference/browser-harness-report-helpers.d.ts +7 -0
- package/src/inference/browser-harness-report-helpers.js +42 -0
- package/src/inference/browser-harness-runtime-helpers.d.ts +61 -0
- package/src/inference/browser-harness-runtime-helpers.js +415 -0
- package/src/inference/browser-harness-suite-helpers.d.ts +28 -0
- package/src/inference/browser-harness-suite-helpers.js +268 -0
- package/src/inference/browser-harness-text-helpers.d.ts +27 -0
- package/src/inference/browser-harness-text-helpers.js +788 -0
- package/src/inference/browser-harness.d.ts +6 -0
- package/src/inference/browser-harness.js +130 -1996
- package/src/inference/kv-cache/base.js +140 -94
- package/src/inference/kv-cache/tiered.js +5 -3
- package/src/inference/moe-router.js +88 -56
- package/src/inference/multi-model-network.js +5 -3
- package/src/inference/network-evolution.d.ts +11 -2
- package/src/inference/network-evolution.js +20 -21
- package/src/inference/pipelines/context.d.ts +3 -0
- package/src/inference/pipelines/context.js +142 -2
- package/src/inference/pipelines/diffusion/helpers.js +7 -2
- package/src/inference/pipelines/diffusion/pipeline.js +2 -1
- package/src/inference/pipelines/diffusion/sd3-transformer.js +10 -10
- package/src/inference/pipelines/diffusion/vae.js +3 -7
- package/src/inference/pipelines/energy/pipeline.js +27 -21
- package/src/inference/pipelines/energy/quintel.d.ts +5 -0
- package/src/inference/pipelines/energy/quintel.js +11 -0
- package/src/inference/pipelines/energy-head/row-head-pipeline.js +17 -13
- package/src/inference/pipelines/structured/json-head-pipeline.js +26 -11
- package/src/inference/pipelines/text/attention/projections.js +151 -101
- package/src/inference/pipelines/text/attention/record.js +62 -8
- package/src/inference/pipelines/text/attention/run.js +62 -8
- package/src/inference/pipelines/text/config.js +3 -4
- package/src/inference/pipelines/text/embed.js +2 -8
- package/src/inference/pipelines/text/execution-plan.js +41 -19
- package/src/inference/pipelines/text/execution-v0-contract-helpers.d.ts +59 -0
- package/src/inference/pipelines/text/execution-v0-contract-helpers.js +937 -0
- package/src/inference/pipelines/text/execution-v0-runtime-builders.d.ts +15 -0
- package/src/inference/pipelines/text/execution-v0-runtime-builders.js +279 -0
- package/src/inference/pipelines/text/execution-v0.js +62 -1013
- package/src/inference/pipelines/text/generator-steps.d.ts +46 -0
- package/src/inference/pipelines/text/generator-steps.js +298 -207
- package/src/inference/pipelines/text/generator.js +6 -23
- package/src/inference/pipelines/text/init.js +78 -20
- package/src/inference/pipelines/text/kernel-path-auto-select.js +2 -0
- package/src/inference/pipelines/text/kernel-trace.d.ts +2 -0
- package/src/inference/pipelines/text/kernel-trace.js +6 -0
- package/src/inference/pipelines/text/layer.js +3 -9
- package/src/inference/pipelines/text/linear-attention.d.ts +10 -0
- package/src/inference/pipelines/text/linear-attention.js +80 -6
- package/src/inference/pipelines/text/logits/gpu.js +10 -5
- package/src/inference/pipelines/text/logits/index.js +10 -11
- package/src/inference/pipelines/text/logits/utils.d.ts +7 -0
- package/src/inference/pipelines/text/logits/utils.js +9 -0
- package/src/inference/pipelines/text/lora-apply.js +50 -32
- package/src/inference/pipelines/text/model-load.js +279 -104
- package/src/inference/pipelines/text/moe-cache.js +5 -4
- package/src/inference/pipelines/text/moe-cpu-gptoss.js +74 -69
- package/src/inference/pipelines/text/moe-cpu.js +42 -38
- package/src/inference/pipelines/text/moe-gpu.js +110 -86
- package/src/inference/pipelines/text/ops.js +90 -90
- package/src/inference/pipelines/text/probes.js +9 -9
- package/src/inference/pipelines/text/weights.js +17 -7
- package/src/inference/pipelines/text.js +13 -1
- package/src/inference/speculative.d.ts +2 -2
- package/src/inference/speculative.js +4 -18
- package/src/inference/test-harness.d.ts +1 -1
- package/src/inference/test-harness.js +15 -5
- package/src/inference/tokenizer.d.ts +0 -5
- package/src/inference/tokenizer.js +4 -23
- package/src/inference/tokenizers/bpe.js +9 -0
- package/src/inference/tokenizers/bundled.js +20 -0
- package/src/inference/tokenizers/sentencepiece.js +12 -0
- package/src/loader/doppler-loader.js +38 -22
- package/src/loader/dtype-utils.js +3 -44
- package/src/loader/embedding-loader.js +7 -3
- package/src/loader/experts/expert-cache.js +13 -6
- package/src/loader/experts/expert-loader.js +10 -6
- package/src/loader/final-weights-loader.js +8 -4
- package/src/loader/layer-loader.js +2 -1
- package/src/loader/loader-state.js +2 -2
- package/src/loader/memory-monitor.js +8 -0
- package/src/loader/multi-model-loader.d.ts +14 -0
- package/src/loader/multi-model-loader.js +70 -24
- package/src/loader/shard-cache.js +81 -12
- package/src/loader/shard-resolver.js +25 -3
- package/src/loader/tensors/tensor-loader.js +209 -144
- package/src/loader/tensors/tensor-reader.js +76 -19
- package/src/loader/weight-downcast.js +1 -1
- package/src/memory/buffer-pool.d.ts +9 -1
- package/src/memory/buffer-pool.js +109 -44
- package/src/memory/unified-detect.js +1 -1
- package/src/rules/inference/kernel-path.rules.json +24 -8
- package/src/rules/rule-registry.js +25 -1
- package/src/storage/backends/opfs-store.js +68 -24
- package/src/storage/downloader.js +364 -83
- package/src/storage/index.d.ts +3 -0
- package/src/storage/index.js +3 -0
- package/src/storage/preflight.d.ts +2 -2
- package/src/storage/preflight.js +24 -2
- package/src/storage/quickstart-downloader.js +11 -5
- package/src/storage/registry.js +10 -4
- package/src/storage/reports.js +1 -1
- package/src/storage/shard-manager.d.ts +15 -1
- package/src/storage/shard-manager.js +51 -3
- package/src/storage/source-artifact-store.d.ts +52 -0
- package/src/storage/source-artifact-store.js +234 -0
- package/src/tooling/command-api-constants.d.ts +9 -0
- package/src/tooling/command-api-constants.js +9 -0
- package/src/tooling/command-api-family-normalizers.d.ts +9 -0
- package/src/tooling/command-api-family-normalizers.js +343 -0
- package/src/tooling/command-api-helpers.d.ts +25 -0
- package/src/tooling/command-api-helpers.js +262 -0
- package/src/tooling/command-api.js +16 -602
- package/src/tooling/command-envelope.js +4 -1
- package/src/tooling/command-runner-shared.js +52 -18
- package/src/tooling/lean-execution-contract.js +150 -3
- package/src/tooling/node-browser-command-runner.js +161 -271
- package/src/tooling/node-command-runner.js +29 -3
- package/src/tooling/node-converter.js +27 -1
- package/src/tooling/node-source-runtime.d.ts +1 -1
- package/src/tooling/node-source-runtime.js +84 -3
- package/src/tooling/node-webgpu.js +24 -21
- package/src/tooling/opfs-cache.js +21 -4
- package/src/tooling/runtime-input-composition.d.ts +38 -0
- package/src/tooling/runtime-input-composition.js +86 -0
- package/src/tooling/source-runtime-bundle.d.ts +40 -5
- package/src/tooling/source-runtime-bundle.js +261 -34
- package/src/tooling/source-runtime-materializer.d.ts +6 -0
- package/src/tooling/source-runtime-materializer.js +93 -0
- package/src/training/attention-backward.js +32 -17
- package/src/training/autograd.js +80 -52
- package/src/training/checkpoint-watch.d.ts +2 -1
- package/src/training/checkpoint-watch.js +39 -6
- package/src/training/checkpoint.js +40 -11
- package/src/training/clip.js +2 -1
- package/src/training/datasets/token-batch.js +20 -8
- package/src/training/distillation/checkpoint-watch.js +1 -0
- package/src/training/distillation/student-fixture.d.ts +22 -0
- package/src/training/distillation/student-fixture.js +846 -0
- package/src/training/distillation/suite-data.d.ts +45 -0
- package/src/training/distillation/suite-data.js +189 -0
- package/src/training/lora-pipeline.js +4 -7
- package/src/training/lora.js +26 -12
- package/src/training/loss.js +5 -6
- package/src/training/objectives/cross_entropy.js +2 -5
- package/src/training/objectives/distill_kd.js +4 -8
- package/src/training/objectives/distill_triplet.js +4 -8
- package/src/training/objectives/ul_stage2_base.js +4 -8
- package/src/training/operator-command.js +2 -0
- package/src/training/optimizer.js +19 -7
- package/src/training/runner.js +2 -1
- package/src/training/suite.js +18 -978
- package/src/training/tensor-factory.d.ts +9 -0
- package/src/training/tensor-factory.js +13 -0
- package/src/training/trainer.js +3 -5
- package/src/training/ul_dataset.js +3 -5
- package/src/training/workloads.js +70 -79
- package/src/version.js +1 -1
- package/tools/convert-safetensors-node.js +22 -16
- package/tools/doppler-cli.js +44 -25
package/src/gpu/profiler.js
CHANGED
|
@@ -179,6 +179,8 @@ export class GPUProfiler {
|
|
|
179
179
|
|
|
180
180
|
if (!this.#device || !this.#querySet || !this.#queryBuffer || !this.#readbackBuffer) {
|
|
181
181
|
log.warn('GPUProfiler', 'Missing required resources for resolve');
|
|
182
|
+
this.#pendingResolves = [];
|
|
183
|
+
this.#nextQueryIndex = 0;
|
|
182
184
|
return;
|
|
183
185
|
}
|
|
184
186
|
|
|
@@ -199,34 +201,35 @@ export class GPUProfiler {
|
|
|
199
201
|
|
|
200
202
|
this.#device.queue.submit([encoder.finish()]);
|
|
201
203
|
|
|
202
|
-
|
|
203
|
-
return;
|
|
204
|
-
}
|
|
204
|
+
let mapped = false;
|
|
205
205
|
|
|
206
|
-
|
|
207
|
-
|
|
208
|
-
|
|
206
|
+
try {
|
|
207
|
+
if (!allowReadback('GPUProfiler.resolve')) {
|
|
208
|
+
return;
|
|
209
|
+
}
|
|
209
210
|
|
|
210
|
-
|
|
211
|
-
|
|
212
|
-
const
|
|
213
|
-
const endNs = timestamps[pending.endIndex];
|
|
211
|
+
await this.#readbackBuffer.mapAsync(GPUMapMode.READ);
|
|
212
|
+
mapped = true;
|
|
213
|
+
const timestamps = new BigUint64Array(this.#readbackBuffer.getMappedRange());
|
|
214
214
|
|
|
215
|
-
|
|
216
|
-
|
|
215
|
+
for (const pending of this.#pendingResolves) {
|
|
216
|
+
const startNs = timestamps[pending.startIndex];
|
|
217
|
+
const endNs = timestamps[pending.endIndex];
|
|
218
|
+
const durationMs = Number(endNs - startNs) / 1_000_000;
|
|
217
219
|
|
|
218
|
-
|
|
219
|
-
|
|
220
|
-
|
|
221
|
-
|
|
222
|
-
|
|
223
|
-
|
|
220
|
+
if (durationMs < 0 || durationMs > this.#maxDurationMs) {
|
|
221
|
+
this.#recordResult(pending.label, pending.cpuEndTime - pending.cpuStartTime);
|
|
222
|
+
} else {
|
|
223
|
+
this.#recordResult(pending.label, durationMs);
|
|
224
|
+
}
|
|
225
|
+
}
|
|
226
|
+
} finally {
|
|
227
|
+
if (mapped) {
|
|
228
|
+
this.#readbackBuffer.unmap();
|
|
224
229
|
}
|
|
230
|
+
this.#pendingResolves = [];
|
|
231
|
+
this.#nextQueryIndex = 0;
|
|
225
232
|
}
|
|
226
|
-
|
|
227
|
-
this.#readbackBuffer.unmap();
|
|
228
|
-
this.#pendingResolves = [];
|
|
229
|
-
this.#nextQueryIndex = 0;
|
|
230
233
|
}
|
|
231
234
|
|
|
232
235
|
|
|
@@ -350,6 +353,8 @@ export class GPUProfiler {
|
|
|
350
353
|
}
|
|
351
354
|
this.#results.clear();
|
|
352
355
|
this.#activeLabels.clear();
|
|
356
|
+
this.#pendingResolves = [];
|
|
357
|
+
this.#nextQueryIndex = 0;
|
|
353
358
|
}
|
|
354
359
|
}
|
|
355
360
|
|
|
@@ -0,0 +1,16 @@
|
|
|
1
|
+
export interface ReadbackBufferEntry {
|
|
2
|
+
buffer: GPUBuffer;
|
|
3
|
+
destroy?: boolean;
|
|
4
|
+
offset?: number;
|
|
5
|
+
size?: number;
|
|
6
|
+
}
|
|
7
|
+
|
|
8
|
+
export function withMappedReadBuffer<T>(
|
|
9
|
+
buffer: GPUBuffer,
|
|
10
|
+
read: (range: ArrayBuffer) => T | Promise<T>,
|
|
11
|
+
): Promise<T>;
|
|
12
|
+
|
|
13
|
+
export function withMappedReadBuffers<T>(
|
|
14
|
+
entries: ReadonlyArray<ReadbackBufferEntry>,
|
|
15
|
+
read: (ranges: Array<ArrayBuffer>) => T | Promise<T>,
|
|
16
|
+
): Promise<T>;
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
export async function withMappedReadBuffer(buffer, read) {
|
|
2
|
+
let mapped = false;
|
|
3
|
+
try {
|
|
4
|
+
await buffer.mapAsync(GPUMapMode.READ);
|
|
5
|
+
mapped = true;
|
|
6
|
+
return await read(buffer.getMappedRange());
|
|
7
|
+
} finally {
|
|
8
|
+
if (mapped) {
|
|
9
|
+
buffer.unmap();
|
|
10
|
+
}
|
|
11
|
+
}
|
|
12
|
+
}
|
|
13
|
+
|
|
14
|
+
export async function withMappedReadBuffers(entries, read) {
|
|
15
|
+
const mappedEntries = [];
|
|
16
|
+
try {
|
|
17
|
+
await Promise.all(entries.map(async (entry) => {
|
|
18
|
+
await entry.buffer.mapAsync(GPUMapMode.READ);
|
|
19
|
+
mappedEntries.push(entry);
|
|
20
|
+
}));
|
|
21
|
+
const ranges = entries.map((entry) => {
|
|
22
|
+
if (entry.offset != null || entry.size != null) {
|
|
23
|
+
return entry.buffer.getMappedRange(
|
|
24
|
+
entry.offset ?? 0,
|
|
25
|
+
entry.size ?? (entry.buffer.size - (entry.offset ?? 0))
|
|
26
|
+
);
|
|
27
|
+
}
|
|
28
|
+
return entry.buffer.getMappedRange();
|
|
29
|
+
});
|
|
30
|
+
return await read(ranges);
|
|
31
|
+
} finally {
|
|
32
|
+
for (let index = mappedEntries.length - 1; index >= 0; index -= 1) {
|
|
33
|
+
mappedEntries[index].buffer.unmap();
|
|
34
|
+
}
|
|
35
|
+
for (const entry of entries) {
|
|
36
|
+
if (entry.destroy === true) {
|
|
37
|
+
entry.buffer.destroy();
|
|
38
|
+
}
|
|
39
|
+
}
|
|
40
|
+
}
|
|
41
|
+
}
|
|
@@ -21,6 +21,8 @@ let submitSources = new Map();
|
|
|
21
21
|
|
|
22
22
|
let currentPhase = 'other';
|
|
23
23
|
|
|
24
|
+
const WRAPPED_QUEUE_SENTINEL = Symbol.for('doppler.submitTrackerWrapped');
|
|
25
|
+
|
|
24
26
|
|
|
25
27
|
const phaseStats = {
|
|
26
28
|
prefill: { count: 0, times: [], totalMs: 0, maxMs: 0, minMs: Infinity, sources: new Map() },
|
|
@@ -194,6 +196,10 @@ function extractSourceFromStack() {
|
|
|
194
196
|
|
|
195
197
|
|
|
196
198
|
export function wrapQueueForTracking(queue) {
|
|
199
|
+
if (!queue || queue[WRAPPED_QUEUE_SENTINEL] === true) {
|
|
200
|
+
return queue;
|
|
201
|
+
}
|
|
202
|
+
|
|
197
203
|
const originalSubmit = queue.submit.bind(queue);
|
|
198
204
|
|
|
199
205
|
(queue).submit = function( commandBuffers) {
|
|
@@ -210,6 +216,13 @@ export function wrapQueueForTracking(queue) {
|
|
|
210
216
|
return result;
|
|
211
217
|
};
|
|
212
218
|
|
|
219
|
+
Object.defineProperty(queue, WRAPPED_QUEUE_SENTINEL, {
|
|
220
|
+
value: true,
|
|
221
|
+
configurable: true,
|
|
222
|
+
enumerable: false,
|
|
223
|
+
writable: false,
|
|
224
|
+
});
|
|
225
|
+
|
|
213
226
|
return queue;
|
|
214
227
|
}
|
|
215
228
|
|
package/src/gpu/uniform-cache.js
CHANGED
|
@@ -17,6 +17,18 @@ function hashArrayBuffer(data) {
|
|
|
17
17
|
return (hash >>> 0).toString(16).padStart(8, '0');
|
|
18
18
|
}
|
|
19
19
|
|
|
20
|
+
function copyUniformBytes(data) {
|
|
21
|
+
return new Uint8Array(data.slice(0));
|
|
22
|
+
}
|
|
23
|
+
|
|
24
|
+
function equalUniformBytes(a, b) {
|
|
25
|
+
if (a.length !== b.length) return false;
|
|
26
|
+
for (let i = 0; i < a.length; i++) {
|
|
27
|
+
if (a[i] !== b[i]) return false;
|
|
28
|
+
}
|
|
29
|
+
return true;
|
|
30
|
+
}
|
|
31
|
+
|
|
20
32
|
|
|
21
33
|
export class UniformBufferCache {
|
|
22
34
|
|
|
@@ -50,14 +62,22 @@ export class UniformBufferCache {
|
|
|
50
62
|
|
|
51
63
|
|
|
52
64
|
getOrCreate(data, label) {
|
|
53
|
-
const
|
|
54
|
-
const
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
65
|
+
const baseKey = `${data.byteLength}:${hashArrayBuffer(data)}`;
|
|
66
|
+
const dataBytes = copyUniformBytes(data);
|
|
67
|
+
let key = baseKey;
|
|
68
|
+
let suffix = 0;
|
|
69
|
+
let existing = this.#cache.get(key);
|
|
70
|
+
|
|
71
|
+
while (existing) {
|
|
72
|
+
if (equalUniformBytes(existing.bytes, dataBytes)) {
|
|
73
|
+
existing.lastUsed = performance.now();
|
|
74
|
+
existing.refCount++;
|
|
75
|
+
this.#stats.hits++;
|
|
76
|
+
return existing.buffer;
|
|
77
|
+
}
|
|
78
|
+
suffix += 1;
|
|
79
|
+
key = `${baseKey}#${suffix}`;
|
|
80
|
+
existing = this.#cache.get(key);
|
|
61
81
|
}
|
|
62
82
|
|
|
63
83
|
// Cache miss - create new buffer
|
|
@@ -80,8 +100,9 @@ export class UniformBufferCache {
|
|
|
80
100
|
this.#evictLRU();
|
|
81
101
|
}
|
|
82
102
|
|
|
83
|
-
this.#cache.set(
|
|
103
|
+
this.#cache.set(key, {
|
|
84
104
|
buffer,
|
|
105
|
+
bytes: dataBytes,
|
|
85
106
|
lastUsed: performance.now(),
|
|
86
107
|
refCount: 1,
|
|
87
108
|
});
|
|
@@ -42,6 +42,9 @@ export async function verifyIntentBundle(bundle, context) {
|
|
|
42
42
|
reasons.push('Missing payload.expectedOutputHash');
|
|
43
43
|
}
|
|
44
44
|
|
|
45
|
+
if (baseModelHash && !context?.manifest) {
|
|
46
|
+
reasons.push('Missing verification context manifest');
|
|
47
|
+
}
|
|
45
48
|
if (context?.manifest && baseModelHash) {
|
|
46
49
|
const manifestHash = await computeManifestHash(context.manifest);
|
|
47
50
|
if (manifestHash !== baseModelHash.replace('sha256:', '')) {
|
|
@@ -49,6 +52,9 @@ export async function verifyIntentBundle(bundle, context) {
|
|
|
49
52
|
}
|
|
50
53
|
}
|
|
51
54
|
|
|
55
|
+
if (kernelRegistryVersion && context?.kernelRegistryVersion == null) {
|
|
56
|
+
reasons.push('Missing verification context kernelRegistryVersion');
|
|
57
|
+
}
|
|
52
58
|
if (context?.kernelRegistryVersion && kernelRegistryVersion) {
|
|
53
59
|
if (context.kernelRegistryVersion !== kernelRegistryVersion) {
|
|
54
60
|
reasons.push('Kernel registry version mismatch');
|
|
@@ -23,11 +23,20 @@ export interface HotSwapVerificationResult {
|
|
|
23
23
|
signerId?: string;
|
|
24
24
|
}
|
|
25
25
|
|
|
26
|
+
export interface HotSwapVerificationContext {
|
|
27
|
+
source?: {
|
|
28
|
+
kind?: 'local' | 'remote' | string | null;
|
|
29
|
+
isLocal?: boolean | null;
|
|
30
|
+
url?: string | null;
|
|
31
|
+
} | null;
|
|
32
|
+
}
|
|
33
|
+
|
|
26
34
|
export declare function fetchHotSwapManifest(url: string): Promise<HotSwapManifest>;
|
|
27
35
|
|
|
28
36
|
export declare function verifyHotSwapManifest(
|
|
29
37
|
manifest: HotSwapManifest,
|
|
30
|
-
policy: HotSwapConfigSchema
|
|
38
|
+
policy: HotSwapConfigSchema,
|
|
39
|
+
context?: HotSwapVerificationContext
|
|
31
40
|
): Promise<HotSwapVerificationResult>;
|
|
32
41
|
|
|
33
42
|
export declare function serializeHotSwapManifest(manifest: HotSwapManifest): string;
|
package/src/hotswap/manifest.js
CHANGED
|
@@ -14,13 +14,23 @@ export async function fetchHotSwapManifest(url) {
|
|
|
14
14
|
return response.json();
|
|
15
15
|
}
|
|
16
16
|
|
|
17
|
-
|
|
17
|
+
function isExplicitLocalSource(source) {
|
|
18
|
+
if (!source || typeof source !== 'object') {
|
|
19
|
+
return false;
|
|
20
|
+
}
|
|
21
|
+
if (source.isLocal === true) {
|
|
22
|
+
return true;
|
|
23
|
+
}
|
|
24
|
+
return source.kind === 'local';
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
export async function verifyHotSwapManifest(manifest, policy, context = {}) {
|
|
18
28
|
if (!policy.enabled) {
|
|
19
29
|
return { ok: false, reason: 'Hot-swap disabled' };
|
|
20
30
|
}
|
|
21
31
|
|
|
22
32
|
if (!manifest.signature) {
|
|
23
|
-
if (policy.localOnly && policy.allowUnsignedLocal) {
|
|
33
|
+
if (policy.localOnly && policy.allowUnsignedLocal && isExplicitLocalSource(context.source)) {
|
|
24
34
|
return { ok: true, reason: 'Local-only unsigned manifest accepted' };
|
|
25
35
|
}
|
|
26
36
|
return { ok: false, reason: 'Signature required' };
|
package/src/hotswap/runtime.js
CHANGED
|
@@ -6,18 +6,40 @@ function normalizeRolloutPolicy(policy) {
|
|
|
6
6
|
? policy.rollout
|
|
7
7
|
: {};
|
|
8
8
|
const rawMode = String(rollout.mode || 'shadow').trim().toLowerCase().replace(/_/g, '-');
|
|
9
|
-
|
|
10
|
-
|
|
11
|
-
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
|
|
9
|
+
if (rawMode !== 'default' && rawMode !== 'canary' && rawMode !== 'opt-in' && rawMode !== 'shadow') {
|
|
10
|
+
throw new Error(
|
|
11
|
+
`hotswap.rollout.mode must be one of default, canary, opt-in, shadow (received "${rollout.mode}")`
|
|
12
|
+
);
|
|
13
|
+
}
|
|
14
|
+
let canaryPercent = 0;
|
|
15
|
+
if (rollout.canaryPercent !== undefined && rollout.canaryPercent !== null) {
|
|
16
|
+
const parsedCanaryPercent = Number(rollout.canaryPercent);
|
|
17
|
+
if (!Number.isFinite(parsedCanaryPercent) || parsedCanaryPercent < 0 || parsedCanaryPercent > 100) {
|
|
18
|
+
throw new Error('hotswap.rollout.canaryPercent must be a number between 0 and 100 when provided.');
|
|
19
|
+
}
|
|
20
|
+
canaryPercent = parsedCanaryPercent;
|
|
21
|
+
}
|
|
22
|
+
if (rollout.cohortSalt !== undefined && rollout.cohortSalt !== null && typeof rollout.cohortSalt !== 'string') {
|
|
23
|
+
throw new Error('hotswap.rollout.cohortSalt must be a string when provided.');
|
|
24
|
+
}
|
|
15
25
|
const cohortSalt = String(rollout.cohortSalt || 'doppler-hotswap-v1').trim() || 'doppler-hotswap-v1';
|
|
26
|
+
if (rollout.optInAllowlist !== undefined && rollout.optInAllowlist !== null && !Array.isArray(rollout.optInAllowlist)) {
|
|
27
|
+
throw new Error('hotswap.rollout.optInAllowlist must be an array of strings when provided.');
|
|
28
|
+
}
|
|
16
29
|
const optInAllowlist = Array.isArray(rollout.optInAllowlist)
|
|
17
|
-
? rollout.optInAllowlist.map((entry) =>
|
|
30
|
+
? rollout.optInAllowlist.map((entry, index) => {
|
|
31
|
+
if (typeof entry !== 'string') {
|
|
32
|
+
throw new Error(`hotswap.rollout.optInAllowlist[${index}] must be a string.`);
|
|
33
|
+
}
|
|
34
|
+
const normalized = entry.trim();
|
|
35
|
+
if (!normalized) {
|
|
36
|
+
throw new Error(`hotswap.rollout.optInAllowlist[${index}] must not be empty.`);
|
|
37
|
+
}
|
|
38
|
+
return normalized;
|
|
39
|
+
})
|
|
18
40
|
: [];
|
|
19
41
|
return {
|
|
20
|
-
mode,
|
|
42
|
+
mode: rawMode,
|
|
21
43
|
canaryPercent,
|
|
22
44
|
cohortSalt,
|
|
23
45
|
optInAllowlist,
|
package/src/index-browser.d.ts
CHANGED
|
@@ -9,6 +9,20 @@ export {
|
|
|
9
9
|
export { MultiModelLoader } from './loader/multi-model-loader.js';
|
|
10
10
|
|
|
11
11
|
export { InferencePipeline, EmbeddingPipeline, createPipeline } from './generation/index.js';
|
|
12
|
+
export {
|
|
13
|
+
StructuredJsonHeadPipeline,
|
|
14
|
+
isStructuredJsonHeadModelType,
|
|
15
|
+
createStructuredJsonHeadPipeline,
|
|
16
|
+
DreamStructuredPipeline,
|
|
17
|
+
isDreamStructuredModelType,
|
|
18
|
+
createDreamStructuredPipeline,
|
|
19
|
+
} from './generation/index.js';
|
|
20
|
+
export {
|
|
21
|
+
EnergyRowHeadPipeline,
|
|
22
|
+
createEnergyRowHeadPipeline,
|
|
23
|
+
DreamEnergyHeadPipeline,
|
|
24
|
+
createDreamEnergyHeadPipeline,
|
|
25
|
+
} from './inference/pipelines/energy-head/row-head-pipeline.js';
|
|
12
26
|
export { KVCache } from './inference/kv-cache.js';
|
|
13
27
|
export { Tokenizer } from './inference/tokenizer.js';
|
|
14
28
|
export { SpeculativeDecoder } from './inference/speculative.js';
|
|
@@ -25,6 +39,22 @@ export {
|
|
|
25
39
|
mergeMultipleLogits,
|
|
26
40
|
} from './gpu/kernels/logit-merge.js';
|
|
27
41
|
|
|
42
|
+
export type { RDRRManifest, ShardInfo } from './formats/rdrr/index.js';
|
|
43
|
+
export type { TensorLocation, LoadProgress, LoadOptions, LoaderStats } from './loader/doppler-loader.js';
|
|
44
|
+
export type { AdapterSource } from './loader/multi-model-loader.js';
|
|
45
|
+
export type { ParsedModelConfig } from './generation/index.js';
|
|
46
|
+
export type { SamplingOptions } from './generation/index.js';
|
|
47
|
+
export type {
|
|
48
|
+
GenerateOptions,
|
|
49
|
+
GenerationResult,
|
|
50
|
+
KVCacheSnapshot,
|
|
51
|
+
LayerWeights,
|
|
52
|
+
ExpertWeights,
|
|
53
|
+
RouterWeights,
|
|
54
|
+
} from './generation/index.js';
|
|
55
|
+
export type { LoRAAdapter, LoRAModuleName } from './generation/index.js';
|
|
56
|
+
export type { ExpertNode, ExpertTask } from './inference/multi-model-network.js';
|
|
57
|
+
|
|
28
58
|
export {
|
|
29
59
|
ADAPTER_MANIFEST_SCHEMA,
|
|
30
60
|
validateManifest as validateAdapterManifest,
|
|
@@ -45,4 +75,18 @@ export {
|
|
|
45
75
|
createMemoryRegistry,
|
|
46
76
|
} from './adapters/index.js';
|
|
47
77
|
|
|
78
|
+
export type {
|
|
79
|
+
AdapterManifest,
|
|
80
|
+
AdapterMetadata,
|
|
81
|
+
AdapterTensorSpec,
|
|
82
|
+
LoRALoadOptions,
|
|
83
|
+
LoRAWeightsResult,
|
|
84
|
+
AdapterState,
|
|
85
|
+
EnableAdapterOptions,
|
|
86
|
+
AdapterStackOptions,
|
|
87
|
+
AdapterManagerEvents,
|
|
88
|
+
AdapterRegistryEntry,
|
|
89
|
+
AdapterQueryOptions,
|
|
90
|
+
} from './adapters/index.js';
|
|
91
|
+
|
|
48
92
|
export * from './tooling-exports.browser.js';
|
package/src/index-browser.js
CHANGED
|
@@ -11,6 +11,20 @@ export { MultiModelLoader } from './loader/multi-model-loader.js';
|
|
|
11
11
|
|
|
12
12
|
// Inference pipeline
|
|
13
13
|
export { InferencePipeline, EmbeddingPipeline, createPipeline } from './generation/index.js';
|
|
14
|
+
export {
|
|
15
|
+
StructuredJsonHeadPipeline,
|
|
16
|
+
isStructuredJsonHeadModelType,
|
|
17
|
+
createStructuredJsonHeadPipeline,
|
|
18
|
+
DreamStructuredPipeline,
|
|
19
|
+
isDreamStructuredModelType,
|
|
20
|
+
createDreamStructuredPipeline,
|
|
21
|
+
} from './generation/index.js';
|
|
22
|
+
export {
|
|
23
|
+
EnergyRowHeadPipeline,
|
|
24
|
+
createEnergyRowHeadPipeline,
|
|
25
|
+
DreamEnergyHeadPipeline,
|
|
26
|
+
createDreamEnergyHeadPipeline,
|
|
27
|
+
} from './inference/pipelines/energy-head/row-head-pipeline.js';
|
|
14
28
|
export { KVCache } from './inference/kv-cache.js';
|
|
15
29
|
export { Tokenizer } from './inference/tokenizer.js';
|
|
16
30
|
export { SpeculativeDecoder } from './inference/speculative.js';
|
|
@@ -0,0 +1,28 @@
|
|
|
1
|
+
import { getInferenceLayerPatternContractArtifact } from '../rules/rule-registry.js';
|
|
2
|
+
import { isPlainObject } from '../utils/plain-object.js';
|
|
3
|
+
import { validateBrowserSuiteMetrics } from '../config/schema/browser-suite-metrics.schema.js';
|
|
4
|
+
import { buildExecutionContractArtifact } from '../config/execution-contract-check.js';
|
|
5
|
+
import { buildManifestRequiredInferenceFieldsArtifact } from '../config/required-inference-fields-contract-check.js';
|
|
6
|
+
|
|
7
|
+
export function buildSuiteContractMetrics(suite, baseMetrics, manifest) {
|
|
8
|
+
const executionContractArtifact = buildExecutionContractArtifact(manifest);
|
|
9
|
+
const executionV0GraphContractArtifact = executionContractArtifact?.executionV0?.graph ?? null;
|
|
10
|
+
const layerPatternContractArtifact = getInferenceLayerPatternContractArtifact();
|
|
11
|
+
const requiredInferenceFieldsArtifact = manifest?.modelType === 'transformer'
|
|
12
|
+
&& isPlainObject(manifest?.inference?.attention)
|
|
13
|
+
? buildManifestRequiredInferenceFieldsArtifact(
|
|
14
|
+
manifest?.inference ?? null,
|
|
15
|
+
`${manifest?.modelId ?? 'unknown'}.inference`
|
|
16
|
+
)
|
|
17
|
+
: null;
|
|
18
|
+
return validateBrowserSuiteMetrics({
|
|
19
|
+
...baseMetrics,
|
|
20
|
+
schemaVersion: 1,
|
|
21
|
+
source: 'doppler',
|
|
22
|
+
suite,
|
|
23
|
+
...(executionContractArtifact ? { executionContractArtifact } : {}),
|
|
24
|
+
executionV0GraphContractArtifact,
|
|
25
|
+
layerPatternContractArtifact,
|
|
26
|
+
requiredInferenceFieldsArtifact,
|
|
27
|
+
});
|
|
28
|
+
}
|